[{"key": "pr#49822", "n": 49822, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "docs", "type": "docs", "title": "Add EntroPack quantization support", "action": "\u2014", "comment": "", "author": "mi804", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49822"}, {"key": "pr#49820", "n": 49820, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "processors (image/video/audio)", "type": "fix", "title": "Fix ProcessorMixin.apply_chat_template dropping explicit processor_kwargs", "action": "\u2014", "comment": "", "author": "codeitwithaayush", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49820"}, {"key": "pr#49818", "n": 49818, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "quantization", "type": "feature / refactor", "title": "\ud83d\udea8 Fine-grained quantization: a host quantizer that mixes methods per module", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49818"}, {"key": "issue#49817", "n": 49817, "kind": "issue", "group": "Open issues", "category": "question / other", "area": "trainer", "type": "question / other", "title": "RT-DETR / RT-DETRv2: main loss is computed on the denoising queries too, so real queries train as background", "action": "\u2014", "comment": "", "author": "hagai-psi", "idle_days": -1, "models": "rt_detr, rt_detr_v2", "url": "https://github.com/huggingface/transformers/issues/49817"}, {"key": "pr#49816", "n": 49816, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "fix model parallel bug for molmo2 model", "action": "\u2014", "comment": "", "author": "kaixuanliu", "idle_days": -1, "models": "molmo2", "url": "https://github.com/huggingface/transformers/pull/49816"}, {"key": "pr#49815", "n": 49815, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "CI / tooling", "title": "add XPU Expectations for 4 new model tests", "action": "\u2014", "comment": "", "author": "kaixuanliu", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49815"}, {"key": "pr#49813", "n": 49813, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "distributed", "type": "feature / refactor", "title": "Add a Dist-MoE expert-parallel strategy", "action": "\u2014", "comment": "", "author": "3outeille", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49813"}, {"key": "pr#49812", "n": 49812, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "Processor unnecessary warnings", "action": "\u2014", "comment": "", "author": "zucchini-nlp", "idle_days": -1, "models": "smolvlm", "url": "https://github.com/huggingface/transformers/pull/49812"}, {"key": "pr#49811", "n": 49811, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "attention / kernels", "type": "feature / refactor", "title": "Warn when sdpa or a flash kernel drops the attention softcap", "action": "\u2014", "comment": "", "author": "qgallouedec", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49811"}, {"key": "issue#49810", "n": 49810, "kind": "issue", "group": "Open issues", "category": "question / other", "area": "attention / kernels", "type": "question / other", "title": "Gemma 2's attention soft-capping is silently dropped with sdpa (the default) and flash kernels without softcap", "action": "\u2014", "comment": "", "author": "qgallouedec", "idle_days": -1, "models": "gemma2", "url": "https://github.com/huggingface/transformers/issues/49810"}, {"key": "pr#49809", "n": 49809, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "CI / tooling", "title": "skip `sdpa_can_dispatch_on_flash` tests for 3 models", "action": "\u2014", "comment": "", "author": "kaixuanliu", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49809"}, {"key": "pr#49808", "n": 49808, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "CI / tooling", "title": "The shared test resets the seed before eager and compiled runs, then \u2026", "action": "\u2014", "comment": "", "author": "sywangyi", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49808"}, {"key": "pr#49806", "n": 49806, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "fix failed test cases for unlimited_ocr model", "action": "\u2014", "comment": "", "author": "kaixuanliu", "idle_days": -1, "models": "unlimited_ocr", "url": "https://github.com/huggingface/transformers/pull/49806"}, {"key": "issue#49805", "n": 49805, "kind": "issue", "group": "Open issues", "category": "bug", "area": "other", "type": "bug", "title": "[serge] integration failure triage - 2026-10-20", "action": "\u2014", "comment": "", "author": "github-actions", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49805"}, {"key": "pr#49804", "n": 49804, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Return distribution parameters from time series generation", "action": "\u2014", "comment": "", "author": "kashif", "idle_days": 0, "models": "autoformer, informer, time_series_transformer", "url": "https://github.com/huggingface/transformers/pull/49804"}, {"key": "pr#49803", "n": 49803, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Support empty lags in TimeSeriesTransformer", "action": "\u2014", "comment": "", "author": "kashif", "idle_days": -1, "models": "informer, time_series_transformer", "url": "https://github.com/huggingface/transformers/pull/49803"}, {"key": "pr#49802", "n": 49802, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "CI / tooling", "title": "Pin the shared Hub MIC fleet dataset", "action": "\u2014", "comment": "", "author": "tarekziade", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49802"}, {"key": "pr#49799", "n": 49799, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "Convert numpy scalars inside lists/tuples in to_py_obj", "action": "review requested: @Rocketknight1", "comment": "", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/49799"}, {"key": "pr#49797", "n": 49797, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "CI / tooling", "title": "Bump datasets from 1.8.0 to 5.0.1 in /tests/sagemaker/scripts/pytorch", "action": "\u2014", "comment": "", "author": "dependabot", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/pull/49797"}, {"key": "pr#49796", "n": 49796, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "CI / tooling", "title": "Add AudioFlamingo3 to the prototype MIC fleet", "action": "\u2014", "comment": "", "author": "tarekziade", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/pull/49796"}, {"key": "pr#49794", "n": 49794, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "Apply EXIF orientation in load_image_as_tensor with torchvision", "action": "review requested: @ydshieh", "comment": "", "author": "digit50", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49794"}, {"key": "issue#49793", "n": 49793, "kind": "issue", "group": "Open issues", "category": "bug", "area": "generation", "type": "bug", "title": "[CB] `generate_batch` with flash attention gets slower as the KV cache gets bigger, even when the cache is mostly empty", "action": "labelled: bug", "comment": "", "author": "albertvillanova", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49793"}, {"key": "pr#49789", "n": 49789, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "Add PR triage bot to enforce contribution rules", "action": "\u2014", "comment": "", "author": "Rocketknight1", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49789"}, {"key": "pr#49787", "n": 49787, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "core / loading", "type": "feature / refactor", "title": "Shard MuseSpark's experts whole under tensor parallelism", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": -1, "models": "muse_spark", "url": "https://github.com/huggingface/transformers/pull/49787"}, {"key": "pr#49786", "n": 49786, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / loading", "type": "feature / refactor", "title": "Rename only MiniMax-M3-VL's MoE block, so a save keeps the dense and vision mlp keys", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": -1, "models": "minimax_m3_vl", "url": "https://github.com/huggingface/transformers/pull/49786"}, {"key": "pr#49784", "n": 49784, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "pipelines", "type": "fix", "title": "Fix duplicate code_revision when passed through model_kwargs", "action": "review requested: @Rocketknight1", "comment": "", "author": "Huang-404-Q", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49784"}, {"key": "pr#49783", "n": 49783, "kind": "PR", "group": "Open PRs", "category": "changes requested", "area": "CI / tooling", "type": "CI / tooling", "title": "Draft updates for contribution/agent rules", "action": "review requested: @vasqu; maintainer pinged", "comment": "@vasqu this was approved by @tarekziade on 2026-10-20; could you give it the core-maintainer review so it can be merged?", "author": "Rocketknight1", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49783"}, {"key": "pr#49780", "n": 49780, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "docs", "type": "docs", "title": "MXFP8 on the fine-grained machinery (`Mxfp8Config`)", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49780"}, {"key": "pr#49779", "n": 49779, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "docs", "type": "docs", "title": "NVFP4 on the fine-grained machinery", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": -1, "models": "muse_spark", "url": "https://github.com/huggingface/transformers/pull/49779"}, {"key": "pr#49778", "n": 49778, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "docs", "type": "docs", "title": "\ud83d\udea8 MXFP4 on the fine-grained machinery, and DeepSeek-V4's FP8 + MXFP4 plan", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": -1, "models": "deepseek_v4", "url": "https://github.com/huggingface/transformers/pull/49778"}, {"key": "pr#49777", "n": 49777, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "docs", "type": "docs", "title": "FP8 on the fine-grained machinery (`Fp8Config`)", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": -1, "models": "deepseek_v4, ministral3, mistral4, qwen4_exp", "url": "https://github.com/huggingface/transformers/pull/49777"}, {"key": "pr#49776", "n": 49776, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "quantization", "type": "feature / refactor", "title": "Fine-grained block-scaled quantization: the shared machinery", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49776"}, {"key": "pr#49775", "n": 49775, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "distributed", "type": "feature / refactor", "title": "Experts and TP: per-instance experts flags, load-time-only experts, scalars and per-expert scales in TP and EP", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49775"}, {"key": "pr#49774", "n": 49774, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "quantization", "type": "feature / refactor", "title": "Quantizer API: checkpoint renames, skip-lists in checkpoint names, dequantized saves", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49774"}, {"key": "pr#49773", "n": 49773, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / loading", "type": "fix", "title": "Weight loading: conversion ops declare many-to-many support, and four small loader fixes", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49773"}, {"key": "pr#49772", "n": 49772, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / loading", "type": "feature / refactor", "title": "Keep a capturing group's escapes when reversing a conversion pattern", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49772"}, {"key": "pr#49771", "n": 49771, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "quantization", "type": "feature / refactor", "title": "Match a skip-list entry from the start only up to a module boundary", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49771"}, {"key": "pr#49770", "n": 49770, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "new model", "title": "CTC loss refactor", "action": "\u2014", "comment": "", "author": "ebezzam", "idle_days": 0, "models": "data2vec, granite_speech5, hubert, lasr, omniasr, omniasr_ctc, parakeet, sew, unispeech, unispeech_sat, vibevoice_asr, wav2vec2, wav2vec2_bert, wav2vec2_conformer, wavlm", "url": "https://github.com/huggingface/transformers/pull/49770"}, {"key": "pr#49769", "n": 49769, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "trainer", "type": "fix", "title": "Fix MoE auxiliary loss scaling in Trainer", "action": "review requested: @SunMarc", "comment": "", "author": "priyanshu0463", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49769"}, {"key": "pr#49764", "n": 49764, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "CI / tooling", "type": "CI / tooling", "title": "feat(tokenizers): allow install of v1.0.0-rc.x", "action": "\u2014", "comment": "", "author": "SBrandeis", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/pull/49764"}, {"key": "pr#49763", "n": 49763, "kind": "PR", "group": "Closed PRs", "category": "duplicate", "area": "", "type": "", "title": "fix(parsing): decode escaped string delimiters in json content parser", "action": "closed", "comment": "Thanks for the fix! @Rocketknight1 opened #49741 a day earlier for the same issue (#49674), changing the same `json` content parser and also updating the docs. To keep review in one place, could you check whether #49741 covers your LFM2.5 cases (escaped `'`, `\\\\`, newlines)? If it does we'll close this one; if not, a review or comment on #49741 with the missing cases would be very welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/49763"}, {"key": "issue#49762", "n": 49762, "kind": "issue", "group": "Open issues", "category": "bug", "area": "tokenization", "type": "bug", "title": "ProcessorMixin.apply_chat_template treats template kwargs unused by the template as processor kwargs", "action": "\u2014", "comment": "", "author": "albertvillanova", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/49762"}, {"key": "issue#49761", "n": 49761, "kind": "issue", "group": "Open issues", "category": "bug", "area": "tokenization", "type": "bug", "title": "SmolVLMProcessor.apply_chat_template always warns \"Kwargs passed to processor.__call__ have to be in processor_kwargs\"", "action": "\u2014", "comment": "", "author": "albertvillanova", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/49761"}, {"key": "issue#49760", "n": 49760, "kind": "issue", "group": "Open issues", "category": "bug", "area": "tokenization", "type": "bug", "title": "ProcessorMixin.apply_chat_template drops explicit processor_kwargs when an extra kwarg is passed", "action": "\u2014", "comment": "", "author": "albertvillanova", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/49760"}, {"key": "pr#49756", "n": 49756, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "docs", "type": "docs", "title": "[Atlas Inference] Add GB10/SM121 Hub-kernel path for NVFP4Config", "action": "review requested: @SunMarc", "comment": "", "author": "AzeezIsh", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49756"}, {"key": "pr#49754", "n": 49754, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "Modular suffix overlap - allow non-model classes", "action": "\u2014", "comment": "", "author": "zucchini-nlp", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49754"}, {"key": "pr#49753", "n": 49753, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "generation", "type": "feature / refactor", "title": "Keep image inputs on the first chunk of a chunked prefill", "action": "\u2014", "comment": "", "author": "digit50", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/pull/49753"}, {"key": "pr#49749", "n": 49749, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "cache / masks", "type": "fix", "title": "fix(cache): clear request position_ids on reset (fixes #49745)", "action": "review requested: @ydshieh", "comment": "", "author": "Omarinho12345", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49749"}, {"key": "pr#49748", "n": 49748, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "processors (image/video/audio)", "type": "fix", "title": "Fix library import failing without `torch` when we have `PIL` installed", "action": "\u2014", "comment": "", "author": "Sai-Suraj-27", "idle_days": 0, "models": "gemma4, gemma4_unified", "url": "https://github.com/huggingface/transformers/pull/49748"}, {"key": "pr#49747", "n": 49747, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "other", "type": "new model", "title": "[jev-likes] have the decision heads set the number of options the model can score", "action": "\u2014", "comment": "", "author": "apolinario", "idle_days": 1, "models": "jev", "url": "https://github.com/huggingface/transformers/pull/49747"}, {"key": "pr#49746", "n": 49746, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "docs", "type": "docs", "title": "[docs] Quantizer integration", "action": "\u2014", "comment": "", "author": "stevhliu", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49746"}, {"key": "issue#49745", "n": 49745, "kind": "issue", "group": "Open issues", "category": "bug", "area": "cache / masks", "type": "bug", "title": "Qwen4-Exp: cache.reset() leaves stale position_ids when reusing DynamicCache or StaticCache", "action": "labelled: bug", "comment": "", "author": "taking-lying-flat", "idle_days": -1, "models": "qwen4_exp", "url": "https://github.com/huggingface/transformers/issues/49745"}, {"key": "pr#49744", "n": 49744, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "fix", "title": "Fix grouped RMSNorm in Mamba2's non-fused path", "action": "\u2014", "comment": "", "author": "taking-lying-flat", "idle_days": 1, "models": "bamba, falcon_h1, granitemoehybrid, mamba2", "url": "https://github.com/huggingface/transformers/pull/49744"}, {"key": "pr#49743", "n": 49743, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "attention / kernels", "type": "feature / refactor", "title": "Fold GQA query heads onto their KV head when decoding instead of repeating K/V", "action": "\u2014", "comment": "", "author": "Yooniel", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49743"}, {"key": "pr#49741", "n": 49741, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "docs", "type": "docs", "title": "JSON backslash escapes in response parsing", "action": "\u2014", "comment": "", "author": "Rocketknight1", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49741"}, {"key": "pr#49737", "n": 49737, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "new model", "title": "Add native Ling 3.1 Flash support to BailingHybrid", "action": "\u2014", "comment": "", "author": "icavan", "idle_days": -1, "models": "bailing_hybrid", "url": "https://github.com/huggingface/transformers/pull/49737"}, {"key": "issue#49734", "n": 49734, "kind": "issue", "group": "Open issues", "category": "bug", "area": "generation", "type": "bug", "title": "[serve] Sampling not enabled for positive temperature", "action": "\u2014", "comment": "", "author": "AbhilakshSinghReen", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49734"}, {"key": "pr#49733", "n": 49733, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tokenization", "type": "fix", "title": "fix(tokenizers): stop treating a missing transformers_version as Mistral", "action": "\u2014", "comment": "", "author": "cdcailf", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/pull/49733"}, {"key": "pr#49732", "n": 49732, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "serving / continuous batching", "type": "fix", "title": "[serve] Fix sampling not enabled for positive temperature", "action": "\u2014", "comment": "", "author": "AbhilakshSinghReen", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49732"}, {"key": "pr#49730", "n": 49730, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "serving / continuous batching", "type": "feature / refactor", "title": "[serve ] update attn for mac users ", "action": "\u2014", "comment": "", "author": "SunMarc", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49730"}, {"key": "pr#49729", "n": 49729, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "serving / continuous batching", "type": "feature / refactor", "title": "[serve] set attn text config", "action": "\u2014", "comment": "", "author": "SunMarc", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49729"}, {"key": "pr#49728", "n": 49728, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "CI / tooling", "title": "test: add cpu parity test for continuous batching generation", "action": "review requested: @remi-or", "comment": "", "author": "nrenvo", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49728"}, {"key": "issue#49727", "n": 49727, "kind": "issue", "group": "Open issues", "category": "bug", "area": "trainer", "type": "bug", "title": "[Bug] Qwen3-MoE router auxiliary loss is multiplied by the DDP world size in Trainer", "action": "labelled: bug", "comment": "", "author": "gss10282025", "idle_days": -1, "models": "qwen3_moe", "url": "https://github.com/huggingface/transformers/issues/49727"}, {"key": "pr#49723", "n": 49723, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "CI / tooling", "title": "Make distributed Trainer tests actually resume from checkpoints", "action": "\u2014", "comment": "", "author": "Titus-von-Koeller", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49723"}, {"key": "pr#49721", "n": 49721, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "Fix inverted eager attention mask in NemotronAsrStreaming encoder", "action": "review requested: @ydshieh", "comment": "", "author": "Alex-Wengg", "idle_days": -1, "models": "nemotron_asr_streaming", "url": "https://github.com/huggingface/transformers/pull/49721"}, {"key": "pr#49714", "n": 49714, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "docs", "type": "docs", "title": "Add NeoBERT checkpoint conversion to EuroBERT", "action": "\u2014", "comment": "", "author": "capemox", "idle_days": 2, "models": "eurobert", "url": "https://github.com/huggingface/transformers/pull/49714"}, {"key": "pr#49709", "n": 49709, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "docs", "type": "docs", "title": "[Atlas Inference] Strix Halo (gfx1151) Mamba-2 + conv1d Hub kernel path", "action": "\u2014", "comment": "", "author": "AzeezIsh", "idle_days": 1, "models": "mamba2", "url": "https://github.com/huggingface/transformers/pull/49709"}, {"key": "pr#49705", "n": 49705, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "[serge] Fix 2 integration tests regressed by commit c48824e84024 (PR #49404)", "action": "review requested: @vasqu; maintainer pinged", "comment": "@vasqu this was approved by @IlyasMoutawwakil on 2026-10-18; could you give it the core-maintainer review so it can be merged?", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/49705"}, {"key": "pr#49703", "n": 49703, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "fix", "title": "fix(mask2former): register input_projections as nn.ModuleList (fixes #49698)", "action": "review requested: @vasqu; maintainer pinged", "comment": "@vasqu this was approved by @Rocketknight1 on 2026-10-19; could you give it the core-maintainer review so it can be merged?", "author": "Omarinho12345", "idle_days": -1, "models": "mask2former", "url": "https://github.com/huggingface/transformers/pull/49703"}, {"key": "issue#49701", "n": 49701, "kind": "issue", "group": "Open issues", "category": "bug", "area": "processors (image/video/audio)", "type": "bug", "title": "Torchvision backend ignores EXIF orientation for path, URL and base64 images", "action": "labelled: bug", "comment": "", "author": "digit50", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49701"}, {"key": "issue#49700", "n": 49700, "kind": "issue", "group": "Open issues", "category": "bug", "area": "generation", "type": "bug", "title": "prefill_chunk_size drops image inputs, VLMs generate without the image", "action": "labelled: bug", "comment": "", "author": "digit50", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49700"}, {"key": "issue#49698", "n": 49698, "kind": "issue", "group": "Open issues", "category": "bug", "area": "other", "type": "bug", "title": "input_projections are not registered as PyTorch modules, causing missing parameters and CUDA device mismatch", "action": "\u2014", "comment": "", "author": "bruno2009", "idle_days": 3, "models": "", "url": "https://github.com/huggingface/transformers/issues/49698"}, {"key": "pr#49696", "n": 49696, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "generation", "type": "fix", "title": "Fix encoder logits processors when precomputed encoder_outputs lack token IDs", "action": "\u2014", "comment": "", "author": "kunalsanga", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49696"}, {"key": "issue#49695", "n": 49695, "kind": "issue", "group": "Open issues", "category": "bug", "area": "generation", "type": "bug", "title": "Generation crashes with encoder_repetition_penalty when encoder_outputs are provided without input_ids", "action": "\u2014", "comment": "", "author": "kunalsanga", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49695"}, {"key": "pr#49694", "n": 49694, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Qwen4-Exp: skip QSA block scoring when the budget covers every visible key", "action": "review requested: @ydshieh", "comment": "", "author": "crisostomi", "idle_days": -1, "models": "qwen4_exp", "url": "https://github.com/huggingface/transformers/pull/49694"}, {"key": "pr#49689", "n": 49689, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "processors (image/video/audio)", "type": "fix", "title": "Fix multi-zoom crops in the SAM mask-generation pipeline and related bugs (Fixes #35530)", "action": "\u2014", "comment": "", "author": "stlukyanenko", "idle_days": 0, "models": "sam, sam2, sam3", "url": "https://github.com/huggingface/transformers/pull/49689"}, {"key": "pr#49685", "n": 49685, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "new model", "title": "Add UltraVBERT", "action": "\u2014", "comment": "", "author": "athrael-soju", "idle_days": -1, "models": "ultravbert", "url": "https://github.com/huggingface/transformers/pull/49685"}, {"key": "pr#49681", "n": 49681, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "serving / continuous batching", "type": "feature / refactor", "title": "Add v1/audio/speech endpoint ", "action": "\u2014", "comment": "", "author": "SunMarc", "idle_days": -1, "models": "vibevoice", "url": "https://github.com/huggingface/transformers/pull/49681"}, {"key": "pr#49679", "n": 49679, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "distributed", "type": "feature / refactor", "title": "\ud83d\udea8 Separate EP logic into its own file", "action": "\u2014", "comment": "", "author": "3outeille", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49679"}, {"key": "pr#49676", "n": 49676, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "generation", "type": "feature / refactor", "title": "Make TextStreamer linear on long lines by decoding a bounded token window", "action": "review requested: @Cyrilvallez", "comment": "", "author": "deadstrobe5", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49676"}, {"key": "issue#49675", "n": 49675, "kind": "issue", "group": "Open issues", "category": "feature", "area": "generation", "type": "feature", "title": "TextStreamer re-decodes the whole line on every token (quadratic on long lines)", "action": "labelled: Feature request", "comment": "", "author": "deadstrobe5", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49675"}, {"key": "issue#49674", "n": 49674, "kind": "issue", "group": "Open issues", "category": "bug", "area": "other", "type": "bug", "title": "Chat parsing: `json` parser's `string_delims` doesn't handle escaped delimiters", "action": "labelled: bug", "comment": "", "author": "qgallouedec", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49674"}, {"key": "pr#49673", "n": 49673, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "pipelines", "type": "feature / refactor", "title": "[Pipelines] Reject format specs in zero-shot hypothesis_template", "action": "review requested: @vasqu; maintainer pinged", "comment": "@vasqu this was approved by @Rocketknight1 on 2026-10-19; could you give it the core-maintainer review so it can be merged?", "author": "errmakov", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49673"}, {"key": "issue#49672", "n": 49672, "kind": "issue", "group": "Open issues", "category": "question / other", "area": "other", "type": "question / other", "title": "Quick survey: your experience as a PyTorch Ecosystem maintainer", "action": "\u2014", "comment": "", "author": "michellearoth", "idle_days": 4, "models": "", "url": "https://github.com/huggingface/transformers/issues/49672"}, {"key": "pr#49663", "n": 49663, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "generation", "type": "feature / refactor", "title": "Defer scikit-learn, GenerationMixin, watermark processors and the gguf loader out of config/tokenizer loading", "action": "review requested: @Cyrilvallez", "comment": "", "author": "jayzuccarelli", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49663"}, {"key": "pr#49658", "n": 49658, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "docs", "type": "docs", "title": "Integrate finegrained kernels: SOTA performance and per-module quantization", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": -1, "models": "deepseek_v4, ministral3, mistral4, muse_spark, qwen4_exp", "url": "https://github.com/huggingface/transformers/pull/49658"}, {"key": "pr#49650", "n": 49650, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "docs", "type": "docs", "title": "[docs] NVIDIA/Intel/CPU", "action": "\u2014", "comment": "", "author": "stevhliu", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49650"}, {"key": "pr#49649", "n": 49649, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "gguf", "type": "feature / refactor", "title": "Read add_bos_token and add_eos_token from GGUF files", "action": "review requested: @SunMarc", "comment": "", "author": "francip", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49649"}, {"key": "issue#49647", "n": 49647, "kind": "issue", "group": "Open issues", "category": "bug", "area": "gguf", "type": "bug", "title": "GGUF tokenizer ignores tokenizer.ggml.add_bos_token, so the default encoding drops BOS", "action": "labelled: bug", "comment": "", "author": "francip", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49647"}, {"key": "issue#49645", "n": 49645, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "BatchEncoding.convert_to_tensors copies the whole batch to check if it's empty (~2x slower collation)", "action": "labelled: Feature request", "comment": "", "author": "jayzuccarelli", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49645"}, {"key": "pr#49640", "n": 49640, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "[Gemma4] Stop prefill at last KV-owning layer when logits_to_keep=1", "action": "\u2014", "comment": "", "author": "pculliton", "idle_days": 0, "models": "gemma4, gemma4_unified", "url": "https://github.com/huggingface/transformers/pull/49640"}, {"key": "pr#49639", "n": 49639, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "docs", "type": "docs", "title": "[i18n-FR] Translate philosophy.md to French", "action": "\u2014", "comment": "", "author": "Zaynabealb", "idle_days": 5, "models": "", "url": "https://github.com/huggingface/transformers/pull/49639"}, {"key": "pr#49638", "n": 49638, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "new model", "title": "Add the `deepep` expert-parallel strategy", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": 1, "models": "diffusion_gemma, gemma4, granitemoe_swa, hy_v4, inkling, llama4, mellum, mimo_v2_flash, qwen3_moe", "url": "https://github.com/huggingface/transformers/pull/49638"}, {"key": "issue#49630", "n": 49630, "kind": "issue", "group": "Open issues", "category": "bug", "area": "trainer", "type": "bug", "title": "Gemma4-12B Training Failed", "action": "\u2014", "comment": "", "author": "kezoy", "idle_days": 5, "models": "", "url": "https://github.com/huggingface/transformers/issues/49630"}, {"key": "pr#49629", "n": 49629, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "tests", "type": "CI / tooling", "title": "Tests: tell MoE models apart by their experts modules, not their config", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49629"}, {"key": "issue#49626", "n": 49626, "kind": "issue", "group": "Open issues", "category": "docs", "area": "docs", "type": "docs", "title": "[i18n-ur] Translating docs to Urdu (\u0627\u0631\u062f\u0648)", "action": "\u2014", "comment": "", "author": "mwzkhalil", "idle_days": 4, "models": "", "url": "https://github.com/huggingface/transformers/issues/49626"}, {"key": "issue#49625", "n": 49625, "kind": "issue", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "Janus / DeepSeek-VL image processors ignore `background_color`; deepseek-vl-7b-chat's low-res branch pads with 127 instead of its saved (122, 116, 104)", "action": "labelled: bug", "comment": "", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/49625"}, {"key": "pr#49624", "n": 49624, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add Cellpose (SAM + DINOv3)", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 0, "models": "cellpose, cellpose_dino, dinov3_vit, sapiens2", "url": "https://github.com/huggingface/transformers/pull/49624"}, {"key": "issue#49623", "n": 49623, "kind": "issue", "group": "Open issues", "category": "bug", "area": "tokenization", "type": "bug", "title": "_patch_mistral_regex treats a missing transformers_version as \"Mistral\": false positive on local DeBERTa-v3, and following its advice breaks the tokenizer", "action": "labelled: bug", "comment": "", "author": "cdcailf", "idle_days": -1, "models": "mistral", "url": "https://github.com/huggingface/transformers/issues/49623"}, {"key": "issue#49622", "n": 49622, "kind": "issue", "group": "Open issues", "category": "bug", "area": "tokenization", "type": "bug", "title": "DebertaV2Tokenizer drops SentencePiece NFKC normalization in v5: ligatures, fullwidth and circled characters encode to [UNK]", "action": "labelled: bug", "comment": "", "author": "tritsystem", "idle_days": -1, "models": "deberta_v2", "url": "https://github.com/huggingface/transformers/issues/49622"}, {"key": "pr#49621", "n": 49621, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "feature / refactor", "title": "Object Detection Losses refactor - from dedicated folder to respective modeling files", "action": "review requested: @molbap", "comment": "", "author": "sbucaille", "idle_days": -1, "models": "conditional_detr, d_fine, dab_detr, deformable_detr, deimv2, detr, grounding_dino, lw_detr, maskformer, mm_grounding_dino, pp_doclayout_v2, pp_doclayout_v3, rf_detr, rt_detr, rt_detr_v2, sam2", "url": "https://github.com/huggingface/transformers/pull/49621"}, {"key": "pr#49619", "n": 49619, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "processors (image/video/audio)", "type": "fix", "title": "fix(auto): handle legacy positional exist_ok in AutoVideoProcessor.register", "action": "review requested: @ydshieh", "comment": "", "author": "xingzihai", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49619"}, {"key": "pr#49617", "n": 49617, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "trainer", "type": "feature / refactor", "title": "[Trainer] Compile with static shapes when the backend is neuron", "action": "review requested: @Cyrilvallez; maintainer pinged", "comment": "@Cyrilvallez this was approved by @SunMarc on 2026-10-15; could you give it the core-maintainer review so it can be merged?", "author": "qgallouedec", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49617"}, {"key": "pr#49616", "n": 49616, "kind": "PR", "group": "Closed PRs", "category": "duplicate", "area": "", "type": "", "title": "Fix MusicGen forward(labels=...) failing when decoder_start_token_id is None", "action": "closed", "comment": "Thanks for the fix! #49118 (opened earlier) implements the same `bos_token_id` fallback for #49095, also with a test. To avoid splitting review, we'll keep the earlier PR and close this one; feel free to leave review comments on #49118.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/49616"}, {"key": "pr#49615", "n": 49615, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "distributed", "type": "feature / refactor", "title": "Remove the legacy `ep_router` expert-parallel plan", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": 1, "models": "llama4", "url": "https://github.com/huggingface/transformers/pull/49615"}, {"key": "pr#49614", "n": 49614, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "\ud83d\udea8 fix(sam3): invoke geometry encoder for text-only prompts to include learned cls token (#49592)", "action": "review requested: @ydshieh", "comment": "", "author": "devtyagi3909", "idle_days": -1, "models": "sam3, sam3_lite_text", "url": "https://github.com/huggingface/transformers/pull/49614"}, {"key": "pr#49613", "n": 49613, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "quantization", "type": "feature / refactor", "title": "Run MXFP4 GPT-OSS experts through `ep_dispatch_experts`", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": 1, "models": "gpt_oss", "url": "https://github.com/huggingface/transformers/pull/49613"}, {"key": "pr#49609", "n": 49609, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "Support chat sanitization on processors", "action": "\u2014", "comment": "", "author": "Rocketknight1", "idle_days": 5, "models": "idefics2, lfm2_vl, minicpmv4_6, minicpmv4_7, mllama", "url": "https://github.com/huggingface/transformers/pull/49609"}, {"key": "pr#49607", "n": 49607, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "feature / refactor", "title": "Migrate remaining router-masked MoE models to `ep_dispatch_experts`", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": 1, "models": "diffusion_gemma, gemma4, granitemoe_swa, hy_v4, inkling, mellum, mimo_v2_flash, qwen3_moe", "url": "https://github.com/huggingface/transformers/pull/49607"}, {"key": "pr#49606", "n": 49606, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "distributed", "type": "feature / refactor", "title": "`deepgemm_megamoe` EP implementation for DeepGEMM Mega MoE", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49606"}, {"key": "pr#49602", "n": 49602, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Full-attention cross-attention cache and export-safe masks", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": 6, "models": "t5gemma2", "url": "https://github.com/huggingface/transformers/pull/49602"}, {"key": "pr#49601", "n": 49601, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "feature / refactor", "title": "Cache: a cross-attention layer type, and shaped empty tensors on lazy init", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": 6, "models": "idefics", "url": "https://github.com/huggingface/transformers/pull/49601"}, {"key": "pr#49594", "n": 49594, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Capture only the projected hidden states in VLM feature extraction", "action": "\u2014", "comment": "", "author": "hmellor", "idle_days": 1, "models": "aria, aya_vision, clipseg, deepseek_vl_hybrid, granite4_vision, internvl, llava, llava_next, llava_next_video, llava_onevision, mistral3, pe_audio, pe_audio_video, pe_video, phi4_multimodal, qianfan_ocr, video_llava, vipllava", "url": "https://github.com/huggingface/transformers/pull/49594"}, {"key": "issue#49592", "n": 49592, "kind": "issue", "group": "Open issues", "category": "bug", "area": "models", "type": "bug", "title": "SAM 3: text-only prompts drop the geometry CLS token that Meta's implementation always adds", "action": "labelled: bug", "comment": "", "author": "ju-ius", "idle_days": -1, "models": "sam3", "url": "https://github.com/huggingface/transformers/issues/49592"}, {"key": "issue#49591", "n": 49591, "kind": "issue", "group": "Open issues", "category": "bug", "area": "other", "type": "bug", "title": "import transformers spends 1-3 s in an eager importlib.metadata.packages_distributions() when torch >= 2.14 is installed (Python 3.12)", "action": "labelled: bug", "comment": "", "author": "MakazhanAlpamys", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49591"}, {"key": "pr#49587", "n": 49587, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "distributed", "type": "feature / refactor", "title": "remove ep_router ", "action": "\u2014", "comment": "", "author": "3outeille", "idle_days": 5, "models": "", "url": "https://github.com/huggingface/transformers/pull/49587"}, {"key": "pr#49586", "n": 49586, "kind": "PR", "group": "Open PRs", "category": "changes requested", "area": "CI / tooling", "type": "CI / tooling", "title": "\ud83d\udd12 Pin GitHub Actions to commit SHAs", "action": "\u2014", "comment": "", "author": "paulinebm", "idle_days": 6, "models": "", "url": "https://github.com/huggingface/transformers/pull/49586"}, {"key": "pr#49584", "n": 49584, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "distributed", "type": "fix", "title": "Fix redundant checkpoint loading across ranks with DeepSpeed ZeRO-3", "action": "review requested: @ydshieh", "comment": "", "author": "xvxuopop", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49584"}, {"key": "pr#49581", "n": 49581, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 1 integration test for model `deepseek_vl_hybrid` failing with `output_mismatch` (list output differs (1))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 5, "models": "deepseek_vl, deepseek_vl_hybrid", "url": "https://github.com/huggingface/transformers/pull/49581"}, {"key": "pr#49579", "n": 49579, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tokenization", "type": "feature / refactor", "title": "Preserve tokenizer.json pipeline stages when a class rebuilds the backend", "action": "\u2014", "comment": "", "author": "OrangerXu", "idle_days": 7, "models": "", "url": "https://github.com/huggingface/transformers/pull/49579"}, {"key": "issue#49578", "n": 49578, "kind": "issue", "group": "Open issues", "category": "bug", "area": "tokenization", "type": "bug", "title": "Tokenizer loses its normalizer from tokenizer.json in 5.18", "action": "\u2014", "comment": "", "author": "AdnanElAssadi", "idle_days": 7, "models": "", "url": "https://github.com/huggingface/transformers/issues/49578"}, {"key": "pr#49576", "n": 49576, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "CI / tooling", "title": "[EmbeddingGemma2] update integration tests", "action": "\u2014", "comment": "", "author": "eustlb", "idle_days": 7, "models": "embedding_gemma2", "url": "https://github.com/huggingface/transformers/pull/49576"}, {"key": "pr#49574", "n": 49574, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "fix", "title": "fix(ci): harden workflow triggers, secrets and gates (#49357)", "action": "\u2014", "comment": "", "author": "hf-security-analysis", "idle_days": 7, "models": "", "url": "https://github.com/huggingface/transformers/pull/49574"}, {"key": "pr#49567", "n": 49567, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "feature / refactor", "title": "handle local sink in TP", "action": "\u2014", "comment": "", "author": "3outeille", "idle_days": 6, "models": "granite_swa, granitemoe_swa, hy_v4, mimo_v2_flash", "url": "https://github.com/huggingface/transformers/pull/49567"}, {"key": "pr#49566", "n": 49566, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "new model", "title": "add support for jev like models", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 0, "models": "clef, jev, kev, qwen3_5, qwen3_5_moe, qwen4_exp, von", "url": "https://github.com/huggingface/transformers/pull/49566"}, {"key": "issue#49564", "n": 49564, "kind": "issue", "group": "Open issues", "category": "bug", "area": "attention / kernels", "type": "bug", "title": "SDPA: enable_gqa=True silently falls back to the math kernel at batch 1 on torch < 2.14 without Flash Attention", "action": "\u2014", "comment": "", "author": "Jose-Ribeir", "idle_days": 7, "models": "", "url": "https://github.com/huggingface/transformers/issues/49564"}, {"key": "pr#49559", "n": 49559, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "generation", "type": "fix", "title": "Fix Whisper timestamp processor with single-element EOS collections (#49556)", "action": "review requested: @Cyrilvallez", "comment": "", "author": "kunalsanga", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49559"}, {"key": "pr#49554", "n": 49554, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "cache / masks", "type": "feature / refactor", "title": "Allocate the StaticCache values with the value states' dtype", "action": "review requested: @Cyrilvallez; maintainer pinged", "comment": "@Cyrilvallez this was approved by @dacorvo on 2026-10-14; could you give it the core-maintainer review so it can be merged?", "author": "qgallouedec", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49554"}, {"key": "issue#49553", "n": 49553, "kind": "issue", "group": "Open issues", "category": "bug", "area": "tokenization", "type": "bug", "title": "`/v1/systemone`: 500s for `chat_template_kwargs` overrides and for models without a chat template", "action": "labelled: bug", "comment": "", "author": "rycerzes", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49553"}, {"key": "pr#49552", "n": 49552, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Only tie MM Grounding DINO decoder heads when the config shares them", "action": "\u2014", "comment": "", "author": "lu123bing", "idle_days": -1, "models": "grounding_dino, mm_grounding_dino", "url": "https://github.com/huggingface/transformers/pull/49552"}, {"key": "pr#49551", "n": 49551, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "feature / refactor", "title": "[PP-DocLayoutV4] Add page geometry prediction", "action": "\u2014", "comment": "", "author": "maxiaolong03", "idle_days": 0, "models": "pp_doclayout_v4", "url": "https://github.com/huggingface/transformers/pull/49551"}, {"key": "issue#49546", "n": 49546, "kind": "issue", "group": "Open issues", "category": "feature", "area": "cache / masks", "type": "feature", "title": "CPU SDPA could use native GQA with a mask on torch >= 2.9 (1.2x decode, no repeat_kv)", "action": "labelled: Feature request", "comment": "", "author": "ToukoUrsin", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49546"}, {"key": "pr#49542", "n": 49542, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "feature / refactor", "title": "\ud83d\udea8 Keep static sliding-window decode on one graph", "action": "review requested: @vasqu; maintainer pinged", "comment": "@vasqu this was approved by @dacorvo on 2026-10-20; could you give it the core-maintainer review so it can be merged?", "author": "IlyasMoutawwakil", "idle_days": -1, "models": "diffusion_gemma, gemma3, gemma4, gemma4_unified, paligemma, unlimited_ocr", "url": "https://github.com/huggingface/transformers/pull/49542"}, {"key": "issue#49540", "n": 49540, "kind": "issue", "group": "Open issues", "category": "bug", "area": "models", "type": "bug", "title": "Gemma 3/4 multimodal models fail with tp_plan=\"auto\" because of mixed Tensor/DTensor usage", "action": "\u2014", "comment": "", "author": "gbharada-aws", "idle_days": 7, "models": "gemma3", "url": "https://github.com/huggingface/transformers/issues/49540"}, {"key": "issue#49536", "n": 49536, "kind": "issue", "group": "Open issues", "category": "bug", "area": "trainer", "type": "bug", "title": "Mamba2: non-fused gated RMSNorm ignores n_groups and disagrees with fused training", "action": "labelled: bug", "comment": "", "author": "taking-lying-flat", "idle_days": -1, "models": "mamba2", "url": "https://github.com/huggingface/transformers/issues/49536"}, {"key": "pr#49534", "n": 49534, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "serving / continuous batching", "type": "feature / refactor", "title": "v1/systemone chat template kwargs + unit tests", "action": "\u2014", "comment": "", "author": "SunMarc", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/pull/49534"}, {"key": "pr#49533", "n": 49533, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "fix(ci): harden GitHub Actions workflows (#49532)", "action": "closed", "comment": "Thanks for the workflow hardening! This targets `contracts-proof`, whose PR #49532 (Model Integration Contracts CI) was closed without merging, so `.github/workflows/contracts.yml` never reached `main` and there's nothing left to harden. Closing for that reason.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/49533"}, {"key": "issue#49529", "n": 49529, "kind": "issue", "group": "Open issues", "category": "bug", "area": "models", "type": "bug", "title": "MM Grounding DINO and LLMDet: decoder box and class heads are tied to layer 0 although the config says they are not shared", "action": "labelled: bug", "comment": "", "author": "MustafaOrhon", "idle_days": -1, "models": "grounding_dino, mm_grounding_dino", "url": "https://github.com/huggingface/transformers/issues/49529"}, {"key": "pr#49528", "n": 49528, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Subconfig creation in testers", "action": "\u2014", "comment": "", "author": "zucchini-nlp", "idle_days": -1, "models": "siglip", "url": "https://github.com/huggingface/transformers/pull/49528"}, {"key": "pr#49527", "n": 49527, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "distributed", "type": "feature / refactor", "title": "\ud83d\udea8 build and expand once the tp/ep_plan/fsdp_plan", "action": "\u2014", "comment": "", "author": "3outeille", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49527"}, {"key": "pr#49526", "n": 49526, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "[Qwen2Audio] Fix dtype mismatch in legacy audio merge", "action": "\u2014", "comment": "", "author": "syc632", "idle_days": 6, "models": "qwen2_audio", "url": "https://github.com/huggingface/transformers/pull/49526"}, {"key": "pr#49521", "n": 49521, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "attention / kernels", "type": "feature / refactor", "title": "Gate enable_gqa on fused SDPA kernel support", "action": "\u2014", "comment": "", "author": "Jose-Ribeir", "idle_days": 7, "models": "", "url": "https://github.com/huggingface/transformers/pull/49521"}, {"key": "pr#49520", "n": 49520, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tokenization", "type": "feature / refactor", "title": "Dispatch BarthezTokenizer on vocabulary shape for BPE checkpoints", "action": "\u2014", "comment": "", "author": "Irfanwani", "idle_days": 5, "models": "barthez", "url": "https://github.com/huggingface/transformers/pull/49520"}, {"key": "issue#49519", "n": 49519, "kind": "issue", "group": "Open issues", "category": "bug", "area": "other", "type": "bug", "title": "Assigning rope_scaling drops existing RoPE base parameters", "action": "\u2014", "comment": "", "author": "jinwukong", "idle_days": 8, "models": "", "url": "https://github.com/huggingface/transformers/issues/49519"}, {"key": "pr#49512", "n": 49512, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / utils", "type": "fix", "title": "[CB] Fix the CPU memory budget used to size the paged cache", "action": "\u2014", "comment": "", "author": "jiqing-feng", "idle_days": 4, "models": "", "url": "https://github.com/huggingface/transformers/pull/49512"}, {"key": "pr#49507", "n": 49507, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 2 integration tests for model `pegasus` failing with `other` (other (2))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 9, "models": "pegasus", "url": "https://github.com/huggingface/transformers/pull/49507"}, {"key": "pr#49503", "n": 49503, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "pipelines", "type": "fix", "title": "[fix][ASR] Fix Qwen3 ASR pipeline transcription", "action": "\u2014", "comment": "", "author": "Sunt-ing", "idle_days": 8, "models": "qwen3_asr", "url": "https://github.com/huggingface/transformers/pull/49503"}, {"key": "pr#49501", "n": 49501, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "distributed", "type": "feature / refactor", "title": "Shard the parameters kept in float32 in their own FSDP2 group", "action": "\u2014", "comment": "", "author": "qgallouedec", "idle_days": 6, "models": "", "url": "https://github.com/huggingface/transformers/pull/49501"}, {"key": "issue#49498", "n": 49498, "kind": "issue", "group": "Open issues", "category": "bug", "area": "generation", "type": "bug", "title": "SynthID-Text: CPU and MPS produce different sampling tables with the same seed", "action": "labelled: bug", "comment": "", "author": "sk8ordie84", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49498"}, {"key": "issue#49497", "n": 49497, "kind": "issue", "group": "Open issues", "category": "bug", "area": "core / loading", "type": "bug", "title": "ZeRO-3: from_pretrained reads and converts the full checkpoint on every rank", "action": "\u2014", "comment": "", "author": "xvxuopop", "idle_days": 9, "models": "", "url": "https://github.com/huggingface/transformers/issues/49497"}, {"key": "pr#49495", "n": 49495, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "processors (image/video/audio)", "type": "fix", "title": "Request to a vision model fails with a 500 - Fix group_images_by_shape to key on full shape, not spatial dims only", "action": "\u2014", "comment": "", "author": "bornmw", "idle_days": 5, "models": "", "url": "https://github.com/huggingface/transformers/pull/49495"}, {"key": "pr#49490", "n": 49490, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "trainer", "type": "fix", "title": "Fix Trainer prediction handling with context parallelism", "action": "review requested: @ydshieh", "comment": "", "author": "DimensionSTP", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49490"}, {"key": "pr#49489", "n": 49489, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "layer execution", "type": "feature / refactor", "title": "[RFC] Add configurable decoder layer execution (reference implementation)", "action": "\u2014", "comment": "", "author": "bebetterest", "idle_days": 10, "models": "recurrent_gemma", "url": "https://github.com/huggingface/transformers/pull/49489"}, {"key": "issue#49488", "n": 49488, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "Discussion: configurable decoder loops with shared parameters and independent execution caches", "action": "labelled: Feature request", "comment": "", "author": "bebetterest", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49488"}, {"key": "pr#49485", "n": 49485, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "docs: clarify GPU memory requirements in run_clm example", "action": "closed", "comment": "We removed the legacy `examples/` scripts from the repository (#49357) since they were no longer maintained, so this change no longer has a home. Thanks for the contribution!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/49485"}, {"key": "pr#49480", "n": 49480, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 2 integration tests for model `qwen2_moe` failing with `other` (other (2))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 10, "models": "qwen2_moe", "url": "https://github.com/huggingface/transformers/pull/49480"}, {"key": "pr#49478", "n": 49478, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix uniform frame indices 49410 upstream", "action": "closed", "comment": "Thanks for the fix! Issue #49410 was resolved on main by #49464, which deprecates `get_uniform_frame_indices` (now emits a FutureWarning, to be removed in v5.26) in favor of `default_sample_indices_fn`, so it is no longer worth patching. Closing as superseded.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/49478"}, {"key": "issue#49469", "n": 49469, "kind": "issue", "group": "Open issues", "category": "bug", "area": "processors (image/video/audio)", "type": "bug", "title": "AutoVideoProcessor.register treats positional exist_ok as a backend dictionary", "action": "\u2014", "comment": "", "author": "DimensionSTP", "idle_days": 11, "models": "", "url": "https://github.com/huggingface/transformers/issues/49469"}, {"key": "pr#49467", "n": 49467, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "quantization", "type": "feature / refactor", "title": "Update metal quant kernels", "action": "\u2014", "comment": "", "author": "SunMarc", "idle_days": 5, "models": "", "url": "https://github.com/huggingface/transformers/pull/49467"}, {"key": "issue#49465", "n": 49465, "kind": "issue", "group": "Open issues", "category": "bug", "area": "models", "type": "bug", "title": "Mamba2 / Zamba2: cached decoding does not respect the time-step limits used by sequence scan", "action": "labelled: bug", "comment": "", "author": "taking-lying-flat", "idle_days": -1, "models": "mamba2, zamba2", "url": "https://github.com/huggingface/transformers/issues/49465"}, {"key": "pr#49459", "n": 49459, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "attention / kernels", "type": "feature / refactor", "title": "Add `force_gqa` flag in `sdpa_attention_forward`", "action": "\u2014", "comment": "", "author": "jakobknauer", "idle_days": 11, "models": "", "url": "https://github.com/huggingface/transformers/pull/49459"}, {"key": "pr#49449", "n": 49449, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "trainer", "type": "fix", "title": "Fix Trainer checkpoint rotation matching backup directories", "action": "review requested: @SunMarc", "comment": "", "author": "SrijanRoy123-github", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49449"}, {"key": "issue#49448", "n": 49448, "kind": "issue", "group": "Open issues", "category": "bug", "area": "trainer", "type": "bug", "title": "Trainer checkpoint rotation can delete valid checkpoints when backup directories match the checkpoint prefix", "action": "\u2014", "comment": "", "author": "SrijanRoy123-github", "idle_days": 10, "models": "", "url": "https://github.com/huggingface/transformers/issues/49448"}, {"key": "pr#49444", "n": 49444, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "attention / kernels", "type": "feature / refactor", "title": "Use 32-row flex attention blocks for head_dim > 256", "action": "\u2014", "comment": "", "author": "qgallouedec", "idle_days": 11, "models": "", "url": "https://github.com/huggingface/transformers/pull/49444"}, {"key": "pr#49438", "n": 49438, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "new model", "title": "Add GLiNER2", "action": "\u2014", "comment": "", "author": "MattThomas-fastino", "idle_days": 5, "models": "gliner2", "url": "https://github.com/huggingface/transformers/pull/49438"}, {"key": "pr#49435", "n": 49435, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "[vLLM] Add Omni token counting and text replacement offsets for the Transformers backend", "action": "\u2014", "comment": "", "author": "harshaljanjani", "idle_days": 0, "models": "embedding_gemma2, gemma3n, gemma4, gemma4_unified, qwen2_5_omni, qwen3_omni_moe", "url": "https://github.com/huggingface/transformers/pull/49435"}, {"key": "pr#49430", "n": 49430, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "Add rerun-failed CPU/GPU callers for the CI dashboard", "action": "\u2014", "comment": "", "author": "tarekziade", "idle_days": 12, "models": "", "url": "https://github.com/huggingface/transformers/pull/49430"}, {"key": "pr#49429", "n": 49429, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "distributed", "type": "feature / refactor", "title": "Expert-parallel implementations for `ep_dispatch_experts`: all-reduce at inference, all-to-all in training", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49429"}, {"key": "pr#49423", "n": 49423, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "serving / continuous batching", "type": "fix", "title": "Fix Gemma 4 tool call parsing for names containing - or .", "action": "\u2014", "comment": "", "author": "Varun-51", "idle_days": 11, "models": "gemma4", "url": "https://github.com/huggingface/transformers/pull/49423"}, {"key": "issue#49422", "n": 49422, "kind": "issue", "group": "Open issues", "category": "bug", "area": "models", "type": "bug", "title": "Gemma 4 `response_template` fails on tool names containing `-` or `.`", "action": "\u2014", "comment": "", "author": "Varun-51", "idle_days": 12, "models": "gemma4", "url": "https://github.com/huggingface/transformers/issues/49422"}, {"key": "pr#49415", "n": 49415, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "docs", "type": "docs", "title": "Support static cache in assisted decoding", "action": "\u2014", "comment": "", "author": "jiqing-feng", "idle_days": 11, "models": "minimax_m3_vl", "url": "https://github.com/huggingface/transformers/pull/49415"}, {"key": "pr#49407", "n": 49407, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 2 integration tests for model `utils` failing with `output_mismatch` (list output differs (2))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 12, "models": "", "url": "https://github.com/huggingface/transformers/pull/49407"}, {"key": "pr#49405", "n": 49405, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "other", "type": "fix", "title": "[`GradientCkpting`] Fix non grad arg inputs", "action": "\u2014", "comment": "", "author": "vasqu", "idle_days": 8, "models": "", "url": "https://github.com/huggingface/transformers/pull/49405"}, {"key": "pr#49402", "n": 49402, "kind": "PR", "group": "Closed PRs", "category": "duplicate", "area": "", "type": "", "title": "Fix Parakeet TDT zero-duration stall with max_symbols_per_step guard", "action": "closed", "comment": "Thanks for the fix! #49401 implements the same max_symbols_per_step guard for Parakeet TDT and also adds a regression test, so we'll go ahead with that one and close this to avoid a duplicate review. Feel free to review #49401 if you see a difference in behaviour (e.g. your blank-mask condition).", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/49402"}, {"key": "pr#49401", "n": 49401, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "Fix Parakeet TDT ignoring max_symbols_per_step during greedy decoding", "action": "\u2014", "comment": "", "author": "ayush3781", "idle_days": 13, "models": "parakeet", "url": "https://github.com/huggingface/transformers/pull/49401"}, {"key": "pr#49392", "n": 49392, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "fix", "title": "fix(ci): harden GitHub Actions workflows (#49254)", "action": "\u2014", "comment": "", "author": "hf-security-analysis", "idle_days": 13, "models": "", "url": "https://github.com/huggingface/transformers/pull/49392"}, {"key": "pr#49390", "n": 49390, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "exporters", "type": "feature / refactor", "title": "Enable in-place cache updates for Whisper XNNPACK exports", "action": "\u2014", "comment": "", "author": "JakeStevens", "idle_days": 13, "models": "", "url": "https://github.com/huggingface/transformers/pull/49390"}, {"key": "issue#49388", "n": 49388, "kind": "issue", "group": "Open issues", "category": "bug", "area": "other", "type": "bug", "title": "Parakeet TDT greedy decoding ignores `max_symbols_per_step`: unbounded token repeats on one encoder frame", "action": "\u2014", "comment": "", "author": "tomschott", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/49388"}, {"key": "pr#49386", "n": 49386, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Initialize masked_spec_embed in _init_weights for 7 audio models", "action": "\u2014", "comment": "", "author": "tarekziade", "idle_days": 8, "models": "data2vec, sew, unispeech, unispeech_sat, wav2vec2, wav2vec2_conformer, wavlm", "url": "https://github.com/huggingface/transformers/pull/49386"}, {"key": "issue#49385", "n": 49385, "kind": "issue", "group": "Open issues", "category": "bug", "area": "cache / masks", "type": "bug", "title": "Wav2Vec2 and 6 other audio models leave masked_spec_embed uninitialized when the checkpoint lacks it", "action": "\u2014", "comment": "", "author": "tarekziade", "idle_days": 13, "models": "", "url": "https://github.com/huggingface/transformers/issues/49385"}, {"key": "pr#49382", "n": 49382, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "Audio from video in jinja ", "action": "\u2014", "comment": "", "author": "zucchini-nlp", "idle_days": 0, "models": "embedding_gemma2, gemma4, gemma4_unified, qwen2_5_omni, qwen3_omni_moe", "url": "https://github.com/huggingface/transformers/pull/49382"}, {"key": "issue#49378", "n": 49378, "kind": "issue", "group": "Open issues", "category": "bug", "area": "tokenization", "type": "bug", "title": "Loading a tokenizer imports torch and scikit-learn when they are installed (about 2 s of a 3 s import)", "action": "labelled: bug", "comment": "", "author": "AshishKumar4", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49378"}, {"key": "pr#49376", "n": 49376, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "cache / masks", "type": "feature / refactor", "title": "Store sliding layers of the static cache at full size on Neuron/TPU", "action": "\u2014", "comment": "", "author": "dacorvo", "idle_days": 8, "models": "", "url": "https://github.com/huggingface/transformers/pull/49376"}, {"key": "issue#49372", "n": 49372, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "set_attn_implementation() does not reach T5 encoder/decoder: model reports SDPA but still runs eager", "action": "closed", "comment": "This was fixed by #49138, which is now on main. Closing, but feel free to reopen if you still see it on the latest version!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/49372"}, {"key": "pr#49351", "n": 49351, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "Remove the GitHub API cache-bust from the AMD docker images", "action": "\u2014", "comment": "", "author": "tarekziade", "idle_days": 14, "models": "", "url": "https://github.com/huggingface/transformers/pull/49351"}, {"key": "pr#49341", "n": 49341, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "tokenization", "type": "feature / refactor", "title": "Warn on pre_tokenizer mismatch between tokenizer.json and tokenizer c\u2026", "action": "\u2014", "comment": "", "author": "tejachavali", "idle_days": 14, "models": "", "url": "https://github.com/huggingface/transformers/pull/49341"}, {"key": "issue#49337", "n": 49337, "kind": "issue", "group": "Open issues", "category": "bug", "area": "other", "type": "bug", "title": "MetalLinear returns wrong outputs for every batch row after the first (3D inputs)", "action": "\u2014", "comment": "", "author": "francip", "idle_days": 13, "models": "", "url": "https://github.com/huggingface/transformers/issues/49337"}, {"key": "pr#49336", "n": 49336, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 2 integration tests for model `ministral3` failing with `other` (other (2))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 14, "models": "ministral3", "url": "https://github.com/huggingface/transformers/pull/49336"}, {"key": "pr#49335", "n": 49335, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 2 integration tests for model `deepseek_vl_hybrid` failing with `other` (other (2))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 14, "models": "deepseek_vl, deepseek_vl_hybrid", "url": "https://github.com/huggingface/transformers/pull/49335"}, {"key": "pr#49332", "n": 49332, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "fix", "title": "fix(dsa): Chunk the lightning indexer scores over the query dimension", "action": "\u2014", "comment": "", "author": "harshaljanjani", "idle_days": 14, "models": "axk2, deepseek_v32, glm5_next, glm_moe_dsa, hy_v4", "url": "https://github.com/huggingface/transformers/pull/49332"}, {"key": "pr#49331", "n": 49331, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "fix", "title": "Fix create_dummy_models CI failures for new models", "action": "\u2014", "comment": "", "author": "ydshieh", "idle_days": 14, "models": "", "url": "https://github.com/huggingface/transformers/pull/49331"}, {"key": "pr#49327", "n": 49327, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tokenization", "type": "fix", "title": "fix(tokenizers): prefer `vocab_file` over `tokenizer.json` vocab in `convert_to_native_format`", "action": "\u2014", "comment": "", "author": "shwetangsinha0509", "idle_days": 14, "models": "", "url": "https://github.com/huggingface/transformers/pull/49327"}, {"key": "pr#49321", "n": 49321, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Run MiniMax lightning-attention decay recurrence in fp32", "action": "review requested: @ydshieh", "comment": "", "author": "tritsystem", "idle_days": -1, "models": "minimax", "url": "https://github.com/huggingface/transformers/pull/49321"}, {"key": "pr#49316", "n": 49316, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Trainer: `save_pretrained` on every rank for PEFT model", "action": "closed", "comment": "Thanks @michaelbenayoun! The PEFT saving issue this targets was fixed by #49500, merged on 2026-10-14. In its current form this PR removes that code path, and every rank calls save_pretrained, which gathers the full sharded model per rank until huggingface/peft#3963 lands. Could you rescope it to just the should_save cleanup on top of #49500 (or close it if that's no longer needed)?", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/49316"}, {"key": "issue#49310", "n": 49310, "kind": "issue", "group": "Open issues", "category": "bug", "area": "models", "type": "bug", "title": "[PERF] DeepSeek-V3.2 / GLM-5 lightning indexer overallocates memory", "action": "\u2014", "comment": "", "author": "harshaljanjani", "idle_days": 15, "models": "deepseek_v3, deepseek_v32", "url": "https://github.com/huggingface/transformers/issues/49310"}, {"key": "pr#49304", "n": 49304, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 6 integration tests for model `plbart` failing with `other` (other (6))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 16, "models": "plbart", "url": "https://github.com/huggingface/transformers/pull/49304"}, {"key": "issue#49297", "n": 49297, "kind": "issue", "group": "Open issues", "category": "bug", "area": "trainer", "type": "bug", "title": "[FSDP2] cpu_ram_efficient_loading materializes full model on CPU for every non-rank-0, defeating rank-0 broadcast (host OOM)", "action": "\u2014", "comment": "", "author": "rjchao2021", "idle_days": 15, "models": "", "url": "https://github.com/huggingface/transformers/issues/49297"}, {"key": "pr#49290", "n": 49290, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "new model", "title": "Add RTDetrForSegmentation and RTDetrV2ForSegmentation", "action": "\u2014", "comment": "", "author": "ranykamel", "idle_days": 16, "models": "rt_detr, rt_detr_v2", "url": "https://github.com/huggingface/transformers/pull/49290"}, {"key": "issue#49289", "n": 49289, "kind": "issue", "group": "Open issues", "category": "bug", "area": "models", "type": "bug", "title": "MiniMax lightning attention: decay recurrence runs in bf16, so slow-decay heads never decay and decode drifts with length", "action": "labelled: bug", "comment": "", "author": "tritsystem", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49289"}, {"key": "pr#49287", "n": 49287, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 2 integration tests for model `videomae` failing with `other` (other (2))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 16, "models": "videomae", "url": "https://github.com/huggingface/transformers/pull/49287"}, {"key": "pr#49284", "n": 49284, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "processors (image/video/audio)", "type": "fix", "title": "[processors] fix processors silently ignoring audio wrong sampling_rate", "action": "\u2014", "comment": "", "author": "eustlb", "idle_days": 15, "models": "audioflamingo3, csm, dia, fun_asr_nano, gemma3n, gemma4, gemma4_unified, glmasr, granite_speech5, higgs_audio_v2, musicflamingo, parakeet, qwen2_5_omni, qwen3_asr, qwen3_omni_moe, vibevoice, vibevoice_asr", "url": "https://github.com/huggingface/transformers/pull/49284"}, {"key": "pr#49283", "n": 49283, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "[Nemotron3Diarization] add left context support", "action": "\u2014", "comment": "", "author": "eustlb", "idle_days": 17, "models": "nemotron3_diarization", "url": "https://github.com/huggingface/transformers/pull/49283"}, {"key": "issue#49272", "n": 49272, "kind": "issue", "group": "Open issues", "category": "bug", "area": "pipelines", "type": "bug", "title": "pipeline(..., model_kwargs={\"code_revision\": ...}) raises TypeError, and top-level code_revision does not reach the model load", "action": "\u2014", "comment": "", "author": "KiSung-Park", "idle_days": 7, "models": "", "url": "https://github.com/huggingface/transformers/issues/49272"}, {"key": "issue#49267", "n": 49267, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Gemma 4 `response_template` parses reasoning into `thinking`, which its chat template does not read back", "action": "closed", "comment": "Thanks for the detailed report! This is fixed by #49535: `apply_chat_template` now aliases `thinking` (and `reasoning`) to `reasoning_content`, so a parsed Gemma 4 response re-renders with its reasoning. It's on main and will be in the next release.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/49267"}, {"key": "pr#49266", "n": 49266, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "feature / refactor", "title": "[WIP] modular cleanup", "action": "\u2014", "comment": "", "author": "itazap", "idle_days": -1, "models": "afmoe, aimv2, altclip, apertus, aria, audio_spectrogram_transformer, audioflamingo3, axk1, axk2, beit, biogpt, bitnet, blip, blip_2, blt, canary, chinese_clip, chmv2, clip, clipseg, clvp, cohere, cohere2, cohere2_moe, cohere2_vision, cohere_asr, cohere_compass, colmodernvbert, colqwen2, conditional_detr, cosmos3_edge, cosmos3_omni, cwm", "url": "https://github.com/huggingface/transformers/pull/49266"}, {"key": "pr#49258", "n": 49258, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tokenization", "type": "fix", "title": "Fix CJK decoding for LlamaTokenizerFast by loading full pipeline from tokenizer.json  ", "action": "\u2014", "comment": "", "author": "MUKUL-PRASAD-SIGH", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49258"}, {"key": "pr#49254", "n": 49254, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "Add Intel XPU CI", "action": "\u2014", "comment": "", "author": "regisss", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49254"}, {"key": "issue#49252", "n": 49252, "kind": "issue", "group": "Open issues", "category": "bug", "area": "tokenization", "type": "bug", "title": "AutoTokenizer with tokenizer_class=LlamaTokenizerFast silently builds a sentencepiece pipeline for a byte-level BPE tokenizer.json, dropping all CJK input", "action": "\u2014", "comment": "", "author": "kangzhengyang", "idle_days": 14, "models": "", "url": "https://github.com/huggingface/transformers/issues/49252"}, {"key": "pr#49251", "n": 49251, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "docs", "type": "docs", "title": "Upgrade Python Min version to `3.11`", "action": "\u2014", "comment": "", "author": "Sai-Suraj-27", "idle_days": 0, "models": "clvp, gpt_neox_japanese", "url": "https://github.com/huggingface/transformers/pull/49251"}, {"key": "pr#49249", "n": 49249, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "feature / refactor", "title": "Rename GradientCheckpointingLayer and supports_gradient_checkpointing to activation checkpointing", "action": "\u2014", "comment": "", "author": "qgallouedec", "idle_days": 18, "models": "afmoe, aimv2, align, altclip, apertus, arcee, aria, audio_spectrogram_transformer, audioflamingo3, autoformer, axk1, axk2, aya_vision, bamba, bark, bart, beit, bert, bert_generation, big_bird, bigbird_pegasus, biogpt, bitnet, blenderbot, blenderbot_small, blip, blip_2, bloom, blt, bridgetower, bros, camembert, canary, canine, chameleon, chinese_clip, chmv2, clap", "url": "https://github.com/huggingface/transformers/pull/49249"}, {"key": "pr#49248", "n": 49248, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "feature / refactor", "title": "Rename gradient_checkpointing to activation_checkpointing in the public API", "action": "\u2014", "comment": "", "author": "qgallouedec", "idle_days": 18, "models": "dia, mbart", "url": "https://github.com/huggingface/transformers/pull/49248"}, {"key": "pr#49244", "n": 49244, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 2 integration tests for model `fuyu` failing with `other` (other (2))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 18, "models": "fuyu", "url": "https://github.com/huggingface/transformers/pull/49244"}, {"key": "pr#49243", "n": 49243, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 2 integration tests for model `pvt_v2` failing with `output_mismatch` (tensor values differ (2))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 18, "models": "pvt_v2", "url": "https://github.com/huggingface/transformers/pull/49243"}, {"key": "pr#49239", "n": 49239, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "trainer", "type": "fix", "title": "Fix Trainer step accounting when GradScaler skips an optimizer update", "action": "review requested: @SunMarc", "comment": "", "author": "gitedmond", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49239"}, {"key": "issue#49237", "n": 49237, "kind": "issue", "group": "Open issues", "category": "bug", "area": "trainer", "type": "bug", "title": "Trainer counts fp16 steps skipped by GradScaler as completed update steps (`global_step`, `max_steps`, `on_optimizer_step`)", "action": "labelled: bug", "comment": "", "author": "SeverinVisionary", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49237"}, {"key": "pr#49223", "n": 49223, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "fix", "title": "fix(ci): harden GitHub Actions workflows (#49220)", "action": "\u2014", "comment": "", "author": "hf-security-analysis", "idle_days": 19, "models": "", "url": "https://github.com/huggingface/transformers/pull/49223"}, {"key": "pr#49218", "n": 49218, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "cache / masks", "type": "fix", "title": "[serge] Fix 5 integration tests for model `utils` failing with `other` (other (5))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49218"}, {"key": "pr#49215", "n": 49215, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "core / loading", "type": "feature / refactor", "title": "feat: update safetensors with arbitrary slice in prefetch plan", "action": "\u2014", "comment": "", "author": "McPatate", "idle_days": 20, "models": "", "url": "https://github.com/huggingface/transformers/pull/49215"}, {"key": "pr#49213", "n": 49213, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "[Nemotron3Diarization] Pad the last streaming chunk by n_fft // 2, so a short one without look-ahead still has a window", "action": "\u2014", "comment": "", "author": "happyarts", "idle_days": 19, "models": "nemotron3_diarization, nemotron_asr_streaming", "url": "https://github.com/huggingface/transformers/pull/49213"}, {"key": "pr#49204", "n": 49204, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "distributed", "type": "feature / refactor", "title": "Perf MoE: Use a GEMM for the expert-bias gather backward in grouped_mm experts", "action": "\u2014", "comment": "", "author": "AmineDiro", "idle_days": 4, "models": "", "url": "https://github.com/huggingface/transformers/pull/49204"}, {"key": "pr#49202", "n": 49202, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "Fix SAM3 video reconditioning masks", "action": "\u2014", "comment": "", "author": "jrile018", "idle_days": 5, "models": "sam3_video", "url": "https://github.com/huggingface/transformers/pull/49202"}, {"key": "pr#49197", "n": 49197, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "[Fix] Fix CSM codebook embedding tying", "action": "\u2014", "comment": "", "author": "orrangetabby17", "idle_days": 19, "models": "csm", "url": "https://github.com/huggingface/transformers/pull/49197"}, {"key": "issue#49196", "n": 49196, "kind": "issue", "group": "Open issues", "category": "bug", "area": "other", "type": "bug", "title": "CSM codebook embedding tying is ignored when `tie_word_embeddings=False`", "action": "\u2014", "comment": "", "author": "orrangetabby17", "idle_days": 20, "models": "", "url": "https://github.com/huggingface/transformers/issues/49196"}, {"key": "issue#49195", "n": 49195, "kind": "issue", "group": "Open issues", "category": "bug", "area": "cache / masks", "type": "bug", "title": "SAM3 Video reconditioning causes mask to grow onto the edges and make the bounding boxes balloon", "action": "\u2014", "comment": "", "author": "Ryan-Xue2", "idle_days": 20, "models": "sam3_video", "url": "https://github.com/huggingface/transformers/issues/49195"}, {"key": "pr#49192", "n": 49192, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "Fix AutoTokenizer backend selection for CohereLabs tiny-aya checkpoints", "action": "review requested: @itazap; maintainer pinged", "comment": "@itazap could you take a look when you get a chance? This is waiting on a maintainer: itazap re-review.", "author": "Subodh-17", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49192"}, {"key": "issue#49190", "n": 49190, "kind": "issue", "group": "Open issues", "category": "bug", "area": "other", "type": "bug", "title": "[eager MOE] Selected expert IDs should probably not be treated as 0-d GPU tensors", "action": "\u2014", "comment": "", "author": "fxmarty-amd", "idle_days": 18, "models": "", "url": "https://github.com/huggingface/transformers/issues/49190"}, {"key": "pr#49189", "n": 49189, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "distributed", "type": "feature / refactor", "title": "Clip grad norm dtensor", "action": "\u2014", "comment": "", "author": "michaelbenayoun", "idle_days": 6, "models": "", "url": "https://github.com/huggingface/transformers/pull/49189"}, {"key": "pr#49188", "n": 49188, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Restore BLT's 512-byte sliding windows (entropy patcher + local encoder/decoder)", "action": "\u2014", "comment": "", "author": "hyunW3", "idle_days": 8, "models": "blt, encoder_decoder", "url": "https://github.com/huggingface/transformers/pull/49188"}, {"key": "pr#49186", "n": 49186, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "new model", "title": "Add a linear attention interface, starting with Mamba2's SSD core", "action": "\u2014", "comment": "", "author": "hmellor", "idle_days": 6, "models": "bamba, falcon_h1, granitemoehybrid, mamba2, nemotron_h, zamba2", "url": "https://github.com/huggingface/transformers/pull/49186"}, {"key": "issue#49185", "n": 49185, "kind": "issue", "group": "Open issues", "category": "bug", "area": "other", "type": "bug", "title": "[BLT, Byte-Latent Transformer] Entropy patcher is missing BLT's 512-token sliding-window attention \u2014 patch rate collapses on inputs longer than 512 bytes", "action": "\u2014", "comment": "", "author": "hyunW3", "idle_days": 11, "models": "", "url": "https://github.com/huggingface/transformers/issues/49185"}, {"key": "issue#49176", "n": 49176, "kind": "issue", "group": "Open issues", "category": "bug", "area": "tokenization", "type": "bug", "title": "tiny-aya models are loaded as CohereTokenizer in v5, and digits are split differently from tokenizer.json", "action": "\u2014", "comment": "", "author": "Yinhan-Lu", "idle_days": 21, "models": "", "url": "https://github.com/huggingface/transformers/issues/49176"}, {"key": "issue#49166", "n": 49166, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Rename `gradient_checkpointing` to `activation_checkpointing` in the next major", "action": "labelled: Feature request", "comment": "", "author": "qgallouedec", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49166"}, {"key": "pr#49162", "n": 49162, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "feature / refactor", "title": "[`distributed`] remaining ep_plan to update", "action": "\u2014", "comment": "", "author": "3outeille", "idle_days": 14, "models": "afmoe, axk1, axk2, cohere2_moe, deepseek_ocr2, deepseek_v2, deepseek_v3, deepseek_v32, deepseek_v4, diffusion_gemma, dots1, ernie4_5_moe, ernie4_5_vl_moe, exaone_moe, flex_olmo, gemma4, glm4_moe, glm4_moe_lite, glm4v_moe, glm5_next, glm_moe_dsa, gpt_oss, granitemoe_swa, hunyuan_v1_moe, hy_v3, hy_v4, inkling, kimi_linear, laguna, lfm2_moe, mellum, mimo_v2_flash, minimax, minimax_m2, minimax_m3_vl", "url": "https://github.com/huggingface/transformers/pull/49162"}, {"key": "pr#49154", "n": 49154, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / loading", "type": "fix", "title": "Serialize MPS transfers in async model loading (fix #48029)", "action": "\u2014", "comment": "", "author": "Irfanwani", "idle_days": 3, "models": "", "url": "https://github.com/huggingface/transformers/pull/49154"}, {"key": "pr#49150", "n": 49150, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "feat(ci): scan the docker images with Wiz before they are pushed", "action": "\u2014", "comment": "", "author": "rtrompier", "idle_days": 22, "models": "", "url": "https://github.com/huggingface/transformers/pull/49150"}, {"key": "pr#49143", "n": 49143, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 2 integration tests for model `jais2` failing with `output_mismatch` (tensor values differ (2))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 19, "models": "jais2", "url": "https://github.com/huggingface/transformers/pull/49143"}, {"key": "pr#49142", "n": 49142, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 2 integration tests for model `cohere2` failing with `other` (other (2))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 19, "models": "cohere2", "url": "https://github.com/huggingface/transformers/pull/49142"}, {"key": "pr#49141", "n": 49141, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Skip the query/key head repeat for grouped value heads in gated DeltaNet", "action": "\u2014", "comment": "", "author": "manav-simile", "idle_days": 15, "models": "olmo_hybrid, qwen3_5, qwen3_5_moe, qwen3_next, qwen4_exp", "url": "https://github.com/huggingface/transformers/pull/49141"}, {"key": "pr#49132", "n": 49132, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "[New model] Add K2 Horizon", "action": "\u2014", "comment": "", "author": "boyangaaaaa", "idle_days": 1, "models": "k2_horizon", "url": "https://github.com/huggingface/transformers/pull/49132"}, {"key": "pr#49125", "n": 49125, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "dummy: trigger CI on gemma3n test_model_4b_batch", "action": "closed", "comment": "Closing this CI-trigger PR, which its author titled \"dummy: trigger CI\". Thanks @ydshieh!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/49125"}, {"key": "pr#49120", "n": 49120, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "exporters", "type": "feature / refactor", "title": "[HF Exporters] Adds opt-in ExecuTorch off-graph KV cache export", "action": "\u2014", "comment": "", "author": "metascroy", "idle_days": 24, "models": "", "url": "https://github.com/huggingface/transformers/pull/49120"}, {"key": "pr#49118", "n": 49118, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "Fix MusicGen forward(labels=...) when decoder.decoder_start_token_id is None", "action": "\u2014", "comment": "", "author": "Omarinho12345", "idle_days": 11, "models": "musicgen", "url": "https://github.com/huggingface/transformers/pull/49118"}, {"key": "pr#49114", "n": 49114, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "pass pretokenizer regex charsmap to qwen2 tokenizer", "action": "\u2014", "comment": "", "author": "itazap", "idle_days": 19, "models": "qwen2", "url": "https://github.com/huggingface/transformers/pull/49114"}, {"key": "issue#49108", "n": 49108, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "Return hidden states in VLMs", "action": "\u2014", "comment": "", "author": "zucchini-nlp", "idle_days": 25, "models": "", "url": "https://github.com/huggingface/transformers/issues/49108"}, {"key": "pr#49107", "n": 49107, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "\ud83d\udea8 Deprecate low use models", "action": "\u2014", "comment": "", "author": "molbap", "idle_days": 12, "models": "", "url": "https://github.com/huggingface/transformers/pull/49107"}, {"key": "issue#49106", "n": 49106, "kind": "issue", "group": "Open issues", "category": "bug", "area": "models", "type": "bug", "title": "ModernBERT: local-attention window is computed from sequence index, not `position_ids`, so non-monotonic positions get a silently different window", "action": "labelled: bug", "comment": "", "author": "johnmofficial16-prog", "idle_days": -1, "models": "modernbert", "url": "https://github.com/huggingface/transformers/issues/49106"}, {"key": "pr#49103", "n": 49103, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "CI / tooling", "title": "Add FastIntegrationTestMixin and fast integration test classes (17 models)", "action": "\u2014", "comment": "", "author": "ydshieh", "idle_days": 15, "models": "", "url": "https://github.com/huggingface/transformers/pull/49103"}, {"key": "pr#49098", "n": 49098, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / utils", "type": "fix", "title": "fix(hf_argparser): support bare list dataclass fields", "action": "review requested: @Cyrilvallez", "comment": "", "author": "aniketkrs", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49098"}, {"key": "issue#49095", "n": 49095, "kind": "issue", "group": "Open issues", "category": "bug", "area": "generation", "type": "bug", "title": "MusicGen: forward(labels=...) raises \"set the decoder_start_token_id\" on every released checkpoint, while generate() and prepare_decoder_input_ids_from_labels() find the token fine", "action": "\u2014", "comment": "", "author": "KhalidIbrahem", "idle_days": 21, "models": "musicgen", "url": "https://github.com/huggingface/transformers/issues/49095"}, {"key": "issue#49094", "n": 49094, "kind": "issue", "group": "Open issues", "category": "bug", "area": "trainer", "type": "bug", "title": "MusicGen: hub configs set decoder dropout=0.1 but the checkpoints were trained with 0 \u2014 train() doubles the loss and silently breaks fine-tuning", "action": "\u2014", "comment": "", "author": "KhalidIbrahem", "idle_days": 22, "models": "musicgen", "url": "https://github.com/huggingface/transformers/issues/49094"}, {"key": "pr#49089", "n": 49089, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "Move processing defaults to class", "action": "\u2014", "comment": "", "author": "guarin", "idle_days": -1, "models": "align, aria, audioflamingo3, aya_vision, blip, blip_2, bridgetower, bros, canary, chameleon, clipseg, clvp, cohere2_vision, cohere_asr, cohere_compass, colmodernvbert, colpali, colqwen2, cosmos3_edge, csm, deepseek_ocr2, deepseek_vl, deepseek_vl_hybrid, dia, donut, embedding_gemma2, emu3, ernie4_5_vl_moe, exaone4_5, florence2, fun_asr_nano, fuyu, gemma3, gemma3n, gemma4, gemma4_unified, glm46v, glm4v, glm5_next, glm_image, glmasr", "url": "https://github.com/huggingface/transformers/pull/49089"}, {"key": "pr#49087", "n": 49087, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "[encodec] separate encoder and decoder", "action": "\u2014", "comment": "", "author": "eustlb", "idle_days": 5, "models": "bark, encodec", "url": "https://github.com/huggingface/transformers/pull/49087"}, {"key": "pr#49085", "n": 49085, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "[WIP] Modality tester auto-parameterize based on model class", "action": "\u2014", "comment": "", "author": "zucchini-nlp", "idle_days": 26, "models": "cohere_compass", "url": "https://github.com/huggingface/transformers/pull/49085"}, {"key": "pr#49079", "n": 49079, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "Composite transparency in shared RGB conversion and image processors", "action": "\u2014", "comment": "", "author": "Nikhi00718", "idle_days": -1, "models": "chameleon, idefics2, mllama", "url": "https://github.com/huggingface/transformers/pull/49079"}, {"key": "pr#49078", "n": 49078, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "Scope GITHUB_TOKEN permissions per job", "action": "\u2014", "comment": "", "author": "hf-security-analysis", "idle_days": 26, "models": "", "url": "https://github.com/huggingface/transformers/pull/49078"}, {"key": "pr#49077", "n": 49077, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix Qwen2Tokenizer ignoring pretokenize_regex (Qwen3.6 / Qwen3.8 wrong token IDs)", "action": "closed", "comment": "Thanks @tanvir-ux! As @itazap explained in #49066, the hardcoded pre-tokenizer regex in `tokenization_qwen2.py` is intentional, and the wrong IDs were fixed by mapping those checkpoints to the right tokenizer class in #49116. Closing as superseded. If you still hit wrong token IDs on current main (e.g. the case @magnuslundstedt described, where the checkpoint declares `Qwen2Tokenizer` itself), please follow up on #49066 with the checkpoint name.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/49077"}, {"key": "pr#49064", "n": 49064, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "dont merge - i'm checking smth", "action": "\u2014", "comment": "", "author": "zucchini-nlp", "idle_days": 26, "models": "gemma3", "url": "https://github.com/huggingface/transformers/pull/49064"}, {"key": "pr#49048", "n": 49048, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "fix: treat empty protobuf bytes as absent in SentencePiece charsmap", "action": "closed", "comment": "Thanks for working on this! The empty-charsmap case from #48942 was fixed on main by #49211: `convert_slow_tokenizer.py` now turns an empty `precompiled_charsmap` into `None` at the source (line 195), so the per-tokenizer `is not None` checks behave correctly without per-model changes. Closing as superseded; please reopen if you still see the failure on main.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/49048"}, {"key": "pr#49043", "n": 49043, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "cache / masks", "type": "fix", "title": "fix(cache): roll back recurrent states in LinearAttentionCacheLayerMi\u2026", "action": "review requested: @Cyrilvallez; maintainer pinged", "comment": "@Cyrilvallez could you take a look when you get a chance? This is waiting on a maintainer: design review by cache owners (crop on recurrent layers is explicitly unsupported on main; PR adds per-forward snapshots).", "author": "Varun-51", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49043"}, {"key": "pr#49041", "n": 49041, "kind": "PR", "group": "Closed PRs", "category": "already fixed", "area": "", "type": "", "title": "[serge] Fix 4 integration tests for model `minicpm3` failing with `other` (other (4))", "action": "closed", "comment": "Thanks for the fix! The `KeyError: 'factor'` is already fixed on main by #49555: `yarn_apply_mscale` now reads `factor` only inside the `mscale_all_dim` branch (`modeling_minicpm3.py:188-189`), so the minicpm3 integration tests should no longer hit it. Closing as already fixed; feel free to reopen if the failure persists on main.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/49041"}, {"key": "issue#49036", "n": 49036, "kind": "issue", "group": "Open issues", "category": "bug", "area": "generation", "type": "bug", "title": "Cache.crop does not roll back linear-attention recurrent states \u2014 block/speculative verification on GDN hybrids (Qwen3-Next / Qwen3.5) is not lossless with DynamicCache", "action": "labelled: bug", "comment": "", "author": "0xWhiteMage", "idle_days": -1, "models": "qwen3, qwen3_5, qwen3_next", "url": "https://github.com/huggingface/transformers/issues/49036"}, {"key": "pr#49030", "n": 49030, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "fix", "title": "Fix : Move pyctcdecode and kenlm from audio to speech-lm extra", "action": "\u2014", "comment": "", "author": "priyanshu0463", "idle_days": 3, "models": "", "url": "https://github.com/huggingface/transformers/pull/49030"}, {"key": "issue#49028", "n": 49028, "kind": "issue", "group": "Open issues", "category": "bug", "area": "other", "type": "bug", "title": "HfArgumentParser crashes on a dataclass field typed as a bare list", "action": "labelled: bug", "comment": "", "author": "caiotheodoro", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/49028"}, {"key": "pr#49025", "n": 49025, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "docs", "type": "docs", "title": "[docs] CB cache behavior", "action": "\u2014", "comment": "", "author": "stevhliu", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/pull/49025"}, {"key": "pr#49024", "n": 49024, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "processors (image/video/audio)", "type": "fix", "title": "Fix/asr batched multichannel mono v3", "action": "review requested: @ydshieh", "comment": "", "author": "kamran-nizamani", "idle_days": -1, "models": "cohere_asr, lasr, nemotron_asr_streaming, parakeet, phi4_multimodal, voxtral_realtime", "url": "https://github.com/huggingface/transformers/pull/49024"}, {"key": "pr#49022", "n": 49022, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "trainer", "type": "feature / refactor", "title": "Trainer resume with weight conversion", "action": "\u2014", "comment": "", "author": "SunMarc", "idle_days": 25, "models": "", "url": "https://github.com/huggingface/transformers/pull/49022"}, {"key": "pr#49020", "n": 49020, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "generation", "type": "feature / refactor", "title": "Maybe just load a tokenizer when loading `gen_config`", "action": "\u2014", "comment": "", "author": "zucchini-nlp", "idle_days": 28, "models": "", "url": "https://github.com/huggingface/transformers/pull/49020"}, {"key": "pr#49015", "n": 49015, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "quantization", "type": "feature / refactor", "title": "Finegrained kernels: unified quantization kernels api (separate quantizers)", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": 4, "models": "glm4v_moe, ministral3, mistral4", "url": "https://github.com/huggingface/transformers/pull/49015"}, {"key": "pr#49012", "n": 49012, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "new model", "title": "Granite for Docling", "action": "\u2014", "comment": "", "author": "nassarofficial", "idle_days": 0, "models": "granite_for_docling", "url": "https://github.com/huggingface/transformers/pull/49012"}, {"key": "pr#49010", "n": 49010, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "feat(perf)L optimisation for pp-ocrv5/6 detection data movement gpu -> cpu", "action": "review requested: @molbap", "comment": "", "author": "FrancescoSaverioZuppichini", "idle_days": -1, "models": "pp_ocrv5_server_rec, pp_ocrv6_small_rec", "url": "https://github.com/huggingface/transformers/pull/49010"}, {"key": "pr#49004", "n": 49004, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "cache / masks", "type": "feature / refactor", "title": "[cache] Make StaticCache buckets", "action": "\u2014", "comment": "", "author": "Cyrilvallez", "idle_days": 28, "models": "", "url": "https://github.com/huggingface/transformers/pull/49004"}, {"key": "pr#48989", "n": 48989, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "fix", "title": "fix(nemotron_h): honor configured Mamba time-step limits", "action": "\u2014", "comment": "", "author": "yuhezhang-ai", "idle_days": 4, "models": "nemotron_h", "url": "https://github.com/huggingface/transformers/pull/48989"}, {"key": "pr#48986", "n": 48986, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "feature / refactor", "title": "\ud83d\udea8 Exporters Runtimes : Export and run any Transformers model (Dynamo/OpenVINO/OnnxRuntime/ExecuTorch)", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": 13, "models": "ctrl, ernie4_5_vl_moe, falcon_h1, falcon_mamba, granite_speech, granite_speech_plus, idefics, internvl, jamba, kimi_k25, kosmos2_5, llava, longcat_flash, mamba, mimi, minimax, minimax_m3_vl, mllama, openai, opt, paddleocr_vl, prophetnet, qianfan_ocr", "url": "https://github.com/huggingface/transformers/pull/48986"}, {"key": "pr#48976", "n": 48976, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "tests", "type": "fix", "title": "QA: Check for a leaked class fixture at the class boundary", "action": "\u2014", "comment": "", "author": "tarekziade", "idle_days": 20, "models": "", "url": "https://github.com/huggingface/transformers/pull/48976"}, {"key": "pr#48970", "n": 48970, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "trainer", "type": "feature / refactor", "title": "docs(optimization): standardize 'Return:' to 'Returns:' in schedule d\u2026", "action": "review requested: @Cyrilvallez; maintainer pinged", "comment": "@Cyrilvallez this was approved by @stevhliu on 2026-09-21; could you give it the core-maintainer review so it can be merged?", "author": "Shreyas-cpu", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/48970"}, {"key": "pr#48969", "n": 48969, "kind": "PR", "group": "Closed PRs", "category": "already fixed", "area": "", "type": "", "title": "Fix module aliases breaking unittest via __warningregistry__ probes", "action": "review requested: @ydshieh; maintainer pinged; closed", "comment": "Thanks for digging into this! The module aliases were reworked in #48992: they're now served through a sys.meta_path finder rather than a module-level __getattr__, and issue #48966 was closed as fixed. The code this PR patches no longer exists on main, so I'm closing it. If you still see the unittest __warningregistry__ problem on current main, please reopen the issue with a repro.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48969"}, {"key": "issue#48966", "n": 48966, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Module-level getattr breaks unittest functionality", "action": "closed", "comment": "This was fixed by #48992, which is now on main. Closing, but feel free to reopen if you still see it on the latest version!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/48966"}, {"key": "pr#48955", "n": 48955, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 2 integration tests for model `llava_next_video` failing with `other` (other (2))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 19, "models": "llava_next, llava_next_video", "url": "https://github.com/huggingface/transformers/pull/48955"}, {"key": "pr#48952", "n": 48952, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "fix", "title": "Fix Windows path handling in repo checkers (check_auto, check_noisy_comments)", "action": "\u2014", "comment": "", "author": "dltsum", "idle_days": 29, "models": "", "url": "https://github.com/huggingface/transformers/pull/48952"}, {"key": "pr#48951", "n": 48951, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "other", "type": "fix", "title": "Fix Windows path separators breaking relative imports of Hub custom-generate modules", "action": "review requested: @Rocketknight1", "comment": "", "author": "dltsum", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/48951"}, {"key": "pr#48950", "n": 48950, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "Fix SuperPoint torch.export by replacing data-dependent shapes", "action": "\u2014", "comment": "", "author": "shreyescodes", "idle_days": 19, "models": "superpoint", "url": "https://github.com/huggingface/transformers/pull/48950"}, {"key": "pr#48949", "n": 48949, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Pin utf-8 in add-new-model-like file IO (Windows non-UTF-8 locale fix)", "action": "closed", "comment": "Thanks for the fix! The same change landed in #48990, which applied ruff's PLW1514 rule repo-wide and pinned `encoding=\"utf-8\"` on every `open()` in `src/transformers/cli/add_new_model_like.py`, so main now has all of these. Closing as superseded, but thanks again for reporting the Windows locale issue.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48949"}, {"key": "pr#48948", "n": 48948, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Pin utf-8 in test source reads (Windows non-UTF-8 locale fix)", "action": "closed", "comment": "Thanks for this, it led directly to #48990, which enabled ruff's PLW1514 and fixed these reads repo-wide. Closing as superseded.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48948"}, {"key": "pr#48938", "n": 48938, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add lm_head tp_plan to tied Gemma multimodal ForConditionalGeneration", "action": "\u2014", "comment": "", "author": "gbharada-aws", "idle_days": 4, "models": "gemma3, gemma4, gemma4_unified", "url": "https://github.com/huggingface/transformers/pull/48938"}, {"key": "pr#48936", "n": 48936, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tokenization", "type": "feature / refactor", "title": "support byte fallback for loading SPM Backend tok", "action": "\u2014", "comment": "", "author": "itazap", "idle_days": 22, "models": "", "url": "https://github.com/huggingface/transformers/pull/48936"}, {"key": "pr#48930", "n": 48930, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / loading", "type": "feature / refactor", "title": "feat: use safetensors prefetch api", "action": "\u2014", "comment": "", "author": "McPatate", "idle_days": 20, "models": "", "url": "https://github.com/huggingface/transformers/pull/48930"}, {"key": "pr#48919", "n": 48919, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "serving / continuous batching", "type": "feature / refactor", "title": "Restrict remote media URLs in serve", "action": "review requested: @ydshieh; maintainer pinged", "comment": "@ydshieh could you take a look when you get a chance? This is waiting on a maintainer: never reviewed (security-relevant).", "author": "Snakinya", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/48919"}, {"key": "pr#48918", "n": 48918, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "serving / continuous batching", "type": "feature / refactor", "title": "Use unique request IDs for continuous batching", "action": "\u2014", "comment": "", "author": "Snakinya", "idle_days": 32, "models": "", "url": "https://github.com/huggingface/transformers/pull/48918"}, {"key": "pr#48915", "n": 48915, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "[perf] Qwen3.5: fused RMSNorm kernels from flash-linear-attention, 0.605 s to 0.486 s per training step on 8x B200", "action": "closed", "comment": "Thanks for the speedup! Fused kernels like this should come through the kernels hub (`use_kernel_forward_from_hub`) rather than living in model files, so we're closing this.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48915"}, {"key": "issue#48913", "n": 48913, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "MtpModel/MtpLayer topology is too rigid to represent non-uniform MTP architectures", "action": "labelled: Feature request", "comment": "", "author": "soyr-redhat", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/48913"}, {"key": "pr#48909", "n": 48909, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "attention / kernels", "type": "feature / refactor", "title": "[`DSA`] Add native FA4 support", "action": "\u2014", "comment": "", "author": "vasqu", "idle_days": 33, "models": "deepseek_v3, glm_moe_dsa", "url": "https://github.com/huggingface/transformers/pull/48909"}, {"key": "pr#48901", "n": 48901, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "new model", "title": "Add BerryLM", "action": "\u2014", "comment": "", "author": "DanilSmorchkov", "idle_days": 14, "models": "berrylm", "url": "https://github.com/huggingface/transformers/pull/48901"}, {"key": "pr#48893", "n": 48893, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Translate audio classification docs to Chinese", "action": "closed", "comment": "Thanks a lot for the translation @v6582374-netizen! As @stevhliu explained (https://github.com/huggingface/transformers/pull/48893#issuecomment-5716181323), translations are moving to an automated doc-builder feature, so closing this one as suggested.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48893"}, {"key": "pr#48891", "n": 48891, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "Add SGLang compatibility tests to Transformers CI", "action": "\u2014", "comment": "", "author": "adarshxs", "idle_days": 33, "models": "", "url": "https://github.com/huggingface/transformers/pull/48891"}, {"key": "pr#48890", "n": 48890, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Feat: Add Kimi K3 architecture support", "action": "review requested: @Cyrilvallez", "comment": "", "author": "priyanshu0463", "idle_days": -1, "models": "kimi_linear", "url": "https://github.com/huggingface/transformers/pull/48890"}, {"key": "issue#48884", "n": 48884, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "Please remove dependency on the unmantained module pyctcdecode: it wasn't updated for years and is still at numpy-1.x", "action": "labelled: Feature request", "comment": "", "author": "yurivict", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/48884"}, {"key": "pr#48880", "n": 48880, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "feature / refactor", "title": "[`MLA`] Add native FA4 support", "action": "\u2014", "comment": "", "author": "vasqu", "idle_days": 32, "models": "axk1, axk2, deepseek_v2, deepseek_v3, deepseek_v32, glm4_moe_lite, glm5_next, glm_moe_dsa, hy_v4, kimi_linear, longcat_flash, minicpm3, mistral4, youtu", "url": "https://github.com/huggingface/transformers/pull/48880"}, {"key": "pr#48871", "n": 48871, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "fix", "title": "Fix `base_model_prefix` in the PP-OCR detectors, CLIPSeg, TimesFM, CLIP and VideoPrism heads so base checkpoints load", "action": "\u2014", "comment": "", "author": "blipbyte", "idle_days": -1, "models": "clip, clipseg, metaclip_2, pp_ocrv5_mobile_det, pp_ocrv5_server_det, pp_ocrv6_medium_det, pp_ocrv6_small_det, timesfm, uvdoc, videoprism", "url": "https://github.com/huggingface/transformers/pull/48871"}, {"key": "pr#48870", "n": 48870, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "Add native CogVLM2-Video support", "action": "closed", "comment": "Thanks for the CogVLM2-Video port! As @zucchini-nlp explained (https://github.com/huggingface/transformers/pull/48870#pullrequestreview-5224380441), we're not adding CogVLM2 to the core library given its maintenance cost; sharing it on the Hub with `trust_remote_code` is the best way to make it available. Closing.\n\nUnless the quality / simplicity of the code is majorely improved!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48870"}, {"key": "pr#48863", "n": 48863, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "\ud83d\udea8Allow training for DS3.2 lightning indexer", "action": "\u2014", "comment": "", "author": "Rocketknight1", "idle_days": 0, "models": "axk2, deepseek_v32, glm_moe_dsa, hy_v4, longcat_flash", "url": "https://github.com/huggingface/transformers/pull/48863"}, {"key": "issue#48862", "n": 48862, "kind": "issue", "group": "Open issues", "category": "bug", "area": "models", "type": "bug", "title": "CLIPSeg, TimesFM, PP-OCR detector and VideoPrism base/head loads still give every weight randomly initialized (#48722 follow-up)", "action": "\u2014", "comment": "", "author": "blipbyte", "idle_days": 3, "models": "clipseg, timesfm", "url": "https://github.com/huggingface/transformers/issues/48862"}, {"key": "pr#48855", "n": 48855, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "[`distributed`] Use unit sharding for attention projections", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 34, "models": "apertus, arcee, aria, cohere, cohere2, cohere2_moe, cohere_compass, cosmos3_edge, cwm, deepseek_ocr2, diffusion_gemma, doge, dots1, ernie4_5, ernie4_5_moe, ernie4_5_vl_moe, esmc, eurobert, exaone4, exaone_moe, gemma, gemma2, gemma3, gemma3n, gemma4, gemma4_unified, glm, glm4, glm4_moe, glm4v, glm4v_moe, glm_image, glm_ocr, granite, granite4_vision, granite_swa, granitemoe_swa, helium, higgs_audio_v2, hy_v3", "url": "https://github.com/huggingface/transformers/pull/48855"}, {"key": "issue#48851", "n": 48851, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Need Kimi K3 model architecture support", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [moonshotai/Kimi-K3](https://huggingface.co/moonshotai/Kimi-K3)\n- Downloads (last 30d): 998,301 \u00b7 Likes: 11,777 \u00b7 Params: ~2.8T \u00b7 License: other (modified MIT)\n- Released: 2026-06 \u00b7 Remote code on the Hub: yes\n- Open PR: #48890\n- The `khtsly/Luau-Coder-*` checkpoints linked in the issue declare `model_type: kimi_linear`, which is already supported natively as `kimi_linear`.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "imezx", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/48851"}, {"key": "pr#48848", "n": 48848, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "distributed", "type": "feature / refactor", "title": "[`distributed`] Add a new parallel plan for unit sharding", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 8, "models": "", "url": "https://github.com/huggingface/transformers/pull/48848"}, {"key": "pr#48844", "n": 48844, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "Add CODEOWNERS for core loading, modeling and generation files", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 34, "models": "", "url": "https://github.com/huggingface/transformers/pull/48844"}, {"key": "pr#48843", "n": 48843, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "attention / kernels", "type": "fix", "title": "Fix NPU SDPA path silently discarding additive attention bias masks", "action": "review requested: @IlyasMoutawwakil; maintainer pinged", "comment": "@IlyasMoutawwakil could you take a look when you get a chance? This is waiting on a maintainer: IlyasMoutawwakil to evaluate the author's no-NPU-needed correctness argument (no Ascend hardware internally).", "author": "junli-gong", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/48843"}, {"key": "issue#48842", "n": 48842, "kind": "issue", "group": "Open issues", "category": "bug", "area": "trainer", "type": "bug", "title": "Context parallelism: Trainer.predict returns duplicated rows, unlabeled path skips the CP context", "action": "labelled: bug", "comment": "", "author": "qgallouedec", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/48842"}, {"key": "issue#48841", "n": 48841, "kind": "issue", "group": "Open issues", "category": "bug", "area": "trainer", "type": "bug", "title": "FSDP2: evaluate/predict on a fresh Trainer raises in Accelerator.prepare", "action": "labelled: bug", "comment": "", "author": "qgallouedec", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/48841"}, {"key": "pr#48840", "n": 48840, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "Fix skip reason for MoE padding-free test (gpt_oss, longcat_flash)", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 1, "models": "gpt_oss, longcat_flash", "url": "https://github.com/huggingface/transformers/pull/48840"}, {"key": "pr#48828", "n": 48828, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add native ZGCM model support", "action": "\u2014", "comment": "", "author": "relic-yuexi", "idle_days": 34, "models": "zgcm", "url": "https://github.com/huggingface/transformers/pull/48828"}, {"key": "issue#48826", "n": 48826, "kind": "issue", "group": "Open issues", "category": "bug", "area": "models", "type": "bug", "title": "[Qwen3.8-flash-next] QSAIndexer can not train", "action": "\u2014", "comment": "", "author": "cxiaolong", "idle_days": 14, "models": "qwen3", "url": "https://github.com/huggingface/transformers/issues/48826"}, {"key": "pr#48825", "n": 48825, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "feature / refactor", "title": "logits_to_keep bool mask for all ForXXX classes", "action": "\u2014", "comment": "", "author": "jp1924", "idle_days": -1, "models": "afmoe, apertus, arcee, aria, audioflamingo3, axk1, axk2, aya_vision, bailing_hybrid, bamba, bart, beam, bert, bert_generation, big_bird, bigbird_pegasus, biogpt, bitnet, blenderbot, blenderbot_small, blip, bloom, blt, camembert, chameleon, codegen, cohere, cohere2, cohere2_moe, cohere2_vision, cohere_compass, cosmos3_edge, cosmos3_omni, cpmant, csm, ctrl, cwm, data2vec, dbrx, deepseek_ocr2, deepseek_v2, deepseek_v3", "url": "https://github.com/huggingface/transformers/pull/48825"}, {"key": "pr#48822", "n": 48822, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "trainer", "type": "fix", "title": "fix(trainer): Apply weight conversions when resuming from a checkpoint", "action": "\u2014", "comment": "", "author": "harshaljanjani", "idle_days": 2, "models": "", "url": "https://github.com/huggingface/transformers/pull/48822"}, {"key": "issue#48818", "n": 48818, "kind": "issue", "group": "Open issues", "category": "bug", "area": "cache / masks", "type": "bug", "title": "Gemma4 audio tower: attention mask inverted under eager attention (uniform attention regardless of content)", "action": "labelled: bug", "comment": "", "author": "dltsum", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/48818"}, {"key": "pr#48808", "n": 48808, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "fix", "title": "[serge] Fix 6 integration tests for model `layoutlmv2` failing with `other` (other (6))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 4, "models": "layoutlmv2", "url": "https://github.com/huggingface/transformers/pull/48808"}, {"key": "issue#48805", "n": 48805, "kind": "issue", "group": "Open issues", "category": "bug", "area": "core / loading", "type": "bug", "title": "Only on saving a checkpoint are weight conversions used", "action": "\u2014", "comment": "", "author": "PashkovD", "idle_days": 5, "models": "", "url": "https://github.com/huggingface/transformers/issues/48805"}, {"key": "pr#48801", "n": 48801, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "[Gemma Quant] Preserve Gemma 4 MoE router projection in full precision during quantization", "action": "\u2014", "comment": "", "author": "pculliton", "idle_days": 35, "models": "gemma4", "url": "https://github.com/huggingface/transformers/pull/48801"}, {"key": "issue#48791", "n": 48791, "kind": "issue", "group": "Open issues", "category": "question / other", "area": "other", "type": "question / other", "title": "Reporting a vulnerability", "action": "\u2014", "comment": "", "author": "S3DFX-CYBER", "idle_days": 5, "models": "", "url": "https://github.com/huggingface/transformers/issues/48791"}, {"key": "pr#48790", "n": 48790, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "docs", "type": "docs", "title": "Audio processors in processors call", "action": "\u2014", "comment": "", "author": "eustlb", "idle_days": 21, "models": "afmoe, altclip, aria, axk1, axk2", "url": "https://github.com/huggingface/transformers/pull/48790"}, {"key": "issue#48784", "n": 48784, "kind": "issue", "group": "Open issues", "category": "feature", "area": "generation", "type": "feature", "title": "Support 2D bool mask for logits_to_keep, expose last_hidden_state", "action": "\u2014", "comment": "", "author": "jp1924", "idle_days": 32, "models": "", "url": "https://github.com/huggingface/transformers/issues/48784"}, {"key": "pr#48783", "n": 48783, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add tensor parallel support for GPT-J", "action": "closed", "comment": "Thanks for this! Main has moved under you: embedding sharding for TP was reworked in #49570 (`distributed/mixin.py` now adds `embedding_rowwise` for the input embeddings by default), so the `configuration_utils.py` / `mixin.py` changes here conflict and are mostly no longer needed. The GPT-J `base_model_tp_plan` itself is still missing on main, though. Closing this one for now; feel free to open a fresh PR against `main` if it's still needed. Thanks!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48783"}, {"key": "pr#48782", "n": 48782, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add tensor parallel support for GPT-Neo", "action": "closed", "comment": "Thanks for this! #49570 just reworked TP embeddings (input embeddings are sharded by default in distributed/mixin.py) and handling for widths that don't divide by the TP size, so the configuration_utils/mixin changes here no longer apply and the PR conflicts. Closing this one for now; feel free to open a fresh PR against `main` if it's still needed. Thanks!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48782"}, {"key": "pr#48781", "n": 48781, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add tensor parallel support for OPT", "action": "review requested: @3outeille; maintainer pinged", "comment": "@3outeille could you take a look when you get a chance? This is waiting on a maintainer: it has been waiting for a maintainer review.", "author": "Siddhant231xyz", "idle_days": -1, "models": "opt", "url": "https://github.com/huggingface/transformers/pull/48781"}, {"key": "pr#48780", "n": 48780, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add tensor parallel plan for Nemotron", "action": "review requested: @Cyrilvallez", "comment": "", "author": "Siddhant231xyz", "idle_days": -1, "models": "nemotron", "url": "https://github.com/huggingface/transformers/pull/48780"}, {"key": "pr#48779", "n": 48779, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Key the tied-embedding TP rule on the model's input embedding attribute", "action": "closed", "comment": "Thanks! #49570 just reworked this: the tied-embedding injection was removed from `configuration_utils.py` and the input-embedding rowwise rule now lives in `distributed/mixin.py` (around line 96), still keyed on the `embed_tokens` attribute name. Closing this one for now; feel free to open a fresh PR against `main` if it's still needed. Thanks!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48779"}, {"key": "pr#48775", "n": 48775, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "docs: fix prompt formatting and code alignment in Gemma model doc", "action": "review requested: @stevhliu", "comment": "", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48775"}, {"key": "pr#48774", "n": 48774, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "docs: fix grammar typo in ALBERT model card", "action": "review requested: @stevhliu", "comment": "", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48774"}, {"key": "pr#48773", "n": 48773, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "fix(wavlm): replace fused F.mha path with explicit submodule calls to\u2026", "action": "review requested: @Cyrilvallez", "comment": "", "author": "p-doshi", "idle_days": -1, "models": "wavlm", "url": "https://github.com/huggingface/transformers/pull/48773"}, {"key": "pr#48772", "n": 48772, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "docs", "type": "docs", "title": "[i18n-TR] Translate run_scripts.md to Turkish", "action": "\u2014", "comment": "", "author": "yuefdev", "idle_days": 20, "models": "", "url": "https://github.com/huggingface/transformers/pull/48772"}, {"key": "pr#48770", "n": 48770, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix DogeForCausalLM attending to future tokens under sdpa", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #48821, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48770"}, {"key": "pr#48765", "n": 48765, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "tests", "type": "CI / tooling", "title": "[DSA tests] Remove stale skips in GLM-5 and DeepSeek-V3.2 tests", "action": "\u2014", "comment": "", "author": "malaiwah", "idle_days": 36, "models": "deepseek_v3, deepseek_v32", "url": "https://github.com/huggingface/transformers/pull/48765"}, {"key": "pr#48759", "n": 48759, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "[FP8] Rewrite TP / EP plans on the text config of composite models", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #49548, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48759"}, {"key": "pr#48758", "n": 48758, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "distributed", "type": "feature / refactor", "title": "Exclude _no_placement_params from get_balanced_memory's per-device budget", "action": "review requested: @SunMarc", "comment": "", "author": "malaiwah", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/48758"}, {"key": "issue#48756", "n": 48756, "kind": "issue", "group": "Open issues", "category": "bug", "area": "distributed", "type": "bug", "title": "get_balanced_memory ignores _no_placement_params, inflating every device's budget (device_map=\"auto\" OOM)", "action": "labelled: bug", "comment": "", "author": "malaiwah", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/48756"}, {"key": "issue#48755", "n": 48755, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add DeepSeek-V4.1-Flash (deepseek_v41)", "action": "labelled: New model", "comment": "", "author": "malaiwah", "idle_days": -1, "models": "deepseek_v4", "url": "https://github.com/huggingface/transformers/issues/48755"}, {"key": "pr#48751", "n": 48751, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "feature / refactor", "title": "Fuse the MoE top-k router into one kernel, served from the Hub", "action": "\u2014", "comment": "", "author": "qgallouedec", "idle_days": 37, "models": "qwen2_moe, qwen3_moe", "url": "https://github.com/huggingface/transformers/pull/48751"}, {"key": "pr#48746", "n": 48746, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "other", "type": "feature / refactor", "title": "[Docs] Add Marathi (\u092e\u0930\u093e\u0920\u0940) translation of README", "action": "review requested: @stevhliu", "comment": "", "author": "shreeyadav06", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/48746"}, {"key": "issue#48736", "n": 48736, "kind": "issue", "group": "Open issues", "category": "question / other", "area": "processors (image/video/audio)", "type": "question / other", "title": "Follow-up: clean up video_duration handling in hyperclovax_vision_v2 (#44314)", "action": "\u2014", "comment": "", "author": "jp1924", "idle_days": 2, "models": "hyperclovax_vision_v2", "url": "https://github.com/huggingface/transformers/issues/48736"}, {"key": "pr#48734", "n": 48734, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "new model", "title": "Add sliding-window local attention to Parakeet for long-form audio", "action": "review requested: @ydshieh; maintainer pinged", "comment": "@ydshieh could you take a look when you get a chance? This is waiting on a maintainer: never reviewed.", "author": "penguinwang96825", "idle_days": -1, "models": "granite_speech5, lasr, nemotron_asr_streaming, parakeet", "url": "https://github.com/huggingface/transformers/pull/48734"}, {"key": "pr#48733", "n": 48733, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "Fix Parakeet TDT conversion for nvidia/parakeet-tdt-1.1b (no <pad> in the NeMo vocab)", "action": "review requested: @Cyrilvallez", "comment": "", "author": "penguinwang96825", "idle_days": -1, "models": "parakeet", "url": "https://github.com/huggingface/transformers/pull/48733"}, {"key": "pr#48731", "n": 48731, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "docs: fix 'allows to' grammar in LiLT docs", "action": "review requested: @stevhliu", "comment": "", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48731"}, {"key": "pr#48730", "n": 48730, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "docs: fix 'allows to' grammar in Vision Encoder Decoder docs", "action": "review requested: @stevhliu", "comment": "", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48730"}, {"key": "pr#48729", "n": 48729, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "docs: remove duplicate 'and' in docs README", "action": "closed", "comment": "Thanks @mrchatam! Per our contribution guidelines (`.ai/AGENTS.md`: no one-off PRs for tiny edits such as a single typo), I'm closing this; it can be folded into a broader docs pass.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48729"}, {"key": "pr#48728", "n": 48728, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "docs: fix 'allows to' grammar in DPT docs", "action": "review requested: @stevhliu", "comment": "", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48728"}, {"key": "pr#48727", "n": 48727, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "docs: fix 'allows to' grammar in Speech Encoder Decoder docs", "action": "review requested: @stevhliu", "comment": "", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48727"}, {"key": "pr#48726", "n": 48726, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "docs: fix 'allows to' grammar in Depth Anything V2 docs", "action": "closed", "comment": "Thanks @mrchatam! Per `CONTRIBUTING.md` on main (\"Avoid one-off busywork PRs (single typo, isolated style cleanup, ...)\"), we bundle typo-only fixes into broader cleanups rather than merging them one by one, so I'm closing this.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48726"}, {"key": "issue#48718", "n": 48718, "kind": "issue", "group": "Open issues", "category": "bug", "area": "trainer", "type": "bug", "title": "Qwen3.5-9B fine-tuning with run_clm.py on 8x B200 ran the gated DeltaNet layers on the fp32 torch chunk loop because flash-linear-attention was not installed", "action": "labelled: bug", "comment": "", "author": "TarzanZhao", "idle_days": -1, "models": "qwen3, qwen3_5", "url": "https://github.com/huggingface/transformers/issues/48718"}, {"key": "pr#48713", "n": 48713, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "trainer", "type": "feature / refactor", "title": "Support mass-preserving label smoothing for multi-label classification", "action": "review requested: @SunMarc", "comment": "", "author": "nandan2003", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/48713"}, {"key": "pr#48711", "n": 48711, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "feature / refactor", "title": "[generate] modularize generate", "action": "\u2014", "comment": "", "author": "eustlb", "idle_days": 36, "models": "csm, dia, higgs_audio_v2, nemotron3_5_asr, nemotron_asr_streaming, parakeet, vibevoice", "url": "https://github.com/huggingface/transformers/pull/48711"}, {"key": "pr#48710", "n": 48710, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "quantization", "type": "fix", "title": "Fix torchao MoE custom-parameter quantization during load", "action": "\u2014", "comment": "", "author": "eryk-roch", "idle_days": 22, "models": "", "url": "https://github.com/huggingface/transformers/pull/48710"}, {"key": "pr#48707", "n": 48707, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "CI / tooling", "type": "CI / tooling", "title": "Key the doc build concurrency group on the pull request number", "action": "\u2014", "comment": "", "author": "albertvillanova", "idle_days": 39, "models": "", "url": "https://github.com/huggingface/transformers/pull/48707"}, {"key": "pr#48706", "n": 48706, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "trainer", "type": "fix", "title": "Fix Qwen3-MoE auxiliary loss being multiplied by DDP world size in Trainer", "action": "review requested: @SunMarc; maintainer pinged", "comment": "@SunMarc could you take a look when you get a chance? This is waiting on a maintainer: SunMarc review + choose between competing PRs #48691 and #48706.", "author": "devtyagi3909", "idle_days": -1, "models": "qwen3_moe", "url": "https://github.com/huggingface/transformers/pull/48706"}, {"key": "pr#48696", "n": 48696, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Add files via uploAdd SCQOS governed model-transition proof for Transformersad", "action": "closed", "comment": "Thanks for the PR, but this page isn't documentation for Transformers itself, so it doesn't belong in `docs/source/en`. Closing; a Hub model card or your own repository would be a better place for it.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48696"}, {"key": "pr#48694", "n": 48694, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "[ImageGPT] Fix cross-attention crashing, dropping the encoder mask, and doubling its cache", "action": "\u2014", "comment": "", "author": "blipbyte", "idle_days": 33, "models": "imagegpt", "url": "https://github.com/huggingface/transformers/pull/48694"}, {"key": "pr#48691", "n": 48691, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "trainer", "type": "fix", "title": "[Bug] Preserve MoE router auxiliary loss weighting in Trainer DDP", "action": "\u2014", "comment": "", "author": "gss10282025", "idle_days": 40, "models": "", "url": "https://github.com/huggingface/transformers/pull/48691"}, {"key": "pr#48686", "n": 48686, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Support converting the unified Parakeet checkpoints (nvidia/parakeet-unified-en-0.6b)", "action": "review requested: @Cyrilvallez", "comment": "", "author": "penguinwang96825", "idle_days": -1, "models": "parakeet", "url": "https://github.com/huggingface/transformers/pull/48686"}, {"key": "pr#48682", "n": 48682, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "fix: respect partial_rotary_factor in GPTNeoXJapaneseRotaryEmbedding", "action": "closed", "comment": "Thanks @truongsontung! The same `partial_rotary_factor` fix for GPTNeoXJapanese landed in #48652, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48682"}, {"key": "pr#48677", "n": 48677, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "docs", "type": "docs", "title": "\ud83d\udea8Add an FP8 backend to `QuantizedCache`", "action": "\u2014", "comment": "", "author": "jiqing-feng", "idle_days": -1, "models": "vibevoice", "url": "https://github.com/huggingface/transformers/pull/48677"}, {"key": "pr#48673", "n": 48673, "kind": "PR", "group": "Closed PRs", "category": "duplicate", "area": "", "type": "", "title": "fix: use sp_model.decode in SentencePieceBackend.convert_tokens_to_string", "action": "closed", "comment": "Thanks for looking into this! @itazap is handling byte-fallback decoding for the SentencePiece backend in #48936 (same file, issue #47473). Decoding each `<0xNN>` token with `chr()` on its own also breaks multi-byte UTF-8 characters (e.g. emoji or CJK split across several byte tokens), which #48936 avoids. Closing in favour of that PR. Feel free to review there if a case is missing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48673"}, {"key": "pr#48666", "n": 48666, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "[Gemma Quant] Fused Triton low-bit kernels (INT2/4/8) for Gemma models & Fused Grouped GEMM for MoE experts", "action": "closed", "comment": "Thanks for this work! We're unifying low-bit kernels through the kernels hub (kernels-community) rather than adding model-specific quantization integrations, so we're closing this; contributions there are very welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48666"}, {"key": "pr#48662", "n": 48662, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "feature / refactor", "title": "[Kernels] Enable rotary_pos_emb hub kernel dispatch for partial RoPE models (GPT-NeoX, Qwen3-Next)", "action": "\u2014", "comment": "", "author": "piyushumate", "idle_days": 40, "models": "gpt_neo, gpt_neox, qwen3_next", "url": "https://github.com/huggingface/transformers/pull/48662"}, {"key": "pr#48642", "n": 48642, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "[generate] Opt-in normalized repetition penalty", "action": "closed", "comment": "Thanks for the PR! We're trying hard to avoid new `GenerationConfig` flags, and new sampling variants are best shared as a Hub custom generation method or custom logits processor. Closing here.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48642"}, {"key": "pr#48623", "n": 48623, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "Fix Whisper beam_indices so compute_transition_scores does not select the beam twice", "action": "\u2014", "comment": "", "author": "speedsharmaai", "idle_days": 8, "models": "whisper", "url": "https://github.com/huggingface/transformers/pull/48623"}, {"key": "issue#48622", "n": 48622, "kind": "issue", "group": "Open issues", "category": "feature", "area": "attention / kernels", "type": "feature", "title": "Kernels for partial RoPE", "action": "\u2014", "comment": "", "author": "zucchini-nlp", "idle_days": 41, "models": "", "url": "https://github.com/huggingface/transformers/issues/48622"}, {"key": "pr#48617", "n": 48617, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "distributed", "type": "fix", "title": "Trainer fix for DTensor API", "action": "\u2014", "comment": "", "author": "michaelbenayoun", "idle_days": 4, "models": "", "url": "https://github.com/huggingface/transformers/pull/48617"}, {"key": "issue#48608", "n": 48608, "kind": "issue", "group": "Open issues", "category": "feature", "area": "quantization", "type": "feature", "title": "[Feature request] FP8 KV cache + flash-attention: is this in scope?", "action": "\u2014", "comment": "", "author": "jiqing-feng", "idle_days": 39, "models": "", "url": "https://github.com/huggingface/transformers/issues/48608"}, {"key": "pr#48604", "n": 48604, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "distributed", "type": "feature / refactor", "title": "Prefetch only the byte spans each rank will read", "action": "\u2014", "comment": "", "author": "qgallouedec", "idle_days": 18, "models": "", "url": "https://github.com/huggingface/transformers/pull/48604"}, {"key": "pr#48599", "n": 48599, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "fix", "title": "[WIP] [MiniCPMV4_6] Fix vit_merger crash when video frames are sliced (non-uniform target_sizes)", "action": "\u2014", "comment": "", "author": "ydshieh", "idle_days": 42, "models": "minicpmv4_6", "url": "https://github.com/huggingface/transformers/pull/48599"}, {"key": "pr#48583", "n": 48583, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add image-text contrastive loss to Blip2ForImageTextRetrieval", "action": "review requested: @ydshieh; maintainer pinged", "comment": "@ydshieh could you take a look when you get a chance? This is waiting on a maintainer: never reviewed.", "author": "xquantize", "idle_days": -1, "models": "blip_2", "url": "https://github.com/huggingface/transformers/pull/48583"}, {"key": "pr#48577", "n": 48577, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "integrations (other)", "type": "fix", "title": "fix(integrations): print the base model architecture for PEFT-wrapped models", "action": "review requested: @SunMarc", "comment": "", "author": "simpleqt", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/48577"}, {"key": "pr#48569", "n": 48569, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "distributed", "type": "feature / refactor", "title": "Trainer: make checkpoints work under tensor parallelism", "action": "\u2014", "comment": "", "author": "qgallouedec", "idle_days": 35, "models": "", "url": "https://github.com/huggingface/transformers/pull/48569"}, {"key": "pr#48566", "n": 48566, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "serving / continuous batching", "type": "feature / refactor", "title": "[CB] Scale the per-token memory estimate by the TP size", "action": "\u2014", "comment": "", "author": "qgallouedec", "idle_days": 14, "models": "", "url": "https://github.com/huggingface/transformers/pull/48566"}, {"key": "pr#48554", "n": 48554, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Normalize repository text file line endings", "action": "closed", "comment": "Thanks @ZikrieHR. Line-ending-only changes across test configs don't fix a concrete issue for us, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48554"}, {"key": "pr#48553", "n": 48553, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 2 integration tests for model `seamless_m4t_v2` failing with `output_mismatch` (list output differs (2))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 19, "models": "seamless_m4t, seamless_m4t_v2", "url": "https://github.com/huggingface/transformers/pull/48553"}, {"key": "pr#48549", "n": 48549, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / utils", "type": "feature / refactor", "title": "feat(hf_argparser): support dict-typed dataclass fields from CLI (#48030)", "action": "\u2014", "comment": "", "author": "mohmedmm", "idle_days": 18, "models": "", "url": "https://github.com/huggingface/transformers/pull/48549"}, {"key": "issue#48548", "n": 48548, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "[RFC] Zero-Copy In-Place MoE Dynamic Token Routing for Mixtral/DeepSeek (1.77x Speedup, 0 Bytes Aux VRAM)", "action": "\u2014", "comment": "", "author": "aemre-cetin", "idle_days": 39, "models": "", "url": "https://github.com/huggingface/transformers/issues/48548"}, {"key": "pr#48545", "n": 48545, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix LukeTokenizer dropping <s>/</s> on the plain-text path", "action": "closed", "comment": "Thanks @Mahnoor-Zaffar! LukeTokenizer now sets a `RobertaProcessing` post-processor on main (#48408), which fixes #48544, so closing. Please reopen if it still reproduces.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48545"}, {"key": "pr#48543", "n": 48543, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "tests", "type": "CI / tooling", "title": "tests: skip the symlinked hub-cache test when symlinks are unavailable", "action": "\u2014", "comment": "", "author": "iam-kira", "idle_days": 12, "models": "", "url": "https://github.com/huggingface/transformers/pull/48543"}, {"key": "pr#48542", "n": 48542, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tokenization", "type": "fix", "title": "Fix GPT-NeoX family tokenization when config flags and serialized post-processor disagree", "action": "review requested: @ydshieh", "comment": "", "author": "PranshulSoni", "idle_days": -1, "models": "gpt_neo, gpt_neox", "url": "https://github.com/huggingface/transformers/pull/48542"}, {"key": "issue#48541", "n": 48541, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Feature request: add NEAT as an optional Trainer optimizer integration", "action": "\u2014", "comment": "", "author": "ItCodinTime", "idle_days": 45, "models": "", "url": "https://github.com/huggingface/transformers/issues/48541"}, {"key": "issue#48527", "n": 48527, "kind": "issue", "group": "Open issues", "category": "bug", "area": "core / loading", "type": "bug", "title": "test_get_cached_module_file_local_handles_symlinked_hub_cache fails on Windows without symlink privilege", "action": "\u2014", "comment": "", "author": "iam-kira", "idle_days": 7, "models": "", "url": "https://github.com/huggingface/transformers/issues/48527"}, {"key": "pr#48525", "n": 48525, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "fix", "title": "fix(audio): Route the ALM conversion scripts through from_pretrained", "action": "review requested: @vasqu; maintainer pinged", "comment": "@vasqu this was approved by @eustlb on 2026-09-07; could you give it the core-maintainer review so it can be merged?", "author": "harshaljanjani", "idle_days": -1, "models": "audioflamingo3, glmasr, musicflamingo, vibevoice_asr, voxtral, voxtral_realtime", "url": "https://github.com/huggingface/transformers/pull/48525"}, {"key": "issue#48512", "n": 48512, "kind": "issue", "group": "Open issues", "category": "question / other", "area": "core / loading", "type": "question / other", "title": "Decide whether the checkpoint shard prefetch should be on by default", "action": "\u2014", "comment": "", "author": "qgallouedec", "idle_days": 14, "models": "", "url": "https://github.com/huggingface/transformers/issues/48512"}, {"key": "pr#48494", "n": 48494, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / utils", "type": "fix", "title": "Fix HfArgumentParser mutating dataclass annotation types", "action": "review requested: @SunMarc; maintainer pinged", "comment": "@SunMarc could you take a look when you get a chance? This is waiting on a maintainer: it has been waiting for a maintainer review.", "author": "urbanyl", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/48494"}, {"key": "pr#48483", "n": 48483, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "fix: Handle tuple and set types in HfArgumentParser", "action": "closed", "comment": "Thanks @dajiaohuang! This duplicates #48432, which fixes the same issue (#48419) and includes tests, so let's consolidate there; closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48483"}, {"key": "pr#48480", "n": 48480, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "docs(zh): add text classification quickstart section", "action": "review requested: @stevhliu", "comment": "", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48480"}, {"key": "pr#48472", "n": 48472, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "[Parakeet] Isolate concurrent TDT generate state", "action": "review requested: @ydshieh", "comment": "", "author": "lorenzozanee", "idle_days": -1, "models": "parakeet", "url": "https://github.com/huggingface/transformers/pull/48472"}, {"key": "pr#48471", "n": 48471, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "[i18n-KO] Translate models_timeline to Korean", "action": "review requested: @stevhliu", "comment": "", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48471"}, {"key": "pr#48470", "n": 48470, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Expose `use_associative_scan` in ZambaConfig", "action": "closed", "comment": "Thanks @lorenzozanee! `use_associative_scan` was added to `ZambaConfig` in #48331, so closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48470"}, {"key": "pr#48468", "n": 48468, "kind": "PR", "group": "Open PRs", "category": "changes requested", "area": "serving / continuous batching", "type": "feature / refactor", "title": "Support minimum-length logits processing in continuous batching", "action": "\u2014", "comment": "", "author": "Robertboy18", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/pull/48468"}, {"key": "pr#48461", "n": 48461, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "docs", "type": "docs", "title": "act offload with full overlap", "action": "\u2014", "comment": "", "author": "stas00", "idle_days": 19, "models": "", "url": "https://github.com/huggingface/transformers/pull/48461"}, {"key": "pr#48453", "n": 48453, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "[i18n-es] Translate token_classification.md to Spanish", "action": "review requested: @stevhliu", "comment": "", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48453"}, {"key": "pr#48452", "n": 48452, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "[i18n-es] Translate custom_tokenizers.md to Spanish", "action": "review requested: @stevhliu", "comment": "", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48452"}, {"key": "pr#48451", "n": 48451, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "CI / tooling", "type": "CI / tooling", "title": "CI: improve gated repo failures flow", "action": "\u2014", "comment": "", "author": "tarekziade", "idle_days": 34, "models": "", "url": "https://github.com/huggingface/transformers/pull/48451"}, {"key": "pr#48432", "n": 48432, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / utils", "type": "fix", "title": "fix(hf_argparser): support tuple and set typed dataclass fields", "action": "review requested: @ydshieh; maintainer pinged", "comment": "@ydshieh could you take a look when you get a chance? This is waiting on a maintainer: never reviewed.", "author": "shenhuaqingshi", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/48432"}, {"key": "pr#48418", "n": 48418, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "avoid repeated vision attention synchronization", "action": "\u2014", "comment": "", "author": "Marchematics", "idle_days": 49, "models": "cohere_compass, ernie4_5_vl_moe, glm4v, glm4v_moe, qwen2_5_vl, qwen2_vl, qwen3_5, qwen3_5_moe, qwen3_omni_moe, qwen3_vl, qwen3_vl_moe, qwen4_exp", "url": "https://github.com/huggingface/transformers/pull/48418"}, {"key": "pr#48417", "n": 48417, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Enable associative scan for Zamba", "action": "closed", "comment": "Thanks for the PR! The same change (a use_associative_scan config flag, wired into the Zamba mamba mixer) landed in #48331, so the Zamba part is already on main. The Japanese docs typo fix isn't related to this PR's topic, so feel free to send it separately as part of a broader docs cleanup. Closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48417"}, {"key": "pr#48412", "n": 48412, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "serving / continuous batching", "type": "fix", "title": "Fix transformers env when requests is not installed", "action": "review requested: @Cyrilvallez; maintainer pinged", "comment": "@Cyrilvallez this was approved by @qgallouedec on 2026-10-05; could you give it the core-maintainer review so it can be merged?", "author": "panshularora", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/48412"}, {"key": "pr#48398", "n": 48398, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Enforce TRF057 everywhere: drop its allowlist and cutoff date", "action": "\u2014", "comment": "", "author": "tarekziade", "idle_days": 50, "models": "afmoe, altclip, audioflamingo3, autoformer, bark, bart, big_bird, bigbird_pegasus, blenderbot, blenderbot_small, blip, blt, bridgetower, bros, clap, cohere_asr, cohere_compass, conditional_detr, convnext, convnextv2, cosmos3_omni, cwm, d_fine, dab_detr, dbrx, decision_transformer, deepseek_ocr2, deepseek_vl, deepseek_vl_hybrid, deformable_detr, deimv2, detr, diffusion_gemma, dinov2, dinov2_with_registers, dpr, edgetam, emu3, ernie4_5_vl_moe", "url": "https://github.com/huggingface/transformers/pull/48398"}, {"key": "pr#48397", "n": 48397, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "other", "type": "feature / refactor", "title": "Neuron distributed training validation", "action": "\u2014", "comment": "", "author": "michaelbenayoun", "idle_days": 42, "models": "", "url": "https://github.com/huggingface/transformers/pull/48397"}, {"key": "pr#48395", "n": 48395, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "Bump the actions group across 1 directory with 13 updates", "action": "\u2014", "comment": "", "author": "dependabot", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/pull/48395"}, {"key": "issue#48390", "n": 48390, "kind": "issue", "group": "Open issues", "category": "feature", "area": "generation", "type": "feature", "title": "Vectorized (batched) logits processors for assisted decoding", "action": "labelled: Feature request", "comment": "", "author": "VaggelisGian", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/48390"}, {"key": "pr#48389", "n": 48389, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "Fix flaky `MellumModelTest.test_load_balancing_loss`", "action": "review requested: @ydshieh", "comment": "", "author": "subin9", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/48389"}, {"key": "pr#48378", "n": 48378, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Modernize ConvNextV2, ResNet, RegNet, RTDetrResnet, SwinV2, Swin2SR", "action": "\u2014", "comment": "", "author": "sbucaille", "idle_days": 7, "models": "bit, convnextv2, efficientnet, mobilenet_v1, mobilenet_v2, mobilevit, mobilevitv2, poolformer, regnet, resnet, rt_detr, swin2sr, swinv2", "url": "https://github.com/huggingface/transformers/pull/48378"}, {"key": "pr#48373", "n": 48373, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "handle transpose during model loading with TP (Qwen3-VL-MoE)", "action": "\u2014", "comment": "", "author": "3outeille", "idle_days": 8, "models": "qwen3_vl, qwen3_vl_moe", "url": "https://github.com/huggingface/transformers/pull/48373"}, {"key": "issue#48371", "n": 48371, "kind": "issue", "group": "Open issues", "category": "bug", "area": "quantization", "type": "bug", "title": "NVFP4 MoE experts are dequantized without weight_global_scale, silently", "action": "labelled: bug", "comment": "", "author": "Bluear7878", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/48371"}, {"key": "issue#48369", "n": 48369, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "\u9002\u914d\u9e3f\u8499", "action": "\u2014", "comment": "", "author": "Hao-bigMan", "idle_days": 54, "models": "", "url": "https://github.com/huggingface/transformers/issues/48369"}, {"key": "pr#48362", "n": 48362, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Unpin the bert copy markers, and delete the ones CI never enforced", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 54, "models": "align, blip, blip_2, bridgetower, bros, clap, data2vec, deberta_v2, fnet, instructblip, layoutlm, layoutlmv2, layoutlmv3, lilt, longformer, luke, markuplm, maskformer, moshi, mra, nystromformer, roberta_prelayernorm, roc_bert, splinter, swin2sr, swinv2, videomae, xmod, yoso", "url": "https://github.com/huggingface/transformers/pull/48362"}, {"key": "pr#48361", "n": 48361, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "serving / continuous batching", "type": "feature / refactor", "title": "Add --experts-implementation flag to transformers serve", "action": "review requested: @remi-or; maintainer pinged", "comment": "@remi-or could you take a look when you get a chance? This is waiting on a maintainer: never reviewed.", "author": "NathanHB", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/48361"}, {"key": "pr#48360", "n": 48360, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Standardise head_dim onto the config and drop the guards that stood in for it", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 54, "models": "albert, align, altclip, bert, bert_generation, big_bird, bigbird_pegasus, blip, blip_2, bridgetower, bros, camembert, canine, chinese_clip, clap, convbert, data2vec, deberta, deberta_v2, dinov2, dinov2_with_registers, dpt, electra, ernie, esm, evolla, flava, gemma2, git, glpn, grounding_dino, hyperclovax, ibert, instructblip, instructblipvideo, layoutlm, layoutlmv2, layoutlmv3, led, lilt, longformer, luke, lxmert, markuplm, megatron_bert, mm_grounding_dino, mobilevit, modernbert, modernbert_decoder, mpnet, mra, nystromformer, omdet_turbo, pp_doclayout_v2", "url": "https://github.com/huggingface/transformers/pull/48360"}, {"key": "pr#48357", "n": 48357, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Modernize FSMT: shared attention interface and remove legacy fairseq code", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 54, "models": "fsmt", "url": "https://github.com/huggingface/transformers/pull/48357"}, {"key": "pr#48356", "n": 48356, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Shape cross-attention keys and values by the states they come from", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 54, "models": "esm, evolla, lightglue, megatron_bert, tapas", "url": "https://github.com/huggingface/transformers/pull/48356"}, {"key": "pr#48353", "n": 48353, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "attention / kernels", "type": "feature / refactor", "title": "Restore .item() on the packed-path max_seqlen (third site missed by #47134)", "action": "\u2014", "comment": "", "author": "foreverYoungGitHub", "idle_days": 49, "models": "", "url": "https://github.com/huggingface/transformers/pull/48353"}, {"key": "pr#48352", "n": 48352, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "quantization", "type": "fix", "title": "Fix fp8 dtype", "action": "\u2014", "comment": "", "author": "SunMarc", "idle_days": 54, "models": "", "url": "https://github.com/huggingface/transformers/pull/48352"}, {"key": "pr#48347", "n": 48347, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "exporters", "type": "feature / refactor", "title": "Branch aware export", "action": "\u2014", "comment": "", "author": "ebezzam", "idle_days": 54, "models": "", "url": "https://github.com/huggingface/transformers/pull/48347"}, {"key": "pr#48335", "n": 48335, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "[`Kernels`] Add swiglu and geglu MLP across models", "action": "\u2014", "comment": "", "author": "vasqu", "idle_days": 50, "models": "afmoe, aimv2, aria, axk1, axk2, bamba, bitnet, cohere, cohere2, cohere2_moe, cohere_compass, csm, cwm, deepseek_ocr2, deepseek_v2, deepseek_v3, deepseek_v32, deepseek_v4, deimv2, diffllama, diffusion_gemma, dinov3_vit, doge, dots1, emu3, ernie4_5, ernie4_5_moe, ernie4_5_vl_moe, esmc, eurobert, evolla, exaone4, exaone_moe, falcon_h1, flex_olmo, gemma, gemma2, gemma3, gemma3n, gemma4, gemma4_unified, glm4_moe, glm4_moe_lite, glm4v_moe, glm5_next, glm_moe_dsa, granite, granite4_vision", "url": "https://github.com/huggingface/transformers/pull/48335"}, {"key": "pr#48333", "n": 48333, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Use the shared attention interface in encoder-decoder attention blocks", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 54, "models": "dab_detr, encoder_decoder, grounding_dino, led, mask2former, mm_grounding_dino, mvp, omdet_turbo, oneformer, prophetnet, table_transformer, trocr, tvp, xglm, zoedepth", "url": "https://github.com/huggingface/transformers/pull/48333"}, {"key": "pr#48317", "n": 48317, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "feature / refactor", "title": "Byte-level tokenizer.json + Hub tokenizer_class LlamaTokenizer/Fast", "action": "\u2014", "comment": "", "author": "itazap", "idle_days": 13, "models": "", "url": "https://github.com/huggingface/transformers/pull/48317"}, {"key": "pr#48311", "n": 48311, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "fix", "title": "Fix Moshi output fields", "action": "\u2014", "comment": "", "author": "ryota-komatsu", "idle_days": 49, "models": "moshi", "url": "https://github.com/huggingface/transformers/pull/48311"}, {"key": "issue#48308", "n": 48308, "kind": "issue", "group": "Open issues", "category": "bug", "area": "distributed", "type": "bug", "title": "Native TP does not shard qwen3_vl_moe experts (weights fully replicated per rank)", "action": "labelled: bug", "comment": "", "author": "qgallouedec", "idle_days": -1, "models": "qwen3_vl, qwen3_vl_moe", "url": "https://github.com/huggingface/transformers/issues/48308"}, {"key": "issue#48307", "n": 48307, "kind": "issue", "group": "Open issues", "category": "bug", "area": "models", "type": "bug", "title": "Native TP has no KV-head replication: tp_size > num_key_value_heads fails mid-forward", "action": "labelled: bug", "comment": "", "author": "qgallouedec", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/48307"}, {"key": "issue#48306", "n": 48306, "kind": "issue", "group": "Open issues", "category": "bug", "area": "distributed", "type": "bug", "title": "Native TP breaks on Qwen3.5/3.6 hybrids: reshape error in linear-attention layers", "action": "labelled: bug", "comment": "", "author": "qgallouedec", "idle_days": -1, "models": "qwen3_5", "url": "https://github.com/huggingface/transformers/issues/48306"}, {"key": "issue#48305", "n": 48305, "kind": "issue", "group": "Open issues", "category": "bug", "area": "serving / continuous batching", "type": "bug", "title": "Continuous batching does not support hybrid linear-attention models (all Qwen3.5/3.6/3.8)", "action": "\u2014", "comment": "", "author": "qgallouedec", "idle_days": 12, "models": "qwen3_5", "url": "https://github.com/huggingface/transformers/issues/48305"}, {"key": "issue#48283", "n": 48283, "kind": "issue", "group": "Open issues", "category": "bug", "area": "serving / continuous batching", "type": "bug", "title": "CLI: transformers env fails with ModuleNotFoundError: No module named 'requests' when run in clean environment", "action": "\u2014", "comment": "", "author": "Mobocop", "idle_days": 6, "models": "", "url": "https://github.com/huggingface/transformers/issues/48283"}, {"key": "pr#48271", "n": 48271, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "feature / refactor", "title": "draft modular v2 + transformers format", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 55, "models": "align, altclip, audioflamingo3, bamba, bert, bert_generation, big_bird, blenderbot_small, blip, blip_2, bridgetower, bros, camembert, chinese_clip, clap, clvp, cpmant, dab_detr, data2vec, deberta, deberta_v2, deepseek_ocr2, distilbert, doge, donut, electra, ernie, esm, esmfold2, evolla, flaubert, flava, fnet, fsmt, gemma3, git, glm, glm4_moe, glm4_moe_lite, glmasr, glpn, grounding_dino", "url": "https://github.com/huggingface/transformers/pull/48271"}, {"key": "pr#48257", "n": 48257, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Add forward method to DeepseekV4RMSNorm class to fix potential dtype mismatch", "action": "closed", "comment": "Thanks @gameofdimension! As @Rocketknight1 noted, the same DeepSeek-V4 norm fix landed in #47486, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48257"}, {"key": "pr#48249", "n": 48249, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "new model", "title": "Add native LFM2-Audio support", "action": "\u2014", "comment": "", "author": "kadirnar", "idle_days": 18, "models": "lfm2_audio", "url": "https://github.com/huggingface/transformers/pull/48249"}, {"key": "issue#48243", "n": 48243, "kind": "issue", "group": "Open issues", "category": "feature", "area": "exporters", "type": "feature", "title": "Make the keypoint models exportable (SuperPoint, SuperGlue, LightGlue)", "action": "labelled: Feature request", "comment": "", "author": "Om-singhaI", "idle_days": -1, "models": "lightglue, superglue, superpoint", "url": "https://github.com/huggingface/transformers/issues/48243"}, {"key": "pr#48237", "n": 48237, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / loading", "type": "fix", "title": "Handle empty FSDP shards in sharded loading: crash and hang fix", "action": "\u2014", "comment": "", "author": "qgallouedec", "idle_days": 38, "models": "", "url": "https://github.com/huggingface/transformers/pull/48237"}, {"key": "pr#48234", "n": 48234, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "other", "type": "feature / refactor", "title": "Avoid shadowing pipeline function in README examples", "action": "\u2014", "comment": "", "author": "dhruvatr", "idle_days": 27, "models": "", "url": "https://github.com/huggingface/transformers/pull/48234"}, {"key": "pr#48227", "n": 48227, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "distributed", "type": "feature / refactor", "title": "Warm the page cache before sharded loading (opt-in via `HF_SHARD_PREFETCH`)", "action": "\u2014", "comment": "", "author": "qgallouedec", "idle_days": 18, "models": "", "url": "https://github.com/huggingface/transformers/pull/48227"}, {"key": "issue#48225", "n": 48225, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "[BUG] MLukeTokenizer emits `<s>` instead of `<ent>` on all entity tasks", "action": "closed", "comment": "This was fixed by #48408, which is now on main. Closing, but feel free to reopen if you still see it on the latest version!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/48225"}, {"key": "pr#48219", "n": 48219, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "Fix LlamaTokenizer overwriting pre_tokenizer and decoder from tokenizer.json", "action": "closed", "comment": "Thanks for the PR! As explained in https://github.com/huggingface/transformers/pull/48219#pullrequestreview-5005851873 and https://github.com/huggingface/transformers/pull/48219#issuecomment-5463396181, `LlamaTokenizer` intentionally defines the original Llama pre-tokenizer/decoder in v5, and generalising the mapping would break Qwen2/DeepSeek tokenizers. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48219"}, {"key": "pr#48215", "n": 48215, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "pipelines", "type": "fix", "title": "fix(pipelines): handle empty container inputs and validate batch_size in Pipeline.__call__", "action": "review requested: @Rocketknight1", "comment": "", "author": "Xayar145", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/48215"}, {"key": "pr#48214", "n": 48214, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "generation", "type": "fix", "title": "fix(generation): validate non-negative bounds for min_length, min_new_tokens, and temperature", "action": "review requested: @Cyrilvallez", "comment": "", "author": "Xayar145", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/48214"}, {"key": "pr#48213", "n": 48213, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tokenization", "type": "fix", "title": "fix(tokenization): prioritize slice indexing in BatchEncoding.__getitem__", "action": "review requested: @ydshieh", "comment": "", "author": "Xayar145", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/48213"}, {"key": "issue#48193", "n": 48193, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add native support for TR-HASH deterministic token-routed MoE", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [AETHORIA-AI/TR-HASH-MoE-200M-160B-SFT](https://huggingface.co/AETHORIA-AI/TR-HASH-MoE-200M-160B-SFT), [AETHORIA-AI/TR-HASH-MOE-500M-HF](https://huggingface.co/AETHORIA-AI/TR-HASH-MOE-500M-HF)\n- Downloads (last 30d): 252 \u00b7 Likes: 5 \u00b7 Params: ~203M \u00b7 License: cc-by-nc-4.0\n- Released: 2026-08 \u00b7 Remote code on the Hub: yes\n- Note: usage is still low and the license is non-commercial, so native support will depend on adoption; remote code works in the meantime.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "Complexity-ML", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/48193"}, {"key": "pr#48192", "n": 48192, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "Simplify group images by shape", "action": "\u2014", "comment": "", "author": "guarin", "idle_days": 14, "models": "beit, bridgetower, chmv2, cohere2_vision, cohere_compass, convnext, cosmos3_edge, deepseek_vl, deepseek_vl_hybrid, depth_pro, dinov3_vit, donut, dpt, efficientloftr, efficientnet, eomt, ernie4_5_vl_moe, flava, fuyu, gemma3, glm46v, glm4v, glm_image, glmga, glpn, got_ocr2, hunyuan_vl, idefics2, idefics3, imagegpt, janus, kimi_k25, kosmos2_5, layoutlmv2, layoutlmv3, lfm2_vl, lightglue, llama4, llava, llava_next, llava_onevision, mask2former, maskformer, minicpmv4_6, minimax_m3_vl, mllama, mobilenet_v2, mobilevit, muse_glimmer, nougat, oneformer, ovis2", "url": "https://github.com/huggingface/transformers/pull/48192"}, {"key": "pr#48158", "n": 48158, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "Deprecate backend=\"torchaudio\" for `load_audio`", "action": "\u2014", "comment": "", "author": "eustlb", "idle_days": 57, "models": "", "url": "https://github.com/huggingface/transformers/pull/48158"}, {"key": "pr#48151", "n": 48151, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "exporters", "type": "feature / refactor", "title": "[PoC] End to end vision exports", "action": "\u2014", "comment": "", "author": "guarin", "idle_days": 28, "models": "mask2former", "url": "https://github.com/huggingface/transformers/pull/48151"}, {"key": "pr#48149", "n": 48149, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "fix: make auto_docstring import-time performance test robust to timing noise.", "action": "review requested: @tarekziade; maintainer pinged", "comment": "@tarekziade could you take a look when you get a chance? This is waiting on a maintainer: needs a maintainer approval/merge.", "author": "deepresearcher08", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/48149"}, {"key": "pr#48146", "n": 48146, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "quantization", "type": "feature / refactor", "title": "Feat/laguna quantized checkpoints support", "action": "\u2014", "comment": "", "author": "lulmer", "idle_days": 21, "models": "", "url": "https://github.com/huggingface/transformers/pull/48146"}, {"key": "pr#48144", "n": 48144, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "QA: Fall back to the import name when a distribution is missing", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #48615, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48144"}, {"key": "pr#48143", "n": 48143, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Persist the Whisper spelling normalizer across save_pretrained", "action": "\u2014", "comment": "", "author": "Bluear7878", "idle_days": -1, "models": "whisper", "url": "https://github.com/huggingface/transformers/pull/48143"}, {"key": "pr#48137", "n": 48137, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "quantization", "type": "feature / refactor", "title": "[Quantization] Add MXFP4 support", "action": "\u2014", "comment": "", "author": "remi-or", "idle_days": 54, "models": "", "url": "https://github.com/huggingface/transformers/pull/48137"}, {"key": "pr#48126", "n": 48126, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "mtp support for setp3p7", "action": "\u2014", "comment": "", "author": "itazap", "idle_days": 1, "models": "step3p7", "url": "https://github.com/huggingface/transformers/pull/48126"}, {"key": "pr#48119", "n": 48119, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "new model", "title": "Add configurable Laguna router scoring", "action": "\u2014", "comment": "", "author": "lulmer", "idle_days": 49, "models": "laguna", "url": "https://github.com/huggingface/transformers/pull/48119"}, {"key": "pr#48115", "n": 48115, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "[AudioProcessor] deprecate SequenceFeatureExtractor", "action": "\u2014", "comment": "", "author": "eustlb", "idle_days": 43, "models": "", "url": "https://github.com/huggingface/transformers/pull/48115"}, {"key": "pr#48114", "n": 48114, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "feature / refactor", "title": "[AudioProcessor] migrate all audio models", "action": "\u2014", "comment": "", "author": "eustlb", "idle_days": 43, "models": "audio_spectrogram_transformer, audioflamingo3, clap, clvp, cohere_asr, csm, dac, dia, encodec, gemma3n, gemma4, gemma4_unified, glmasr, granite_speech, inkling, kyutai_speech_to_text, lasr", "url": "https://github.com/huggingface/transformers/pull/48114"}, {"key": "pr#48113", "n": 48113, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "[AudioProcessor] add AutoAudioProcessor", "action": "\u2014", "comment": "", "author": "eustlb", "idle_days": 43, "models": "", "url": "https://github.com/huggingface/transformers/pull/48113"}, {"key": "pr#48112", "n": 48112, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "core / utils", "type": "feature / refactor", "title": "[AudioProcessor] add numpy and torch backends", "action": "\u2014", "comment": "", "author": "eustlb", "idle_days": 43, "models": "", "url": "https://github.com/huggingface/transformers/pull/48112"}, {"key": "pr#48111", "n": 48111, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "other", "type": "feature / refactor", "title": "[AudioProcessor] add BaseAudioProcessor and SpectrogramConfig", "action": "\u2014", "comment": "", "author": "eustlb", "idle_days": 43, "models": "", "url": "https://github.com/huggingface/transformers/pull/48111"}, {"key": "pr#48110", "n": 48110, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "[AudioProcessor] add PreprocessingMixin", "action": "\u2014", "comment": "", "author": "eustlb", "idle_days": 43, "models": "", "url": "https://github.com/huggingface/transformers/pull/48110"}, {"key": "pr#48104", "n": 48104, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "chore: update pr-ci-post-dashboard-link.yml", "action": "\u2014", "comment": "", "author": "hf-security-analysis", "idle_days": 62, "models": "", "url": "https://github.com/huggingface/transformers/pull/48104"}, {"key": "pr#48063", "n": 48063, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "DO NOT LAND Test hfh 1.28.0.rc0", "action": "closed", "comment": "Closing this do-not-land test branch; `setup.py` on main already requires `huggingface-hub>=1.31.0`. Thanks @tarekziade!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/48063"}, {"key": "pr#47990", "n": 47990, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "quantization", "type": "feature / refactor", "title": "[Quantization] Anchor should_convert_module regex to stop prefix-matching siblings", "action": "\u2014", "comment": "", "author": "Ilya-a-sergeyev-ger", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/pull/47990"}, {"key": "pr#47984", "n": 47984, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "docs: fix malformed link to custom models section in CONTRIBUTING", "action": "closed", "comment": "Thanks @Agentic-JJ-Web3! This link was fixed in #48160, so closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47984"}, {"key": "pr#47983", "n": 47983, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 6 integration tests for model `starcoder2` failing with `output_mismatch` (list output differs (6))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 19, "models": "starcoder2", "url": "https://github.com/huggingface/transformers/pull/47983"}, {"key": "pr#47978", "n": 47978, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "fix(deps): bump tokenizers upper bound to <0.24.0 (#47429)", "action": "closed", "comment": "Thanks @Prakhar00001! The tokenizers bound was already raised to `<0.24.0` in #46381, so closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47978"}, {"key": "pr#47977", "n": 47977, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "fix(cache): clamp DynamicLayer.crop on oversized negative crops (#47433)", "action": "closed", "comment": "Thanks for the fix! `DynamicLayer.crop` was reworked in #47720 (negative values only) and the deprecated cache utils removed in #49605, so this diff no longer applies. Closing; please open an issue if oversized crops still misbehave on `main`.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47977"}, {"key": "pr#47974", "n": 47974, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Unified multimodal mixin (mrope)", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": 39, "models": "cohere_compass, cosmos3_edge, cosmos3_omni, ernie4_5_vl_moe, exaone4_5, glm46v, glm4v, glm4v_moe, glm5_next, glm_image, glm_ocr, hunyuan_vl, kimi_k25, minimax_m3_vl, muse_glimmer, paddleocr_vl, qwen2_5_omni, qwen2_5_vl, qwen2_vl, qwen3_5, qwen3_5_moe, qwen3_omni_moe, qwen3_vl, qwen3_vl_moe, qwen4_exp, vibevoice, video_llama_3", "url": "https://github.com/huggingface/transformers/pull/47974"}, {"key": "pr#47952", "n": 47952, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "docs", "type": "docs", "title": "Docs: cuda windows arm64 update installation guidelines", "action": "\u2014", "comment": "", "author": "mfuntowicz", "idle_days": 48, "models": "", "url": "https://github.com/huggingface/transformers/pull/47952"}, {"key": "pr#47941", "n": 47941, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "exporters", "type": "feature / refactor", "title": "[Exporters] varlen vision attention using native aten operator and test using bf16", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": 68, "models": "", "url": "https://github.com/huggingface/transformers/pull/47941"}, {"key": "pr#47929", "n": 47929, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add support for Lumma Model", "action": "\u2014", "comment": "", "author": "ramsha-khann", "idle_days": 6, "models": "depth_anything, efficientloftr, lumma, prompt_depth_anything", "url": "https://github.com/huggingface/transformers/pull/47929"}, {"key": "pr#47919", "n": 47919, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / loading", "type": "fix", "title": "Fix/bitsandbytes caching allocator warmup", "action": "review requested: @SunMarc; maintainer pinged", "comment": "@SunMarc could you take a look when you get a chance? This is waiting on a maintainer: never reviewed.", "author": "LiaoBill", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/47919"}, {"key": "pr#47908", "n": 47908, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 2 integration tests for model `swinv2` failing with `output_mismatch` (tensor values differ (2))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 70, "models": "swinv2", "url": "https://github.com/huggingface/transformers/pull/47908"}, {"key": "pr#47891", "n": 47891, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "Fix NeMo conversion for Parakeet TDT checkpoints without a <pad> token (parakeet-tdt-0.6b-v2)", "action": "review requested: @Cyrilvallez; maintainer pinged", "comment": "@Cyrilvallez could you take a look when you get a chance? This is waiting on a maintainer: it has been waiting for a maintainer review.", "author": "RahulRachuri", "idle_days": -1, "models": "parakeet", "url": "https://github.com/huggingface/transformers/pull/47891"}, {"key": "pr#47874", "n": 47874, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add WeatherNext 2 weather forecasting model", "action": "\u2014", "comment": "", "author": "kashif", "idle_days": -1, "models": "weathernext2", "url": "https://github.com/huggingface/transformers/pull/47874"}, {"key": "issue#47869", "n": 47869, "kind": "issue", "group": "Open issues", "category": "feature", "area": "pipelines", "type": "feature", "title": "[Enhancement] Add early device validation in pipeline() to prevent downloading/loading models before crashing on typos", "action": "\u2014", "comment": "", "author": "AryanSharma21", "idle_days": 70, "models": "", "url": "https://github.com/huggingface/transformers/issues/47869"}, {"key": "pr#47862", "n": 47862, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "tokenization", "type": "feature / refactor", "title": "Strip the metaspace prefix space only when it is unambiguous", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 54, "models": "code_llama, llama", "url": "https://github.com/huggingface/transformers/pull/47862"}, {"key": "pr#47855", "n": 47855, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "[Tokenizer] Fix silent data loss when migrating additional_special_tokens (Fixes #47838)", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #47848, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47855"}, {"key": "pr#47845", "n": 47845, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "fix mllama cross-attention masking", "action": "review requested: @zucchini-nlp; maintainer pinged", "comment": "@zucchini-nlp could you take a look when you get a chance? This is waiting on a maintainer: maintainer decision on whether the reorder is correct (zucchini-nlp saw gibberish generations with the same reorder in #39979 and said HF matched Meta's impl).", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47845"}, {"key": "pr#47844", "n": 47844, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add dots3-note Preview model support", "action": "\u2014", "comment": "", "author": "miraclezqc", "idle_days": 6, "models": "dots3_note", "url": "https://github.com/huggingface/transformers/pull/47844"}, {"key": "pr#47841", "n": 47841, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 2 integration tests for model `m2m_100` failing with `other` (other (2))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 74, "models": "m2m_100", "url": "https://github.com/huggingface/transformers/pull/47841"}, {"key": "pr#47840", "n": 47840, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 2 integration tests for model `efficientnet` failing with `output_mismatch` (tensor values differ (2))", "action": "review requested: @ydshieh; maintainer pinged", "comment": "@ydshieh could you take a look when you get a chance? This is waiting on a maintainer: never reviewed; needs a maintainer to confirm the new expected slice on CI hardware.", "author": "sergereview", "idle_days": -1, "models": "efficientnet", "url": "https://github.com/huggingface/transformers/pull/47840"}, {"key": "pr#47834", "n": 47834, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "[Mistral] Add native checkpoint support to PixtralProcessor", "action": "closed", "comment": "Thanks for the work on this! As explained in https://github.com/huggingface/transformers/pull/47834#pullrequestreview-4949002697, we can't set a precedent for a model-specific backend in `PixtralProcessor`; a conversion script to the standard format is the supported path. Closing, but happy to review a conversion-script PR.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47834"}, {"key": "issue#47805", "n": 47805, "kind": "issue", "group": "Open issues", "category": "bug", "area": "other", "type": "bug", "title": "Whisper large-v3 and large-v3-turbo: config declares float16 but WhisperFeatureExtractor emits float32, so the manual path raises", "action": "labelled: bug", "comment": "", "author": "Kayvan-Zahiri", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/47805"}, {"key": "pr#47802", "n": 47802, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Update EXAONE MoE for K-EXAONE 2.0", "action": "\u2014", "comment": "", "author": "nuxlear", "idle_days": 49, "models": "exaone_moe", "url": "https://github.com/huggingface/transformers/pull/47802"}, {"key": "issue#47787", "n": 47787, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add RWKV-7 (Goose)", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [RWKV/RWKV7-G1j-2.9B-20260831](https://huggingface.co/RWKV/RWKV7-G1j-2.9B-20260831), [fla-hub/rwkv7-0.1B-g1](https://huggingface.co/fla-hub/rwkv7-0.1B-g1), [BlinkDL/rwkv7-g1](https://huggingface.co/BlinkDL/rwkv7-g1)\n- Downloads (last 30d): 1.2k \u00b7 Likes: 2 \u00b7 Params: ~2.9B \u00b7 License: apache-2.0\n- Released: 2026-09 \u00b7 Remote code on the Hub: yes\n- Note: `rwkv` on main is RWKV-4; RWKV-7 is a different architecture (the `fla-hub` repos run through flash-linear-attention remote code).\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "Hakureirm", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/47787"}, {"key": "pr#47782", "n": 47782, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "new model", "title": "Add tiny model", "action": "\u2014", "comment": "", "author": "sushmanthreddy", "idle_days": 61, "models": "tiny_model", "url": "https://github.com/huggingface/transformers/pull/47782"}, {"key": "pr#47780", "n": 47780, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add RWKV-7 (Goose)", "action": "\u2014", "comment": "", "author": "Hakureirm", "idle_days": 0, "models": "rwkv7", "url": "https://github.com/huggingface/transformers/pull/47780"}, {"key": "pr#47778", "n": 47778, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix broken links in troubleshooting guide", "action": "closed", "comment": "Thanks @seedaylight! As @stevhliu mentioned, these links were fixed in #47748, so closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47778"}, {"key": "pr#47776", "n": 47776, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "new model", "title": "Add minimax vl 01", "action": "\u2014", "comment": "", "author": "sushmanthreddy", "idle_days": 76, "models": "minimax_vl_01", "url": "https://github.com/huggingface/transformers/pull/47776"}, {"key": "pr#47773", "n": 47773, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "feature / refactor", "title": ":rotating_light: Kernels=True ?!", "action": "\u2014", "comment": "", "author": "vasqu", "idle_days": 1, "models": "bark, beam, falcon, flex_olmo, gpt_neo, gpt_oss, gptj, llama4, molmo2, olmo, olmo2, olmo3, olmo_hybrid", "url": "https://github.com/huggingface/transformers/pull/47773"}, {"key": "pr#47771", "n": 47771, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "fix device-to-host sync in flash attention", "action": "closed", "comment": "Thanks for working on this! The decode-time device sync in `_is_packed_sequence` was removed in #49610, so closing this as superseded. If you still see a sync in another path, please open an issue with a profile.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47771"}, {"key": "pr#47758", "n": 47758, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "chore: update ssh-runner.yml", "action": "\u2014", "comment": "", "author": "hf-security-analysis", "idle_days": 52, "models": "", "url": "https://github.com/huggingface/transformers/pull/47758"}, {"key": "pr#47756", "n": 47756, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "new model", "title": "support for VoxCPM2", "action": "\u2014", "comment": "", "author": "sushmanthreddy", "idle_days": 1, "models": "minicpm4, voxcpm2", "url": "https://github.com/huggingface/transformers/pull/47756"}, {"key": "pr#47750", "n": 47750, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 4 integration tests for model `pvt` failing with `output_mismatch` (tensor values differ (4))", "action": "review requested: @ydshieh; maintainer pinged", "comment": "@ydshieh could you take a look when you get a chance? This is waiting on a maintainer: it has been waiting for a maintainer review.", "author": "sergereview", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/47750"}, {"key": "pr#47746", "n": 47746, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-ZH] Translate accelerator_selection.md to Chinese", "action": "review requested: @stevhliu", "comment": "", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47746"}, {"key": "pr#47736", "n": 47736, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Minicpm4", "action": "\u2014", "comment": "", "author": "sushmanthreddy", "idle_days": -1, "models": "minicpm4", "url": "https://github.com/huggingface/transformers/pull/47736"}, {"key": "pr#47734", "n": 47734, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Trainer: fix optimizer flat mesh for MoE for native DistributedConfig", "action": "closed", "comment": "The `distributed` branch this targets is no longer used (#46269 was closed). On `main`, clipping gradients across a mix of DTensor and plain tensors was fixed in #48205, dense and expert device meshes were added in #48857, and FSDP2 + EP was enabled in #48516. If the flat-mesh optimizer issue still reproduces on `main`, a fresh PR against `main` would be very welcome. Closing this one, thanks!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47734"}, {"key": "issue#47732", "n": 47732, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add native support for OpenBMB MiniCPM4", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [openbmb/MiniCPM4-8B](https://huggingface.co/openbmb/MiniCPM4-8B), [openbmb/MiniCPM4.1-8B](https://huggingface.co/openbmb/MiniCPM4.1-8B), [openbmb/MiniCPM4-0.5B](https://huggingface.co/openbmb/MiniCPM4-0.5B)\n- Downloads (last 30d): 6,626 \u00b7 Likes: 289 \u00b7 Params: ~8.2B (0.43B for MiniCPM4-0.5B) \u00b7 License: apache-2.0\n- Released: 2025-06 (4.1: 2025-09) \u00b7 Remote code on the Hub: yes\n- Open PR: #47736\n- Closest existing model: `minicpm3`.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "sushmanthreddy", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/47732"}, {"key": "pr#47712", "n": 47712, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / loading", "type": "feature / refactor", "title": "config: raise on missing local config file instead of silently falling back to defaults", "action": "review requested: @ydshieh", "comment": "", "author": "Ahmetshbzz", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/47712"}, {"key": "pr#47703", "n": 47703, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "Fix D2H sync in Qwen VL vision encoder", "action": "\u2014", "comment": "", "author": "Mahnoor-Zaffar", "idle_days": 1, "models": "cosmos3_edge, cosmos3_omni, exaone4_5, glm46v, glm4v, glm4v_moe, glm_ocr, minicpmv4_6, paddleocr_vl, qwen2_5_vl, qwen2_vl, qwen3_5, qwen3_5_moe, qwen3_omni_moe, qwen3_vl, qwen3_vl_moe", "url": "https://github.com/huggingface/transformers/pull/47703"}, {"key": "pr#47702", "n": 47702, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "pipelines", "type": "fix", "title": "Fix text-generation pipeline not forwarding tokenizer to generate", "action": "\u2014", "comment": "", "author": "alwaysprince05", "idle_days": 77, "models": "", "url": "https://github.com/huggingface/transformers/pull/47702"}, {"key": "pr#47699", "n": 47699, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Remove unused DPT fusion residual parameters", "action": "closed", "comment": "Thanks for your interest, @VictorMaxWang. As @Rocketknight1 mentioned in https://github.com/huggingface/transformers/pull/47699#issuecomment-5181127913 and per the policy in `CONTRIBUTING.md`, we're not accepting code-agent PRs from first-time contributors, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47699"}, {"key": "pr#47698", "n": 47698, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 2 integration tests for model `dab_detr` failing with `output_mismatch` (tensor values differ (2))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 80, "models": "dab_detr", "url": "https://github.com/huggingface/transformers/pull/47698"}, {"key": "pr#47697", "n": 47697, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 2 integration tests for model `phi3` failing with `other` (other (2))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 80, "models": "phi3", "url": "https://github.com/huggingface/transformers/pull/47697"}, {"key": "issue#47695", "n": 47695, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add native support for OpenBMB VoxCPM2", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [openbmb/VoxCPM2](https://huggingface.co/openbmb/VoxCPM2)\n- Downloads (last 30d): 873,199 \u00b7 Likes: 1702 \u00b7 Params: ~2.3B \u00b7 License: apache-2.0\n- Released: 2026-04 \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "sushmanthreddy", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/47695"}, {"key": "pr#47693", "n": 47693, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "added the chronous 2 first revision code", "action": "closed", "comment": "Closing this draft for now since it hasn't moved since July. Feel free to reopen (or open a fresh PR) once Chronos-2 is ready for review: build it with a `modular_chronos2.py` and run `transformers-mlinter` before asking for review. Guides: https://huggingface.co/docs/transformers/main/en/modular_transformers and https://huggingface.github.io/transformers-mlinter/", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47693"}, {"key": "issue#47692", "n": 47692, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add native support for Amazon Chronos-2", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [amazon/chronos-2](https://huggingface.co/amazon/chronos-2)\n- Downloads (last 30d): 15.6M \u00b7 Likes: 544 \u00b7 Params: ~119M \u00b7 License: apache-2.0\n- Released: 2025-10 \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "sushmanthreddy", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/47692"}, {"key": "pr#47675", "n": 47675, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "Fix Grounding DINO text position embeddings being truncated to int64", "action": "\u2014", "comment": "", "author": "Yigtwxx", "idle_days": 5, "models": "grounding_dino, mm_grounding_dino", "url": "https://github.com/huggingface/transformers/pull/47675"}, {"key": "pr#47670", "n": 47670, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "new model", "title": "Add K3", "action": "maintainer pinged", "comment": "@remi-or this one has been waiting since 2026-09-23 (backburner (remi-or); competing community PR #48890). Still planned, or should we close it?", "author": "remi-or", "idle_days": -1, "models": "kimi_k3, olmo_hybrid, qwen3_5, qwen3_5_moe, qwen3_next", "url": "https://github.com/huggingface/transformers/pull/47670"}, {"key": "issue#47667", "n": 47667, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add native Transformers support for Nanbeige4.2-3B", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [Nanbeige/Nanbeige4.2-3B](https://huggingface.co/Nanbeige/Nanbeige4.2-3B)\n- Downloads (last 30d): 119,966 \u00b7 Likes: 834 \u00b7 Params: ~4.2B \u00b7 License: apache-2.0\n- Released: 2026-07 \u00b7 Remote code on the Hub: yes\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "sushmanthreddy", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/47667"}, {"key": "pr#47662", "n": 47662, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "new model", "title": "Apertus 1.5: add model", "action": "\u2014", "comment": "", "author": "RaphaelKreft", "idle_days": 1, "models": "apertus1p5, wavtokenizer", "url": "https://github.com/huggingface/transformers/pull/47662"}, {"key": "pr#47661", "n": 47661, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "quantization", "type": "feature / refactor", "title": "Allow bnb_4bit_compute_dtype override for pre-quantized bnb checkpoints", "action": "review requested: @ydshieh; maintainer pinged", "comment": "@ydshieh could you take a look when you get a chance? This is waiting on a maintainer: it has been waiting for a maintainer review.", "author": "Zohaibarif69", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/47661"}, {"key": "pr#47659", "n": 47659, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "new model", "title": "Add Ovis2.5 model support", "action": "\u2014", "comment": "", "author": "sushmanthreddy", "idle_days": -1, "models": "ovis2_5", "url": "https://github.com/huggingface/transformers/pull/47659"}, {"key": "pr#47658", "n": 47658, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "docs", "type": "docs", "title": "Add detection training support to OWL-ViT (OwlViTHungarianMatcher + OwlViTLoss)", "action": "review requested: @ydshieh; maintainer pinged", "comment": "@ydshieh could you take a look when you get a chance? This is waiting on a maintainer: never reviewed.", "author": "anidoesdev", "idle_days": -1, "models": "owlv2, owlvit", "url": "https://github.com/huggingface/transformers/pull/47658"}, {"key": "pr#47657", "n": 47657, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "core / loading", "type": "fix", "title": "Fix dtype mismatch when loading", "action": "maintainer pinged", "comment": "@SunMarc this one has been waiting since 2026-08-10 (design discussion between SunMarc/ArthurZucker/Cyrilvallez (gate on WeightConverter); CONFLICTING). Still planned, or should we close it?", "author": "SunMarc", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/47657"}, {"key": "pr#47654", "n": 47654, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-FR] Translate `training.md` to French", "action": "review requested: @stevhliu", "comment": "", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47654"}, {"key": "pr#47653", "n": 47653, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "FlashAttention decode packed-sequence sync fix", "action": "closed", "comment": "Thanks for the fix! The same decode-time sync was removed in #49610, which returns early from `_is_packed_sequence` when decoding, so closing this as superseded.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47653"}, {"key": "pr#47643", "n": 47643, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "trainer", "type": "fix", "title": "Fix TDT loss failing on torch below 2.8", "action": "review requested: @ydshieh", "comment": "", "author": "Ar4ikov", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/47643"}, {"key": "pr#47629", "n": 47629, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "docs", "type": "docs", "title": "[docs] Audio backends", "action": "\u2014", "comment": "", "author": "stevhliu", "idle_days": 61, "models": "", "url": "https://github.com/huggingface/transformers/pull/47629"}, {"key": "pr#47616", "n": 47616, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "[serge] Fix 2 integration tests for model `regnet` failing with `output_mismatch` (tensor values differ (2))", "action": "\u2014", "comment": "", "author": "sergereview", "idle_days": 84, "models": "regnet", "url": "https://github.com/huggingface/transformers/pull/47616"}, {"key": "pr#47600", "n": 47600, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Potential fix for code scanning alert no. 265: Environment variable built from user-controlled sources", "action": "closed", "comment": "Closing: the workflow hardening in #49059 (and #49046) addressed these code-scanning alerts in `check_failed_tests.yml`.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47600"}, {"key": "pr#47596", "n": 47596, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "Fix dtype casting when dequantizing", "action": "closed", "comment": "Closing: #49548 moved the pre-quantized dtype choice into `hf_quantizer.get_new_dtype`, so the condition changed here no longer exists. If the dequantization dtype issue from #47486 still reproduces on main, it should be fixed in the relevant quantizer's `get_new_dtype` (cc @SunMarc).", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47596"}, {"key": "pr#47594", "n": 47594, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "exporters", "type": "feature / refactor", "title": "[Experimental / Do not Review] Add multi-method export + QNN (Qualcomm HTP) backend to the ExecuTorch exporter", "action": "\u2014", "comment": "", "author": "psiddh", "idle_days": 83, "models": "", "url": "https://github.com/huggingface/transformers/pull/47594"}, {"key": "pr#47588", "n": 47588, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "feature / refactor", "title": "feat[vLLM]: Make VoxtralProcessor compatible with the Transformers modeling backend for vLLM", "action": "\u2014", "comment": "", "author": "eustlb", "idle_days": 81, "models": "voxtral", "url": "https://github.com/huggingface/transformers/pull/47588"}, {"key": "pr#47585", "n": 47585, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Support solar-open2 model", "action": "\u2014", "comment": "", "author": "oesni", "idle_days": 48, "models": "solar_open, solar_open2", "url": "https://github.com/huggingface/transformers/pull/47585"}, {"key": "pr#47510", "n": 47510, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "fix", "title": "[Quantization] Fix CT `apply_quantization_config`", "action": "\u2014", "comment": "", "author": "kylesayrs", "idle_days": 83, "models": "hunyuan_vl", "url": "https://github.com/huggingface/transformers/pull/47510"}, {"key": "pr#47478", "n": 47478, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "fix", "title": "Fix CLVP `c_proj` residual scaled-init silently never applied", "action": "review requested: @vasqu; maintainer pinged", "comment": "@vasqu this was approved by @qgallouedec on 2026-10-17; could you give it the core-maintainer review so it can be merged?", "author": "ShamSaleem", "idle_days": -1, "models": "clvp", "url": "https://github.com/huggingface/transformers/pull/47478"}, {"key": "pr#47472", "n": 47472, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "Fix ESMFold error in fp16 precision", "action": "\u2014", "comment": "", "author": "crackalamoo", "idle_days": 89, "models": "esm", "url": "https://github.com/huggingface/transformers/pull/47472"}, {"key": "pr#47468", "n": 47468, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "core / utils", "type": "fix", "title": "Fix torch_float casting to int instead of float", "action": "review requested: @Rocketknight1", "comment": "", "author": "rahulrshetty45", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/47468"}, {"key": "pr#47462", "n": 47462, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "add MTP for GLM OCR and other MROPEs", "action": "\u2014", "comment": "", "author": "molbap", "idle_days": 32, "models": "glm_ocr", "url": "https://github.com/huggingface/transformers/pull/47462"}, {"key": "pr#47441", "n": 47441, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "Fix Whisper integration tests dtype mismatch and max_length off-by-one", "action": "review requested: @ydshieh", "comment": "", "author": "zjin-lcf", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/47441"}, {"key": "pr#47438", "n": 47438, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "cache / masks", "type": "feature / refactor", "title": "[`DSA`] Less materialized masks", "action": "\u2014", "comment": "", "author": "vasqu", "idle_days": 92, "models": "deepseek_v32", "url": "https://github.com/huggingface/transformers/pull/47438"}, {"key": "pr#47430", "n": 47430, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "quantization", "type": "fix", "title": "fix(compressed_tensors): collect zero_point for asymmetric MoE expert decompression", "action": "review requested: @SunMarc", "comment": "", "author": "egesabanci", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/47430"}, {"key": "pr#47421", "n": 47421, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "gguf", "type": "feature / refactor", "title": "Add GGUF loading support for DeepSeek-V3 (deepseek2 arch)", "action": "review requested: @stevhliu", "comment": "", "author": "gideonphilip", "idle_days": -1, "models": "deepseek_v3", "url": "https://github.com/huggingface/transformers/pull/47421"}, {"key": "pr#47401", "n": 47401, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add LoMa keypoint matching model", "action": "\u2014", "comment": "", "author": "Gitanaskhan26", "idle_days": 0, "models": "loma", "url": "https://github.com/huggingface/transformers/pull/47401"}, {"key": "pr#47400", "n": 47400, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "gguf", "type": "feature / refactor", "title": "Add GGUF loading support for Phi-2 (#33260)", "action": "review requested: @stevhliu", "comment": "", "author": "vibhorxpandey", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/47400"}, {"key": "pr#47391", "n": 47391, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "fix(trainer): fix resume_from_checkpoint logic bug (issue #47375)", "action": "closed", "comment": "We removed the `examples/pytorch` scripts from the repository in #49357, so this fix no longer applies. Thanks for the contribution!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47391"}, {"key": "pr#47387", "n": 47387, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "core / loading", "type": "feature / refactor", "title": "[WIP] Support one-to-many conversions", "action": "\u2014", "comment": "", "author": "kylesayrs", "idle_days": 92, "models": "", "url": "https://github.com/huggingface/transformers/pull/47387"}, {"key": "pr#47371", "n": 47371, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "serving / continuous batching", "type": "feature / refactor", "title": "Make non-blocking serve uvicorn thread a daemon with force_exit", "action": "review requested: @ydshieh; maintainer pinged", "comment": "@ydshieh could you take a look when you get a chance? This is waiting on a maintainer: it has been waiting for a maintainer review.", "author": "SebTardif", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/47371"}, {"key": "pr#47370", "n": 47370, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "attention / kernels", "type": "feature / refactor", "title": "Preserve exception cause when hub kernel loading fails", "action": "review requested: @SunMarc", "comment": "", "author": "SebTardif", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/47370"}, {"key": "pr#47368", "n": 47368, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "ci: Handle missing summaries in collated reports", "action": "closed", "comment": "Closing: the missing-summary handling (`get_missing_summary_result`) landed in #47364.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47368"}, {"key": "pr#47358", "n": 47358, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add beats model", "action": "\u2014", "comment": "", "author": "eslam-ahmed43", "idle_days": 91, "models": "beats", "url": "https://github.com/huggingface/transformers/pull/47358"}, {"key": "pr#47342", "n": 47342, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "attention / kernels", "type": "fix", "title": "Fix _is_packed_sequence for 1D and 3D position_ids", "action": "\u2014", "comment": "", "author": "winglian", "idle_days": 83, "models": "", "url": "https://github.com/huggingface/transformers/pull/47342"}, {"key": "pr#47329", "n": 47329, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Add LocalMode to `awesome-transformers.md`", "action": "closed", "comment": "Thanks for sharing LocalMode, @MuathJ! `awesome-transformers.md` lists projects built on the Python library, so we're not adding this entry; Transformers.js-based projects are a better fit for the Transformers.js community channels. Closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47329"}, {"key": "pr#47327", "n": 47327, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "[docs] Improve CLI copy", "action": "maintainer pinged", "comment": "@stevhliu this one has been waiting since 2026-07-14 (CONFLICTING; never reviewed). Still planned, or should we close it?", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47327"}, {"key": "pr#47306", "n": 47306, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "other", "type": "feature / refactor", "title": "Experiments with Neuron", "action": "\u2014", "comment": "", "author": "michaelbenayoun", "idle_days": 55, "models": "", "url": "https://github.com/huggingface/transformers/pull/47306"}, {"key": "pr#47298", "n": 47298, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "processors (image/video/audio)", "type": "fix", "title": "Fix important model failures on MI450", "action": "\u2014", "comment": "", "author": "Abdennacer-Badaoui", "idle_days": 92, "models": "conditional_detr, deformable_detr, detr, grounding_dino, yolos", "url": "https://github.com/huggingface/transformers/pull/47298"}, {"key": "pr#47297", "n": 47297, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "tests", "type": "CI / tooling", "title": "[DeepSpeed] Composite models are not initialized correctly under ZeRO-3", "action": "\u2014", "comment": "", "author": "kashif", "idle_days": 42, "models": "", "url": "https://github.com/huggingface/transformers/pull/47297"}, {"key": "pr#47294", "n": 47294, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Kernel native causal_conv1d functions", "action": "closed", "comment": "Closing as this was superseded by #47450 (and the kernels fallback in #47630). Thanks!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47294"}, {"key": "pr#47290", "n": 47290, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "[Improvement] Formalize KV sharing", "action": "\u2014", "comment": "", "author": "remi-or", "idle_days": 1, "models": "diffusion_gemma, gemma3n, gemma4, gemma4_unified", "url": "https://github.com/huggingface/transformers/pull/47290"}, {"key": "pr#47275", "n": 47275, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "new model", "title": "Add new model `EfficientVit-SAM` to transformers", "action": "review requested: @molbap; maintainer pinged", "comment": "@molbap could you take a look when you get a chance? This is waiting on a maintainer: molbap re-review of the restructured PR; the stray edit to utils/modular_model_converter.py must be reverted.", "author": "abhishekkapoorx", "idle_days": -1, "models": "efficientvitsam, sam", "url": "https://github.com/huggingface/transformers/pull/47275"}, {"key": "pr#47271", "n": 47271, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "Add repeated n-gram stopping criterion", "action": "closed", "comment": "Thanks for the PR! Since `generate` already accepts custom `stopping_criteria`, and new decoding behaviours now go to [custom generate](https://huggingface.co/docs/transformers/main/en/generation_strategies#custom-generation-methods) repos on the Hub rather than core, we won't add this built-in. A Hub repo with your criterion would be a great way to share it.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47271"}, {"key": "pr#47262", "n": 47262, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "generation", "type": "feature / refactor", "title": "Respect min_tokens_to_keep in TopHLogitsWarper", "action": "review requested: @Cyrilvallez", "comment": "", "author": "Incheonkirin", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/47262"}, {"key": "pr#47257", "n": 47257, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / loading", "type": "feature / refactor", "title": "Avoid ~2x transient host memory when loading to a unified-memory device", "action": "\u2014", "comment": "", "author": "cgu2022", "idle_days": 32, "models": "", "url": "https://github.com/huggingface/transformers/pull/47257"}, {"key": "pr#47224", "n": 47224, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "fix: resolve serge integration test failures (#47182)", "action": "closed", "comment": "Thanks @maslinedwin! The EdgeTAM failure was fixed in #48322 and the Whisper `max_length` handling was reworked in #46795, so closing this. If one of the tokenizer issues still reproduces on main, a small focused PR would be easiest to review.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47224"}, {"key": "pr#47222", "n": 47222, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "other", "type": "fix", "title": "pytorch 2.13 is released , updating transformer to handle the fix tf32 issue", "action": "\u2014", "comment": "", "author": "khushali9", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/pull/47222"}, {"key": "issue#47217", "n": 47217, "kind": "issue", "group": "Open issues", "category": "bug", "area": "other", "type": "bug", "title": "VL models fail when their image placeholder is in user text.", "action": "\u2014", "comment": "", "author": "samuelstevens", "idle_days": 6, "models": "", "url": "https://github.com/huggingface/transformers/issues/47217"}, {"key": "pr#47215", "n": 47215, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix FA2 long-context perf regression (#46693)", "action": "closed", "comment": "Closing as the FA long-context regression was fixed by #47134, as discussed with @vasqu. Thanks @ammmanism!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47215"}, {"key": "pr#47210", "n": 47210, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "attention / kernels", "type": "feature / refactor", "title": "Route packed sdpa rows through torch.nn.attention.varlen.varlen_attn", "action": "\u2014", "comment": "", "author": "winglian", "idle_days": 99, "models": "", "url": "https://github.com/huggingface/transformers/pull/47210"}, {"key": "pr#47204", "n": 47204, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "torch-light: add executorch (+onnx) for CPU exporter tests", "action": "\u2014", "comment": "", "author": "tarekziade", "idle_days": 103, "models": "", "url": "https://github.com/huggingface/transformers/pull/47204"}, {"key": "pr#47202", "n": 47202, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "CI / tooling", "type": "CI / tooling", "title": "Drop CPU/light docker image builder (moved to transformers-ci)", "action": "maintainer pinged", "comment": "@tarekziade this one has been waiting since 2026-07-09 (conflicts with #46887 in build-ci-docker-images.yml; needs ydshieh's sign-off). Still planned, or should we close it?", "author": "tarekziade", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/47202"}, {"key": "pr#47171", "n": 47171, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "integrations (other)", "type": "feature / refactor", "title": "nCompass triton-moe experts integration", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": 104, "models": "", "url": "https://github.com/huggingface/transformers/pull/47171"}, {"key": "pr#47167", "n": 47167, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "fix", "title": "chore(qa): fix Windows compatibility for type checker excludes", "action": "review requested: @ydshieh", "comment": "", "author": "Partha-Shankar", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/47167"}, {"key": "pr#47162", "n": 47162, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "Fix off-by-one in MoonshineStreaming sliding_window_mask_function right window", "action": "\u2014", "comment": "", "author": "namantiwari2002", "idle_days": 103, "models": "gemma4, moonshine_streaming", "url": "https://github.com/huggingface/transformers/pull/47162"}, {"key": "pr#47140", "n": 47140, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix chat CLI management URLs with /v1 base", "action": "closed", "comment": "Thanks @VectorPeak! The `/v1` routing issue (#47139) was fixed in #47303 (`get_service_root_url` in `cli/chat.py`), so closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47140"}, {"key": "pr#47108", "n": 47108, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "Keep doc-builder pin comment dependabot-compatible", "action": "\u2014", "comment": "", "author": "mishig25", "idle_days": 106, "models": "", "url": "https://github.com/huggingface/transformers/pull/47108"}, {"key": "pr#47093", "n": 47093, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add OpenPangu-V2 Model", "action": "\u2014", "comment": "", "author": "YuChangHui", "idle_days": 1, "models": "hunyuan_vl, openpangu_v2", "url": "https://github.com/huggingface/transformers/pull/47093"}, {"key": "pr#47082", "n": 47082, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add RoMa keypoint matcher", "action": "review requested: @molbap; maintainer pinged", "comment": "@molbap could you take a look when you get a chance? This is waiting on a maintainer: never reviewed; author waiting on a maintainer to settle Hub repo ids for the converted weights.", "author": "srishtiii28", "idle_days": -1, "models": "roma", "url": "https://github.com/huggingface/transformers/pull/47082"}, {"key": "pr#47078", "n": 47078, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Improve Japanese documentation wording", "action": "closed", "comment": "Thanks for improving the Japanese docs, @shirokuma-xv! As @stevhliu noted, this overlaps with #47077 and we didn't hear back, so I'm closing it. Feel free to open a focused PR with only the non-overlapping changes.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47078"}, {"key": "pr#47077", "n": 47077, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "docs", "type": "docs", "title": "Add Japanese translation for DistilBERT docs", "action": "\u2014", "comment": "", "author": "shirokuma-xv", "idle_days": 106, "models": "", "url": "https://github.com/huggingface/transformers/pull/47077"}, {"key": "pr#47075", "n": 47075, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "fix(tokenizers): accept dict Unigram vocabs in slow tokenizer init", "action": "closed", "comment": "Thanks for the work! As explained in #47020 this behavior is by design for the v5 tokenizers, so we're closing this.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/47075"}, {"key": "pr#47052", "n": 47052, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "gguf", "type": "feature / refactor", "title": "Add GGUF support for Cohere2 (Command R7B / Command A)", "action": "review requested: @SunMarc", "comment": "", "author": "Madhav-000-s", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/47052"}, {"key": "pr#47047", "n": 47047, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Support loading original InternVL2 checkpoints natively", "action": "\u2014", "comment": "", "author": "malakazlan", "idle_days": 41, "models": "internvl, internvl2, qianfan_ocr", "url": "https://github.com/huggingface/transformers/pull/47047"}, {"key": "pr#47046", "n": 47046, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "fix", "title": "[xLSTM] Fix train mode crashes with return_last_states=False and use_cache", "action": "\u2014", "comment": "", "author": "Heatdh", "idle_days": 78, "models": "xlstm", "url": "https://github.com/huggingface/transformers/pull/47046"}, {"key": "pr#47017", "n": 47017, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Route byte-level llama tokenizers to TokenizersBackend", "action": "\u2014", "comment": "", "author": "subin9", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/pull/47017"}, {"key": "issue#47004", "n": 47004, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add LoMa keypoint detector & matcher", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: no checkpoint on the Hub yet; weights are published as GitHub release assets in [davnords/LoMa](https://github.com/davnords/LoMa) (`loma_B`, `loma_L`, `loma_G`, `loma_R`, ...)\n- Released: 2026-04 \u00b7 Closest existing model: `lightglue`\n- Open PR: #47401\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "tcourat", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/47004"}, {"key": "pr#46997", "n": 46997, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "integrations (other)", "type": "feature / refactor", "title": "Add after-load fusion for static quantized MLPs", "action": "\u2014", "comment": "", "author": "LiangSu8899", "idle_days": 98, "models": "", "url": "https://github.com/huggingface/transformers/pull/46997"}, {"key": "pr#46958", "n": 46958, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix CohereASR training loss double shift", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #46895, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46958"}, {"key": "pr#46957", "n": 46957, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "feature / refactor", "title": "Finegrained FP8 v4: fused grouepd/batched", "action": "commented", "comment": "Hi! The fused grouped/batched FP8 path isn't on main yet, but 13 of the 18 touched files now conflict: `_optimize_model_for_decode` became an argument-free context manager, and `finegrained_fp8`/`deepgemm`/the FP8 quantizer were reworked in #49548, #49428 and #47579. Is this still planned? If so, a rebase onto the current kernels/quantization layout is needed before review.", "author": "IlyasMoutawwakil", "idle_days": -1, "models": "deepseek_v4, gpt_oss, higgs_audio_v2, minimax_m3_vl, openai_privacy_filter", "url": "https://github.com/huggingface/transformers/pull/46957"}, {"key": "pr#46942", "n": 46942, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "fix", "title": "Fix dtype cast in TimesFM 2.5 ResidualBlock for quantized weights", "action": "review requested: @vasqu; maintainer pinged", "comment": "@vasqu this was approved by @SunMarc on 2026-09-17; could you give it the core-maintainer review so it can be merged?", "author": "JOhnsonKC201", "idle_days": -1, "models": "timesfm2_5", "url": "https://github.com/huggingface/transformers/pull/46942"}, {"key": "pr#46930", "n": 46930, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Bump minimum torch version to 2.5 and clean up torch<2.5 workarounds", "action": "closed", "comment": "Thanks @cyyever! The minimum torch version was bumped to 2.6 in #49395, which supersedes this experiment; closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46930"}, {"key": "pr#46924", "n": 46924, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "feature / refactor", "title": "Replace VideoMAE sinusoid encoding helper with PyTorch implementation", "action": "\u2014", "comment": "", "author": "praful-srinivasan-027", "idle_days": 116, "models": "videomae", "url": "https://github.com/huggingface/transformers/pull/46924"}, {"key": "pr#46922", "n": 46922, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix yolos", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #46886, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46922"}, {"key": "pr#46913", "n": 46913, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / utils", "type": "feature / refactor", "title": "Validate shard filenames in checkpoint index to prevent path traversal (silent weight injection)", "action": "review requested: @Rocketknight1", "comment": "", "author": "Snakinya", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/46913"}, {"key": "pr#46896", "n": 46896, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "trainer", "type": "feature / refactor", "title": "Update EvalPrediction annotations for batch_eval_metrics", "action": "review requested: @SunMarc", "comment": "", "author": "praful-srinivasan-027", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/46896"}, {"key": "pr#46887", "n": 46887, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "CI / tooling", "type": "CI / tooling", "title": "feat: build CI docker images for linux/arm64 (Graviton)", "action": "\u2014", "comment": "", "author": "glegendre01", "idle_days": 117, "models": "", "url": "https://github.com/huggingface/transformers/pull/46887"}, {"key": "pr#46848", "n": 46848, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "integrations (other)", "type": "fix", "title": "fix(integrations): prevent device mismatch in msa_attention", "action": "\u2014", "comment": "", "author": "leaderofARS", "idle_days": 98, "models": "", "url": "https://github.com/huggingface/transformers/pull/46848"}, {"key": "pr#46843", "n": 46843, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "Enhance return type annotation in apply_chat_template in processing_u\u2026", "action": "\u2014", "comment": "", "author": "bartjellema", "idle_days": 118, "models": "", "url": "https://github.com/huggingface/transformers/pull/46843"}, {"key": "issue#46822", "n": 46822, "kind": "issue", "group": "Open issues", "category": "bug", "area": "distributed", "type": "bug", "title": "ZeRO-3 zero.Init does not partition composite minimax_m3_vl language submodule -> OOM on multi-GPU load", "action": "\u2014", "comment": "", "author": "trevorgordon981", "idle_days": 7, "models": "minimax_m3_vl", "url": "https://github.com/huggingface/transformers/issues/46822"}, {"key": "pr#46820", "n": 46820, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "trainer", "type": "feature / refactor", "title": "Migrate ALBERT task heads to self.loss_function and add ForMultipleChoiceLoss", "action": "review requested: @SunMarc", "comment": "", "author": "Sahith59", "idle_days": -1, "models": "albert", "url": "https://github.com/huggingface/transformers/pull/46820"}, {"key": "pr#46813", "n": 46813, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "[docs] Fix grammar in BLIP visual question answering example", "action": "closed", "comment": "Thanks @Charuhasini30! This sentence was fixed on main in #47679, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46813"}, {"key": "pr#46792", "n": 46792, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "serving / continuous batching", "type": "feature / refactor", "title": "Enable CB paged decode fast path for MPS hub kernels + sensible block_size default", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 116, "models": "", "url": "https://github.com/huggingface/transformers/pull/46792"}, {"key": "pr#46780", "n": 46780, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "[RT-DETRv2] Fix MPS crash in build_2d_sinusoidal_position_embedding (float64 \u2192 dtype)", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #47897, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46780"}, {"key": "pr#46776", "n": 46776, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add LocateAnything", "action": "\u2014", "comment": "", "author": "molbap", "idle_days": 90, "models": "locateanything", "url": "https://github.com/huggingface/transformers/pull/46776"}, {"key": "pr#46775", "n": 46775, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "trainer", "type": "feature / refactor", "title": "New SP support in Trainer", "action": "maintainer pinged", "comment": "@SunMarc this one has been waiting since 2026-06-24 (depends on huggingface/accelerate#4084 (closed unmerged 2026-08-28; successor #4209 open) + qgallouedec's loss-scaling bug report). Still planned, or should we close it?", "author": "SunMarc", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/46775"}, {"key": "pr#46771", "n": 46771, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "Support custom kernels for processing ops", "action": "\u2014", "comment": "", "author": "molbap", "idle_days": 1, "models": "cohere_compass, ernie4_5_vl_moe, glm_image, hunyuan_vl, minimax_m3_vl, paddleocr_vl, qwen2_vl, qwen3_vl, sam2, sam3, video_llama_3", "url": "https://github.com/huggingface/transformers/pull/46771"}, {"key": "pr#46761", "n": 46761, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "[Whisper] Add Unpack[TransformersKwargs] to forward() and set use_cache=False when labels provided", "action": "\u2014", "comment": "", "author": "Sahith59", "idle_days": 123, "models": "whisper", "url": "https://github.com/huggingface/transformers/pull/46761"}, {"key": "pr#46730", "n": 46730, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "Fix hardcoded torch.float64 crashing on MPS (#46723)", "action": "review requested: @Cyrilvallez", "comment": "", "author": "Abhijeet1815", "idle_days": -1, "models": "musicflamingo, udop", "url": "https://github.com/huggingface/transformers/pull/46730"}, {"key": "pr#46722", "n": 46722, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "tokenization", "type": "feature / refactor", "title": "Reflect baked post-processor in `add_bos_token`/`add_eos_token`", "action": "review requested: @vasqu; maintainer pinged", "comment": "@vasqu this was approved by @itazap on 2026-06-18; could you give it the core-maintainer review so it can be merged?", "author": "ArthurZucker", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/46722"}, {"key": "pr#46718", "n": 46718, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "gguf", "type": "feature / refactor", "title": "Add GGUF loading support for GLM-4.5 MoE (glm4moe)", "action": "review requested: @SunMarc", "comment": "", "author": "mironable", "idle_days": -1, "models": "glm4_moe", "url": "https://github.com/huggingface/transformers/pull/46718"}, {"key": "pr#46716", "n": 46716, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "chore: update pr-ci-security-gate.yml", "action": "closed", "comment": "Closing since `secrets: inherit` was already removed from the security gate (#46560).", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46716"}, {"key": "pr#46713", "n": 46713, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "new model", "title": "feat(model): add bailing v2.6 model", "action": "review requested: @vasqu; maintainer pinged", "comment": "@vasqu could you take a look when you get a chance? This is waiting on a maintainer: vasqu's re-review after the linear-attention refactors he asked the author to wait for; naming clash: the PR's class is `BailingHybrid*`, and main now has a different `bailing_hybrid` (Ling 3.0, KDA) from #49446.", "author": "kemuxiaozi000", "idle_days": -1, "models": "bailing2_5_moe", "url": "https://github.com/huggingface/transformers/pull/46713"}, {"key": "pr#46642", "n": 46642, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "\ud83d\udea8 [DiffusionGemma] Return router logits and load balancing loss", "action": "closed", "comment": "@kashif #48920 now returns DiffusionGemma's router logits (and Gemma 4 / Gemma 4 Unified's) through the output recorder, so that half of this PR is on main. Like @qgallouedec said, the aux-loss part is better left to the training libraries (#48886). OK to close this, or do you want to keep a reduced version?", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46642"}, {"key": "pr#46623", "n": 46623, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Made SAM3 to return loss in outputs", "action": "\u2014", "comment": "", "author": "ash01825", "idle_days": 48, "models": "sam3, sam3_lite_text", "url": "https://github.com/huggingface/transformers/pull/46623"}, {"key": "pr#46580", "n": 46580, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix DeepSeek-Coder v1 / DeepSeek-LLM v1 loading as wrong tokenizer class (#46489)", "action": "closed", "comment": "Thanks for the fix, @kpal002! DeepSeek-Coder/LLM v1 now route through `TokenizersBackend` on main via #46091 and #46489 was closed, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46580"}, {"key": "pr#46578", "n": 46578, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "fix", "title": "Fix Wav2Vec2 word delimiter special token handling", "action": "review requested: @ydshieh; maintainer pinged", "comment": "@ydshieh could you take a look when you get a chance? This is waiting on a maintainer: never reviewed.", "author": "nightcityblade", "idle_days": -1, "models": "wav2vec2", "url": "https://github.com/huggingface/transformers/pull/46578"}, {"key": "issue#46577", "n": 46577, "kind": "issue", "group": "Open issues", "category": "feature", "area": "models", "type": "feature", "title": "Add RTDetrForSegmentation", "action": "\u2014", "comment": "", "author": "ranykamel", "idle_days": 16, "models": "rt_detr", "url": "https://github.com/huggingface/transformers/issues/46577"}, {"key": "pr#46575", "n": 46575, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "feature / refactor", "title": "[`Kernels`] Add new kernels to models: CE, Swiglu, Geglu(, Linear)", "action": "\u2014", "comment": "", "author": "vasqu", "idle_days": 57, "models": "gemma2, llama", "url": "https://github.com/huggingface/transformers/pull/46575"}, {"key": "pr#46563", "n": 46563, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "chore: update pr-ci-security-gate.yml", "action": "closed", "comment": "Closing since the security gate now delegates to the `transformers-ci` reusable workflow (#47040).", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46563"}, {"key": "pr#46558", "n": 46558, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "chore: update pr-ci-security-gate.yml", "action": "\u2014", "comment": "", "author": "hf-security-analysis", "idle_days": 131, "models": "", "url": "https://github.com/huggingface/transformers/pull/46558"}, {"key": "pr#46554", "n": 46554, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "chore: update pr-ci-security-gate.yml", "action": "closed", "comment": "Closing since the security gate now delegates to the `transformers-ci` reusable workflow (#47040).", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46554"}, {"key": "pr#46550", "n": 46550, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Remove position_ids documentation from function", "action": "closed", "comment": "Thanks for the fix! The `position_ids` entry is already gone from `apply_multimodal_rotary_pos_emb`'s docstring in `src/transformers/models/qwen2_vl/modeling_qwen2_vl.py` on `main`, so closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46550"}, {"key": "pr#46547", "n": 46547, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "docs: Update NeMo Automodel getting-started docs for updated user API", "action": "closed", "comment": "Thanks for keeping the NeMo AutoModel docs current! The examples were updated in #46857, so closing this draft as superseded.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46547"}, {"key": "pr#46504", "n": 46504, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "chore: update build-ci-docker-images.yml", "action": "\u2014", "comment": "", "author": "hf-security-analysis", "idle_days": 134, "models": "", "url": "https://github.com/huggingface/transformers/pull/46504"}, {"key": "pr#46503", "n": 46503, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "chore: update benchmark_v2.yml", "action": "review requested: @ydshieh; maintainer pinged", "comment": "@ydshieh could you take a look when you get a chance? This is waiting on a maintainer: it has been waiting for a maintainer review.", "author": "hf-security-analysis", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/46503"}, {"key": "pr#46502", "n": 46502, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "chore: update assign-reviewers.yml", "action": "closed", "comment": "Closing since `assign-reviewers.yml` was rewritten on main in #48085 and no longer runs `actions/checkout` on the PR code.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46502"}, {"key": "pr#46501", "n": 46501, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "chore: update anti-slop.yml", "action": "closed", "comment": "Closing as per-job permission scoping landed in #49046.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46501"}, {"key": "pr#46467", "n": 46467, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "gguf", "type": "feature / refactor", "title": "feat: update GGUF model configuration parameters for Gemma 4", "action": "\u2014", "comment": "", "author": "voharaa", "idle_days": 90, "models": "", "url": "https://github.com/huggingface/transformers/pull/46467"}, {"key": "pr#46453", "n": 46453, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "fix(integrations): correct typo in ClearMLCallback config overrides variable.", "action": "review requested: @eustlb; maintainer pinged", "comment": "@eustlb could you take a look when you get a chance? This is waiting on a maintainer: eustlb input on the HiggsAudioV2 part (SunMarc deferred to him 2026-06-08).", "author": "leaderofARS", "idle_days": -1, "models": "higgs_audio_v2", "url": "https://github.com/huggingface/transformers/pull/46453"}, {"key": "pr#46447", "n": 46447, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "new model", "title": "Add MOSS-TTS-v1.5 (delay) and MOSS-Audio-Tokenizer", "action": "\u2014", "comment": "", "author": "ted-mosi", "idle_days": 49, "models": "moss_audio_tokenizer, moss_tts_delay", "url": "https://github.com/huggingface/transformers/pull/46447"}, {"key": "pr#46432", "n": 46432, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "chore: update ssh-runner.yml", "action": "closed", "comment": "Closing: the pinned commit `6e9333e8...` does not exist in this repository and the checksum does not match any file, so the workflow would fail at the `sha256sum -c` step. If we want to pin `memory_monitor.py`, it should point at a real commit on `utility_scripts` (currently `1a037a1864`) with its actual sha256. cc @ydshieh", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46432"}, {"key": "pr#46420", "n": 46420, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "CI / tooling", "type": "CI / tooling", "title": "Run repository consistency checkers in parallel", "action": "\u2014", "comment": "", "author": "ydshieh", "idle_days": 137, "models": "", "url": "https://github.com/huggingface/transformers/pull/46420"}, {"key": "pr#46413", "n": 46413, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix finegrained-fp8 offline loading bug", "action": "review requested: @SunMarc; maintainer pinged; closed", "comment": "Thanks for the report! This is a real bug, but the fp8 kernel loader was reworked in #47334 and this patch no longer applies (there is no local finegrained-fp8 fallback for offline loading yet). Closing this PR as outdated; a fresh PR against current `main` adding that fallback would be very welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46413"}, {"key": "pr#46408", "n": 46408, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "feature / refactor", "title": "Remove unused shared_expert_combination_strategy from cohere2_moe", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 138, "models": "cohere2_moe", "url": "https://github.com/huggingface/transformers/pull/46408"}, {"key": "pr#46395", "n": 46395, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "new model", "title": "Add support for BharatGen's Param2MoE model architecture", "action": "\u2014", "comment": "", "author": "bhargav-patel-29", "idle_days": 11, "models": "param2moe", "url": "https://github.com/huggingface/transformers/pull/46395"}, {"key": "pr#46386", "n": 46386, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "trainer", "type": "feature / refactor", "title": "HF Buckets integration to Trainer", "action": "\u2014", "comment": "", "author": "SunMarc", "idle_days": 124, "models": "", "url": "https://github.com/huggingface/transformers/pull/46386"}, {"key": "pr#46384", "n": 46384, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "feature / refactor", "title": "DSv4 static cache", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": 123, "models": "deepseek_v4", "url": "https://github.com/huggingface/transformers/pull/46384"}, {"key": "issue#46382", "n": 46382, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "Propagation of output_hidden_states to sub_configs", "action": "\u2014", "comment": "", "author": "julien-gadonneix", "idle_days": 138, "models": "", "url": "https://github.com/huggingface/transformers/issues/46382"}, {"key": "pr#46380", "n": 46380, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "generation", "type": "fix", "title": "Fix cis", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 139, "models": "whisper", "url": "https://github.com/huggingface/transformers/pull/46380"}, {"key": "pr#46376", "n": 46376, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "serving / continuous batching", "type": "feature / refactor", "title": "[CB] [Major] Add multimodal support to continous batching", "action": "\u2014", "comment": "", "author": "remi-or", "idle_days": 89, "models": "", "url": "https://github.com/huggingface/transformers/pull/46376"}, {"key": "issue#46370", "n": 46370, "kind": "issue", "group": "Open issues", "category": "feature", "area": "exporters", "type": "feature", "title": "Deprecate CompileableContextVar in favor of plain ContextVar with upcoming Dynamo native support", "action": "\u2014", "comment": "", "author": "ayushsatyam146", "idle_days": 127, "models": "", "url": "https://github.com/huggingface/transformers/issues/46370"}, {"key": "pr#46364", "n": 46364, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / utils", "type": "feature / refactor", "title": "chore(linter): update DataClass type definitions to TypeAlias", "action": "review requested: @Rocketknight1; maintainer pinged", "comment": "@Rocketknight1 could you take a look when you get a chance? This is waiting on a maintainer: never reviewed.", "author": "cheparity", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/46364"}, {"key": "pr#46355", "n": 46355, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "Test models instantiate from default config on meta + fix failing models", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 140, "models": "aimv2, autoformer, aya_vision, bridgetower, chameleon, convnext, convnextv2, dbrx, deepseek_ocr2, dots1, emu3, esm, focalnet, granite4_vision, hiera, hunyuan_v1_dense, hunyuan_v1_moe, idefics3, informer, lfm2_moe, ministral, moonshine_streaming, nemotron, perceiver, perception_lm, pp_ocrv5_server_det, pp_ocrv6_medium_det, reformer, smolvlm, time_series_transformer, timm_backbone, vilt, vitmatte, wav2vec2_bert, wav2vec2_conformer", "url": "https://github.com/huggingface/transformers/pull/46355"}, {"key": "pr#46341", "n": 46341, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix tokenizer keep_accents defaults and update integration tests", "action": "closed", "comment": "@itazap it looks like #45936 already covered the albert/camembert/rembert `keep_accents` changes here. Close this, or rebase whatever test updates are still missing?", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46341"}, {"key": "pr#46330", "n": 46330, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Fix: strip EOS token when continue_final_message=True in MistralCommo\u2026", "action": "closed", "comment": "Thanks for the fix, @ranjithchinnanan! As @Rocketknight1 discussed in #46326, we'd rather not strip the EOS on our side until `mistral-common` supports this properly, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46330"}, {"key": "pr#46324", "n": 46324, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "quantization", "type": "feature / refactor", "title": "Quantize models on `__init__` / `from_config`", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 140, "models": "", "url": "https://github.com/huggingface/transformers/pull/46324"}, {"key": "pr#46319", "n": 46319, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "docs", "type": "docs", "title": "docs: customizing models.md, better comments for resnetconfig class", "action": "\u2014", "comment": "", "author": "ArjunSrivastava1", "idle_days": 9, "models": "", "url": "https://github.com/huggingface/transformers/pull/46319"}, {"key": "pr#46303", "n": 46303, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix arbitrary code execution via compressed-tensors sparsity format", "action": "closed", "comment": "Thanks for reporting this! As @SunMarc suggested (https://github.com/huggingface/transformers/pull/46303#pullrequestreview-4409844045), we followed compressed-tensors in dropping sparsity support in #46342, and the `sparsity_config` path no longer exists in `quantization_config.py` on main. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46303"}, {"key": "pr#46300", "n": 46300, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "LocateAnything 3B", "action": "\u2014", "comment": "", "author": "SangbumChoi", "idle_days": 87, "models": "locateanything", "url": "https://github.com/huggingface/transformers/pull/46300"}, {"key": "issue#46295", "n": 46295, "kind": "issue", "group": "Open issues", "category": "feature", "area": "quantization", "type": "feature", "title": "Training Support for FP8", "action": "\u2014", "comment": "", "author": "Banaxi-Tech", "idle_days": 143, "models": "", "url": "https://github.com/huggingface/transformers/issues/46295"}, {"key": "pr#46281", "n": 46281, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / utils", "type": "fix", "title": "Fix path traversal in sharded-checkpoint loader via the index weight_map", "action": "review requested: @ydshieh; maintainer pinged", "comment": "@ydshieh could you take a look when you get a chance? This is waiting on a maintainer: never reviewed (security-relevant).", "author": "LinZiyuu", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/46281"}, {"key": "pr#46257", "n": 46257, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Update deepseek v4 snippet: model id, quant config", "action": "\u2014", "comment": "", "author": "pcuenca", "idle_days": 145, "models": "deepseek_v4", "url": "https://github.com/huggingface/transformers/pull/46257"}, {"key": "pr#46236", "n": 46236, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "I18n bn", "action": "closed", "comment": "Thanks for the Bengali translation! The review points in https://github.com/huggingface/transformers/pull/46236#pullrequestreview-4385744069 haven't been addressed (no commits since), so I'm closing for now. A smaller PR (e.g. `index.md` and `_toctree.yml` first) would be easier to review and merge.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46236"}, {"key": "issue#46222", "n": 46222, "kind": "issue", "group": "Open issues", "category": "question / other", "area": "other", "type": "question / other", "title": "\u0e04\u0e34\u0e14\u0e14\u0e35 \u0e17\u0e33\u0e14\u0e35 \u0e40\u0e08\u0e2d\u0e41\u0e15\u0e48\u0e2a\u0e34\u0e48\u0e07\u0e14\u0e35 \u0e2a\u0e32\u0e18\u0e38 365.ai.th", "action": "\u2014", "comment": "", "author": "imgoingbacktheonlything239-ops", "idle_days": 101, "models": "", "url": "https://github.com/huggingface/transformers/issues/46222"}, {"key": "pr#46211", "n": 46211, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "fix(llama): allow explicit head_dim when hidden_size not divisible by num_attention_heads", "action": "review requested: @Rocketknight1; maintainer pinged", "comment": "@Rocketknight1 could you take a look when you get a chance? This is waiting on a maintainer: it has been waiting for a maintainer review.", "author": "Sriniketh24", "idle_days": -1, "models": "arcee, aria, cwm, deepseek_v2, eurobert, higgs_audio_v2, hrm_text, jais2, llama", "url": "https://github.com/huggingface/transformers/pull/46211"}, {"key": "pr#46177", "n": 46177, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tokenization", "type": "feature / refactor", "title": "Tighten exception handling in find_sentencepiece_model_file", "action": "review requested: @itazap", "comment": "", "author": "Aditya-ad48", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/46177"}, {"key": "pr#46174", "n": 46174, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "fix(mps): build_2d_sinusoidal_position_embedding crashes on Apple Silicon due to float64 on MPS device", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #47897, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46174"}, {"key": "pr#46169", "n": 46169, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "feature / refactor", "title": "[LED][Longformer] Replace for-loop with unfold in _chunk ONNX-export path", "action": "review requested: @Cyrilvallez; maintainer pinged", "comment": "@Cyrilvallez could you take a look when you get a chance? This is waiting on a maintainer: never reviewed.", "author": "guinik", "idle_days": -1, "models": "led, longformer", "url": "https://github.com/huggingface/transformers/pull/46169"}, {"key": "pr#46161", "n": 46161, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Fix tokenizer subfolder fallback", "action": "closed", "comment": "Thanks @Prachi-kushwaha! Silently falling back to a subfolder changes loading semantics, and this didn't get traction or a linked issue, so I'm closing it. If you have a concrete Hub repo that fails to load, please open an issue with it so we can discuss the right fix.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46161"}, {"key": "pr#46160", "n": 46160, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "[ROCm] Fix gpt_oss test suite: AITER attention fallback, SonicMoE guard, and distributed test fixes", "action": "closed", "comment": "Thanks for the ROCm fixes, @Abdennacer-Badaoui! gpt-oss on ROCm now routes FA to `kernels-community/aiter-flash-attn` with regenerated fixtures (#46837), in line with @vasqu's hub-kernels suggestion, so I'm closing this one. Anything still failing can go in a smaller follow-up.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46160"}, {"key": "pr#46152", "n": 46152, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / utils", "type": "feature / refactor", "title": "Avoid reading self source code via file open()", "action": "\u2014", "comment": "", "author": "rasmi", "idle_days": 145, "models": "", "url": "https://github.com/huggingface/transformers/pull/46152"}, {"key": "pr#46148", "n": 46148, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "[Qwen3Next] preserve linear-attn-mask optimization under torch.compile/export", "action": "closed", "comment": "Thanks for the export fix, @yuvrajsharma9981! The linear-attention mask code this changes was rewritten in #47630 (`_update_linear_attn_mask` is gone from Qwen3-Next), so I'm closing this one. If `torch.export` still fails on main, please open an issue with the trace.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46148"}, {"key": "issue#46144", "n": 46144, "kind": "issue", "group": "Open issues", "category": "feature", "area": "attention / kernels", "type": "feature", "title": "RoFormer attention implementation does not use attention interface", "action": "\u2014", "comment": "", "author": "ir2718", "idle_days": 140, "models": "", "url": "https://github.com/huggingface/transformers/issues/46144"}, {"key": "pr#46138", "n": 46138, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "chore: update self-comment-ci.yml", "action": "\u2014", "comment": "", "author": "hf-security-analysis", "idle_days": 152, "models": "", "url": "https://github.com/huggingface/transformers/pull/46138"}, {"key": "pr#46128", "n": 46128, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add SAM 3.1 to Transformers", "action": "maintainer pinged", "comment": "@yonigozlan this one has been waiting since 2026-06-25. Still planned, or should we close it?", "author": "yonigozlan", "idle_days": -1, "models": "sam3, sam3_1_tracker_video, sam3_1_video, sam3_tracker_video, sam3_video", "url": "https://github.com/huggingface/transformers/pull/46128"}, {"key": "pr#46127", "n": 46127, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "[deepseek_v4] Add DeepseekV4NextNPredictor class for MTP draft-head support", "action": "closed", "comment": "Thanks for the detailed work on DeepSeek-V4 MTP, @pasta-paul! MTP is now supported generically in `generate` (#46229) with `MtpModel` helpers for external drafting loops (#49544), so a model-specific predictor class isn't the direction we want. Closing this; wiring DeepSeek-V4 into that generic MTP path would be very welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46127"}, {"key": "pr#46113", "n": 46113, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "clip_grad_norm avoid O(global_params) replication", "action": "closed", "comment": "Thanks for the careful profiling, @AmineDiro! The `distributed` branch this targets is no longer used (#46269 was closed), and the native distributed stack landed on `main` directly. The DTensor `clip_grad_norm` from #45028 was reverted in #46246, and mixed-mesh gradient norms are now computed per mesh in `Trainer._mixed_mesh_grad_norm` (#48205), which avoids the full replication. If you still see O(global_params) memory on `main`, please open a new PR against `main` with your Qwen3-30B-A3B numbers. Closing this one, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46113"}, {"key": "pr#46048", "n": 46048, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "new model", "title": "Add TDT loss kernel", "action": "\u2014", "comment": "", "author": "ebezzam", "idle_days": 6, "models": "parakeet", "url": "https://github.com/huggingface/transformers/pull/46048"}, {"key": "pr#46036", "n": 46036, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Widen tols for float16/bfloat16", "action": "closed", "comment": "Hey @Rocketknight1, #48815 relaxed the bf16/fp16 tolerances in both test_generate_with_static_cache tests (generation common + gemma3n), so this now conflicts. If the flakes still show up at 5e-5, could you rebase with the extra widening? Otherwise let's close this.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46036"}, {"key": "pr#46031", "n": 46031, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Support Granite Speech NAR (NLE)", "action": "\u2014", "comment": "", "author": "avihu111", "idle_days": 1, "models": "granite_speech, granite_speech_nar, granite_speech_plus", "url": "https://github.com/huggingface/transformers/pull/46031"}, {"key": "pr#46024", "n": 46024, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "fix(llama4): align MoE interface for EP/TP compatibility", "action": "review requested: @3outeille; maintainer pinged", "comment": "@3outeille could you take a look when you get a chance? This is waiting on a maintainer: ArthurZucker re-review; author addressed both inline points 2026-06-22 and proposed expert_implementation as a follow-up.", "author": "Chao1Han", "idle_days": -1, "models": "llama4", "url": "https://github.com/huggingface/transformers/pull/46024"}, {"key": "pr#46013", "n": 46013, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "attention / kernels", "type": "feature / refactor", "title": "docs: document flash attention supports_mapping", "action": "\u2014", "comment": "", "author": "nightcityblade", "idle_days": 155, "models": "", "url": "https://github.com/huggingface/transformers/pull/46013"}, {"key": "pr#46011", "n": 46011, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "new model", "title": "[new model] Add zaya1 vl", "action": "\u2014", "comment": "", "author": "JJJYmmm", "idle_days": 156, "models": "zaya, zaya1_vl", "url": "https://github.com/huggingface/transformers/pull/46011"}, {"key": "pr#46007", "n": 46007, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "docs: fix OOM/hang for Qwen dynamic resolution models by setting min_pixels and max_pixels", "action": "review requested: @stevhliu; maintainer pinged", "comment": "@stevhliu you approved this earlier; could you (or the right owner) take it to merge? .", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46007"}, {"key": "pr#46004", "n": 46004, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[Image Processor] Speed up image processors by casting to array before BatchFeature", "action": "closed", "comment": "Thanks for looking at image processor speed, @Apeksha23-hub! @zucchini-nlp asked for benchmarks and noted `BatchFeature` handles lists on purpose for some models (https://github.com/huggingface/transformers/pull/46004#pullrequestreview-4306795141); with no follow-up, I'm closing this. Feel free to open a new PR with benchmarks.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/46004"}, {"key": "pr#45996", "n": 45996, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "[Weight converter] Account for base model prefix in scoped weight conversion", "action": "closed", "comment": "Thanks @yonigozlan! This was carried over and merged in #46067, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45996"}, {"key": "pr#45985", "n": 45985, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Fix decoder_attention_mask None handling in generation utils", "action": "closed", "comment": "Thanks @damodharg6! @Rocketknight1 ran the provided snippet on main without hitting the issue (https://github.com/huggingface/transformers/pull/45985#issuecomment-4460727164), so I'm closing this. If you find a reproducer, please open an issue with it.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45985"}, {"key": "pr#45982", "n": 45982, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "no empty label when Grounding Dino detects nothing", "action": "\u2014", "comment": "", "author": "catwell", "idle_days": 155, "models": "grounding_dino", "url": "https://github.com/huggingface/transformers/pull/45982"}, {"key": "pr#45977", "n": 45977, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "GgufLinear: inference-time GGUF matmul on Apple Silicon \u2014 llama.cpp parity", "action": "closed", "comment": "Closing this one: inference-time GGUF matmul landed through #47779 (`GgufLinear` in `integrations/gguf/utils.py`, kernels in `integrations/gguf/kernels.py`, `quantizer_gguf.py`), then #48529 for MoE. Anything still missing from here (e.g. llama.cpp parity tests or arch coverage) can go in a small follow-up on top of that code.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45977"}, {"key": "pr#45975", "n": 45975, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "GGUF: optional Metal dequant fast path via kernels-community", "action": "closed", "comment": "Closing: the Metal GGUF fast path landed with the GGUF refactor in #47779, and this PR's base branch (#44794) was closed without merging.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45975"}, {"key": "issue#45938", "n": 45938, "kind": "issue", "group": "Open issues", "category": "bug", "area": "models", "type": "bug", "title": "[DeepSeekV4] Compressor does not seem to account for padding tokens when forming compressed KV blocks", "action": "labelled: bug", "comment": "", "author": "WKQ9411", "idle_days": -1, "models": "deepseek_v4", "url": "https://github.com/huggingface/transformers/issues/45938"}, {"key": "pr#45917", "n": 45917, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "fix", "title": "Prevent Whisper long-form crashes after Trainer.train()", "action": "\u2014", "comment": "", "author": "1fanwang", "idle_days": 20, "models": "", "url": "https://github.com/huggingface/transformers/pull/45917"}, {"key": "pr#45883", "n": 45883, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix Gemma4 inputs_embeds OOM during per-layer lookup", "action": "closed", "comment": "Thanks for the report and fix! On `main`, Gemma4 now skips the reverse lookup whenever `per_layer_inputs` is passed, which covers the OOM case, and we'd like to keep `inputs_embeds`-only generation working, so closing this.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45883"}, {"key": "pr#45868", "n": 45868, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "feat(t5gemma2): add Flash Attention 2 support", "action": "closed", "comment": "Thanks for working on this, @AjAyrAo43! As @vasqu explained, T5Gemma2's merged attention can't actually run with flash attention, so marking it as FA-supported while silently falling back to eager would be misleading. I'm closing this for now. If a real FA path for merged attention becomes possible, we'd be happy to look at a new PR.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45868"}, {"key": "pr#45848", "n": 45848, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "docs", "type": "docs", "title": "Add Maximal Update Parametrization (\u03bcP)", "action": "\u2014", "comment": "", "author": "qgallouedec", "idle_days": 131, "models": "", "url": "https://github.com/huggingface/transformers/pull/45848"}, {"key": "pr#45840", "n": 45840, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "other", "type": "fix", "title": "fix: ModuleNotFoundError caused by distributed race condition in ci/cd", "action": "\u2014", "comment": "", "author": "rigen1048", "idle_days": 135, "models": "", "url": "https://github.com/huggingface/transformers/pull/45840"}, {"key": "pr#45825", "n": 45825, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "CI / tooling", "type": "CI / tooling", "title": "[CI] Replace PAT with GitHub App token in repo-consistency-bot", "action": "\u2014", "comment": "", "author": "paulinebm", "idle_days": 165, "models": "", "url": "https://github.com/huggingface/transformers/pull/45825"}, {"key": "pr#45819", "n": 45819, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "core / utils", "type": "fix", "title": "fix: use per-instance lru_cache in compile_compatible_method_lru_cache to prevent memory leak", "action": "review requested: @yonigozlan; maintainer pinged", "comment": "@yonigozlan could you take a look when you get a chance? This is waiting on a maintainer: yonigozlan said he'd fix it himself after the ViT refactor (#41693, merged 2026-05-08).", "author": "fyhon", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/45819"}, {"key": "pr#45814", "n": 45814, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tests", "type": "CI / tooling", "title": "Add pretrained model-family distributed smoke matrix", "action": "\u2014", "comment": "", "author": "winglian", "idle_days": 132, "models": "", "url": "https://github.com/huggingface/transformers/pull/45814"}, {"key": "pr#45776", "n": 45776, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "attention / kernels", "type": "feature / refactor", "title": "Gate enable_gqa=True on actual flash-attention eligibility", "action": "\u2014", "comment": "", "author": "dvdimitrov13", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/pull/45776"}, {"key": "pr#45773", "n": 45773, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix UnboundLocalError for `is_updated` in encoder-decoder cross-attention", "action": "closed", "comment": "Thanks for the fix, and for surfacing the speculative-decoding issue behind it! #48000 fixed the same `UnboundLocalError` together with the cache corruption @vasqu mentioned, so closing this as superseded.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45773"}, {"key": "pr#45756", "n": 45756, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "fix attribute access in PermuteForRope._apply", "action": "closed", "comment": "Thanks @CharlieKerfoot! The same fix landed in #47517: `PermuteForRope` in `src/transformers/core_model_loading.py` now uses `getattr(self.config, \"num_attention_heads\", 1)`, so closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45756"}, {"key": "pr#45751", "n": 45751, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add nvidia Conformer-CTC", "action": "\u2014", "comment": "", "author": "jonghwanhyeon", "idle_days": 147, "models": "conformer", "url": "https://github.com/huggingface/transformers/pull/45751"}, {"key": "pr#45745", "n": 45745, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "feat: add crop() to StaticCache layers for assisted generation", "action": "closed", "comment": "Thanks @ArkaD171717 for the detailed work! As @Cyrilvallez explained in https://github.com/huggingface/transformers/pull/45745#pullrequestreview-4545195720, assisted decoding doesn't trigger compilation with a static cache and `crop` can't work with sliding-window layers, so we'll pass on this; closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45745"}, {"key": "pr#45742", "n": 45742, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "feat(llama): add has_weight parameter to LlamaRMSNorm for FlashNorm-folded checkpoints (+12.77% e2e on Llama-3.2-1B at bf16/A100)", "action": "closed", "comment": "Thanks for the clever idea! As discussed, we don't want to add this flag to Llama's modeling code (it's inherited by many models), but fusion mapping (docs/source/en/fusion_mapping.md) is a good fit. Closing here, happy to review a fusion-mapping version.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45742"}, {"key": "pr#45738", "n": 45738, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "fix(musicgen_melody): use DynamicCache instead of EncoderDecoderCache", "action": "closed", "comment": "Thanks @adityachoksi2512 for bisecting this! The MusicgenMelody conditioning bug was fixed in #48679, so closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45738"}, {"key": "pr#45737", "n": 45737, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Add regression test for MusicgenMelody audio conditioning (GH #45647)", "action": "closed", "comment": "Thanks for the regression test, @voodoovampire! The MusicgenMelody conditioning bug from #45647 was fixed in #48679, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45737"}, {"key": "pr#45730", "n": 45730, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "gpt_oss multi-GPU AMD support", "action": "closed", "comment": "Thanks for looking into AMD multi-GPU support! @ArthurZucker wasn't convinced this change addresses the root cause, and it hasn't moved since, so I'm closing it. If it still fails on `main`, an issue with the full traceback and setup would help us find the proper fix.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45730"}, {"key": "pr#45729", "n": 45729, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "fix: Add prefer_auto_map opt-in for registered-class vs auto_map conflicts", "action": "\u2014", "comment": "", "author": "sebastianvlad1", "idle_days": 139, "models": "", "url": "https://github.com/huggingface/transformers/pull/45729"}, {"key": "pr#45720", "n": 45720, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "feature / refactor", "title": "chore(mlinter): implement part of rule TRF003", "action": "maintainer pinged", "comment": "@tarekziade this one has been waiting since 2026-05-04. Still planned, or should we close it?", "author": "tarekziade", "idle_days": -1, "models": "bloom, data2vec, deberta, deberta_v2, falcon, hubert, longformer, longt5, mbart, mt5, sew, sew_d, t5, umt5, unispeech, unispeech_sat, wav2vec2, wav2vec2_bert, wav2vec2_conformer, wavlm, xlnet", "url": "https://github.com/huggingface/transformers/pull/45720"}, {"key": "pr#45718", "n": 45718, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix check_auto.py wiping auto_mappings.py due to missing flush", "action": "closed", "comment": "Thanks for the fix! `utils/check_auto.py` was rewritten in #45774 to write through a `tempfile.TemporaryDirectory` instead of a `NamedTemporaryFile`, so the missing flush is no longer an issue. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45718"}, {"key": "pr#45714", "n": 45714, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "fix: restore vocabulary loading in CamembertTokenizer", "action": "closed", "comment": "Thanks for looking into this! In v5 we enforce the tokenizer template rather than always reading `vocab_file`, so this isn't the direction we want, as noted in review. Closing, but thanks for the report.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45714"}, {"key": "pr#45709", "n": 45709, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Added SWA and Linear Attention to Llama3.2-1b", "action": "closed", "comment": "Thanks for sharing the experiment! Architecture variants don't get added as branches to an existing model's file; they would need to be a new model (via a modular file) or Hub remote code, so we're closing this. Pretty sure SWA + Linear attention is a QWEN that already exists!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45709"}, {"key": "pr#45707", "n": 45707, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "Fix padding calculation in new_conv_state assignment", "action": "closed", "comment": "Thanks @yileld! The `new_conv_state` code this touches no longer exists in `modeling_qwen3_5.py` on main, so the diff no longer applies; closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45707"}, {"key": "pr#45683", "n": 45683, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "CI / tooling", "type": "CI / tooling", "title": "Exclude audio modules from conversion process", "action": "\u2014", "comment": "", "author": "softguy777", "idle_days": 109, "models": "zamba2", "url": "https://github.com/huggingface/transformers/pull/45683"}, {"key": "pr#45682", "n": 45682, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "integrations (other)", "type": "fix", "title": "FIX Restore LoRA hotswapping functionality", "action": "\u2014", "comment": "", "author": "BenjaminBossan", "idle_days": 92, "models": "", "url": "https://github.com/huggingface/transformers/pull/45682"}, {"key": "pr#45670", "n": 45670, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "[nit] glmasr should be in AutoModelForMultimodalLM", "action": "closed", "comment": "Closing since GLM-ASR was moved into `AutoModelForMultimodalLM` as part of #45534.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45670"}, {"key": "pr#45668", "n": 45668, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "[GGUF] Add support for Qwen3.5 MoE (qwen35moe arch)", "action": "closed", "comment": "Thanks @lucaspirola for this, and for your patience while we reworked GGUF. Qwen3.5 MoE (`qwen35moe`) GGUF loading has now landed through the new GGUF API in #48529, so I'm closing this as superseded. Please open an issue if anything in your checkpoints doesn't load as expected!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45668"}, {"key": "pr#45662", "n": 45662, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix EP + FSDP2: experts silently overwritten by rank-0 broadcast", "action": "review requested: @SunMarc; maintainer pinged; closed", "comment": "Thanks for the careful investigation! EP + FSDP2 now has native support (#48516): the model is sharded at load time on a 2-D device mesh, so the rank-0 broadcast path this PR patches is no longer used. Closing; if experts are still overwritten on current `main`, please reopen or open a new PR with a repro.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45662"}, {"key": "pr#45649", "n": 45649, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "core / loading", "type": "fix", "title": "Fix OOM regression for FSDP2 + cpu_ram_efficient_loading on large models", "action": "\u2014", "comment": "", "author": "AmineDiro", "idle_days": 8, "models": "", "url": "https://github.com/huggingface/transformers/pull/45649"}, {"key": "pr#45645", "n": 45645, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "Fix xdist collisions for captured_info artifacts and preserve CI debug logs", "action": "closed", "comment": "Thanks for looking into the xdist `captured_info.txt` collisions, @stationeros! `.github/workflows/model_jobs.yml` was removed when CI moved to the transformers-ci daily workflow (#47360), so this can't land as is; closing. If the collision still happens with the current CI, a focused fix in `testing_utils.py` would be welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45645"}, {"key": "pr#45639", "n": 45639, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "fix", "title": "Make patched testing debug logs xdist-safe", "action": "\u2014", "comment": "", "author": "oleksii-tumanov", "idle_days": 120, "models": "", "url": "https://github.com/huggingface/transformers/pull/45639"}, {"key": "issue#45636", "n": 45636, "kind": "issue", "group": "Open issues", "category": "feature", "area": "attention / kernels", "type": "feature", "title": "Proposal: add sdpa_memeff attn_implementation for shape combinations no fast backend covers", "action": "labelled: Feature request", "comment": "", "author": "dvdimitrov13", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/45636"}, {"key": "pr#45635", "n": 45635, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / loading", "type": "feature / refactor", "title": "qa: speed up dtype regex weight load + reduce dtype tests to 3 random", "action": "\u2014", "comment": "", "author": "tarekziade", "idle_days": 162, "models": "", "url": "https://github.com/huggingface/transformers/pull/45635"}, {"key": "pr#45614", "n": 45614, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "Add missing requests dependency to transformers[serving]", "action": "\u2014", "comment": "", "author": "Oneirag", "idle_days": 8, "models": "", "url": "https://github.com/huggingface/transformers/pull/45614"}, {"key": "pr#45609", "n": 45609, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "make it possible to ser/deser HF MoE models with torchao", "action": "closed", "comment": "Closing since TorchAO save/load for non-`weight` params like MoE experts landed in #46325 and `quantizer_torchao.py` has been reworked since. Thanks @vkuzo! Please reopen against main if a gap remains.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45609"}, {"key": "pr#45604", "n": 45604, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "serving / continuous batching", "type": "feature / refactor", "title": "Agent first cli with skill", "action": "\u2014", "comment": "", "author": "LysandreJik", "idle_days": 137, "models": "", "url": "https://github.com/huggingface/transformers/pull/45604"}, {"key": "pr#45586", "n": 45586, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "new model", "title": "Add Audio-Visual Flamingo model", "action": "review requested: @eustlb; maintainer pinged", "comment": "@eustlb could you take a look when you get a chance? This is waiting on a maintainer: eustlb review pending (promised 2026-09-25); stray root files c.sh/main.py/nvidia.mp4 should be removed.", "author": "lashahub", "idle_days": -1, "models": "audiovisualflamingo", "url": "https://github.com/huggingface/transformers/pull/45586"}, {"key": "pr#45578", "n": 45578, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Remove attribute_map from GptOssConfig", "action": "closed", "comment": "Closing as discussed: the original PEFT/TRL loading issue was fixed independently (huggingface/trl#5622), so removing `attribute_map` isn't needed anymore. Thanks @AmineDiro!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45578"}, {"key": "issue#45571", "n": 45571, "kind": "issue", "group": "Open issues", "category": "feature", "area": "exporters", "type": "feature", "title": "A nice UX for generating dynamic tensors that break torch.compile/torch.export", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": 177, "models": "", "url": "https://github.com/huggingface/transformers/issues/45571"}, {"key": "pr#45570", "n": 45570, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "Fix whisper long-form generation when eos_token_id is a list", "action": "\u2014", "comment": "", "author": "ronansgd", "idle_days": 181, "models": "whisper", "url": "https://github.com/huggingface/transformers/pull/45570"}, {"key": "pr#45569", "n": 45569, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "new model", "title": "Proper nemotron H and 3 and 2", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 138, "models": "nemotron_h, nemotron_h_dense, nemotron_h_sparse", "url": "https://github.com/huggingface/transformers/pull/45569"}, {"key": "pr#45550", "n": 45550, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "CI / tooling", "type": "CI / tooling", "title": "Add runner selection for mi325 GPU type", "action": "review requested: @ydshieh; maintainer pinged", "comment": "@ydshieh could you take a look when you get a chance? This is waiting on a maintainer: needs review from ydshieh/tarekziade.", "author": "glegendre01", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/45550"}, {"key": "pr#45549", "n": 45549, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "processors (image/video/audio)", "type": "fix", "title": "fix: apply channel averaging correctly in audio feature extractors", "action": "\u2014", "comment": "", "author": "jonghwanhyeon", "idle_days": 181, "models": "cohere_asr, lasr, parakeet, phi4_multimodal, voxtral_realtime", "url": "https://github.com/huggingface/transformers/pull/45549"}, {"key": "pr#45548", "n": 45548, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix EP + DeepSpeed ZeRO-3 loading via accelerate launch", "action": "closed", "comment": "Thanks for digging into this! Expert parallelism was reworked since (`ep_size` with separate dense/expert meshes, #48857, #48859, #49633) and both files now conflict, so this can't land as is. Closing; if ZeRO-3 + EP loading still fails on current `main`, please open a new PR or issue with a repro.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45548"}, {"key": "pr#45546", "n": 45546, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "gguf", "type": "feature / refactor", "title": "feat: Add GGUF loading support for Llama 4 (text)", "action": "\u2014", "comment": "", "author": "garybadwal", "idle_days": 4, "models": "", "url": "https://github.com/huggingface/transformers/pull/45546"}, {"key": "pr#45543", "n": 45543, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "ci: OTEL support", "action": "closed", "comment": "Closing this OTEL CI draft: OTEL CI landed in #47360, which also removed `model_jobs.yml` and `self-scheduled.yml` that this PR edits.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45543"}, {"key": "pr#45541", "n": 45541, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "tokenization", "type": "fix", "title": "Fix local_files_only tokenizer fallback when tokenizer files are missing (Issue 45538)", "action": "\u2014", "comment": "", "author": "Brianzhengca", "idle_days": 5, "models": "", "url": "https://github.com/huggingface/transformers/pull/45541"}, {"key": "pr#45523", "n": 45523, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Fix Seq2SeqLM ExecuTorch export: add encoder_attention_mask to decoder and use static encoder shapes", "action": "closed", "comment": "Thanks @duyhv-qualgo! @Cyrilvallez's request for a minimal reproducer (https://github.com/huggingface/transformers/pull/45523#issuecomment-4428375663) was never answered, so closing for now. Please reopen with a repro if T5 ExecuTorch export is still broken on main.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45523"}, {"key": "pr#45512", "n": 45512, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "feature / refactor", "title": "[OutputRecorder] re.search on layer_name", "action": "maintainer pinged", "comment": "@eustlb this one has been waiting since 2026-04-19 (BC question from zvik (2026-04-19): add a separate regex option instead of changing substring matching). Still planned, or should we close it?", "author": "eustlb", "idle_days": -1, "models": "blip, blip_2, decision_transformer, gpt2, instructblip, instructblipvideo, minimax, qwen3_omni_moe, qwen3_vl_moe, switch_transformers, vjepa2", "url": "https://github.com/huggingface/transformers/pull/45512"}, {"key": "pr#45497", "n": 45497, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add V-JEPA 2.1 inference support", "action": "\u2014", "comment": "", "author": "davevanveen", "idle_days": 49, "models": "vjepa2", "url": "https://github.com/huggingface/transformers/pull/45497"}, {"key": "issue#45496", "n": 45496, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "Add V-JEPA 2.1 inference support", "action": "\u2014", "comment": "", "author": "davevanveen", "idle_days": 21, "models": "", "url": "https://github.com/huggingface/transformers/issues/45496"}, {"key": "pr#45490", "n": 45490, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add ctsm model", "action": "\u2014", "comment": "", "author": "kashif", "idle_days": 120, "models": "ctsm, timesfm, timesfm2_5", "url": "https://github.com/huggingface/transformers/pull/45490"}, {"key": "pr#45462", "n": 45462, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "CI / tooling", "type": "CI / tooling", "title": "chore(sec): added a handful of security checks", "action": "maintainer pinged", "comment": "@tarekziade this one has been waiting since 2026-06-11. Still planned, or should we close it?", "author": "tarekziade", "idle_days": -1, "models": "ernie4_5, ernie4_5_vl_moe, hubert, sew, sew_d, videomt", "url": "https://github.com/huggingface/transformers/pull/45462"}, {"key": "pr#45453", "n": 45453, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "core / loading", "type": "feature / refactor", "title": "Draft commit", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 183, "models": "", "url": "https://github.com/huggingface/transformers/pull/45453"}, {"key": "pr#45452", "n": 45452, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "refactor: replace wildcard imports with explicit imports in model __init__.py files", "action": "closed", "comment": "Thanks for the effort here! As @LysandreJik explained in https://github.com/huggingface/transformers/pull/45452#issuecomment-4251890274, each module already declares its exports via `__all__`, and duplicating them in every `__init__.py` adds maintenance cost, so we won't take this change. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45452"}, {"key": "pr#45422", "n": 45422, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "tokenization", "type": "feature / refactor", "title": "Drop `content=None` from messages in `apply_chat_template`", "action": "\u2014", "comment": "", "author": "qgallouedec", "idle_days": 43, "models": "", "url": "https://github.com/huggingface/transformers/pull/45422"}, {"key": "pr#45421", "n": 45421, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "Improve nested `base_model_prefix` handling in weight conversion and loading", "action": "closed", "comment": "Closing per the discussion with @Cyrilvallez and @ArthurZucker: we don't want to support multi-level `base_model_prefix` remapping in general, and RF-DETR landed in #36895 without it. Thanks @yonigozlan!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45421"}, {"key": "issue#45419", "n": 45419, "kind": "issue", "group": "Open issues", "category": "feature", "area": "tokenization", "type": "feature", "title": "Chat template inconsistencies in tool-calling support", "action": "labelled: Feature request", "comment": "", "author": "qgallouedec", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/45419"}, {"key": "pr#45415", "n": 45415, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "Adds type checking to `src/transformers/*py`", "action": "maintainer pinged; commented", "comment": "Hey @tarekziade, since this was opened, ty adoption moved to the per-file allow-list in utils/check_types.py (#45425, #46278 and the follow-ups), and the pyproject rule added here is already on main. This 42-file diff now conflicts. Could you split what's left into the incremental per-file PRs, or close this if those already cover it?", "author": "tarekziade", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/45415"}, {"key": "pr#45401", "n": 45401, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add support for Voxtral-4B-TTS-2603 to transformers", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! It has been inactive for a long time and the codebase has changed a lot since (v5), so closing it. Feel free to open a fresh PR against `main` if it's still relevant.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45401"}, {"key": "pr#45382", "n": 45382, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Add AudioGen (AudioCraft) to MusicGen conversion scripts", "action": "closed", "comment": "Closing this one: no activity since 2026-04-12, 2 of the 3 files it changes no longer exist on `main`. Thanks for the contribution, and feel free to open a fresh PR against `main` if it's still relevant!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45382"}, {"key": "pr#45379", "n": 45379, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "fix(config): add deepstack_visual_indexes to Qwen3_5MoeVisionConfig", "action": "closed", "comment": "Thanks for the report, @hijingsong! @zucchini-nlp couldn't reproduce the dropped `deepstack_visual_indexes` on main (https://github.com/huggingface/transformers/pull/45379#issuecomment-4237648245) and we didn't hear back on the setup, so I'm closing this for now. If you can still reproduce it on the latest release, please share a minimal script in a new issue.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45379"}, {"key": "pr#45378", "n": 45378, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "fix(mistral): guard ReasoningEffort import for older mistral_common versions", "action": "closed", "comment": "Thanks @hijingsong! As @zucchini-nlp mentioned (https://github.com/huggingface/transformers/pull/45378#issuecomment-4236307668), `setup.py` pins `mistral-common[image]>=1.11.7` on main, so upgrading the package is the fix; closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45378"}, {"key": "pr#45360", "n": 45360, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Replace deprecated `huggingface-cli` references with `hf`", "action": "closed", "comment": "Thanks @hanouticelina! These `huggingface-cli` references were replaced in #48988, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45360"}, {"key": "pr#45346", "n": 45346, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix Double Application of Softmax for Router Logits in MoE models", "action": "closed", "comment": "Thanks for working on this! As @vasqu noted, #45131 fixed the router softmax issue, so closing this as superseded.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45346"}, {"key": "pr#45332", "n": 45332, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "integrations (other)", "type": "feature / refactor", "title": "Add generic heterogeneous model support (per-layer config and modeling)", "action": "\u2014", "comment": "", "author": "eladsegal", "idle_days": 1, "models": "", "url": "https://github.com/huggingface/transformers/pull/45332"}, {"key": "pr#45321", "n": 45321, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Remove references to torchao's AffineQuantizedTensor", "action": "closed", "comment": "Thanks for starting this! The AffineQuantizedTensor removal landed in #47797 (which followed your approach), so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45321"}, {"key": "pr#45317", "n": 45317, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "tokenization", "type": "fix", "title": "Add regression test for _patch_mistral_regex with fix_mistral_regex=True", "action": "\u2014", "comment": "", "author": "mohdfaour03", "idle_days": 14, "models": "", "url": "https://github.com/huggingface/transformers/pull/45317"}, {"key": "pr#45300", "n": 45300, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix Nemotron-H: add mlp layer type support", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #44763, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45300"}, {"key": "pr#45296", "n": 45296, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Add GGUF support to Gemma4 (31B & 26B-A4B) text ", "action": "closed", "comment": "Thanks for the Gemma4 GGUF support! The `gemma4` GGUF mapping has since landed on `main` (#46449), so closing this one. If something specific to the 26B-A4B MoE variant is still missing, a focused follow-up PR would be welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45296"}, {"key": "pr#45294", "n": 45294, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "feat: add Gemma4ForSequenceClassification", "action": "closed", "comment": "Thanks for adding this, @jesperschlegel! As @Cyrilvallez noted in https://github.com/huggingface/transformers/pull/45294#issuecomment-4302015703, no checkpoints were released for a Gemma 4 sequence-classification head, so I'm closing this one. If you have a model that needs it, feel free to share it on the Hub with custom code.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45294"}, {"key": "pr#45293", "n": 45293, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "fix", "title": "Fix \"AttributeError: NewTokenizer has no attribute special_attribute_present\" (Remove `REGISTERED_FAST_ALIASES`)", "action": "\u2014", "comment": "", "author": "yonigozlan", "idle_days": 161, "models": "", "url": "https://github.com/huggingface/transformers/pull/45293"}, {"key": "pr#45270", "n": 45270, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[Trainer] Support multi-loss component logging", "action": "closed", "comment": "Closing this one: no activity since 2026-04-06. Thanks for the contribution, and feel free to open a fresh PR against `main` if it's still relevant!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45270"}, {"key": "pr#45267", "n": 45267, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Add docstring to FFN.forward in DistilBERT", "action": "closed", "comment": "Thanks @KetanP1618! As @Rocketknight1 explained (https://github.com/huggingface/transformers/pull/45266#issuecomment-4206688161), we're avoiding small docstring-only changes for now, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45267"}, {"key": "pr#45254", "n": 45254, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Fix more integration tests for important models", "action": "closed", "comment": "Closing: this PR has no changed files left (0 additions, 0 deletions).", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45254"}, {"key": "pr#45244", "n": 45244, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Let's CI go great", "action": "closed", "comment": "Closing this CI-only PR: it has no file changes and was only used to trigger a CI run.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45244"}, {"key": "pr#45233", "n": 45233, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "feat: make timesfm2_5 onnx export compatible", "action": "\u2014", "comment": "", "author": "pdufour", "idle_days": 161, "models": "timesfm, timesfm2_5", "url": "https://github.com/huggingface/transformers/pull/45233"}, {"key": "pr#45218", "n": 45218, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "serving / continuous batching", "type": "feature / refactor", "title": "Proposal: Agent-first CLI", "action": "\u2014", "comment": "", "author": "LysandreJik", "idle_days": 139, "models": "", "url": "https://github.com/huggingface/transformers/pull/45218"}, {"key": "pr#45213", "n": 45213, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "new model", "title": "DO NOT MERGE - model creation skill", "action": "\u2014", "comment": "", "author": "tarekziade", "idle_days": 83, "models": "persimmon, sarvam_moe, stablelm", "url": "https://github.com/huggingface/transformers/pull/45213"}, {"key": "pr#45202", "n": 45202, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix gemma4 has flash-attention incompatbile head-dim=512", "action": "closed", "comment": "Gemma4 already has `_supports_flash_attn = False` on main (see `src/transformers/models/gemma4/modular_gemma4.py`), as agreed in https://github.com/huggingface/transformers/pull/45202#pullrequestreview-4159428337 until FA kernels support head_dim 512, so closing. Thanks @Qubitium for flagging it!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45202"}, {"key": "pr#45193", "n": 45193, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / loading", "type": "feature / refactor", "title": "Config can apply pyndatic validation without torch-dependence", "action": "\u2014", "comment": "", "author": "zucchini-nlp", "idle_days": 176, "models": "", "url": "https://github.com/huggingface/transformers/pull/45193"}, {"key": "pr#45189", "n": 45189, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Add doc test CI workflow reusing existing model job infrastructure", "action": "closed", "comment": "Thanks for putting this together! `.github/workflows/model_jobs.yml` and `.github/workflows/self-scheduled.yml` were removed in #47360 when daily CI moved to the transformers-ci workflow, so these doc-test inputs have nothing to hook into anymore. Closing; if doc tests are still wanted, they would need to be added to the new daily workflow.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45189"}, {"key": "pr#45181", "n": 45181, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Make the cli a top-level package", "action": "closed", "comment": "Thanks @tarekziade for looking into CLI startup time! The draft has been idle for ~200 days and the CLI has changed since; serve handlers are now imported lazily (#46473), but the separate `transformers_cli` package proposed here never landed. If faster `transformers --help` is still a goal, a fresh PR against `main` would be easiest. Closing this one for now.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45181"}, {"key": "pr#45176", "n": 45176, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "added efficietvitsam model to HF", "action": "closed", "comment": "Thanks for this port, @masoudpz! EfficientViT-SAM is being worked on in #47275, which has review going, and this version would need a substantial rework to follow current library standards (a `modular_*.py` file, the attention interface, etc.), so I'm closing it to keep the effort in one place. You're welcome to help review or contribute on #47275.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45176"}, {"key": "issue#45175", "n": 45175, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "Feature request: Add EfficientViT-SAM (efficientvitsam) to Transformers", "action": "\u2014", "comment": "", "author": "masoudpz", "idle_days": 100, "models": "", "url": "https://github.com/huggingface/transformers/issues/45175"}, {"key": "pr#45170", "n": 45170, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "`layrnorm` -> `layernorm`", "action": "maintainer pinged", "comment": "@zucchini-nlp this one has been waiting since 2026-04-01 (was waiting on #44431 (merged 2026-04-09); needs rebase). Still planned, or should we close it?", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45170"}, {"key": "pr#45168", "n": 45168, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "trainer", "type": "feature / refactor", "title": "Update min_lr and max_lr default values to better defaults", "action": "\u2014", "comment": "", "author": "w601sxs", "idle_days": 195, "models": "", "url": "https://github.com/huggingface/transformers/pull/45168"}, {"key": "pr#45167", "n": 45167, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Add anthropic style of function schema", "action": "closed", "comment": "Thanks @xu-song! As @Rocketknight1 explained (https://github.com/huggingface/transformers/pull/45167#issuecomment-4177839815), we're enforcing a single standard tool-schema format for chat templates, so we'll pass on an Anthropic-style output in `get_json_schema`; closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45167"}, {"key": "pr#45153", "n": 45153, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "[`FA`] Native torch integration", "action": "\u2014", "comment": "", "author": "vasqu", "idle_days": 27, "models": "bark, falcon, gemma2, gemma3, gemma3n, gpt_neo, gptj, t5gemma, vaultgemma, videoprism", "url": "https://github.com/huggingface/transformers/pull/45153"}, {"key": "pr#45149", "n": 45149, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "DO NOT MERGE adding SAML3-LiteText with a skill, first pass", "action": "closed", "comment": "Closing this experiment: it was marked do-not-merge and SAM3-LiteText already landed via #44320.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45149"}, {"key": "pr#45147", "n": 45147, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "quantization", "type": "fix", "title": "Fix broken HQQ support", "action": "\u2014", "comment": "", "author": "mobicham", "idle_days": 194, "models": "", "url": "https://github.com/huggingface/transformers/pull/45147"}, {"key": "pr#45134", "n": 45134, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "Optimize Parakeet feature extraction on CUDA", "action": "\u2014", "comment": "", "author": "milesial", "idle_days": 160, "models": "parakeet", "url": "https://github.com/huggingface/transformers/pull/45134"}, {"key": "pr#45133", "n": 45133, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Add sarvam model", "action": "closed", "comment": "Closing this modular experiment: it only adds the config/modeling files plus the committed `sarvam_MODULAR_PROMPT`, with no auto mappings or tests. Thanks @itazap!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45133"}, {"key": "pr#45128", "n": 45128, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Fix: handle future annotations in _process_kwargs_parameters", "action": "closed", "comment": "Thanks for opening this! The diff mostly reformats `src/transformers/utils/auto_docstring.py` and adds an unrelated `all_requirements.txt` and a root-level test script, so it's hard to isolate the actual fix. If `from __future__ import annotations` still breaks `_process_kwargs_parameters` on `main`, please open a minimal PR against `main` with just the fix and a test under `tests/utils/`. Closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45128"}, {"key": "pr#45115", "n": 45115, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "feature / refactor", "title": "Refactor/nemotron h inherit granitemoehybrid", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 204, "models": "nemotron_h", "url": "https://github.com/huggingface/transformers/pull/45115"}, {"key": "pr#45114", "n": 45114, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "docs", "type": "fix", "title": "fix: lets fix all doctests", "action": "maintainer pinged; commented", "comment": "Most of this landed in #47687 (the `auto_docstring.md` and `internal/import_utils.md` doctest fixes), so the branch now conflicts. What is left is the missing `from transformers import pipeline` in `main_classes/pipelines.md` and the small reformat in `tasks/zero_shot_object_detection.md`. Could you rebase down to those two hunks, or close this if they are no longer needed?", "author": "tarekziade", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/45114"}, {"key": "pr#45113", "n": 45113, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "Add GDS support for safetensors loading ", "action": "closed", "comment": "Thanks for the GDS work and benchmarks! As @McPatate said in https://github.com/huggingface/transformers/pull/45113#pullrequestreview-4037725138, GDS support is best added directly in https://github.com/huggingface/safetensors rather than in transformers, so I'm closing this here.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45113"}, {"key": "pr#45110", "n": 45110, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "new model", "title": "Add SAM 3.1", "action": "maintainer pinged", "comment": "@NielsRogge this one has been waiting since 2026-04-23 (CONFLICTING; stray plan.md/progress.md and generation/utils.py edits need cleanup before review). Still planned, or should we close it?", "author": "NielsRogge", "idle_days": -1, "models": "sam3, sam3_1_video", "url": "https://github.com/huggingface/transformers/pull/45110"}, {"key": "pr#45105", "n": 45105, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix @auto_docstring crash with from __future__ import annotations in _process_kwargs_parameters", "action": "closed", "comment": "Thanks @rpathade! The crash from #45103 no longer reproduces on main: since #45139, `_process_kwargs_parameters` skips annotations without `__args__` (which covers string annotations from `from __future__ import annotations`), so I'm closing this as superseded. If you need the kwargs to actually be unrolled into the docstring in that setup, feel free to open a small follow-up on top of current main.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45105"}, {"key": "issue#45099", "n": 45099, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "add HyperCLOVA X SEED Vision Instruct 3B", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B](https://huggingface.co/naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B)\n- Downloads (last 30d): 22.1k \u00b7 Likes: 221 \u00b7 Params: ~3.7B \u00b7 License: other\n- Released: 2025-04 \u00b7 Remote code on the Hub: yes\n- Note: `hyperclovax_vision_v2` is on main; this v1 checkpoint (`hyperclovax_vlm`) still uses remote code.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "bigshanedogg", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/45099"}, {"key": "pr#45086", "n": 45086, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "fix AttributeError in _patch_mistral_regex", "action": "closed", "comment": "Thanks for the fix, @knQzx! The same `_patch_mistral_regex` AttributeError was fixed in #45359, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45086"}, {"key": "pr#45082", "n": 45082, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "[VidEoMT] Update conversion script", "action": "closed", "comment": "Thanks @NielsRogge! VidEoMT landed in #44285, but this update to `convert_videomt_to_hf.py` for the remaining checkpoints never got merged, and the PR also commits a stray `progress.md`. If those checkpoints still need converting, a clean PR against `main` with only the script change would be easy to review. Closing this one for now.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45082"}, {"key": "pr#45073", "n": 45073, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Refactor OwlViT to modular Transformers", "action": "\u2014", "comment": "", "author": "Aravind-11", "idle_days": 186, "models": "owlvit", "url": "https://github.com/huggingface/transformers/pull/45073"}, {"key": "pr#45067", "n": 45067, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "trainer", "type": "feature / refactor", "title": "feat: trainer resume_from_checkpoint support hub downloads (#43375)", "action": "review requested: @SunMarc; maintainer pinged", "comment": "@SunMarc could you take a look when you get a chance? This is waiting on a maintainer: never reviewed.", "author": "aarushisingh04", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/45067"}, {"key": "pr#45064", "n": 45064, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "refactor: shard checkers", "action": "closed", "comment": "Closing as checker grouping already landed on main via #44879 (the `Makefile` targets now run `utils/checkers.py`), as noted in https://github.com/huggingface/transformers/pull/45064#issuecomment-4474902497. Thanks @tarekziade!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45064"}, {"key": "pr#45056", "n": 45056, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "[`auto_docstring`] needs to be only run on __doc__ ", "action": "closed", "comment": "Closing this for now: it's been idle for ~6 months and `src/transformers/utils/auto_docstring.py` has moved a lot since, so the branch would need a full rebase. Lazy class docstrings are still not on `main`, so if the import-time win still holds, a fresh PR against `main` (with the benchmark numbers) is welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45056"}, {"key": "pr#45055", "n": 45055, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Save model config in Trainer checkpoints for non-PreTrainedModel models", "action": "closed", "comment": "Closing this one: no activity since 2026-03-27. Thanks for the contribution, and feel free to open a fresh PR against `main` if it's still relevant!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45055"}, {"key": "pr#45037", "n": 45037, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "add missing colon in custom_attention function signature in attention\u2026", "action": "closed", "comment": "Thanks for spotting the missing colon in `attention_interface.md`, @asuryateja! As @Rocketknight1 explained (https://github.com/huggingface/transformers/pull/45266#issuecomment-4206688161), we're avoiding very small doc/typo PRs right now, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45037"}, {"key": "issue#45027", "n": 45027, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Support Voxtral-4B-TTS-2603 on transformers library", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [mistralai/Voxtral-4B-TTS-2603](https://huggingface.co/mistralai/Voxtral-4B-TTS-2603)\n- Downloads (last 30d): 1,637 \u00b7 Likes: 937 \u00b7 Params: ~4B \u00b7 License: cc-by-nc-4.0\n- Released: 2026-03 \u00b7 Remote code on the Hub: no (Mistral-format weights)\n- A previous attempt (#45401) was closed as stale; feel free to build on it with a fresh PR against `main`.\n- Related models already in transformers: `voxtral`, `voxtral_realtime`.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "MohamedAliRashad", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/45027"}, {"key": "pr#45017", "n": 45017, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "[WIP][Fix] GLM 5 set `apply_rotary_pos_emb` to `is_neox_style=False` && remove `F.relu()`", "action": "closed", "comment": "Thanks @JaredforReal! The interleaved RoPE fix for the GLM-MoE-DSA indexer landed in #46842. For the `F.relu` / FP8 indexer part, as I said it would need to live in a quantization-specific path rather than the default one, so feel free to open a focused PR for that; closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/45017"}, {"key": "pr#44981", "n": 44981, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Trainer: set skip_logits for loss-only eval when liger enabled", "action": "closed", "comment": "Thanks for the fix and the benchmarks! The same change landed in #45273, so closing this one as superseded.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44981"}, {"key": "pr#44973", "n": 44973, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "Fix max_seqlen type in vision attention for torch.compile + FA2", "action": "closed", "comment": "Thanks @andylizf! As @vasqu explained (https://github.com/huggingface/transformers/pull/44973#pullrequestreview-4006974229), main already handles a tensor `max_seqlen` downstream in `modeling_flash_attention_utils.py` and deliberately avoids `.item()` syncs, so we won't add this; closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44973"}, {"key": "issue#44969", "n": 44969, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "Add optional learnable context tokens for CLIP text prompts", "action": "\u2014", "comment": "", "author": "anuj-aj", "idle_days": 210, "models": "", "url": "https://github.com/huggingface/transformers/issues/44969"}, {"key": "pr#44965", "n": 44965, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "try", "action": "closed", "comment": "Thanks @ydshieh! This was a temporary test PR (adding `temp.py`) for the CircleCI API, and it has been idle for over 200 days, so I'm closing it to keep the queue clean.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44965"}, {"key": "issue#44963", "n": 44963, "kind": "issue", "group": "Open issues", "category": "feature", "area": "processors (image/video/audio)", "type": "feature", "title": "Your lazy loading and torchvision dependency in Wav2Vec2/Hubert is breaking PyInstaller builds. Stop forcing torchvision for audio tasks", "action": "\u2014", "comment": "", "author": "vinkenzor001-afk", "idle_days": 209, "models": "", "url": "https://github.com/huggingface/transformers/issues/44963"}, {"key": "issue#44957", "n": 44957, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Add HyperCLOVA X SEED Think 14B", "action": "closed", "comment": "This has been added in #44956 and is available on main. Closing, but feel free to reopen if something is still missing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/44957"}, {"key": "pr#44952", "n": 44952, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix: Add correct return behaviour when output_hidden_states=True for CLIP and SIGLIP vision models", "action": "closed", "comment": "Thanks for the fix! As discussed, #44431 already landed with the same change, so closing as superseded.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44952"}, {"key": "pr#44940", "n": 44940, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Fix tie_weights skipping logic is not tied to model thread scope", "action": "closed", "comment": "Thanks @Qubitium! As @Cyrilvallez noted (https://github.com/huggingface/transformers/pull/44940#pullrequestreview-3990951997), making `from_pretrained` thread-safe needs more than the tie-weights flag (e.g. `guard_torch_init_functions` in `initialization.py`), and multi-processing is the better fit for concurrent loading, so I'm closing this.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44940"}, {"key": "issue#44929", "n": 44929, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "First-class fine-tuning support for Mamba / Mamba-2 SSMs \u2014 architecture is production-ready, but the training path in Transformers isn't", "action": "\u2014", "comment": "", "author": "lochanharishwar", "idle_days": 44, "models": "", "url": "https://github.com/huggingface/transformers/issues/44929"}, {"key": "issue#44928", "n": 44928, "kind": "issue", "group": "Open issues", "category": "bug", "area": "trainer", "type": "bug", "title": "[Bug] Catastrophic gradient explosion (NaN) in RLHF with Qwen3.5 due to 3D position_ids forcing SDPA Math fallback and BF16 collapse", "action": "\u2014", "comment": "", "author": "ouroborosscr", "idle_days": 133, "models": "qwen3_5", "url": "https://github.com/huggingface/transformers/issues/44928"}, {"key": "pr#44923", "n": 44923, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "fix: avoid unconditional model_info call in _patch_mistral_regex", "action": "closed", "comment": "Closing this one: no activity since 2026-03-22, it conflicts with `main`. Thanks for the contribution, and feel free to open a fresh PR against `main` if it's still relevant!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44923"}, {"key": "pr#44893", "n": 44893, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "add `StaticLayer.crop()` to match `DynamicLayer` API", "action": "closed", "comment": "Thanks for the PR! @qgallouedec's requested changes weren't addressed, and the cache `crop` API has since changed (it now only accepts negative values, #47720), so this no longer fits. Closing; a fresh PR against the new API would be welcome if `StaticCache` cropping is still needed.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44893"}, {"key": "pr#44891", "n": 44891, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "docs", "type": "docs", "title": "[Trainer] add MoERouterHealthCallback Callback", "action": "\u2014", "comment": "", "author": "kashif", "idle_days": 184, "models": "", "url": "https://github.com/huggingface/transformers/pull/44891"}, {"key": "pr#44889", "n": 44889, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "trainer", "type": "fix", "title": "[DeepSpeed] Fix evaluate()/predict() before train()", "action": "review requested: @SunMarc", "comment": "", "author": "roycho96", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/44889"}, {"key": "pr#44875", "n": 44875, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "refactor: improved the cli server module code organization", "action": "closed", "comment": "Closing since `cli/serve.py` was split into proper modules in #44796. Thanks @tarekziade!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44875"}, {"key": "pr#44872", "n": 44872, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "generation", "type": "fix", "title": "Fix: Update outdated sampler comment in generation/utils.py", "action": "review requested: @Cyrilvallez; maintainer pinged", "comment": "@Cyrilvallez could you take a look when you get a chance? This is waiting on a maintainer: it has been waiting for a maintainer review.", "author": "IvanFan-Van", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/44872"}, {"key": "pr#44836", "n": 44836, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Add cu_seqlens support to OlmoHybridGatedDeltaNet for packed sequences", "action": "closed", "comment": "Thanks for the packed-sequence fix, @tyler-romero! OlmoHybrid now passes `cu_seqlens` to the linear-attention kernels on main via #47630, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44836"}, {"key": "pr#44815", "n": 44815, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "quantization", "type": "fix", "title": "Dequant fix", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 210, "models": "mistral4", "url": "https://github.com/huggingface/transformers/pull/44815"}, {"key": "pr#44772", "n": 44772, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "quantization", "type": "feature / refactor", "title": "bitsandbytes: Update links and docs", "action": "\u2014", "comment": "", "author": "githubnemo", "idle_days": 217, "models": "", "url": "https://github.com/huggingface/transformers/pull/44772"}, {"key": "pr#44771", "n": 44771, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "wtf", "action": "closed", "comment": "Closing this debugging draft: `src/transformers/models/pi0/image_processing_pi0_fast.py`, the only file it touches, no longer exists on `main`.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44771"}, {"key": "pr#44729", "n": 44729, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "feature / refactor", "title": "Avoid floating point math for ceil operations", "action": "review requested: @molbap; maintainer pinged", "comment": "@molbap could you take a look when you get a chance? This is waiting on a maintainer: never reviewed; many-file rebase.", "author": "xenova", "idle_days": -1, "models": "bit, chmv2, dac, depth_pro, dpt, encodec, eomt, falcon_mamba, fuyu, gemma3, granite_speech, higgs_audio_v2_tokenizer, idefics3, jamba, layoutlmv2, lfm2_vl, llava_onevision, mamba, mamba2, mask2former, maskformer, metaclip_2, mimi, mllama, mobilevit, musicgen_melody, nllb_moe, pe_audio, perceiver, phi4_multimodal, qwen3_vl, smolvlm, voxtral_realtime, xcodec, zamba", "url": "https://github.com/huggingface/transformers/pull/44729"}, {"key": "pr#44724", "n": 44724, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Fix some missing / incorrect entries in auto files", "action": "closed", "comment": "Closing this draft auto-mapping fix since it has been idle for months; feel free to reopen if those entries are still wrong on main.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44724"}, {"key": "pr#44722", "n": 44722, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Refactor gptj output tracing to use standardized decorators", "action": "closed", "comment": "Thanks for picking this up from #43979! That call to contributions has been closed, and the reference refactor (#43590) landed with a slightly different pattern, so this branch is out of date and I'm closing it (there's also a duplicate GPT-J PR, #44066). A fresh PR following #43590 is welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44722"}, {"key": "pr#44713", "n": 44713, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[ColQwen2] Refactor output tracing (issue #43979)", "action": "closed", "comment": "Thanks for picking this up from #43979! That call to contributions has been closed, and the reference refactor (#43590) landed with a slightly different pattern, so this branch is out of date and I'm closing it. A fresh PR following #43590 is welcome if you'd still like to migrate ColQwen2.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44713"}, {"key": "pr#44682", "n": 44682, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "transformers serve + llamacpp", "action": "closed", "comment": "@SunMarc serve has been restructured into `cli/serving/` and GGUF now goes through the ggml-kernels path (#47779, #48991, #49031), so this no longer applies. Close it, or reopen a fresh version on the new serving layout if the llama.cpp backend is still wanted?", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44682"}, {"key": "pr#44676", "n": 44676, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "fix(gpt2): Resolve NaN/Inf issue in lm_head on Python 3.13 with tied weights", "action": "closed", "comment": "Thanks @JokeYoonic! Cloning the tied weight on every forward adds a full copy per step, and as @Rocketknight1 asked, the NaNs more likely come from the tied weight not being loaded; closing. If you can still reproduce on main, an issue with a minimal script would help us find the real cause.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44676"}, {"key": "pr#44662", "n": 44662, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "new model", "title": "[model] Add PenguinVL implementation", "action": "\u2014", "comment": "", "author": "CyrilSterling", "idle_days": 202, "models": "penguinvl", "url": "https://github.com/huggingface/transformers/pull/44662"}, {"key": "pr#44660", "n": 44660, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Fix: avoid late CUDA OOM in load_best_model_at_end with PEFT models", "action": "closed", "comment": "Thanks for looking into the PEFT OOM, @DogWala! As @BenjaminBossan noted, we wanted to settle the approach in #44637 first, and that discussion pointed towards using the existing `release_memory` / `accelerator.clear()` path instead. Since the issue has been closed, I'm closing this one too; feel free to open a smaller PR along those lines if the OOM persists.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44660"}, {"key": "pr#44659", "n": 44659, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "docs: remove outdated use_diff docstring from DistributedConfig.to_js\u2026", "action": "closed", "comment": "Thanks for spotting this! The `use_diff` docstring is no longer in `src/transformers/distributed/configuration_utils.py` on `main` after the `DistributedConfig` rework (#46705), so closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44659"}, {"key": "pr#44650", "n": 44650, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "trainer", "type": "fix", "title": "Fix Seq2SeqTrainer generation path for decoder-only models", "action": "review requested: @SunMarc; maintainer pinged", "comment": "@SunMarc could you take a look when you get a chance? This is waiting on a maintainer: never reviewed (cc SunMarc 2026-03-13); author pinged 2026-04-02 and 2026-04-27.", "author": "shaealh", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/44650"}, {"key": "pr#44642", "n": 44642, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / utils", "type": "feature / refactor", "title": "Clarify that causal LM labels are shifted internally", "action": "review requested: @Rocketknight1", "comment": "", "author": "joshuaswanson", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/44642"}, {"key": "pr#44641", "n": 44641, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "Conditinally passing and_mask_function arg to create_causal_mask ", "action": "closed", "comment": "Thanks for the PR! The Falcon mask creation was rewritten in #47243 (no `and_mask_function` anymore, but always materialising the mask for alibi), so this diff no longer applies. If the extra 4D mask without alibi is still an issue, a fresh PR conditioning `allow_is_causal_skip` on `alibi` would be welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44641"}, {"key": "pr#44635", "n": 44635, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "[Gemma] Modular-friendly buffers", "action": "closed", "comment": "Thanks for kicking this off, @RyanMullins! Modular-friendly `nn.Buffer` buffers landed across models in #47722, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44635"}, {"key": "pr#44626", "n": 44626, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "don't break legacy behavior when enforced!", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 222, "models": "llama", "url": "https://github.com/huggingface/transformers/pull/44626"}, {"key": "pr#44615", "n": 44615, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "Restore is_torch_fx_available for trust_remote_code backwards compatibility", "action": "closed", "comment": "Thanks for the PR! `is_torch_fx_available` went away with #41683, and as @Cyrilvallez said (https://github.com/huggingface/transformers/pull/44615#issuecomment-4045654139), we can't guarantee backward compatibility for remote code, so we're not restoring it. The fix belongs in the affected Hub repos; closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44615"}, {"key": "pr#44606", "n": 44606, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "optionally override tokenizer class with serialized tokenizer ", "action": "closed", "comment": "Closing per the review in https://github.com/huggingface/transformers/pull/44606#pullrequestreview-3961959723: we'd rather fix the specific model types that misbehave than let the serialized tokenizer override the enforced `tokenizer_class`.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44606"}, {"key": "pr#44603", "n": 44603, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "fixed dockerfile for arm64 systems", "action": "closed", "comment": "Closing this one: no activity since 2026-03-11. Thanks for the contribution, and feel free to open a fresh PR against `main` if it's still relevant!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44603"}, {"key": "pr#44601", "n": 44601, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "[Distributed] Add PP support natively", "action": "closed", "comment": "Native pipeline parallelism has since landed as `src/transformers/distributed/pipeline_parallel.py` in #47289, so this draft (`integrations/pipeline_parallel.py`) is superseded. Any follow-ups should go into a new PR against `main` that builds on that module. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44601"}, {"key": "pr#44594", "n": 44594, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "pipelines", "type": "feature / refactor", "title": "[Pipeline] Add top_k, label filtering, box_format and score sorting to ObjectDetectionPipeline", "action": "review requested: @Rocketknight1", "comment": "", "author": "vimal-crypto", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/44594"}, {"key": "pr#44587", "n": 44587, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Fix: Handling fused qkv result tensor slicing for tp sharded qkv weights", "action": "closed", "comment": "Thanks @kmbhattt-aws! As @Cyrilvallez pointed out (https://github.com/huggingface/transformers/pull/44587#issuecomment-4045778245), Falcon has no `tp_plan` and Phi3 replicates the fused qkv, so it's unclear what this fixes on main; closing for now. Happy to reopen with a reproducer showing the failure.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44587"}, {"key": "pr#44569", "n": 44569, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add SarvamMLA model (sarvamai/sarvam-105b)", "action": "\u2014", "comment": "", "author": "aashay-sarvam", "idle_days": 216, "models": "sarvam_mla", "url": "https://github.com/huggingface/transformers/pull/44569"}, {"key": "pr#44550", "n": 44550, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "docs", "type": "docs", "title": "Improve clarity and grammar in Auto Classes documentation", "action": "review requested: @vasqu; maintainer pinged", "comment": "@vasqu this was approved by @stevhliu on 2026-03-09; could you give it the core-maintainer review so it can be merged?", "author": "himani27301", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/44550"}, {"key": "pr#44517", "n": 44517, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "new model", "title": "Add qwen3 tts", "action": "\u2014", "comment": "", "author": "ShahVandit", "idle_days": 1, "models": "qwen3_tts, qwen3_tts_tokenizer", "url": "https://github.com/huggingface/transformers/pull/44517"}, {"key": "pr#44467", "n": 44467, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Placeholder tokens update", "action": "closed", "comment": "Closing as this landed via #44468 (\"Replace placeholder tokens as specified in added_tokens_decoder\").", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44467"}, {"key": "issue#44450", "n": 44450, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Support argumentless loading from Trainer checkpoints", "action": "\u2014", "comment": "", "author": "adosar", "idle_days": 229, "models": "", "url": "https://github.com/huggingface/transformers/issues/44450"}, {"key": "pr#44445", "n": 44445, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "new model", "title": "Adding support for GraniteDoclingHybrid", "action": "\u2014", "comment": "", "author": "gabe-l-hart", "idle_days": 223, "models": "granite_docling_hybrid", "url": "https://github.com/huggingface/transformers/pull/44445"}, {"key": "pr#44438", "n": 44438, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "trainer", "type": "feature / refactor", "title": "Add flashoptim", "action": "\u2014", "comment": "", "author": "SunMarc", "idle_days": 201, "models": "", "url": "https://github.com/huggingface/transformers/pull/44438"}, {"key": "pr#44407", "n": 44407, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "docs", "type": "docs", "title": "docs: add energy efficiency considerations to bitsandbytes quantization guide", "action": "review requested: @SunMarc; maintainer pinged; review requested: @vasqu", "comment": "@vasqu this was approved by @SunMarc on 2026-03-03; could you give it the core-maintainer review so it can be merged?", "author": "hongping-zh", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/44407"}, {"key": "pr#44394", "n": 44394, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "docs", "type": "docs", "title": "\ud83d\udea8\ud83d\udea7 FeatureExtractor \u2192 AudioProcessor", "action": "\u2014", "comment": "", "author": "eustlb", "idle_days": 14, "models": "aria, audio_spectrogram_transformer, audioflamingo3, beit, bridgetower, canary, chmv2, clap, clvp, cohere_asr", "url": "https://github.com/huggingface/transformers/pull/44394"}, {"key": "pr#44385", "n": 44385, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix make check-repo", "action": "closed", "comment": "Closing as superseded: the tokenizer guard in `attention_visualizer.py` landed with the `ty` 0.0.20 bump (#44494), `utils/_typing.py` has since been removed, and `torch_compilable_check` was rewritten, so `make check-repo` no longer needs these changes. If it still fails locally on current main, please share the error.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44385"}, {"key": "pr#44375", "n": 44375, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Add RF-DETR", "action": "closed", "comment": "Closing as RF-DETR landed via #36895. Thanks @NielsRogge!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44375"}, {"key": "pr#44369", "n": 44369, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Feature/integrations docs fix", "action": "closed", "comment": "Thanks for the PR! `src/transformers/integrations/mistral.py` no longer exists on `main`, and @stevhliu's review comments weren't addressed, so I'm closing this. A small fresh PR with only the still-valid docstring fixes is welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44369"}, {"key": "pr#44348", "n": 44348, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "quantization", "type": "feature / refactor", "title": "Enable MetalConfig to load pre-quantized MLX models from HuggingFace Hub", "action": "\u2014", "comment": "", "author": "n0kovo", "idle_days": 232, "models": "", "url": "https://github.com/huggingface/transformers/pull/44348"}, {"key": "pr#44298", "n": 44298, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "fix", "title": "Auto detect wrong mapping models", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 232, "models": "gemma, t5", "url": "https://github.com/huggingface/transformers/pull/44298"}, {"key": "pr#44270", "n": 44270, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "Add correct typing to custom images_kwargs in ProcessorsKwargs", "action": "maintainer pinged", "comment": "@yonigozlan this one has been waiting since 2026-02-25 (rebase (44 files) + core review). Still planned, or should we close it?", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44270"}, {"key": "pr#44264", "n": 44264, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "feature / refactor", "title": "[`Moe`] Enable aux loss automatically when in training + coef is not 0", "action": "\u2014", "comment": "", "author": "vasqu", "idle_days": 237, "models": "dbrx, doge, ernie4_5_moe, ernie4_5_vl_moe, flex_olmo, glm4v_moe, gpt_oss, granitemoe, granitemoehybrid, granitemoeshared, jamba, minimax, minimax_m2, mixtral, nllb_moe, olmoe, phimoe, qwen2_moe, qwen3_5_moe, qwen3_moe, qwen3_next, qwen3_omni_moe, switch_transformers", "url": "https://github.com/huggingface/transformers/pull/44264"}, {"key": "pr#44259", "n": 44259, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Async data producer", "action": "closed", "comment": "Thanks for the proposal! This draft has been inactive since February and `src/transformers/trainer.py` has changed substantially since, so the branch no longer merges, and nothing like a data producer has landed on `main`. If async data production is still something you want for RL, we'd be happy to discuss the design in an issue first and then review a rebased PR against `main`. Closing for now.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44259"}, {"key": "pr#44257", "n": 44257, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "trainer", "type": "feature / refactor", "title": "use nanmean for aggregating loss", "action": "review requested: @Cyrilvallez; maintainer pinged", "comment": "@Cyrilvallez this was approved by @qgallouedec on 2026-02-25; could you give it the core-maintainer review so it can be merged?", "author": "winglian", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/44257"}, {"key": "pr#44252", "n": 44252, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "Timm unification continued", "action": "maintainer pinged", "comment": "@zucchini-nlp this one has been waiting since 2026-02-26. Still planned, or should we close it?", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44252"}, {"key": "issue#44247", "n": 44247, "kind": "issue", "group": "Open issues", "category": "bug", "area": "other", "type": "bug", "title": "[MPS] Silent correctness issue in bidirectional attention", "action": "\u2014", "comment": "", "author": "hvaara", "idle_days": 126, "models": "", "url": "https://github.com/huggingface/transformers/issues/44247"}, {"key": "pr#44228", "n": 44228, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[Quantisation] account for nested tensors from quantisers", "action": "closed", "comment": "Thanks @JonoLF! @SunMarc's request for a quanto test was never addressed and the quantizer utilities were reworked since (#49548), so I'm closing this. Feel free to reopen against current main with a test if the byte count is still wrong for nested tensors.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44228"}, {"key": "pr#44215", "n": 44215, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "new model", "title": "Add sequence classification capability to Granite models", "action": "\u2014", "comment": "", "author": "jmriosal", "idle_days": 238, "models": "granite, granitemoe, granitemoehybrid, granitemoeshared", "url": "https://github.com/huggingface/transformers/pull/44215"}, {"key": "issue#44214", "n": 44214, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "Add sequence classification capabilities to the Granite models", "action": "\u2014", "comment": "", "author": "jmriosal", "idle_days": 239, "models": "", "url": "https://github.com/huggingface/transformers/issues/44214"}, {"key": "pr#44189", "n": 44189, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "trainer", "type": "fix", "title": "fix: don't move model to device under other dist train backends", "action": "review requested: @SunMarc; maintainer pinged", "comment": "@SunMarc could you take a look when you get a chance? This is waiting on a maintainer: never reviewed.", "author": "excepshenal", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/44189"}, {"key": "pr#44159", "n": 44159, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Add SDPA and Flash Attention support for OWL-ViT", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #42136, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44159"}, {"key": "pr#44154", "n": 44154, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Refactored vits to match standardized output collection interface", "action": "closed", "comment": "Thanks for picking this up @Aatman09! The #43979 call for contributions has since been closed and this PR never got reviewed before falling behind main, so I'm closing it. Sorry it didn't land!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44154"}, {"key": "pr#44142", "n": 44142, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "[voxtral-realtime] get more perfs!", "action": "\u2014", "comment": "", "author": "eustlb", "idle_days": 239, "models": "voxtral_realtime", "url": "https://github.com/huggingface/transformers/pull/44142"}, {"key": "pr#44129", "n": 44129, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Refactor SpeechT5 output tracing to standardized output capture", "action": "closed", "comment": "Thanks for the SpeechT5 refactor, @preetam1407! The #43979 call to contributions was closed right after this was opened and we couldn't review all the PRs it produced, so I'm closing this one. Sorry it didn't get looked at in time.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44129"}, {"key": "pr#44123", "n": 44123, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Avoid device sync in training loss accumulation", "action": "closed", "comment": "Thanks @cyyever! As @Rocketknight1 mentioned, we need a profile or benchmark showing the reduced syncs for these changes, and that never came, so closing for now. Happy to reopen with numbers.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44123"}, {"key": "pr#44116", "n": 44116, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[WIP] [Flaubert] Refactor output tracing to decorator-based interface", "action": "closed", "comment": "Thanks for picking this up @dtiourine! The #43979 call for contributions has since been closed and this PR never got reviewed before falling behind main, so I'm closing it. Sorry it didn't land!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44116"}, {"key": "pr#44114", "n": 44114, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Migrate wav2vec2, wav2vec2_conformer, and wav2vec2_bert to standardized output collection decorators", "action": "closed", "comment": "Thanks @23atharvaS for the big effort! The wav2vec2 family has since been migrated to `@capture_outputs` on main (see #46196), so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44114"}, {"key": "pr#44101", "n": 44101, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[XLM] Refactor output tracing to align with capture_outputs standardized architecture", "action": "closed", "comment": "Thanks for the XLM refactor, @aman-coder03! The #43979 call to contributions was closed right after this was opened and we couldn't review all the PRs it produced, so I'm closing this one. Sorry it didn't get looked at in time.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44101"}, {"key": "pr#44098", "n": 44098, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[ViLT] Refactor output handling to align with standardized patterns", "action": "closed", "comment": "Thanks for picking this up from #43979! That call to contributions has been closed, and this version still threads `output_*` flags manually instead of using `@capture_outputs` as in #43590, so I'm closing it. A fresh PR following #43590 is welcome if you'd still like to migrate ViLT.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44098"}, {"key": "pr#44086", "n": 44086, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[MGP-STR] Refactor output tracing to use capture_outputs/can_return_tuple decorators", "action": "closed", "comment": "Thanks for the MGP-STR refactor, @Zephyr-Blessed! The #43979 call to contributions was closed right after this was opened and we couldn't review all the PRs it produced, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44086"}, {"key": "pr#44085", "n": 44085, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Refactor RemBERT to use output tracing decorators", "action": "closed", "comment": "Thanks @Zephyr-Blessed. The title says RemBERT but the diff only changes `gptj/modeling_gptj.py`, and the #43979 call for contributions has been closed, so I'm closing this.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44085"}, {"key": "pr#44076", "n": 44076, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Refectored modeling_imagegpt.py to enable hooks to capture_outputs", "action": "closed", "comment": "Thanks for the ImageGPT refactor, @mmahjoub5! The #43979 call to contributions was closed right after this was opened and we couldn't review all the PRs it produced, so I'm closing this one. Sorry it didn't get looked at in time.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44076"}, {"key": "pr#44074", "n": 44074, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "[TextNet] Refactor output tracing using capture_outputs decorator", "action": "closed", "comment": "Thanks @aman-coder03! TextNet was already migrated as part of #43590, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44074"}, {"key": "pr#44073", "n": 44073, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[VisualBert] Refactor output tracing using capture_outputs and can_return_tuple decorators", "action": "closed", "comment": "Thanks for the VisualBert refactor, @aman-coder03! The #43979 call to contributions was closed right after this was opened and we couldn't review all the PRs it produced, so I'm closing this one. Sorry it didn't get looked at in time.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44073"}, {"key": "pr#44072", "n": 44072, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "refactor efficientnet output tracing with @capture_outputs and @can_r\u2026", "action": "closed", "comment": "Thanks for picking this up @Siddhartha7340! The #43979 call for contributions has since been closed and this PR never got reviewed before falling behind main, so I'm closing it. Sorry it didn't land!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44072"}, {"key": "pr#44071", "n": 44071, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[Refactor] Migrate MPT to standardized output tracing decorators", "action": "closed", "comment": "Thanks for the MPT refactor, @ArivunidhiA! The #43979 call to contributions was closed right after this PR was opened and we couldn't review all the PRs it produced, so I'm closing this one. Sorry it didn't get looked at in time.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44071"}, {"key": "pr#44070", "n": 44070, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "gguf", "type": "feature / refactor", "title": "Add GGUF loading support for Qwen3-Next (qwen3_next) architecture", "action": "\u2014", "comment": "", "author": "rudybear", "idle_days": 186, "models": "qwen3_next", "url": "https://github.com/huggingface/transformers/pull/44070"}, {"key": "pr#44068", "n": 44068, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Refactor GPT-Neo to use `@capture_outputs` and `@can_return_tuple` decorators", "action": "closed", "comment": "Thanks for picking this up from #43979! That call to contributions has been closed, and the reference refactor (#43590) landed afterwards with a slightly different pattern, so this branch is out of date and I'm closing it. If you'd still like to migrate GPT-Neo, a fresh PR following #43590 is welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44068"}, {"key": "pr#44066", "n": 44066, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Refactor GPT-J to use standardized output tracing (#43979)", "action": "closed", "comment": "Thanks for picking this up from #43979! That call to contributions has been closed, and the reference refactor (#43590) landed afterwards with a slightly different pattern, so this branch is out of date and I'm closing it (there's also a parallel GPT-J PR, #44722). A single fresh PR following #43590 is welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44066"}, {"key": "pr#44059", "n": 44059, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "[GPT2] Refactor output tracing to use capture_outputs/can_return_tuple decorators", "action": "closed", "comment": "Thanks @lakprigan! GPT2 was migrated to the output capture decorators in #43983, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44059"}, {"key": "pr#44056", "n": 44056, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[MPNet] Refactor output tracing using capture_outputs decorator", "action": "closed", "comment": "Thanks for picking this up from #43979! That call to contributions has been closed, and the reference refactor (#43590) landed afterwards with a slightly different pattern, so this branch is out of date and I'm closing it. If you'd still like to migrate MPNet, a fresh PR following #43590 is welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44056"}, {"key": "pr#44054", "n": 44054, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "feature / refactor", "title": "Flash mla interface", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 242, "models": "glm_moe_dsa", "url": "https://github.com/huggingface/transformers/pull/44054"}, {"key": "pr#44044", "n": 44044, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Refactor DeBERTa's output tracing interface", "action": "closed", "comment": "Thanks for picking this up from #43979! That call to contributions has been closed, and the reference refactor (#43590) landed afterwards with a slightly different pattern, so this branch is out of date and I'm closing it. If you'd still like to migrate DeBERTa-v2, a fresh PR following #43590 is welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44044"}, {"key": "pr#44030", "n": 44030, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "refactor output tracing in `dpr`", "action": "closed", "comment": "Thanks for picking this up @omkar-334! The #43979 call for contributions has since been closed and this PR never got reviewed before falling behind main, so I'm closing it. Sorry it didn't land!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44030"}, {"key": "pr#44029", "n": 44029, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "refactor output tracing in `rwkv`", "action": "closed", "comment": "Thanks for the RWKV refactor, @omkar-334! The #43979 call to contributions was closed right after this was opened and we couldn't review all the PRs it produced, so I'm closing this one. Sorry it didn't get looked at in time.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44029"}, {"key": "pr#44028", "n": 44028, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "refactor output tracing for `superpoint`", "action": "closed", "comment": "Thanks for the SuperPoint refactor, @omkar-334! The #43979 call to contributions was closed right after this was opened and we couldn't review all the PRs it produced, so I'm closing this one. Sorry it didn't get looked at in time.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44028"}, {"key": "pr#44027", "n": 44027, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "refactor output tracing in `speech_encoder_decoder`", "action": "closed", "comment": "Thanks for the refactor, @omkar-334! The #43979 call to contributions was closed right after this was opened and we couldn't review all the PRs it produced, so I'm closing this one. Sorry it didn't get looked at in time.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44027"}, {"key": "pr#44026", "n": 44026, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "refactor output tracing for `vision_encoder_decoder`", "action": "closed", "comment": "Thanks for picking this up from #43979! That call to contributions has been closed, and the reference refactor (#43590) landed afterwards with a slightly different pattern, so this branch is out of date and I'm closing it. A fresh PR following #43590 is welcome if you'd still like to do it.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44026"}, {"key": "pr#44025", "n": 44025, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "refactor output tracing for `depth_anything`", "action": "closed", "comment": "Thanks for picking this up from #43979! That call to contributions has been closed, and the reference refactor (#43590) landed afterwards with a slightly different pattern, so this branch is out of date and I'm closing it. If you'd still like to migrate Depth Anything, a fresh PR following #43590 is welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44025"}, {"key": "pr#44024", "n": 44024, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Focalnet standardized outputs", "action": "closed", "comment": "Thanks for the FocalNet refactor, @mmahjoub5! The #43979 call to contributions was closed right after this was opened and FocalNet was partly migrated in #43590, so I'm closing this one. Sorry it didn't get looked at in time.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44024"}, {"key": "pr#44019", "n": 44019, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Refactor `resnet` to use `@capture_outputs` / `@can_return_tuple` output tracing", "action": "closed", "comment": "Thanks @Sid-V5! ResNet was already migrated as part of #43590, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44019"}, {"key": "pr#44018", "n": 44018, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Refactor GPT-Neo output tracing to use capture_outputs/can_return_tuple", "action": "closed", "comment": "Thanks for picking this up @yashbora9! The #43979 call for contributions has since been closed and this PR never got reviewed before falling behind main, so I'm closing it. Sorry it didn't land!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44018"}, {"key": "pr#44015", "n": 44015, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Refactor GPT2-based models to standardized output collection interface", "action": "closed", "comment": "Thanks for the migration! GPT-2 and Decision Transformer were migrated in #43983 and GPTBigCode in #43590, so this is already done on `main`. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44015"}, {"key": "pr#44013", "n": 44013, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Ouptut tracing: Standardizing MobileNetv2", "action": "closed", "comment": "Thanks for the MobileNetV2 refactor, @gabrielfruet! The #43979 call to contributions was closed right after this PR was opened and we couldn't review all the PRs it produced, so I'm closing this one. Sorry it didn't get looked at in time.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44013"}, {"key": "pr#44010", "n": 44010, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "feature / refactor", "title": "[SqueezeBert] Migrate to standardized output collection decorators", "action": "\u2014", "comment": "", "author": "preetam1407", "idle_days": 190, "models": "squeezebert", "url": "https://github.com/huggingface/transformers/pull/44010"}, {"key": "pr#44007", "n": 44007, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[ResNet] Refactor output tracing to decorator-based interface", "action": "closed", "comment": "Thanks for the ResNet refactor, @pdwi2020! The #43979 call to contributions was closed shortly after this was opened and ResNet was partly migrated in #43590 / #44638, so I'm closing this one. Sorry it didn't get looked at in time.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44007"}, {"key": "pr#44004", "n": 44004, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "refactor output tracing for `codegen`", "action": "closed", "comment": "Thanks for picking this up @omkar-334! The #43979 call for contributions has since been closed and this PR never got reviewed before falling behind main, so I'm closing it. Sorry it didn't land!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44004"}, {"key": "pr#44003", "n": 44003, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "refactor output tracing in `mamba`", "action": "closed", "comment": "Thanks for picking this up from #43979! That call to contributions has been closed, and the reference refactor (#43590) landed afterwards with a slightly different pattern, so this branch is out of date and I'm closing it. If you'd still like to migrate Mamba/FalconMamba, a fresh PR following #43590 (regenerating copies with `make fix-repo`) is welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44003"}, {"key": "pr#44002", "n": 44002, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "refactor output tracing in `upernet`", "action": "closed", "comment": "Thanks for picking this up from #43979! That call to contributions has been closed, and the reference refactor (#43590) landed afterwards with a slightly different pattern, so this branch is out of date and I'm closing it. If you'd still like to migrate UperNet, a fresh PR following #43590 is welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44002"}, {"key": "pr#44001", "n": 44001, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "refactor output tracing in `univnet`", "action": "closed", "comment": "Thanks for picking this up @omkar-334! The #43979 call for contributions has since been closed and this PR never got reviewed before falling behind main, so I'm closing it. Sorry it didn't land!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44001"}, {"key": "pr#44000", "n": 44000, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "refactor output tracing in `vision_text_dual_encoder`", "action": "closed", "comment": "Thanks for picking this up @omkar-334! The #43979 call for contributions has since been closed and this PR never got reviewed before falling behind main, so I'm closing it. Sorry it didn't land!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/44000"}, {"key": "pr#43999", "n": 43999, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "refactor output tracing in `mobilenet_v1`", "action": "closed", "comment": "Thanks for picking this up from #43979! That call to contributions has been closed, and the reference refactor (#43590) landed afterwards with a slightly different pattern, so this branch is out of date and I'm closing it. If you'd still like to migrate MobileNetV1, a fresh PR following #43590 is welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43999"}, {"key": "pr#43998", "n": 43998, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "refactor output tracing in `timm_backbone`", "action": "closed", "comment": "Thanks for the refactor, @omkar-334! The #43979 call to contributions was closed shortly after this was opened and `timm_backbone` was since updated in #46422, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43998"}, {"key": "pr#43997", "n": 43997, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Migrate RegNet to standardized output tracing", "action": "closed", "comment": "Thanks for picking this up from #43979! That call to contributions has been closed, and the reference refactor (#43590) landed afterwards with a slightly different pattern, so this branch is out of date and I'm closing it. If you'd still like to migrate RegNet, a fresh PR following #43590 is welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43997"}, {"key": "pr#43996", "n": 43996, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Refactor FNet and CVT output tracing", "action": "closed", "comment": "Thanks for the refactor, @beelapranay! The #43979 call to contributions was closed shortly after this was opened and we couldn't review the many PRs it produced, so I'm closing this one. Sorry it didn't get looked at in time.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43996"}, {"key": "pr#43995", "n": 43995, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Refactoring falcon model to match standardized output collection interface", "action": "closed", "comment": "Thanks for picking this up @akeemlh! The #43979 call for contributions has since been closed and this PR never got reviewed before falling behind main, so I'm closing it. Sorry it didn't land!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43995"}, {"key": "pr#43989", "n": 43989, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix AutoVideoProcessor class lookup when torchvision is unavailable", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #48984, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43989"}, {"key": "pr#43973", "n": 43973, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "new model", "title": "Add lfm2.5 audio", "action": "\u2014", "comment": "", "author": "MHRDYN7", "idle_days": 25, "models": "lfm2_audio", "url": "https://github.com/huggingface/transformers/pull/43973"}, {"key": "pr#43967", "n": 43967, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Fix AttributeError in run_classification.py when detecting multi-label data", "action": "closed", "comment": "We removed the legacy `examples/` scripts from the repository (#49357) since they were no longer maintained, so this change no longer has a home. Thanks for the fix!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43967"}, {"key": "pr#43961", "n": 43961, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Replace mutable default arguments with None", "action": "closed", "comment": "Thanks for the PR! As @Rocketknight1 pointed out (https://github.com/huggingface/transformers/pull/43961#issuecomment-3897035327), none of these functions mutate their default argument, so this doesn't fix an actual bug. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43961"}, {"key": "pr#43911", "n": 43911, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "add Llama to mapping names in tokenization_auto.py", "action": "closed", "comment": "Thanks @pavel-esir! As @ArthurZucker mentioned, the root cause was fixed in #44127 (we follow the `tokenizer_class` from the config) and we don't want Llama in the default mapping, so closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43911"}, {"key": "issue#43909", "n": 43909, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add LFM2.5 Audio 1.5B", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [LiquidAI/LFM2.5-Audio-1.5B](https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B), [LiquidAI/LFM2.5-Audio-1.5B-JP](https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B-JP)\n- Downloads (last 30d): 6,901 \u00b7 Likes: 471 \u00b7 Params: ~1.5B \u00b7 License: other\n- Released: 2025-12 \u00b7 Remote code on the Hub: no\n- Open PR: #43973 (also #48249)\n- The text backbone (`lfm2`) is already in transformers.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "MHRDYN7", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/43909"}, {"key": "pr#43875", "n": 43875, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Improve handling of QuantizedLayer.reset", "action": "closed", "comment": "Thanks @cyyever! `QuantizedLayer.reset` now drops the quantized states on main via #48700, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43875"}, {"key": "pr#43863", "n": 43863, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "[whisper] allow to pass text/audio specific kwargs", "action": "\u2014", "comment": "", "author": "eustlb", "idle_days": 252, "models": "whisper", "url": "https://github.com/huggingface/transformers/pull/43863"}, {"key": "pr#43833", "n": 43833, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "fix: ensure dtype consistency in grouped_mm under autocast", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #43839, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43833"}, {"key": "pr#43826", "n": 43826, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "fix: error message of pipeline", "action": "closed", "comment": "Thanks @math-hiyoko! This cleanup was included in #43869, which is now merged, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43826"}, {"key": "pr#43823", "n": 43823, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Add `facebook/MobileLLM-125M`", "action": "closed", "comment": "Thanks for the PR! MobileLLM checkpoints can load with `LlamaForCausalLM` once the Hub configs are updated, so a separate model in the library isn't needed. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43823"}, {"key": "pr#43785", "n": 43785, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix FSDP_CPU_RAM_EFFICIENT_LOADING (#43749)", "action": "closed", "comment": "Thanks for the fix! Following the reviewers' suggestion, the FSDP CPU-RAM-efficient loading case is now handled in the loading path itself (#44473) and #43749 is closed, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43785"}, {"key": "pr#43779", "n": 43779, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "SwanLab: Add support for id and resume arguments in SwanLabCallback", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #43719, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43779"}, {"key": "pr#43757", "n": 43757, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Avoid hard failure for gpt-oss GGUF architecture by falling back to g\u2026", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #45506, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43757"}, {"key": "pr#43751", "n": 43751, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "Fix ruff warnings", "action": "closed", "comment": "Thanks for the cleanup, @cyyever! 17 of the 40 files in this diff (the `examples/pytorch/` scripts) were removed from main in #49357, so this can't land as is; closing. If ruff still reports these warnings on main, a fresh PR against current files would be welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43751"}, {"key": "pr#43747", "n": 43747, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Remove CompressedLinear support for compressed-tensors > 0.13", "action": "closed", "comment": "Thanks for preparing this, @colldata79! The compressed-tensors integration was reworked on main (#47216, #47652) and no longer relies on `CompressedLinear`, so I'm closing this one. If something still breaks with recent compressed-tensors releases, please open an issue with a reproducer.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43747"}, {"key": "pr#43743", "n": 43743, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "CI / tooling", "type": "CI / tooling", "title": "Modular playground", "action": "\u2014", "comment": "", "author": "itazap", "idle_days": 28, "models": "", "url": "https://github.com/huggingface/transformers/pull/43743"}, {"key": "issue#43682", "n": 43682, "kind": "issue", "group": "Open issues", "category": "question / other", "area": "exporters", "type": "question / other", "title": "`torch.compile` and `torch.export` helper function to be moved from `import_utils`", "action": "\u2014", "comment": "", "author": "IlyasMoutawwakil", "idle_days": 229, "models": "", "url": "https://github.com/huggingface/transformers/issues/43682"}, {"key": "issue#43680", "n": 43680, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "Add Nvidia NitroGen to huggingface transformers", "action": "\u2014", "comment": "", "author": "AffanBinFaisal", "idle_days": 237, "models": "", "url": "https://github.com/huggingface/transformers/issues/43680"}, {"key": "issue#43671", "n": 43671, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Proposal to add Qwen3-TTS support", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [Qwen/Qwen3-TTS-12Hz-1.7B-Base](https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-Base), [Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice](https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice), [Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice](https://huggingface.co/Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice)\n- Downloads (last 30d): 3.0M \u00b7 Likes: 567 \u00b7 Params: ~1.9B \u00b7 License: apache-2.0\n- Released: 2026-01 \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "ShahVandit", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/43671"}, {"key": "pr#43665", "n": 43665, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "fix", "action": "closed", "comment": "Closing this CI test PR, it was only used to trigger `run-slow: vit, clip` (https://github.com/huggingface/transformers/pull/43665#issuecomment-3829427951).", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43665"}, {"key": "pr#43663", "n": 43663, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "trainer", "type": "feature / refactor", "title": "Add _get_signature_columns method to allow custom trainers to override column filtering", "action": "\u2014", "comment": "", "author": "abigailllr", "idle_days": 259, "models": "", "url": "https://github.com/huggingface/transformers/pull/43663"}, {"key": "pr#43656", "n": 43656, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix TypeAdapter NameError in transformers CLI", "action": "closed", "comment": "Thanks for the fix! `transformers serve` was refactored into separate modules in #44796, which removed this `TypeAdapter` annotation, and #43576 is closed. Closing as superseded.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43656"}, {"key": "pr#43654", "n": 43654, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tokenization", "type": "fix", "title": "fix(tokenizer): Avert special token property overwrites in batch add_tokens calls", "action": "\u2014", "comment": "", "author": "harshaljanjani", "idle_days": 185, "models": "", "url": "https://github.com/huggingface/transformers/pull/43654"}, {"key": "pr#43649", "n": 43649, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Check new failures reporting 5", "action": "closed", "comment": "Closing this CI-reporting experiment since it has been idle for months; feel free to reopen if still needed.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43649"}, {"key": "pr#43636", "n": 43636, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Add _metrics dict to Trainer for custom metric logging", "action": "closed", "comment": "Thanks @abigailllr! This adds new Trainer state and never got buy-in from the Trainer maintainers, so I'm closing it for now. If you'd still like it, please discuss the API in #43599 with @SunMarc first.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43636"}, {"key": "pr#43613", "n": 43613, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "docs", "type": "docs", "title": "Add Promptable Visual Segmentation pipeline", "action": "\u2014", "comment": "", "author": "yonigozlan", "idle_days": 259, "models": "", "url": "https://github.com/huggingface/transformers/pull/43613"}, {"key": "pr#43612", "n": 43612, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "docs", "type": "docs", "title": "Add Promptable Concept Segmentation pipeline", "action": "maintainer pinged", "comment": "@yonigozlan this one has been waiting since 2026-04-28 (design decision: new pipeline for a single model (Cyrilvallez/Rocketknight1, 2026-02-03)). Still planned, or should we close it?", "author": "yonigozlan", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/43612"}, {"key": "issue#43595", "n": 43595, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Tell Us: What Would Make Trainer Better?", "action": "\u2014", "comment": "", "author": "SunMarc", "idle_days": 185, "models": "", "url": "https://github.com/huggingface/transformers/issues/43595"}, {"key": "pr#43549", "n": 43549, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "attention / kernels", "type": "feature / refactor", "title": "[kernels] exception handling for fa kernels", "action": "maintainer pinged", "comment": "@MekkCyber this one has been waiting since 2026-01-28 (second reviewer requested by danieldk). Still planned, or should we close it?", "author": "MekkCyber", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/43549"}, {"key": "pr#43543", "n": 43543, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix fp16 underflow in MoE load balancing loss by enforcing fp32 softmax", "action": "closed", "comment": "Thanks @ITcarrot! #48131 rewrote `load_balancing_loss_func` to accumulate per-layer router probabilities and token counts in float32 accumulators, which removes the fp16 underflow in the averaged probabilities you described here. Closing as superseded. If you still see `aux_loss` collapsing to 0 in fp16 on current main, please open an issue with a reproducer.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43543"}, {"key": "pr#43542", "n": 43542, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "fix: output router capture wrong router logits in qwen moe models", "action": "closed", "comment": "Thanks for catching this, it was a real bug! The same fix landed in #45131 (routers now keep raw `router_logits` and use a separate `router_probs`), so closing this one as superseded.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43542"}, {"key": "pr#43532", "n": 43532, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[don't merge] Show diff", "action": "closed", "comment": "Closing this CI experiment: as its title says, @ydshieh opened it as \"[don't merge]\".", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43532"}, {"key": "pr#43506", "n": 43506, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Add RishAI model with full transformers integration", "action": "closed", "comment": "Thanks for the PR! As @Rocketknight1 explained (https://github.com/huggingface/transformers/pull/43506#issuecomment-3804954078), we add models to the library once there are public checkpoints with strong performance or usage; until then, Hub custom code (`trust_remote_code`) is the way to share RishAI. Closing for now.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43506"}, {"key": "pr#43498", "n": 43498, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "fix/backward compatibility for tie_weights", "action": "closed", "comment": "Thanks @marcndo! As @ArthurZucker said (https://github.com/huggingface/transformers/pull/43498#pullrequestreview-3706879117), the `tie_weights` change in v5 is an intentional break, so we won't add a BC path; closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43498"}, {"key": "pr#43492", "n": 43492, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "feature / refactor", "title": "Perception Encoder follow up PR", "action": "maintainer pinged", "comment": "@eustlb this one has been waiting since 2026-01-26. Still planned, or should we close it?", "author": "eustlb", "idle_days": -1, "models": "pe_audio, pe_audio_video, pe_video", "url": "https://github.com/huggingface/transformers/pull/43492"}, {"key": "pr#43488", "n": 43488, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[don't merge] bad format to check repo bot", "action": "closed", "comment": "Closing this bot-test PR; as its title says, it was never meant to be merged.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43488"}, {"key": "issue#43472", "n": 43472, "kind": "issue", "group": "Open issues", "category": "feature", "area": "quantization", "type": "feature", "title": "Introduce standardized BatchLinear module for MoE architectures to facilitate PEFT and Quantization", "action": "\u2014", "comment": "", "author": "ITcarrot", "idle_days": 268, "models": "", "url": "https://github.com/huggingface/transformers/issues/43472"}, {"key": "pr#43469", "n": 43469, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "argparser: Allow optional bool flags without values", "action": "closed", "comment": "Closing this one: no activity since 2026-01-26. Thanks for the contribution, and feel free to open a fresh PR against `main` if it's still relevant!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43469"}, {"key": "pr#43448", "n": 43448, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "Add Molmo", "action": "review requested: @merveenoyan; maintainer pinged", "comment": "@merveenoyan could you take a look when you get a chance? This is waiting on a maintainer: merveenoyan asked zucchini-nlp to take over (2026-03-13).", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43448"}, {"key": "pr#43446", "n": 43446, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "feature / refactor", "title": "[`typings`] Automatically type decorator return types as `tuple | X`", "action": "review requested: @vasqu; maintainer pinged", "comment": "@vasqu could you take a look when you get a chance? This is waiting on a maintainer: it has been waiting for a maintainer review.", "author": "tomaarsen", "idle_days": -1, "models": "afmoe, aimv2, albert, align, altclip, aria, audioflamingo3, aya_vision, bert, bert_generation, blip, blip_2, blt, bridgetower, bros, camembert, chameleon, chinese_clip, clap, clip, clipseg, clvp, cohere2_vision, conditional_detr, csm, d_fine, data2vec, deepseek_vl, deepseek_vl_hybrid, deformable_detr, detr, dia, distilbert, edgetam, edgetam_video, electra, emu3, encoder_decoder, ernie, ernie4_5_vl_moe", "url": "https://github.com/huggingface/transformers/pull/43446"}, {"key": "issue#43431", "n": 43431, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Add dataloader_multiprocessing_context to TrainingArguments + Ensure dataloader_prefetch_factor fully supports IterableDataset (follow-up to #34925)", "action": "closed", "comment": "This has been added in #47164 and is available on main. Closing, but feel free to reopen if something is still missing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/43431"}, {"key": "pr#43424", "n": 43424, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "exporters", "type": "feature / refactor", "title": "Add test to ensure executorch exportability with dynamic shapes", "action": "\u2014", "comment": "", "author": "justinchuby", "idle_days": 243, "models": "", "url": "https://github.com/huggingface/transformers/pull/43424"}, {"key": "issue#43386", "n": 43386, "kind": "issue", "group": "Open issues", "category": "feature", "area": "serving / continuous batching", "type": "feature", "title": "Support other types of model inputs for continuous batching", "action": "\u2014", "comment": "", "author": "yhshin11", "idle_days": 230, "models": "", "url": "https://github.com/huggingface/transformers/issues/43386"}, {"key": "pr#43382", "n": 43382, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "Allow Path type in transformers.image_utils.load_image function", "action": "\u2014", "comment": "", "author": "mrdbourke", "idle_days": 272, "models": "", "url": "https://github.com/huggingface/transformers/pull/43382"}, {"key": "pr#43378", "n": 43378, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "feat(models): Make MimiModel encoding padding-aware to ensure batch-to-individual consistency", "action": "\u2014", "comment": "", "author": "harshaljanjani", "idle_days": 185, "models": "mimi", "url": "https://github.com/huggingface/transformers/pull/43378"}, {"key": "issue#43375", "n": 43375, "kind": "issue", "group": "Open issues", "category": "feature", "area": "core / loading", "type": "feature", "title": "Feature request: Support resume_from_checkpoint for checkpoints stored on Hugging Face Hub", "action": "\u2014", "comment": "", "author": "meodbr", "idle_days": 14, "models": "", "url": "https://github.com/huggingface/transformers/issues/43375"}, {"key": "pr#43363", "n": 43363, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "[Improvement] Update `DistributedLengthGroupedSampler` to allow customizing length function", "action": "closed", "comment": "Thanks @Dogacel for this improvement, and sorry it didn't get reviewed in time! Since then the Trainer tests were reorganized and `tests/trainer/test_trainer_utils.py` was removed in #44260, so the branch no longer applies. The sampler itself still lacks a custom length function, so if you're still interested please open a new PR against `main` with the tests in the current layout. Closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43363"}, {"key": "pr#43340", "n": 43340, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Claude code skills for transformers-api", "action": "closed", "comment": "Thanks for the effort here, @gautamvarmadatla! As discussed in the thread, we're not adding such a large skill set, since it's hard to verify and may not help agents on this codebase. Agent setup has since been centralized in `.ai/` (#44489), with a shared `AGENTS.md` and small, targeted skills that local agents link in via a `make` target. If you have a focused skill that clearly helps a specific workflow, a small PR against `main` adding it under `.ai/skills/` is welcome. Closing this one, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43340"}, {"key": "issue#43328", "n": 43328, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "Add support for SNAC model type", "action": "\u2014", "comment": "", "author": "rtATvw", "idle_days": 272, "models": "", "url": "https://github.com/huggingface/transformers/issues/43328"}, {"key": "pr#43310", "n": 43310, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Replace regex with re", "action": "maintainer pinged", "comment": "@Rocketknight1 this one has been waiting since 2026-01-15 (never reviewed; conflicts). Still planned, or should we close it?", "author": "Rocketknight1", "idle_days": -1, "models": "bertweet, blenderbot_small, clvp, ctrl, deepseek_vl, deepseek_vl_hybrid, depth_pro, fastspeech2_conformer, got_ocr2, gpt_oss, mllama, pixtral, recurrent_gemma, sam3, sam3_video, switch_transformers", "url": "https://github.com/huggingface/transformers/pull/43310"}, {"key": "issue#43298", "n": 43298, "kind": "issue", "group": "Open issues", "category": "feature", "area": "processors (image/video/audio)", "type": "feature", "title": "continous batching not support audio Model", "action": "\u2014", "comment": "", "author": "MengLeebin", "idle_days": 221, "models": "", "url": "https://github.com/huggingface/transformers/issues/43298"}, {"key": "pr#43297", "n": 43297, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[Feat] Reduces redundant tokenization of <pad> tags to accelerate Qwen3VL.", "action": "closed", "comment": "Thanks @ZLkanyo009 for the profiling! As @itazap concluded in https://github.com/huggingface/transformers/pull/43297#discussion_r2764471737, this adds maintenance overhead for a path that only covers `bs==1, padding=False`, so we'll pass for now; happy to revisit if more users hit this bottleneck. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43297"}, {"key": "pr#43291", "n": 43291, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix whisper tests", "action": "closed", "comment": "Closing as the whisper test fixes landed via #47150 in the meantime. Thanks @eustlb!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43291"}, {"key": "pr#43271", "n": 43271, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix typo: necesary \u2192 necessary", "action": "closed", "comment": "Thanks for the fix, @Arthi706! These typos were fixed on main in #43949, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43271"}, {"key": "pr#43270", "n": 43270, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "fix _retrieve_segment timestamps offset bug", "action": "review requested: @zucchini-nlp; maintainer pinged", "comment": "@zucchini-nlp could you take a look when you get a chance? This is waiting on a maintainer: it has been waiting for a maintainer review.", "author": "YCmove", "idle_days": -1, "models": "whisper", "url": "https://github.com/huggingface/transformers/pull/43270"}, {"key": "pr#43267", "n": 43267, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Add auto_docstring decorator to Sam3ImageProcessorFast", "action": "closed", "comment": "Thanks for spotting this! After the image-processor refactor, `Sam3ImageProcessor` on `main` is already decorated with `@auto_docstring` (`src/transformers/models/sam3/image_processing_sam3.py`), so closing as resolved.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43267"}, {"key": "pr#43265", "n": 43265, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "new model", "title": "Adding Omnilingual ASR models", "action": "\u2014", "comment": "", "author": "ebezzam", "idle_days": -1, "models": "omniasr, omniasr_ctc, vibevoice_asr", "url": "https://github.com/huggingface/transformers/pull/43265"}, {"key": "pr#43254", "n": 43254, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "trainer", "type": "fix", "title": "Add supported kwargs to fixed_cross_entropy", "action": "\u2014", "comment": "", "author": "Rocketknight1", "idle_days": 264, "models": "", "url": "https://github.com/huggingface/transformers/pull/43254"}, {"key": "pr#43251", "n": 43251, "kind": "PR", "group": "Closed PRs", "category": "duplicate", "area": "", "type": "", "title": "Fix(43240): pass kwargs to nn.functional.cross_entropy", "action": "closed", "comment": "Thanks for the fix! #43254 makes the same change to fixed_cross_entropy (passes weight and label_smoothing through to cross_entropy) and was opened the same day. We'll keep one of the two to avoid duplicate review, and this one will be closed if we go with #43254. Sorry for the overlap!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43251"}, {"key": "pr#43249", "n": 43249, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "[WIP] Processor moves to `device` in `__call__`", "action": "closed", "comment": "Closing this WIP: the image-processor backend refactor removed `image_processing_utils_fast.py` and the `*_fast` processors it edits, so it can't be rebased as it is. Moving outputs to `device` in `__call__` is still a good idea if we want to pick it up again on the new backend layout.", "author": "zucchini-nlp", "idle_days": -1, "models": "llava", "url": "https://github.com/huggingface/transformers/pull/43249"}, {"key": "pr#43246", "n": 43246, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "GptOss slow tests", "action": "closed", "comment": "Closing this draft since gpt_oss GPU tests were fixed in #48118 and the fixtures regenerated in #46837; feel free to reopen if something is still failing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43246"}, {"key": "pr#43238", "n": 43238, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Fix ObjectDetectionPipeline batch processing bug #31356", "action": "closed", "comment": "Thanks for looking into the batching issue! This branch includes a full virtualenv directory (95 of the 97 changed files), which makes it impossible to review, and #31356 has since been closed. If you can still reproduce the bug with `ObjectDetectionPipeline` on `main`, please open a clean PR that only touches `src/transformers/pipelines/object_detection.py` and its test. Closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43238"}, {"key": "pr#43213", "n": 43213, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "feat: allow output_hidden_states and output_attensions to record outputs of specific layers", "action": "review requested: @Rocketknight1; maintainer pinged; closed", "comment": "Thanks for the PR! The output-recording code this edits was rewritten and moved to `src/transformers/utils/output_capturing.py` in #43765, so this diff no longer applies (none of the changed lines exist on main). If selecting specific layers via `output_hidden_states=[...]` is still something you want, could you re-implement it on top of the new `output_capturing.py` hooks? Otherwise we can close this. Closing this one for now; feel free to open a fresh PR against `main` if it's still needed. Thanks!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43213"}, {"key": "pr#43212", "n": 43212, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Add regression test for offline tokenizer loading (fixes #43200)", "action": "closed", "comment": "Thanks for adding a regression test, @Anri-Lombard! The offline `_patch_mistral_regex` path was fixed and tested in #43603, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43212"}, {"key": "pr#43151", "n": 43151, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Make TF32 tests hardware-aware for PyTorch 2.9+", "action": "closed", "comment": "Thanks for investigating #42371! The PR has no changes left, and as discussed the behaviour comes from PyTorch itself, so I'm closing it.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43151"}, {"key": "pr#43149", "n": 43149, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "docs(serving): add minimal Python client examples for chat completion\u2026", "action": "closed", "comment": "Thanks @khajamoddin! The serving docs have since been rewritten under `docs/source/en/serve-cli/` (#42853), so this no longer applies; closing. If something is still missing there (e.g. an `openai` client example as @Wauplin suggested), a small focused PR against the new page is welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43149"}, {"key": "pr#43139", "n": 43139, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "[perf] optimize whisper GPU performance", "action": "review requested: @vasqu; maintainer pinged", "comment": "@vasqu could you take a look when you get a chance? This is waiting on a maintainer: it has been waiting for a maintainer review.", "author": "JaredforReal", "idle_days": -1, "models": "glmasr, whisper", "url": "https://github.com/huggingface/transformers/pull/43139"}, {"key": "pr#43104", "n": 43104, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "docs: clarify tokenizer decoder behavior in v5 (#43066)", "action": "closed", "comment": "Thanks for wanting to clarify this! As @ArthurZucker noted in https://github.com/huggingface/transformers/pull/43104#pullrequestreview-3627373685, the diff adds far more than the fix needs, and there have been no commits since that review, so I'm closing it. A small PR with only the one-line clarification would be welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43104"}, {"key": "pr#43102", "n": 43102, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Add CPU vs GPU performance comparison example", "action": "closed", "comment": "Thanks for contributing! As mentioned, we'd like to keep the pipeline tutorial general, so we're closing this.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43102"}, {"key": "pr#43096", "n": 43096, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix save_pretrained for quantized models with custom serialization", "action": "closed", "comment": "Thanks for the fix, @480284856! mxfp4 saving was fixed on main with a proper `reverse_op` in #43148 (the approach @MekkCyber suggested), so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43096"}, {"key": "pr#43094", "n": 43094, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Avoid inline .item() sync in decoder start token check", "action": "closed", "comment": "Closing this one: no activity since 2026-01-04. Thanks for the contribution, and feel free to open a fresh PR against `main` if it's still relevant!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43094"}, {"key": "pr#43088", "n": 43088, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Skip attention_mask.all() GPU-CPU sync during generation", "action": "closed", "comment": "Thanks for chasing the GPU-CPU sync! @zucchini-nlp suggested dropping an all-ones mask once at the start (https://github.com/huggingface/transformers/pull/43088#discussion_r2660716664), with @vasqu's caveat that eager/flex still need the mask, and the PR hasn't been updated since. Closing; a fresh PR along those lines would be welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43088"}, {"key": "pr#43085", "n": 43085, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "Add async_stopping_criteria flag to reduce GPU-CPU syncs during generation", "action": "closed", "comment": "Thanks for the detailed work on async stopping criteria! As @ArthurZucker mentioned, the current focus for this kind of GPU-CPU overlap is continuous batching rather than adding a new mode to `generate`, so I'm closing this. The underlying sync issue (#43089) is still useful context for that work.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43085"}, {"key": "pr#43077", "n": 43077, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "compileable=>compilable", "action": "closed", "comment": "Thanks @cyyever! Since `compileable` appears in public attributes (e.g. `is_compileable`), renaming would need a deprecation path, and the spelling debate didn't reach a conclusion, so I'm closing this.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43077"}, {"key": "pr#43063", "n": 43063, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Improve documentation for SegFormer image processor", "action": "closed", "comment": "Thanks @sindhukodati! As @stevhliu noted (https://github.com/huggingface/transformers/pull/43063#pullrequestreview-3637255572), the SegFormer image processor is already self-explanatory, so we'll pass on this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43063"}, {"key": "pr#43056", "n": 43056, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Perf: enable pin_memory in DataLoader for CLM no_trainer example", "action": "closed", "comment": "We removed the `examples/pytorch` scripts from the repository in #49357, so this change no longer applies. Thanks!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43056"}, {"key": "issue#43037", "n": 43037, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "DeepSeek v3.2 support", "action": "closed", "comment": "DeepSeek-V3.2 (`deepseek_v32`, with DeepSeek Sparse Attention) is now supported natively, added in #41251. Thanks to everyone who upvoted this. Closing, but feel free to open a new issue if something doesn't load correctly.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/43037"}, {"key": "issue#43025", "n": 43025, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Enable passing block_size and bf16_stochastic_round to torchao optimizers via TrainingArguments", "action": "\u2014", "comment": "", "author": "klae01", "idle_days": 288, "models": "", "url": "https://github.com/huggingface/transformers/issues/43025"}, {"key": "pr#43015", "n": 43015, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "FIX: TF32 warning (#43012)", "action": "closed", "comment": "Thanks @shantanugupta2004! RoPE now uses element-wise multiplication instead of matmul on main via #49265, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/43015"}, {"key": "issue#43011", "n": 43011, "kind": "issue", "group": "Open issues", "category": "feature", "area": "cache / masks", "type": "feature", "title": "`StaticLayer` cache layer to implement `.crop(seq_len)` to match API of `DynamicLayer`", "action": "\u2014", "comment": "", "author": "vadimkantorov", "idle_days": 81, "models": "", "url": "https://github.com/huggingface/transformers/issues/43011"}, {"key": "pr#42978", "n": 42978, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "new model", "title": "Add ViT NEPA", "action": "\u2014", "comment": "", "author": "sbucaille", "idle_days": 251, "models": "vit_nepa", "url": "https://github.com/huggingface/transformers/pull/42978"}, {"key": "issue#42977", "n": 42977, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add ViT NEPA", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [SixAILab/nepa-large-patch14-224](https://huggingface.co/SixAILab/nepa-large-patch14-224), [SixAILab/nepa-base-patch14-224](https://huggingface.co/SixAILab/nepa-base-patch14-224)\n- Downloads (last 30d): 15 \u00b7 Likes: 3 \u00b7 Params: ~303M (base: ~86M) \u00b7 License: apache-2.0\n- Released: 2025-12 \u00b7 Remote code on the Hub: no\n- Open PR: #42978\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "sbucaille", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/42977"}, {"key": "pr#42976", "n": 42976, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Upgrade GitHub Actions to latest versions", "action": "closed", "comment": "Thanks @salmanmkc! The GitHub Actions in these workflows were upgraded and pinned in #46414, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42976"}, {"key": "pr#42975", "n": 42975, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Upgrade GitHub Actions for Node 24 compatibility", "action": "closed", "comment": "Thanks for this! The workflows have since been pinned to commit SHAs in #45180 and #49049, which covers these upgrades, so closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42975"}, {"key": "pr#42944", "n": 42944, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "quantization", "type": "feature / refactor", "title": "[Quantization] From config Quantization for FP8", "action": "maintainer pinged; commented", "comment": "Hi! This overlaps with #46324, which makes `cls(config)` / `from_config` honor `quantization_config` through a generic per-quantizer opt-in, with FP8 as the first user. All three files here also conflict with main now. Should we consolidate on #46324 and close this one, or is there something here (e.g. the FP8 weight init) that should be carried over?", "author": "MekkCyber", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/42944"}, {"key": "pr#42942", "n": 42942, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "Fix result retrieval starvation and terminate request-scoped iteration on completion", "action": "closed", "comment": "Thank you for looking into continuous batching result retrieval, @pythongiant! As @McPatate explained in https://github.com/huggingface/transformers/pull/42942#issuecomment-3670658494, the plan is a different design (tracking running requests by id), so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42942"}, {"key": "pr#42908", "n": 42908, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix gguf tokenizers", "action": "closed", "comment": "Closing since GGUF tokenizer loading was reworked in #48656 and #49008. Please open a new PR if something is still broken.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42908"}, {"key": "pr#42887", "n": 42887, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "core / utils", "type": "fix", "title": "[Quantization] [Compressed Tensors] Support Transforms, Fix Tests", "action": "\u2014", "comment": "", "author": "kylesayrs", "idle_days": 306, "models": "", "url": "https://github.com/huggingface/transformers/pull/42887"}, {"key": "pr#42881", "n": 42881, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "gguf", "type": "feature / refactor", "title": "[GGUF] Add attn_logit_softcapping to Gemma2/Gemma3 config mapping", "action": "\u2014", "comment": "", "author": "kitaekatt", "idle_days": 258, "models": "", "url": "https://github.com/huggingface/transformers/pull/42881"}, {"key": "pr#42876", "n": 42876, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Document tensor parallelism configuration with Trainer", "action": "closed", "comment": "Thanks for writing this up! @stevhliu's suggestions (https://github.com/huggingface/transformers/pull/42876#pullrequestreview-3637117742) weren't pushed (no commits since), so I'm closing for now. You're welcome to open a fresh PR against `docs/source/en/trainer.md` with that change.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42876"}, {"key": "pr#42865", "n": 42865, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Raise explicit error when FP8 is requested via from_config", "action": "closed", "comment": "Thanks @dikshyantacharya for flagging this! The branch is now far behind `main` and `modeling_utils.py` has changed a lot; as far as I can tell `from_config()` still silently ignores `quantization_config`, so the idea may still be worth it. If you'd like to continue, please open a fresh PR against `main` (with the test under the existing `tests/quantization/` layout). Closing this one in the meantime.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42865"}, {"key": "pr#42829", "n": 42829, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "[WIP] End-to-end exportable pipelines (object detection)", "action": "maintainer pinged", "comment": "@yonigozlan this one has been waiting since 2025-12-15. Still planned, or should we close it?", "author": "yonigozlan", "idle_days": -1, "models": "detr, rt_detr, rt_detr_v2", "url": "https://github.com/huggingface/transformers/pull/42829"}, {"key": "pr#42824", "n": 42824, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "processors (image/video/audio)", "type": "fix", "title": "Fix torch only support for fast Processors", "action": "commented", "comment": "Thanks for this! #43514 refactored image processors into backends (so `image_processing_utils_fast.py` is gone), and #48984 now lets you register new processing backends. Could you rework the MLX support on top of that registration API (and the `BatchFeature.convert_to_tensors` part) and rebase on `main`?", "author": "Blaizzy", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/42824"}, {"key": "issue#42821", "n": 42821, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "Support batched torch Tensors in ASTFeatureExtractor", "action": "\u2014", "comment": "", "author": "psu-om", "idle_days": 36, "models": "", "url": "https://github.com/huggingface/transformers/issues/42821"}, {"key": "pr#42816", "n": 42816, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tokenization", "type": "feature / refactor", "title": "validate tokenizer components", "action": "\u2014", "comment": "", "author": "itazap", "idle_days": 312, "models": "", "url": "https://github.com/huggingface/transformers/pull/42816"}, {"key": "issue#42804", "n": 42804, "kind": "issue", "group": "Open issues", "category": "feature", "area": "quantization", "type": "feature", "title": "[`Quantization FP8`] Native `from_config` support", "action": "\u2014", "comment": "", "author": "vasqu", "idle_days": 310, "models": "", "url": "https://github.com/huggingface/transformers/issues/42804"}, {"key": "pr#42793", "n": 42793, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "[Quantization] Fixing last issues for a green 2026 CI hopefully", "action": "closed", "comment": "Closing since the same test fix landed in #42882. Thanks!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42793"}, {"key": "issue#42791", "n": 42791, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add support for GPT_OSS with tp_plan or enable native tensor parallelism", "action": "commented", "comment": "GPT-OSS is supported in transformers as `gpt_oss` (with an expert-parallel plan, `base_model_ep_plan`), and the DTensor mismatch fix from #42906 has been merged. Could you retry on the latest `main` with `tp_plan=\"auto\"` and share a fresh traceback if it still fails?\n\n- Hub: [openai/gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b), [openai/gpt-oss-120b](https://huggingface.co/openai/gpt-oss-120b)\n- Downloads (last 30d): 5,933,604 \u00b7 Likes: 5193 \u00b7 Params: ~20.9B \u00b7 License: apache-2.0", "author": "quic-akuruvil", "idle_days": 0, "models": "gpt_oss", "url": "https://github.com/huggingface/transformers/issues/42791"}, {"key": "pr#42781", "n": 42781, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "new model", "title": "Add VibeVoice Realtime", "action": "\u2014", "comment": "", "author": "ebezzam", "idle_days": 48, "models": "vibevoice_realtime", "url": "https://github.com/huggingface/transformers/pull/42781"}, {"key": "issue#42759", "n": 42759, "kind": "issue", "group": "Open issues", "category": "bug", "area": "other", "type": "bug", "title": "output_hidden_states is not working correctly with SiglipModel class", "action": "\u2014", "comment": "", "author": "hamzaakyildiz", "idle_days": 284, "models": "", "url": "https://github.com/huggingface/transformers/issues/42759"}, {"key": "issue#42758", "n": 42758, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add support for Microsoft VibeVoice models", "action": "commented", "comment": "Good news: this is now supported in transformers as `vibevoice` (added in #40546).", "author": "AliBagheriNejad", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/42758"}, {"key": "pr#42744", "n": 42744, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[FP8 Devstral 24B] Repro PR", "action": "closed", "comment": "Thanks for the repro! As its title says, this draft was a repro of the Devstral FP8 activation-scale issue rather than a fix to merge, so I'm closing it. If the scale issue still reproduces on `main`, an issue pointing to the checkpoint would be the best way to track it.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42744"}, {"key": "pr#42742", "n": 42742, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Remove redundant else in activations.py", "action": "closed", "comment": "Thanks for the cleanup, @leaderofARS! Per our contribution guidelines (`.ai/AGENTS.md`: no one-off PRs for tiny edits such as an isolated lint fix), I'm closing this one. Happy to review contributions on open bugs or features though!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42742"}, {"key": "issue#42738", "n": 42738, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "BERT-like models with RoPE", "action": "commented", "comment": "Status update on this tracking list:\n- NomicBert: supported as `nomic_bert` (#43067)\n- GTE: supported as `gte` (#48416)\n- Jina embeddings: v3 supported as `jina_embeddings_v3` (#44251)\n- Snowflake GTE: still remote code, e.g. [Snowflake/snowflake-arctic-embed-m-v2.0](https://huggingface.co/Snowflake/snowflake-arctic-embed-m-v2.0) (Downloads last 30d: 81,027 \u00b7 Likes: 108 \u00b7 Params: ~305M \u00b7 License: apache-2.0 \u00b7 Released: 2024-11). It may only need a mapping onto the new `gte` model; no open PR found.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "hmellor", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/42738"}, {"key": "pr#42717", "n": 42717, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "processors (image/video/audio)", "type": "fix", "title": "image_transforms: fix tensor annotations", "action": "review requested: @yonigozlan; maintainer pinged", "comment": "@yonigozlan could you take a look when you get a chance? This is waiting on a maintainer: yonigozlan re-review after the 2025-12-12 overloads.", "author": "cebtenzzre", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/42717"}, {"key": "pr#42706", "n": 42706, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Nit parakeet", "action": "maintainer pinged", "comment": "@eustlb this one has been waiting since 2025-12-08. Still planned, or should we close it?", "author": "eustlb", "idle_days": -1, "models": "parakeet", "url": "https://github.com/huggingface/transformers/pull/42706"}, {"key": "pr#42668", "n": 42668, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "More robust processor from pretrained", "action": "\u2014", "comment": "", "author": "yonigozlan", "idle_days": 163, "models": "audioflamingo3, lasr, phi4_multimodal, pix2struct", "url": "https://github.com/huggingface/transformers/pull/42668"}, {"key": "pr#42665", "n": 42665, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "core / loading", "type": "feature / refactor", "title": "Some optimizations for offloading", "action": "maintainer pinged", "comment": "@Cyrilvallez this one has been waiting since 2025-12-05 (never reviewed; core_model_loading.py moved a lot since). Still planned, or should we close it?", "author": "Cyrilvallez", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/42665"}, {"key": "issue#42659", "n": 42659, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "SAM3 Loss", "action": "\u2014", "comment": "", "author": "aselimc", "idle_days": 284, "models": "", "url": "https://github.com/huggingface/transformers/issues/42659"}, {"key": "pr#42655", "n": 42655, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "New Feature: Enabling Speculative Decoding with Batch Size > 1 (If draft and target model share tokenizer)", "action": "closed", "comment": "Thanks for the batched speculative decoding work! As discussed with @zucchini-nlp (https://github.com/huggingface/transformers/pull/42655#issuecomment-3681318123), this is moving to a [custom generate](https://huggingface.co/docs/transformers/main/en/generation_strategies#custom-generation-methods) repo on the Hub, so I'm closing the PR here. Feel free to link the Hub repo once it's published.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42655"}, {"key": "issue#42645", "n": 42645, "kind": "issue", "group": "Open issues", "category": "bug", "area": "processors (image/video/audio)", "type": "bug", "title": "Processor's attribute is mis-detected for child class from #41633", "action": "\u2014", "comment": "", "author": "MilkClouds", "idle_days": 285, "models": "", "url": "https://github.com/huggingface/transformers/issues/42645"}, {"key": "pr#42631", "n": 42631, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "Make GraniteMoeHybridModel compatible with torch.export", "action": "closed", "comment": "Thanks for working on export support for GraniteMoeHybrid, @juanigp! The mamba mask code this PR changes was rewritten for export and linear-attention kernels in #43436 and #47630, so this no longer applies. If export still fails on main, a new issue or PR against the current code is welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42631"}, {"key": "pr#42598", "n": 42598, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "pipelines", "type": "fix", "title": "[pipeline] fix unwanted import failure", "action": "maintainer pinged", "comment": "@eustlb this one has been waiting since 2025-12-03 (never reviewed). Still planned, or should we close it?", "author": "eustlb", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/42598"}, {"key": "pr#42588", "n": 42588, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Document the /v1/models endpoint", "action": "closed", "comment": "Closing as the `/v1/models` endpoint is now documented in `docs/source/en/serve-cli/serving.md` (#42853). Thanks @LysandreJik!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42588"}, {"key": "pr#42572", "n": 42572, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "docs: add doctest for SqueezeBERT", "action": "closed", "comment": "Thanks for the contribution @arrdel! The doctest sprint (#16292) is long closed and SqueezeBERT's examples are now generated by `@auto_docstring`, so hand-written doctests don't fit anymore; closing this.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42572"}, {"key": "issue#42559", "n": 42559, "kind": "issue", "group": "Open issues", "category": "feature", "area": "core / loading", "type": "feature", "title": "Parallelize the model download with the weight loading", "action": "\u2014", "comment": "", "author": "LysandreJik", "idle_days": 316, "models": "", "url": "https://github.com/huggingface/transformers/issues/42559"}, {"key": "pr#42527", "n": 42527, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "Added doctests for SwiftFormer model", "action": "closed", "comment": "Thanks for the contribution @ankitmalakar7! The doctest sprint (#16292) is long closed and model examples are now generated by `@auto_docstring`, so hand-written doctests don't fit anymore; closing this.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42527"}, {"key": "pr#42521", "n": 42521, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Fix FSDP2 defaulting to version 1 in TrainingArguments; add dynamic plugin param passthrough", "action": "closed", "comment": "Thanks a lot for the fix and your patience, @amanzoni1, and sorry this didn't get merged in time! Since then, #45640 rewrote how `fsdp_config` is turned into the Accelerate plugin and made FSDP2 the default (`fsdp_config[\"version\"]` defaults to `2`), so the version being ignored is fixed on `main`. `tests/fsdp/test_fsdp.py` was also replaced by `tests/trainer/distributed/` in #44338. If some plugin options you needed are still not passed through, a small PR against `main` with a test under `tests/trainer/distributed/` would be very welcome. Closing this one, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42521"}, {"key": "pr#42496", "n": 42496, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "feat: allow CP with trainer", "action": "closed", "comment": "Thanks for working on context parallelism in the Trainer, @kmehant! The Trainer already handles CP in `_prepare_context_parallel_inputs` (extended to eval in #48167), and as @SunMarc noted, any changes should go there. Closing this one; a focused PR on top of that function is welcome if something is still missing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42496"}, {"key": "pr#42493", "n": 42493, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "fix: remove trailing os sep in local pretrained model path", "action": "closed", "comment": "Thanks for the fix! As @Rocketknight1 mentioned, the trailing-separator problem from #42490 was already resolved on `main`, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42493"}, {"key": "pr#42461", "n": 42461, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "feature / refactor", "title": "Refactor RMSNorm implementations to use torch.nn.functional.rms_norm", "action": "review requested: @Rocketknight1; maintainer pinged", "comment": "@Rocketknight1 could you take a look when you get a chance? This is waiting on a maintainer: it has been waiting for a maintainer review.", "author": "mstojkovicTT", "idle_days": -1, "models": "aimv2, apertus, arcee, aria, bamba, bitnet, blt, chameleon, clvp, csm, cwm, deepseek_v2, deepseek_v3, dia, diffllama, doge, emu3, ernie4_5, ernie4_5_moe, evolla, exaone4, falcon_h1, glm, glm4, glm4_moe, glm4v, glm4v_moe, gpt_neox, granite, granitemoe, granitemoehybrid, granitemoeshared, hunyuan_v1_dense, hunyuan_v1_moe, idefics2, idefics3, internvl, jamba, jetmoe, lfm2, lfm2_moe, llama, longcat_flash, minimax, ministral3, mistral, mistral3, mixtral, mllama, olmoe, ovis2, paddleocr_vl, phi3, phi4_multimodal, phimoe, pixtral, qwen2_moe, qwen3_moe, qwen3_omni_moe, qwen3_vl_moe", "url": "https://github.com/huggingface/transformers/pull/42461"}, {"key": "pr#42453", "n": 42453, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Add SDPA and FlashAttention support to T5", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #47014, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42453"}, {"key": "pr#42446", "n": 42446, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "Fix DataParallel dtype access in smolvlm", "action": "closed", "comment": "Thanks for the fix! The test it targeted, `test_multi_gpu_data_parallel_forward`, was retired in #48508, so the DataParallel fallback isn't needed anymore. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42446"}, {"key": "pr#42437", "n": 42437, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "One tok typing", "action": "closed", "comment": "Closing this one since the v5 tokenizer unification has landed on main through other PRs and this branch can't be rebased as is. Thanks @itazap!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42437"}, {"key": "pr#42430", "n": 42430, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "\ud83d\udea8 Clean up `image-text-to-text` pipeline", "action": "maintainer pinged", "comment": "@ebezzam this one has been waiting since 2025-12-04 (heavy conflicts + decision on breaking chat-template enforcement). Still planned, or should we close it?", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42430"}, {"key": "pr#42424", "n": 42424, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[WIP] attempt to fix ooms in tests", "action": "closed", "comment": "Closing this WIP: as agreed in https://github.com/huggingface/transformers/pull/42424#pullrequestreview-3515319553, we'd rather find the real root cause of the qwen3 OOMs than add workarounds. Thanks @molbap!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42424"}, {"key": "issue#42422", "n": 42422, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "Performance Improvement: Avoid host\u2194device synchronizations caused by tensor-to-Python conversions and certain tensor ops", "action": "commented", "comment": "Thanks, the first instance was fixed in #42433. To find the rest, run a model under `torch.cuda.set_sync_debug_mode(\"warn\")` (or a profiler trace) and look for `.item()`, `.nonzero()`, `.tolist()` or tensor-valued Python args (e.g. `torch.arange(past_seen_tokens, ...)`) in hot paths. Small, focused PRs per call site (fixing the modular source and running `make fix-repo`) are the easiest to review.", "author": "Dhruv88", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/42422"}, {"key": "pr#42415", "n": 42415, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "initial clean", "action": "closed", "comment": "Closing this early cleanup branch: the v5 tokenization refactor has landed on main and `examples/pytorch` was removed in #49357, so this diff can't be rebased. Anything still missing can go in a fresh, focused PR.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42415"}, {"key": "pr#42413", "n": 42413, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add chatterbox support", "action": "\u2014", "comment": "", "author": "manmay-nakhashi", "idle_days": 285, "models": "chatterbox, s3gen, s3tokenizer", "url": "https://github.com/huggingface/transformers/pull/42413"}, {"key": "pr#42403", "n": 42403, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Sam3 release on v4.57.3", "action": "closed", "comment": "Thanks @yonigozlan for keeping this branch open so people could try SAM3 on a stable install while v5 was being prepared! SAM3 landed on `main` in #42285 and has been in every stable release since v5.0.0, and the `v4.57-release` line ended at v4.57.6. A plain `pip install -U transformers` is all that's needed now, so I'm closing this PR.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42403"}, {"key": "issue#42398", "n": 42398, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "LlamaRMSNorm and equivalent module implementations using torch ops", "action": "\u2014", "comment": "", "author": "mstojkovicTT", "idle_days": 327, "models": "", "url": "https://github.com/huggingface/transformers/issues/42398"}, {"key": "pr#42385", "n": 42385, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix weight tying logic between `_tied_weights_keys` and `tie_word_embeddings`", "action": "closed", "comment": "Thanks for digging into this! The same tying fixes landed in #42523 (kyutai `_tied_weights_keys`, VLM `tie_word_embeddings` handling) and the config rework in #41541, so closing as superseded.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42385"}, {"key": "pr#42345", "n": 42345, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "GPT-OSS Flash Attention and memory-efficient attention via Native PyTorch SDPA", "action": "closed; comment corrected", "comment": "Closing this one: no activity since 2025-11-23, it was written against v4 (v5 changed the codebase a lot). Thanks for the contribution, and feel free to open a fresh PR against `main` if it's still relevant!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42345"}, {"key": "issue#42344", "n": 42344, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "How to fine-tune SAM 3D models?", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [facebook/sam-3d-objects](https://huggingface.co/facebook/sam-3d-objects), [facebook/sam-3d-body-dinov3](https://huggingface.co/facebook/sam-3d-body-dinov3)\n- Downloads (last 30d): 5,934 \u00b7 Likes: 582 \u00b7 License: other\n- Released: 2025-10 \u00b7 Remote code on the Hub: no\n- Weights are gated and ship in the original `sam-3d-objects` format. `sam3` (2D) is already in transformers but SAM 3D is a separate architecture.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "bruno686", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/42344"}, {"key": "pr#42310", "n": 42310, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[WIP] Add moondream3 model", "action": "closed", "comment": "Thanks a lot for starting the Moondream3 port, @nyxkrage! This draft hasn't moved in almost a year, so I'm closing it for now. If you pick it back up, please base it on a `modular_moondream3.py` file ([modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers)) and check it with [mlinter](https://huggingface.github.io/transformers-mlinter/); we'd be happy to review a fresh PR.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42310"}, {"key": "pr#42292", "n": 42292, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "docs: clarify recommended usage of max_new_tokens in generate()", "action": "closed", "comment": "Thanks for the docs suggestion! @stevhliu suggested putting the recommendation in the `GenerationConfig` docstring instead (https://github.com/huggingface/transformers/pull/42292#discussion_r2546735631) and there have been no updates since, so I'm closing it. Happy to review a fresh PR that updates the docstring directly.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42292"}, {"key": "pr#42277", "n": 42277, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "doc(kernels): update kernels integration documentation", "action": "closed", "comment": "Thanks for this! The kernels docs were reworked in #43269 (new `kernel_doc/overview.md` plus `loading_kernels.md`), which covers the integration guide this PR adds and now conflicts with it. Closing as superseded; if something from here is still missing, a small follow-up on top of the current pages would be very welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42277"}, {"key": "pr#42256", "n": 42256, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Integrate Core AutoAWQ Inference Components into Transformers", "action": "closed", "comment": "Thanks for offering to take this on @fanqiNO1! We ended up moving AWQ inference onto GPT-QModel in #41567 (see `src/transformers/integrations/awq.py`), and this PR never got any code, so closing it.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42256"}, {"key": "pr#42244", "n": 42244, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[core] Fix torchao loading", "action": "closed", "comment": "Closing this one: the branch carries an unrelated 100-file diff and torchao loading has been reworked since (#44604, #49548). Please open a fresh PR from main if a torchao loading issue remains.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42244"}, {"key": "pr#42228", "n": 42228, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Support .to(device) or Device Aware Handling for Segmentation Labels in EOMTImageProcessor #42205", "action": "closed", "comment": "Thanks a lot for the fix and for tracking this down! The underlying issue (#42205) was solved generically in #42884: `BatchFeature.to()` now recursively moves tensors inside lists/tuples (see `maybe_to` in `src/transformers/feature_extraction_utils.py`), so `mask_labels`/`class_labels` reach the right device without per-model checks in `forward()`. Closing this one as superseded, but feel free to reopen if you still hit the device mismatch on main.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42228"}, {"key": "pr#42210", "n": 42210, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "[WIP] started adding support for evo2", "action": "\u2014", "comment": "", "author": "McClain-Thiel", "idle_days": 1, "models": "evo2", "url": "https://github.com/huggingface/transformers/pull/42210"}, {"key": "issue#42209", "n": 42209, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add Evo2 Genomic foundation model", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [arcinstitute/evo2_7b](https://huggingface.co/arcinstitute/evo2_7b), [arcinstitute/evo2_1b_base](https://huggingface.co/arcinstitute/evo2_1b_base)\n- Downloads (last 30d): 8.3k \u00b7 Likes: 67 \u00b7 Params: n/a \u00b7 License: apache-2.0\n- Released: 2025-02 \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "McClain-Thiel", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/42209"}, {"key": "issue#42179", "n": 42179, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Add TileLang Kernel Support", "action": "closed", "comment": "A TileLang attention kernel can already be used as an optional backend without changes to the library: register it with `AttentionInterface.register(...)` (`src/transformers/modeling_utils.py`) and select it via `attn_implementation`, see https://huggingface.co/docs/transformers/main/en/attention_interface. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/42179"}, {"key": "pr#42166", "n": 42166, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "add internvl_flash model", "action": "\u2014", "comment": "", "author": "Chenhao-Guan", "idle_days": 269, "models": "internvl_flash", "url": "https://github.com/huggingface/transformers/pull/42166"}, {"key": "pr#42134", "n": 42134, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Add AutoMergeAdapters utility for merging multiple LoRA adapters with\u2026", "action": "closed", "comment": "Thanks for the contribution! Adapter merging is handled by the PEFT library, which transformers integrates with, so we'd rather not add a separate adapters module here; huggingface/peft would be the right place.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42134"}, {"key": "pr#42133", "n": 42133, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Fix qwen moe Load balancing loss calculation outside training", "action": "closed", "comment": "Thanks @diegoakel! On main the load-balancing loss is only computed when `output_router_logits` is enabled (see `modular_qwen3_moe.py`), and @Rocketknight1's question in #42100 about how it got turned on at inference was never resolved, so I'm closing this. Happy to reopen if you have a reproducer on current main.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42133"}, {"key": "issue#42132", "n": 42132, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Support for Megatron", "action": "closed", "comment": "Closing this one: no activity since 2025-11-10, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/42132"}, {"key": "pr#42131", "n": 42131, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Mistral Tokenizer Converter Script - Initialization with Pattern Argument", "action": "closed", "comment": "Thanks for the fix! `src/transformers/integrations/mistral.py` was moved into the `src/transformers/integrations/mistral/` package in #46603, and the Tekken converter there now reads the regex pattern directly from `tekken.json`, so the correct pattern is picked up without this change. Closing as superseded; please open an issue if you still see a wrong pre-tokenizer on `main`.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42131"}, {"key": "pr#42130", "n": 42130, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Try refactoring logging to make type checks pass", "action": "closed", "comment": "Closing this experimental logging refactor since it has been idle for almost a year; feel free to reopen if you pick the typing cleanup of `utils/logging.py` back up.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42130"}, {"key": "pr#42127", "n": 42127, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Standardize conv len function for audio models", "action": "closed", "comment": "Closing this one: no activity since 2025-11-10, it was written against v4 (v5 changed the codebase a lot). Thanks for the contribution, and feel free to open a fresh PR against `main` if it's still relevant!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42127"}, {"key": "pr#42098", "n": 42098, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "fix", "title": "Fix mel length computation in Qwen2-Audio", "action": "\u2014", "comment": "", "author": "anferico", "idle_days": 327, "models": "qwen2_audio", "url": "https://github.com/huggingface/transformers/pull/42098"}, {"key": "pr#42051", "n": 42051, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "tokenization", "type": "fix", "title": "Fix model_input_names singleton issue causing shared state", "action": "review requested: @itazap; maintainer pinged", "comment": "@itazap could you take a look when you get a chance? This is waiting on a maintainer: needs maintainer review (Rocketknight1 LGTM'd and cc'd ArthurZucker/itazap).", "author": "yashwantbezawada", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/42051"}, {"key": "pr#42039", "n": 42039, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "models", "type": "feature / refactor", "title": "[WIP] \ud83d\udea8 clean xcodec \ud83e\uddfc", "action": "\u2014", "comment": "", "author": "eustlb", "idle_days": 348, "models": "dac, xcodec", "url": "https://github.com/huggingface/transformers/pull/42039"}, {"key": "pr#42000", "n": 42000, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix Mixtral: Docstring uses consistent 'top_k_index' and 'top_k_weights' in MixtralSparseMoeBlock", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #42041, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/42000"}, {"key": "issue#41999", "n": 41999, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Add Timestamp Support for Voxtral Models", "action": "closed", "comment": "As @eustlb explained (https://github.com/huggingface/transformers/issues/41999#issuecomment-3527840653), Voxtral's timestamps are only available through Mistral's API and the released model has no alignment heads like Whisper, so we can't add it. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/41999"}, {"key": "pr#41980", "n": 41980, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Correct type hint in config models", "action": "closed", "comment": "Thanks for catching this! The configs were since rewritten and now all declare `rms_norm_eps: float` (see `src/transformers/models/llama/configuration_llama.py` on main), so this is already fixed. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41980"}, {"key": "pr#41977", "n": 41977, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Add Phi3.5 Vision Model", "action": "\u2014", "comment": "", "author": "yaswanth19", "idle_days": 306, "models": "phi3_v", "url": "https://github.com/huggingface/transformers/pull/41977"}, {"key": "pr#41967", "n": 41967, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "feat: RoPE-related typing improvements", "action": "closed", "comment": "Closing this draft since it has been inactive for a while and `modeling_rope_utils.py` has been substantially reworked since (e.g. #48105, #48594, #49555), so it no longer applies. The typing improvement itself (`total=False` + `Required[...]` on `RopeParameters`) is still welcome. #41964 stays open, and a fresh, smaller PR against current main would be easy to review. Thanks @RyanMullins!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41967"}, {"key": "issue#41964", "n": 41964, "kind": "issue", "group": "Open issues", "category": "feature", "area": "core / utils", "type": "feature", "title": "Improve typing support and flexibility for RopeParameters", "action": "commented", "comment": "Still relevant: `RopeParameters` in `src/transformers/modeling_rope_utils.py` is still a total `TypedDict`, and `standardize_rope_params` hasn't gained the typing improvements. The open draft #41967 implements this; it needs a rebase on current main and to be marked ready for review.", "author": "RyanMullins", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/41964"}, {"key": "pr#41928", "n": 41928, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "fix: add clear error message when mistral-common is missing for AutoTokenizer loading Voxtral", "action": "closed", "comment": "Thanks for the fix! As @vasqu noted (https://github.com/huggingface/transformers/pull/41928#issuecomment-3461405988), we wanted a model-agnostic fix rather than a Voxtral-only check, and `tokenization_auto.py` on main now resolves the Mistral backend via `_use_mistral_format` / `resolve_mistral_format`. Closing; if you still hit an unclear error on main, please open an issue with a repro.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41928"}, {"key": "pr#41901", "n": 41901, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "[executorch] Update pytree registration for DynamicCache", "action": "closed", "comment": "Thanks @justinchuby! The DynamicCache pytree registration in `integrations/executorch.py` has since been rewritten around a dict + layout flatten (#47900), so this diff no longer applies. If the ONNX exporter still needs key/value interleaving, a new PR against the current code would be great.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41901"}, {"key": "pr#41899", "n": 41899, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Testing checkpoint limit changes from PR #37196", "action": "closed", "comment": "Thanks for picking this up! Since this was opened, checkpoint rotation moved out of `Trainer` into `trainer_utils.rotate_checkpoints` (#43736), and that function now always keeps the best checkpoint next to the most recent one, which was the main motivation here. The branch targets the old `_rotate_checkpoints` code and now conflicts. I'll close this for now; if separate limits for full checkpoints vs model weights are still needed, please open an issue describing the use case and we can see whether it fits on top of `rotate_checkpoints`. Closing this one for now; feel free to open a fresh PR against `main` if it's still needed. Thanks!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41899"}, {"key": "pr#41895", "n": 41895, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Add Telugu Sentiment Classification Example using DistilBERT", "action": "closed", "comment": "We removed the legacy `examples/` scripts from the repository (#49357) since they were no longer maintained, so this change no longer has a home. A notebook or Space on the Hub would be a great way to share this. Thanks!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41895"}, {"key": "issue#41893", "n": 41893, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add MiniViT: lightweight Vision Transformer for CIFAR-scale image classification", "action": "commented", "comment": "No public weights on the Hub as of today. The proposal describes a new from-scratch model with no released pretrained checkpoint. New architectures need published weights first; in the meantime it can be shared on the Hub with [custom code](https://huggingface.co/docs/transformers/main/en/custom_models).", "author": "justynigam", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/41893"}, {"key": "pr#41886", "n": 41886, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "ADD FG-CLIP2", "action": "\u2014", "comment": "", "author": "binwang777", "idle_days": 251, "models": "fgclip2", "url": "https://github.com/huggingface/transformers/pull/41886"}, {"key": "pr#41880", "n": 41880, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Indonesian Language Support for ReadMe", "action": "closed", "comment": "Thanks for the Indonesian translation! As @stevhliu pointed out in https://github.com/huggingface/transformers/pull/41880#discussion_r2511884861, it was based on an older `README.md`, so it no longer applies cleanly. Closing for now; you're very welcome to open a fresh PR translated from the current `README.md`.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41880"}, {"key": "pr#41879", "n": 41879, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Fix/processor multiple tokenizers", "action": "closed", "comment": "Thanks for tackling this! Saving processors with several tokenizers was since fixed in #42630, and #42667 added support for multiple sub-processors of any kind (building on the refactor in #41633), so this branch is superseded. Closing; please open an issue if you still hit a case that isn't covered on `main`.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41879"}, {"key": "issue#41862", "n": 41862, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Request for InternVL3_5_Flash", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [OpenGVLab/InternVL3_5-8B-Flash](https://huggingface.co/OpenGVLab/InternVL3_5-8B-Flash)\n- Downloads (last 30d): 378 \u00b7 Likes: 8 \u00b7 Params: ~8.8B \u00b7 License: apache-2.0\n- Released: 2025-09 \u00b7 Remote code on the Hub: yes\n- Open PR: #42166\n- The base `internvl` model is already supported; only the Flash variant (with its resolution router) is missing.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "YanxingLiu", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/41862"}, {"key": "pr#41855", "n": 41855, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "tokenization", "type": "feature / refactor", "title": "Add Mistral tokenizer missing methods", "action": "\u2014", "comment": "", "author": "ChrisHughes", "idle_days": 181, "models": "", "url": "https://github.com/huggingface/transformers/pull/41855"}, {"key": "pr#41851", "n": 41851, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Fix deepcopy in ProcessorMixin.to_dict for GemmaTokenizerFast", "action": "closed", "comment": "Thanks @aijadugar for the fix! `ProcessorMixin.to_dict` was since reworked in #44894 to exclude tokenizer attributes before deep-copying, which avoids the fast-tokenizer deepcopy issue from #41837. Closing as this is now covered on `main`; please open a new issue if you still see the error.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41851"}, {"key": "issue#41838", "n": 41838, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "ONNX export vmap functorch\"", "action": "closed; comment corrected", "comment": "The vmap in mask creation was the culprit, and since #41852 masks are built without `torch.vmap` by default (`src/transformers/masking_utils.py`), so `torch.onnx.export` should get past this on v5. For export with a KV cache, see `transformers.exporters` (https://huggingface.co/docs/transformers/main/en/exporters). Closing; please reopen with a traceback if it still fails on main.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/41838"}, {"key": "pr#41827", "n": 41827, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "attention / kernels", "type": "feature / refactor", "title": "[`Flash Attention`] Disable packed sequences with pos ids only during torch compile", "action": "closed", "comment": "Closing this reference draft: `modeling_flash_attention_utils.py` has been substantially reworked since (e.g. #49610 and the `is_tracing` guards), so the idea would need to be redone on top of main anyway. Thanks @vasqu!", "author": "vasqu", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/pull/41827"}, {"key": "issue#41826", "n": 41826, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "Integrating TiledMLP for a much smaller memory footprint", "action": "\u2014", "comment": "", "author": "stas00", "idle_days": 57, "models": "", "url": "https://github.com/huggingface/transformers/issues/41826"}, {"key": "pr#41823", "n": 41823, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Lfm2-VL vllm", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41823"}, {"key": "pr#41807", "n": 41807, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "git commit -m \"Fix: corrected outdated documentation link in README.md\"", "action": "closed", "comment": "Thanks for looking out for broken links! As @Rocketknight1 checked (https://github.com/huggingface/transformers/pull/41807#issuecomment-3436686928), both URLs resolve correctly, so there's nothing to fix. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41807"}, {"key": "issue#41803", "n": 41803, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "torch.compile graph break in `flash_attention_v2` backend", "action": "closed", "comment": "As explained above, a graph break around the flash-attention varlen path can't be avoided in general right now. Use `fullgraph=False` (usually negligible overhead) or prepare the inputs as @vasqu showed; for fully compiled graphs `attn_implementation=\"sdpa\"` compiles cleanly. #41827 has the reference change if someone wants to push it further. Closing as not planned for now.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/41803"}, {"key": "pr#41800", "n": 41800, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Increasing clarity", "action": "closed", "comment": "We removed the legacy `examples/` scripts from the repository (#49357) since they were no longer maintained, so this change no longer has a home. Thanks!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41800"}, {"key": "pr#41798", "n": 41798, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "p-less Sampling:  A Robust Hyperparameter-Free Approach for LLM Decoding", "action": "closed", "comment": "Thanks for contributing p-less sampling! New decoding strategies now live on the Hub as custom `generate` methods, which lets you ship them without waiting on core review, so we're closing this here.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41798"}, {"key": "pr#41797", "n": 41797, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "new model", "title": "Add deepseek ocr", "action": "maintainer pinged", "comment": "@molbap this one has been waiting since 2026-03-01 (solid checkpoint-to-canonical weight mapping (molbap, 2026-01-19)). Still planned, or should we close it?", "author": "molbap", "idle_days": -1, "models": "deepseek_ocr", "url": "https://github.com/huggingface/transformers/pull/41797"}, {"key": "pr#41794", "n": 41794, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Enable flake8-pie rules", "action": "closed", "comment": "Closing this one: no activity since 2026-01-01. Thanks for the contribution, and feel free to open a fresh PR against `main` if it's still relevant!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41794"}, {"key": "pr#41754", "n": 41754, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "exporters", "type": "feature / refactor", "title": "Add pytree registration for static cache", "action": "\u2014", "comment": "", "author": "angelayi", "idle_days": 340, "models": "", "url": "https://github.com/huggingface/transformers/pull/41754"}, {"key": "issue#41753", "n": 41753, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Please port for MPS acceleration on MacOS.", "action": "closed", "comment": "Transformers already runs on Apple Silicon through PyTorch's MPS backend, e.g. `model.to(\"mps\")` or `device_map=\"mps\"`, as documented in `docs/source/en/apple_silicon.md` (https://huggingface.co/docs/transformers/main/en/apple_silicon). Closing; please open a new issue with a reproducer if a specific model fails on MPS.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/41753"}, {"key": "pr#41734", "n": 41734, "kind": "PR", "group": "Closed PRs", "category": "obsolete", "area": "", "type": "", "title": "Fix CUDA errors in sharded generation with Qwen3", "action": "closed", "comment": "Thank you for digging into #41720, @SrijanUpadhyay! Forcing `remove_invalid_values` for all sharded models in the generic generation path would hide numerical issues rather than fix them, and a tester reported the error still occurs with this branch. Since #41720 has been closed, I'm closing this PR as well.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41734"}, {"key": "pr#41724", "n": 41724, "kind": "PR", "group": "Open PRs", "category": "conflicting", "area": "models", "type": "fix", "title": "Fix confusing warning in EncoderDecoderModel when creating decoder_input_ids from labels", "action": "review requested: @NielsRogge; maintainer pinged", "comment": "@NielsRogge could you take a look when you get a chance? This is waiting on a maintainer: it has been waiting for a maintainer review.", "author": "st81", "idle_days": -1, "models": "encoder_decoder", "url": "https://github.com/huggingface/transformers/pull/41724"}, {"key": "pr#41718", "n": 41718, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "AutoTokenizer: clear ImportError when loading Voxtral without mistral-common + unit test", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41718"}, {"key": "pr#41710", "n": 41710, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `main_classes/backbones.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41710"}, {"key": "pr#41701", "n": 41701, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix qwen3_vl mix precision dtype", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41701"}, {"key": "pr#41698", "n": 41698, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Fix tokenizer check script: safe dataset access, default checkpoints, and tested in dry-run mode", "action": "closed", "comment": "Thanks a lot for improving `scripts/check_tokenizers.py`, @aijadugar! As @Rocketknight1 said (https://github.com/huggingface/transformers/pull/41698#issuecomment-3421981803), we're not sure we're keeping that script, so we won't merge fixes to it right now. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41698"}, {"key": "pr#41687", "n": 41687, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix(data): Handle integer labels in DataCollatorWithFlattening", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41687"}, {"key": "pr#41654", "n": 41654, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Improve LLaMA tokenizer error when vocab is missing: suggest installi\u2026", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41654"}, {"key": "pr#41631", "n": 41631, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Incorrect access of dataset field fixed", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41631"}, {"key": "pr#41597", "n": 41597, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Standardize RoBERTa model card following issue #36979", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41597"}, {"key": "pr#41594", "n": 41594, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Add beginner-friendly sentiment analysis example", "action": "closed", "comment": "We removed the legacy `examples/` scripts from the repository (#49357) since they were no longer maintained, so this change no longer has a home. For tutorials, a blog post or a notebook on the Hub is a great place. Thanks!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41594"}, {"key": "pr#41593", "n": 41593, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "examples: add multi-label text classification (BCEWithLogitsLoss, met\u2026", "action": "closed", "comment": "Thanks for putting this multi-label classification example together, @aryaMehta26! The `examples/pytorch/` scripts, including the text-classification folder this builds on, were removed from the repo in #49357 because they were no longer maintained. Multi-label setups are covered by `problem_type=\"multi_label_classification\"` on the sequence-classification models, and a notebook or blog post would be a great place to share this example. Closing this PR, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41593"}, {"key": "pr#41592", "n": 41592, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Improve AutoTokenizer error message for Voxtral models missing mistral-common", "action": "closed", "comment": "Thanks for working on this! The underlying issue is now handled by #42299 (auto-converting tekken.json), so closing this as superseded.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41592"}, {"key": "pr#41584", "n": 41584, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add clear error message for missing SentencePiece model in `get_spm_processor` (fix #41553)", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41584"}, {"key": "pr#41565", "n": 41565, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Updated `perf_train_gpu_many.md`", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41565"}, {"key": "pr#41561", "n": 41561, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Optimize Mamba2 memory usage by replacing broadcast with einsum", "action": "closed", "comment": "Thanks @SohamWalam11 for working on this! The same einsum contraction of the Mamba2 chunk scan was merged in #48978, so I'm closing this as superseded. Thanks for pushing it forward, and for the benchmarking discussion.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41561"}, {"key": "pr#41557", "n": 41557, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update modeling_llama4.py", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41557"}, {"key": "pr#41531", "n": 41531, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `video_processor.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41531"}, {"key": "pr#41528", "n": 41528, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add position encoding interpolation to DeiT", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41528"}, {"key": "pr#41527", "n": 41527, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated selecting.md to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41527"}, {"key": "pr#41524", "n": 41524, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add max_eval_batches argument to TrainingArguments", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41524"}, {"key": "pr#41523", "n": 41523, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add test coverage for ConvNextImageProcessorFast", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41523"}, {"key": "pr#41522", "n": 41522, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Fix _init_weights to safely skip int8 quantized weights", "action": "closed", "comment": "Thanks for the fix! Since then, initialization was reworked for v5 (`src/transformers/initialization.py` skips tensors already loaded from the checkpoint), the related issue #39366 was closed as no longer relevant, and `modeling_qwen2_5_vl.py` is now generated from `modular_qwen2_5_vl.py`, so edits there would be overwritten anyway. If W8A8 Qwen2.5-VL still fails to load on v5, please open an issue with a reproducer. Closing this PR.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41522"}, {"key": "pr#41521", "n": 41521, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Fix forced_bos_token_id not set in generation_config", "action": "closed", "comment": "Thanks @Addyk-24 for the fix and for cleaning up the diff! `examples/pytorch/translation/run_translation.py` was removed together with the rest of the `examples/` folder in #49357, so this change has nothing left to apply to. If you think `forced_bos_token_id` should be propagated to `generation_config` in the library itself, feel free to open an issue or PR against `main`. Closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41521"}, {"key": "pr#41491", "n": 41491, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add skip_unnecessary_grad_clip to TrainingArguments for optimized gradient clipping", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41491"}, {"key": "pr#41490", "n": 41490, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Fix _init_weights to safely skip int8 tensors in Qwen2_5_VL model", "action": "closed", "comment": "Thanks for digging into this, @KaparthyReddy! Weight initialization was reworked in #42191: Qwen2.5-VL no longer has its own `_init_weights`, and weights loaded from the checkpoint are flagged and skipped by the init functions, so quantized int8 weights aren't touched anymore. As a side note, `modeling_qwen2_5_vl.py` is generated from a [modular](https://huggingface.co/docs/transformers/main/en/modular_transformers) file, so fixes go in `modular_qwen2_5_vl.py`. If you still hit an error loading a quantized Qwen2.5-VL on `main`, please open an issue with a reproducer. Closing this one, thanks!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41490"}, {"key": "pr#41488", "n": 41488, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Create 1", "action": "closed", "comment": "Thanks for your interest in Transformers! This PR only adds a file named `1` without any description, so there's nothing for us to review. Closing it; if you meant to propose a change, please open a new PR against `main` describing what it does.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41488"}, {"key": "pr#41485", "n": 41485, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix smolvlm2 dtype mismatch final", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41485"}, {"key": "issue#41462", "n": 41462, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "Support encoder text classification for sequence to sequence models like BART and T5", "action": "commented", "comment": "Not available yet: `T5ForSequenceClassification`/`BartForSequenceClassification` use the full encoder-decoder. `T5ForTokenClassification` in `src/transformers/models/t5/modeling_t5.py` already wraps only `T5EncoderModel`, so a `T5EncoderForSequenceClassification` with mean pooling over non-padded positions could follow that pattern (registered in `models/auto/modeling_auto.py`, with tests). Please open a PR if you'd like to take it, see https://huggingface.co/docs/transformers/main/en/contributing.", "author": "cbhyphen", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/41462"}, {"key": "pr#41458", "n": 41458, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "Adding ScatterMoE kernel support for Granite models. ", "action": "\u2014", "comment": "", "author": "shawntan", "idle_days": 264, "models": "granitemoe, granitemoehybrid, granitemoeshared", "url": "https://github.com/huggingface/transformers/pull/41458"}, {"key": "pr#41441", "n": 41441, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "core / utils", "type": "feature / refactor", "title": "Enhance the handling of Union types in HfArgumentParser", "action": "\u2014", "comment": "", "author": "cyyever", "idle_days": 159, "models": "", "url": "https://github.com/huggingface/transformers/pull/41441"}, {"key": "pr#41419", "n": 41419, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "First QAT for Finegrained FP8", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41419"}, {"key": "pr#41406", "n": 41406, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83d\udea8 [v5] Remove deprecated cache classes", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41406"}, {"key": "issue#41399", "n": 41399, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Filter models by Original Provider", "action": "closed", "comment": "As @Rocketknight1 explained in https://github.com/huggingface/transformers/issues/41399#issuecomment-3376750682, organization pages (e.g. https://huggingface.co/Qwen) list only that org's releases; this is a Hub website request rather than a `transformers` one, so https://github.com/huggingface/hub-docs/issues is the right place. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/41399"}, {"key": "pr#41362", "n": 41362, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Added Hacktoberfest banner image to README.md", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41362"}, {"key": "pr#41356", "n": 41356, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Add DEIMv2 model, image processor, and basic tests", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #44339, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41356"}, {"key": "pr#41349", "n": 41349, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Create (3d_parrallel_v2.py) - Add 3D parallelism training example script", "action": "closed", "comment": "Thanks @Haziq2006 for putting together this 3D-parallel training example! The `examples/` folder was removed from the repository in #49357, so we're no longer accepting new example scripts here; this kind of script is a good fit for a blog post, a Hub repo or the docs instead. Closing for that reason.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41349"}, {"key": "pr#41333", "n": 41333, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Add DeepseekVLV2 Model", "action": "closed", "comment": "Closing this one: no activity since 2025-11-16, 8 of the 12 files it changes no longer exist on `main`, it was written against v4 (v5 changed the codebase a lot). Thanks for the contribution, and feel free to open a fresh PR against `main` if it's still relevant!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41333"}, {"key": "pr#41330", "n": 41330, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Unskip and fix offline mode tests, use HF_HUB_OFFLINE, make hermetic", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41330"}, {"key": "pr#41329", "n": 41329, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Fix GIL=0 segfault and Add GIL=0 compat for regex paths", "action": "closed", "comment": "Thanks for the free-threading work! As mentioned, v5 moved tokenization onto the `tokenizers` backend and these slow per-model files are going away, so we'll close this; the `regex` -> `re` cleanup was tracked separately in #43310.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41329"}, {"key": "pr#41319", "n": 41319, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Use torch._check instead of a test in Gemma3Model", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41319"}, {"key": "pr#41315", "n": 41315, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[model deprecations] Define new version-based model deprecation/deletions with user warnings/exceptions", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41315"}, {"key": "pr#41313", "n": 41313, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Jitter noise PR", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41313"}, {"key": "pr#41312", "n": 41312, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "tests: unskip and fix offline mode test using HF_HUB_OFFLINE + hermetic cache warmup", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41312"}, {"key": "issue#41306", "n": 41306, "kind": "issue", "group": "Open issues", "category": "feature", "area": "generation", "type": "feature", "title": "TiledFusedLogitsLoss (LigerFusedLinearCrossEntropy, FLCE) Flag in `from_pretrained()`", "action": "\u2014", "comment": "", "author": "Tcc0403", "idle_days": 130, "models": "", "url": "https://github.com/huggingface/transformers/issues/41306"}, {"key": "pr#41299", "n": 41299, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "copied changes from soghomon-b:add-eval-step-limit-31561", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41299"}, {"key": "pr#41291", "n": 41291, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Add-Deimv2", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #44339, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41291"}, {"key": "pr#41273", "n": 41273, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix(quantization): Skip weight initialization for quantized models", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41273"}, {"key": "pr#41254", "n": 41254, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "docs: Add Hacktoberfest banner, Contributing and License sections to README", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41254"}, {"key": "pr#41239", "n": 41239, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Add num_hidden_layers to t5gemma's top level config", "action": "closed", "comment": "Thanks @winnie0617! The underlying issue was fixed by #40939, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41239"}, {"key": "pr#41224", "n": 41224, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "new model", "title": "Add DINOv3ViTForImageClassification support", "action": "\u2014", "comment": "", "author": "dimidagd", "idle_days": 222, "models": "dinov3_vit", "url": "https://github.com/huggingface/transformers/pull/41224"}, {"key": "pr#41215", "n": 41215, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix CLIP memory leak causing 600-800MB accumulation per batch", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41215"}, {"key": "pr#41202", "n": 41202, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[WIP] standardize audio kwargs", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41202"}, {"key": "pr#41169", "n": 41169, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix TorchDynamo crash in StaticCache by validating offloading and offload_only_non_sliding arguments", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41169"}, {"key": "pr#41162", "n": 41162, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Add Sinhala (\u0dc3\u0dd2\u0d82\u0dc4\u0dbd) translation of README", "action": "closed", "comment": "Thanks a lot for the Sinhala translation, @Ranjuna120! This PR has sat for over a year and `README.md` has changed a lot since, so the translation and the language links no longer match `main`. We still take README translations (see the Turkish one in #46312), so a fresh PR against `main` that translates the current `README.md` into `i18n/README_si.md` would be very welcome. Closing this one, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41162"}, {"key": "pr#41160", "n": 41160, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[docs] update tips syntax", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41160"}, {"key": "pr#41159", "n": 41159, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Support setting total_train_batch_size.", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41159"}, {"key": "pr#41144", "n": 41144, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Support automatic conversion from zero checkpoint to universal checkpoint.", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41144"}, {"key": "pr#41132", "n": 41132, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix(SpeechT5Config): missing annotation on `inputs_to_logits_ratio` property", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41132"}, {"key": "pr#41121", "n": 41121, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix: resolve the unexpected video frame drop issue of the InternVL model with multiple video inputs", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41121"}, {"key": "pr#41105", "n": 41105, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix is_torch_neuroncore_available", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41105"}, {"key": "pr#41097", "n": 41097, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Delay and probably avoid unnecessary graph breaks in _upad_input of modeling_flash_attention_utils.py", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41097"}, {"key": "pr#41095", "n": 41095, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "new model", "title": "Add LLaVA-OneVision-1.5 model and related configurations", "action": "commented", "comment": "Thanks for the work on LLaVA-OneVision-1.5! This branch predates v5 and now conflicts with main (auto mappings, model init and docs moved a lot since Nov 2025). The model is still not on main, so it is still welcome: could you rebase on current main and regenerate the modeling file from the modular file (ideally reusing the shared Qwen/LLaVA components) so we can do a fresh review?", "author": "g1050", "idle_days": -1, "models": "llava_onevision, llava_onevision1_5", "url": "https://github.com/huggingface/transformers/pull/41095"}, {"key": "issue#41084", "n": 41084, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Set Block Decoding", "action": "closed", "comment": "As @Rocketknight1 noted (https://github.com/huggingface/transformers/issues/41084#issuecomment-3908685819), Set Block Decoding requires retrained models, so it isn't a standard decoding method we can add to `generate()`. Closing for now.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/41084"}, {"key": "issue#41081", "n": 41081, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add support for LLaVA-OneVision-1.5 Multi-Modal Model", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [lmms-lab/LLaVA-OneVision-1.5-8B-Instruct](https://huggingface.co/lmms-lab/LLaVA-OneVision-1.5-8B-Instruct)\n- Downloads (last 30d): 7.0k \u00b7 Likes: 68 \u00b7 Params: ~8.5B \u00b7 License: apache-2.0\n- Released: 2025-09 \u00b7 Remote code on the Hub: yes\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "g1050", "idle_days": 0, "models": "llava_onevision", "url": "https://github.com/huggingface/transformers/issues/41081"}, {"key": "pr#41075", "n": 41075, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix Qwen3 deterministic generation when do_sample=False and num_beams=1 for Greedy Decoding", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41075"}, {"key": "pr#41053", "n": 41053, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Qwen3 moe", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41053"}, {"key": "pr#41041", "n": 41041, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[WIP] Add YuE model", "action": "closed", "comment": "Thanks for the work on YuE @Manalelaidouni! This draft hasn't moved in a long time and the modeling part never landed, so I'm closing it for now. Please feel free to reopen or start a fresh [modular](https://huggingface.co/docs/transformers/main/en/modular_transformers) PR if you get back to it.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41041"}, {"key": "pr#41040", "n": 41040, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Add Keye vl 8b 1.5", "action": "closed", "comment": "Thanks for the work on Keye-VL-1.5 @Kwai-Keye! @zucchini-nlp's request to fix the failing tests (https://github.com/huggingface/transformers/pull/41040#issuecomment-3425587340) is still open and there have been no commits since 2025-10-20, so I'm closing it. If you pick it back up, a fresh modular PR on top of Qwen2-VL would make review much faster.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41040"}, {"key": "pr#41037", "n": 41037, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Tests: Apertus integration tests", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41037"}, {"key": "pr#41035", "n": 41035, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "docs: update speech recognition examples to use modern Common Voice d\u2026", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #41027, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41035"}, {"key": "pr#41033", "n": 41033, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "feat: make audio feature extractors torch.export-able", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41033"}, {"key": "pr#41022", "n": 41022, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `backbones.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41022"}, {"key": "pr#41021", "n": 41021, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `video_processors.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41021"}, {"key": "pr#41009", "n": 41009, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add Lexa-Delta model support", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/41009"}, {"key": "issue#40986", "n": 40986, "kind": "issue", "group": "Open issues", "category": "feature", "area": "processors (image/video/audio)", "type": "feature", "title": "Rewrite audio feature processors to accept torch tensors", "action": "\u2014", "comment": "", "author": "jackzhxng", "idle_days": 33, "models": "", "url": "https://github.com/huggingface/transformers/issues/40986"}, {"key": "pr#40976", "n": 40976, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Better defaults for assisted generation", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40976"}, {"key": "pr#40962", "n": 40962, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "perceptron: Isaac-0.1 implementation", "action": "\u2014", "comment": "", "author": "AkshatSh", "idle_days": 102, "models": "isaac", "url": "https://github.com/huggingface/transformers/pull/40962"}, {"key": "pr#40908", "n": 40908, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix `load_balancing_loss_func` incompatible with `past_key_values`", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40908"}, {"key": "pr#40898", "n": 40898, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Adding [T5/MT5/UMT5]EncoderForSequenceClassification", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40898"}, {"key": "pr#40887", "n": 40887, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Refactor output handling in generate for cleaner decoding methods", "action": "closed", "comment": "Thanks a lot for starting this refactor! It has been idle since @zucchini-nlp asked whether you'd finish it, and `generation/utils.py` has moved a lot since, so I'm closing it. @zucchini-nlp, feel free to pick up the idea in a fresh PR.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40887"}, {"key": "pr#40877", "n": 40877, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Bug #40833: Fix for kv_offset calculation for mixed padding", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40877"}, {"key": "pr#40871", "n": 40871, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Refactor benchmark utils: add type hints, GPU metrics helper, and con\u2026", "action": "closed", "comment": "Thanks for the effort here! `benchmark_v2/benchmark_framework.py` was removed in the benchmark overhaul in #41408, so these changes no longer apply. Closing this one, but if you'd like to improve the new benchmark code, feel free to open a PR against `main` that targets the current `benchmark_v2/` layout.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40871"}, {"key": "pr#40870", "n": 40870, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Reduce vRAM usage during generation by allowing to transfer logits to CPU", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40870"}, {"key": "pr#40861", "n": 40861, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Support n_groups>1 for mamba2 ", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40861"}, {"key": "pr#40857", "n": 40857, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Token", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #41156, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40857"}, {"key": "issue#40841", "n": 40841, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add Support for Ovis2.5 Multi-Modal Model\u200b", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [ATH-MaaS/Ovis2.5-9B](https://huggingface.co/ATH-MaaS/Ovis2.5-9B), [ATH-MaaS/Ovis2.5-2B](https://huggingface.co/ATH-MaaS/Ovis2.5-2B)\n- Downloads (last 30d): 4,452 \u00b7 Likes: 308 \u00b7 Params: ~9.2B (2B: ~2.6B) \u00b7 License: apache-2.0\n- Released: 2025-08 \u00b7 Remote code on the Hub: yes\n- The previous generation is already supported as `ovis2` (#37088), so this is likely a modular extension of it. Note `AIDC-AI/*` now redirects to `ATH-MaaS/*`.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "xschen-beb", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/40841"}, {"key": "pr#40840", "n": 40840, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "feat: add qwen2 pruning support", "action": "closed", "comment": "Thanks for the PR! As discussed in #40759, we don't want to add per-layer structured-pruning support to models that weren't released that way. Pruned variants are best shared as Hub remote code. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40840"}, {"key": "issue#40830", "n": 40830, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "`pipeline` support for backends like `vLLM` and `SGLang`", "action": "closed", "comment": "As @Rocketknight1 explained (https://github.com/huggingface/transformers/issues/40830#issuecomment-3284124728), `pipeline` is meant as a simple high-level API. For large offline jobs, transformers now has continuous batching (`model.generate_batch(...)`, https://huggingface.co/docs/transformers/main/en/continuous_batching) and `transformers serve`. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/40830"}, {"key": "pr#40820", "n": 40820, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Add models to benchmarks", "action": "closed", "comment": "Thanks @ahadnagy for adding these models to the benchmarks! The `benchmark_v2` framework this builds on (`benchmark_framework.py` and the per-model `benches/*.py`) was replaced by the benchmark overhaul in #41408, so these files no longer fit. If more model coverage is still wanted, it can be added on top of the new setup in a PR against `main`. Closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40820"}, {"key": "issue#40798", "n": 40798, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "SigLIP2 loss is incomplete. Add the other loss.", "action": "closed", "comment": "Closing this one: no activity since 2025-12-21. If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/40798"}, {"key": "issue#40794", "n": 40794, "kind": "issue", "group": "Open issues", "category": "feature", "area": "generation", "type": "feature", "title": "Feature Request: Option to transfer logits to CPU during generation", "action": "commented", "comment": "Scores/logits are still accumulated on device in the decoding loops (`scores += (next_token_scores,)` in `src/transformers/generation/utils.py`). #40870 added an offload option but was closed because it targeted v4; a fresh PR on main would be welcome. Meanwhile, a custom `LogitsProcessor` that copies `scores.cpu()` into a list each step (and returns them unchanged), with `output_scores=False`, gives the same result without holding them on GPU.", "author": "YunruiZhang", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/40794"}, {"key": "pr#40790", "n": 40790, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Handle loading non-existent checkpoints or corrupted checkpoints.", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40790"}, {"key": "pr#40783", "n": 40783, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix None quantization_config equivalence with omitted param in AutoModel.from_pretrained", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40783"}, {"key": "issue#40776", "n": 40776, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Add function for reversing chat templates", "action": "closed", "comment": "This landed in #40894: `tokenizer.parse_response(...)` turns a model's generated output back into a structured assistant message (content, thinking, tool calls), see https://huggingface.co/docs/transformers/main/en/chat_response_parsing. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/40776"}, {"key": "issue#40767", "n": 40767, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "3D Object Detection Models", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: no official BEVFormer (or mmdet3d) checkpoints on the Hub; weights are only on the [original GitHub repo](https://github.com/fundamentalvision/BEVFormer)\n- Released: 2022-03 (paper) \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "SeucheAchat9115", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/40767"}, {"key": "pr#40759", "n": 40759, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "feat: add qwen3 pruning support", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40759"}, {"key": "pr#40756", "n": 40756, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "[WIP] Add Canary", "action": "closed", "comment": "Thanks @eustlb! Canary landed in #46825, which built on this PR, so closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40756"}, {"key": "pr#40755", "n": 40755, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "[TimesFM] Add support for forecasting with covariates", "action": "maintainer pinged", "comment": "@kashif this one has been waiting since 2026-03-06. Still planned, or should we close it?", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40755"}, {"key": "issue#40743", "n": 40743, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Support for 4D attention mask for T5", "action": "closed", "comment": "T5 now builds its masks with `create_causal_mask`/`create_bidirectional_mask` (#41804), which accept a pre-built 4D mask and use it as-is (see `src/transformers/masking_utils.py`). Passing 4D `attention_mask`/`decoder_attention_mask` should work on main; note the expected shape is `(batch, 1 or num_heads, q_len, kv_len)`. Please reopen with a repro if it still fails.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/40743"}, {"key": "pr#40740", "n": 40740, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Configure assistant model's generation_config with user parameters", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40740"}, {"key": "pr#40738", "n": 40738, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Docs: Clarify rjieba installation for RoFormerTokenizer", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40738"}, {"key": "pr#40736", "n": 40736, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `jan.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40736"}, {"key": "pr#40714", "n": 40714, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Remove TF and Flax from README", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40714"}, {"key": "issue#40674", "n": 40674, "kind": "issue", "group": "Open issues", "category": "feature", "area": "models", "type": "feature", "title": "Add object detection heads to DINOv3", "action": "commented", "comment": "Still not available: main has `DINOv3ViTBackbone` (`src/transformers/models/dinov3_vit/`) and an EoMT-DINOv3 segmentation model (`eomt_dinov3`), but no detection head. The cleanest path is a new modular model that reuses `DINOv3ViTBackbone` plus DETR components (see `detr`/`deformable_detr`, or how `eomt_dinov3` builds on DINOv3), following https://huggingface.co/docs/transformers/main/en/modular_transformers and https://huggingface.co/docs/transformers/main/en/add_new_model. Please confirm the DINOv3 detection weights' license allows redistribution first.", "author": "ryan-minato", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/40674"}, {"key": "pr#40670", "n": 40670, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add ability to run Gemma 2 models without post layer norm", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40670"}, {"key": "issue#40649", "n": 40649, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Support for HumanMLLM/HumanOmni", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [StarJiaxing/HumanOmni-7B](https://huggingface.co/StarJiaxing/HumanOmni-7B)\n- Downloads (last 30d): 41 \u00b7 Likes: 8 \u00b7 Params: ~9.6B \u00b7 License: n/a\n- Released: 2025-02 \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "Zha-Miku", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/40649"}, {"key": "pr#40648", "n": 40648, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Bump torch from 2.7.1 to 2.8.0 in /examples/flax/vision", "action": "closed", "comment": "Thanks! The Flax examples were removed along with Flax support in v5, so this bump is no longer needed. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40648"}, {"key": "pr#40640", "n": 40640, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Resume training by trained samples to avoid elastic job loss or over-reading of data.", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40640"}, {"key": "issue#40621", "n": 40621, "kind": "issue", "group": "Open issues", "category": "feature", "area": "gguf", "type": "feature", "title": "[Feature Proposal] Add GGUF support for Korean-specific LLM architectures", "action": "\u2014", "comment": "", "author": "Kim-Ju-won", "idle_days": 94, "models": "", "url": "https://github.com/huggingface/transformers/issues/40621"}, {"key": "pr#40587", "n": 40587, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "feat(utils): add vision utils for embedding images and getting the hidden size", "action": "closed", "comment": "Thanks for the PR and for opening #40457, @AmitMY! The review raised questions about scope (it mostly covered image-classification models) and asked for docs on how these helpers should be used, and the branch hasn't moved for about a year. If you still need this, please start a discussion on the issue about where it fits, then open a new PR against `main` that follows the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets). Closing this one for now, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40587"}, {"key": "pr#40524", "n": 40524, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Use begin_of_sequence token in all sliding windows for correct model behaviour", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40524"}, {"key": "pr#40520", "n": 40520, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[generate] add faster `stop_strings` stopping criteria", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40520"}, {"key": "pr#40515", "n": 40515, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Add Context-Aware Tokenizer Selection Utility Based on Corpus Analysis", "action": "closed", "comment": "Thanks for the work! As mentioned in review, this is outside what transformers aims to cover; a standalone package or a Hugging Face community post would be a great home. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40515"}, {"key": "issue#40512", "n": 40512, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Context-Aware Tokenizer Suggestions", "action": "closed", "comment": "Thanks for the proposal! As @Rocketknight1 said (https://github.com/huggingface/transformers/issues/40512#issuecomment-3233210823), we don't want a tokenizer-recommendation utility in Transformers right now; training a tokenizer for your corpus is supported via `tokenizer.train_new_from_iterator(...)`. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/40512"}, {"key": "pr#40505", "n": 40505, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Refactor Siglip-like models", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40505"}, {"key": "pr#40493", "n": 40493, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update dtypes to suit colab bf16 -> fp16 -> fp32.", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40493"}, {"key": "pr#40473", "n": 40473, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "[tests] Unskip DeBERTaV2 tokenizer parity tests; re-enable fast/slow checks", "action": "closed", "comment": "Thanks! In v5 DeBERTa-v2 only has the `tokenizers`-backed tokenizer, so slow/fast parity tests no longer apply. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40473"}, {"key": "pr#40471", "n": 40471, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "DOC: Standardize CodeGen model card (issue #36979)", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40471"}, {"key": "pr#40465", "n": 40465, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `tools.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40465"}, {"key": "pr#40464", "n": 40464, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `agents.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40464"}, {"key": "issue#40457", "n": 40457, "kind": "issue", "group": "Open issues", "category": "feature", "area": "processors (image/video/audio)", "type": "feature", "title": "Image Embedding Models (Feature extractors) should have a `.hidden_size`", "action": "commented", "comment": "#40587 was closed because it mostly covered image-classification models and lacked docs on intended usage. If you'd like to pursue it, please propose here a design that works across vision encoders (e.g. reading `config.get_text_config()`-style accessors for vision configs, and `projection_dim` for CLIP-like models) before opening a new PR against main.", "author": "AmitMY", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/40457"}, {"key": "pr#40448", "n": 40448, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[model] Support MiniCPM-V 4.5", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40448"}, {"key": "pr#40446", "n": 40446, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add convert_segmentation_map_to_binary_masks_sorted function for hand\u2026", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40446"}, {"key": "pr#40438", "n": 40438, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Resolve automatic label name detection when single label provided", "action": "closed", "comment": "Thanks @ndonthi! @SunMarc's request for a test and a reproducer (https://github.com/huggingface/transformers/pull/40438#pullrequestreview-3616208657) wasn't followed up (no commits since), so I'm closing this for now. If you can still reproduce the `label_names` issue on main, feel free to reopen with a test.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40438"}, {"key": "pr#40425", "n": 40425, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix check_quantized_param method when param_value is a safetensors slice", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40425"}, {"key": "pr#40404", "n": 40404, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "update model card for gpt-j", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40404"}, {"key": "pr#40403", "n": 40403, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Customizable Logit Warping Strategies for Generation #40010", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40403"}, {"key": "pr#40400", "n": 40400, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fixed redundant words in readme.md", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40400"}, {"key": "pr#40395", "n": 40395, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Updated `text_generation.md`", "action": "closed", "comment": "Thanks a lot for the Korean translation, @skwh54! This one never got picked up for review and the English `text_generation.md` has changed since, so I'm closing it for now. If you'd like to refresh it against the current English page, a new PR tagging the i18n-KO reviewers in #20179 would be very welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40395"}, {"key": "pr#40392", "n": 40392, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Remove debug print statement from ShieldGemma2 conversion script", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40392"}, {"key": "pr#40390", "n": 40390, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix typo: 'lenght' to 'length'", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40390"}, {"key": "pr#40385", "n": 40385, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix typo: 'seperate' -> 'separate' in mm_grounding_dino conversion sc\u2026", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40385"}, {"key": "pr#40358", "n": 40358, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix MXFP4 mlp_forward to handle 2D and 3D hidden_states shapes for multi-turn chat", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40358"}, {"key": "issue#40301", "n": 40301, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Label smoothing for multi-label classification", "action": "commented", "comment": "As noted above, the crash is fixed and label smoothing now safely no-ops for multi-label. The corrected multi-label smoothing would go in `LabelSmoother` in `src/transformers/trainer_pt_utils.py` (a branch for `problem_type == \"multi_label_classification\"` using BCE with smoothed targets), with a test in `tests/trainer/`; a draft PR is welcome.", "author": "andreabosisio", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/40301"}, {"key": "pr#40299", "n": 40299, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Remove deprecated max_size parameter from ConditionalDetr image processors", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #41284, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40299"}, {"key": "pr#40286", "n": 40286, "kind": "PR", "group": "Commented (now closed by others)", "category": "other", "area": "", "type": "", "title": "Add MOSS-TTSD with XY-Tokenizer", "action": "review requested: @ebezzam; maintainer pinged", "comment": "@ebezzam could you take a look when you get a chance? This is waiting on a maintainer: needs re-review of the Dec 2025 push, then a rebase on v5 main.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40286"}, {"key": "pr#40265", "n": 40265, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Enable PLW and PLE rules", "action": "closed", "comment": "Thanks a lot for this @cyyever! Main has since moved to enabling individual pylint rules one by one (e.g. `PLW1514` in `pyproject.toml`) rather than whole families, and @Cyrilvallez's question about the extra `noqa` noise was never resolved, so I'm closing this. A focused PR enabling a specific rule that catches real bugs would be very welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40265"}, {"key": "issue#40260", "n": 40260, "kind": "issue", "group": "Open issues", "category": "feature", "area": "exporters", "type": "feature", "title": "Add dynamic shapes to ExecuTorch static cache export tests", "action": "commented; comment corrected", "comment": "On `main`, export goes through `transformers.exporters` (https://huggingface.co/docs/transformers/main/en/exporters) (#41992), which supports dynamic shapes; extending it is documented at https://huggingface.co/docs/transformers/main/en/exporters_extend. The legacy helpers in `integrations/executorch.py` are not where new export work should go.", "author": "jackzhxng", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/40260"}, {"key": "pr#40244", "n": 40244, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "add-loftr-keypoints-to-map", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40244"}, {"key": "pr#40225", "n": 40225, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "docs: clarify decoder_input_ids vs decoder_inputs_embeds usage (#39542)", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40225"}, {"key": "pr#40209", "n": 40209, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add fast image processor for  ViViT ", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40209"}, {"key": "pr#40208", "n": 40208, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Save only model sharded sd", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40208"}, {"key": "pr#40180", "n": 40180, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Enable native mxfp4 training support for GPT-OSS models", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40180"}, {"key": "issue#40179", "n": 40179, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "[docs] Running jobs with SkyPilot", "action": "closed", "comment": "Thanks for the suggestion! We'd rather keep orchestrator-specific instructions in the orchestrators' own docs (SkyPilot already has Hugging Face training examples), so the docs PR #44309 was closed and we'll close this too.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/40179"}, {"key": "pr#40171", "n": 40171, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Rename to CamelCase", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40171"}, {"key": "issue#40170", "n": 40170, "kind": "issue", "group": "Open issues", "category": "feature", "area": "quantization", "type": "feature", "title": "Add MXFP4 MoE/attention backward kernels", "action": "labelled: Feature request", "comment": "", "author": "Rocketknight1", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/40170"}, {"key": "pr#40155", "n": 40155, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `t5.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40155"}, {"key": "pr#40149", "n": 40149, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Implemented fast image processor for VitPose", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40149"}, {"key": "pr#40131", "n": 40131, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "add missing Arabic translations", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40131"}, {"key": "pr#40123", "n": 40123, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Lazily import torchao Int4WeightOnlyConfig to avoid side effects", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40123"}, {"key": "pr#40115", "n": 40115, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[layer_types] update layer_types with conv", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40115"}, {"key": "pr#40114", "n": 40114, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix torch.export compatibility for Mixtral MoE models ", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40114"}, {"key": "pr#40102", "n": 40102, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `auto_docstring.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40102"}, {"key": "pr#40092", "n": 40092, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Optimize LlamaAttention by fusing QKV projections", "action": "closed", "comment": "Thanks for the benchmarks! As discussed, we keep the modeling code close to the original checkpoints and leave fusions to compile or kernels, so we're closing this; a cookbook recipe would be a nice place for it.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40092"}, {"key": "pr#40090", "n": 40090, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Fix RuntimeError when loading quantized models with int8 weights (#39366)", "action": "closed", "comment": "Thanks for working on this! The weight-loading and initialization path has since been rewritten for v5 (`_init_weights` now goes through `src/transformers/initialization.py`, which skips tensors already loaded from the checkpoint), and #39366 was closed as no longer relevant. If you still hit the int8 `normal_()` error with transformers v5, please reopen the issue with a reproducer. Closing this PR.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40090"}, {"key": "issue#40070", "n": 40070, "kind": "issue", "group": "Open issues", "category": "feature", "area": "gguf", "type": "feature", "title": "Transformer GGUF support philosophy / naive question", "action": "\u2014", "comment": "", "author": "luke14free", "idle_days": 41, "models": "", "url": "https://github.com/huggingface/transformers/issues/40070"}, {"key": "pr#40065", "n": 40065, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Delay float32 upcast in ForCausalLMLoss after filtering ignore_index", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40065"}, {"key": "pr#40064", "n": 40064, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `videomae.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40064"}, {"key": "pr#40061", "n": 40061, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `vitdet.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40061"}, {"key": "pr#40059", "n": 40059, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix Inefficient GELU implementation in GPT2", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40059"}, {"key": "pr#40058", "n": 40058, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "GGUF Qwen2VL", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40058"}, {"key": "pr#40055", "n": 40055, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Auto-log parallelism info to wandb.config using HF Accelerate", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40055"}, {"key": "issue#40049", "n": 40049, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Please support loading Qwen 2.5 VL from GGUF", "action": "commented; closed", "comment": "Closing as a duplicate of #35282: llama.cpp uses the same `qwen2vl` architecture for Qwen2-VL and Qwen2.5-VL, so the GGUF support (the `GGUF_CONFIG_MAPPING` entry plus the mmproj vision file) is the same piece of work. Please follow and contribute there. The pointers above apply as-is.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/40049"}, {"key": "pr#40047", "n": 40047, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update wavlm.md to match new model card template", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40047"}, {"key": "issue#40042", "n": 40042, "kind": "issue", "group": "Open issues", "category": "feature", "area": "gguf", "type": "feature", "title": "Support loading glm4moe GGUF", "action": "\u2014", "comment": "", "author": "adonishong", "idle_days": -1, "models": "glm4_moe", "url": "https://github.com/huggingface/transformers/issues/40042"}, {"key": "pr#40023", "n": 40023, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add support for SDPA for OWLViT and OWLv2", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40023"}, {"key": "pr#40022", "n": 40022, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix: resolve dropout type error in DogeDecoder", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/40022"}, {"key": "issue#40010", "n": 40010, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Customizable Logit Warping Strategies for Generation", "action": "closed", "comment": "Thanks for the PR (#40403), which was closed since it targeted v4. Custom warping already plugs in via `generate(..., logits_processor=LogitsProcessorList([...]))`, and reusable strategies can be shared as Hub `custom_generate` repos (https://huggingface.co/docs/transformers/main/en/generation_strategies#custom-generation-methods); built-in options such as `bad_words_ids`, `sequence_bias` and `repetition_penalty` cover the examples listed. We don't plan an extra declarative layer on top.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/40010"}, {"key": "pr#39999", "n": 39999, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "allow TP to work in ND-parallel with fsdp cpu ram efficient loading", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39999"}, {"key": "pr#39997", "n": 39997, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "make sure position_ids are passed in for causal mask creation for gpt-oss", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39997"}, {"key": "pr#39987", "n": 39987, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Add a VGGT(Visual Geometry Grounded Transformer) model compatible with huggingface transfromers", "action": "closed", "comment": "Thanks a lot @DDsacu for working on a VGGT port! The PR has been inactive for over a year and its layout (separate `heads/` and `track_modules/` packages) doesn't follow the single-file model structure we use, so it can't be merged as is. If you'd like to pick it back up, please open a fresh PR against `main` using a [modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers) file, following the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets) and checking it with [mlinter](https://huggingface.github.io/transformers-mlinter/). Closing this one for now.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39987"}, {"key": "pr#39962", "n": 39962, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Use torch._check instead of a test to make the model Gemma3 exportable", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39962"}, {"key": "issue#39946", "n": 39946, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Retaining computational graph after using AutoImageProcessor", "action": "closed", "comment": "This works on main: image processors now run on the torchvision/torch backend by default, so tensor inputs keep their autograd graph. E.g. with `google/vit-base-patch16-224`, `processor(x, do_rescale=False, return_tensors=\"pt\")[\"pixel_values\"]` has a `grad_fn` when `x.requires_grad=True`. Closing, but please reopen with the specific processor if you still see it detached.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/39946"}, {"key": "pr#39941", "n": 39941, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fixing image_utils.py todo", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39941"}, {"key": "pr#39931", "n": 39931, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Registers StaticCache serialization functions for torch.export.export", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39931"}, {"key": "pr#39922", "n": 39922, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `attention_interface.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39922"}, {"key": "pr#39920", "n": 39920, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Updated ko/perf_train_special.md", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39920"}, {"key": "pr#39917", "n": 39917, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Updated ko/perf_train_cpu.md", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39917"}, {"key": "pr#39901", "n": 39901, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `fp_quant` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39901"}, {"key": "pr#39886", "n": 39886, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `perf_train_gaudi.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39886"}, {"key": "issue#39882", "n": 39882, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "[Feature Request] Automatically log parallelism configuration from Accelerate to W&B", "action": "closed", "comment": "When the parallelism setup is passed through `TrainingArguments(parallelism_config=ParallelismConfig(...))` (#40205), it is serialized in `args.to_dict()`, which `WandbCallback.setup` already logs to the W&B run config, so TP/DP/CP sizes show up automatically. Closing; if you configure parallelism only through a raw `Accelerator`, please open a new issue with details.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/39882"}, {"key": "pr#39866", "n": 39866, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "make sure model.save_pretrained has the correct is_main_process", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39866"}, {"key": "pr#39859", "n": 39859, "kind": "PR", "group": "Open PRs", "category": "draft", "area": "quantization", "type": "feature / refactor", "title": "WIP: Initial support for bnb 4bit on any nn.Parameter", "action": "review requested: @SunMarc; maintainer pinged", "comment": "@SunMarc could you take a look when you get a chance? This is waiting on a maintainer: SunMarc's request (2025-08-07) for gpt-oss inference/saving tests.", "author": "matthewdouglas", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/pull/39859"}, {"key": "issue#39834", "n": 39834, "kind": "issue", "group": "Open issues", "category": "feature", "area": "generation", "type": "feature", "title": "Allow extra outputs from `GenerationMixin.generate`", "action": "\u2014", "comment": "", "author": "jood-canva", "idle_days": 396, "models": "", "url": "https://github.com/huggingface/transformers/issues/39834"}, {"key": "pr#39831", "n": 39831, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "refactor(modeling_llama): make RotaryEmbedding default path explicit", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39831"}, {"key": "pr#39807", "n": 39807, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `bamba.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39807"}, {"key": "pr#39796", "n": 39796, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[pipelines] text-to-audio pipeline standardization", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39796"}, {"key": "pr#39794", "n": 39794, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix ProphetNet forward to handle tuple encoder_outputs", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39794"}, {"key": "pr#39793", "n": 39793, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix DAC conversion script", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39793"}, {"key": "pr#39792", "n": 39792, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Served models handle with nested content", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #39829, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39792"}, {"key": "pr#39785", "n": 39785, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix mllama integration tests", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39785"}, {"key": "pr#39772", "n": 39772, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix missing initializations for models created in 2022", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39772"}, {"key": "pr#39760", "n": 39760, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[Draft] Add Llasa TTS family of models", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39760"}, {"key": "pr#39756", "n": 39756, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix rope_deltas corruption in Qwen2.5VL during CFG generation", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39756"}, {"key": "issue#39753", "n": 39753, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Inv frequency has not default, going against our philosophy", "action": "closed", "comment": "The default RoPE path is now explicit (`compute_default_rope_parameters` in each RotaryEmbedding) after the v5 RoPE refactor in #39847. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/39753"}, {"key": "pr#39751", "n": 39751, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `text-to-speech.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39751"}, {"key": "pr#39741", "n": 39741, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix HfArgumentParser to filter out dict types from Union", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39741"}, {"key": "pr#39735", "n": 39735, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "handle multimodal models with tp_plan on the text_config", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39735"}, {"key": "pr#39722", "n": 39722, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "new model", "title": "[Feat] Adding Intern-S1", "action": "\u2014", "comment": "", "author": "hhaAndroid", "idle_days": 251, "models": "intern_s1", "url": "https://github.com/huggingface/transformers/pull/39722"}, {"key": "pr#39718", "n": 39718, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix SigLIP2 documentation model/processor mismatch", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39718"}, {"key": "pr#39708", "n": 39708, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10[i18n-bn] Introduce Bengali version of Transformers documentation", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39708"}, {"key": "issue#39705", "n": 39705, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "[i18n-<bn>] Translating docs to <Bengali>", "action": "closed", "comment": "Bengali docs translation is now being tracked in #46222, with an open PR (#46236) adding `docs/source/bn`. Closing this one as a duplicate so the effort stays in one place; please coordinate there.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/39705"}, {"key": "pr#39698", "n": 39698, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix exaone4 layer_types ZeroDivision/TypeError when sliding_window_pattern is None/\"LLLG\"", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39698"}, {"key": "pr#39697", "n": 39697, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "use untyped storage for dtensors due to deprecation", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39697"}, {"key": "issue#39684", "n": 39684, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Add multi-candidate & tree search for assisted decoding (speculative decoding)", "action": "closed", "comment": "As @zucchini-nlp suggested (https://github.com/huggingface/transformers/issues/39684#issuecomment-3126495163), new decoding strategies like this can be shipped as custom generation methods on the Hub, loadable with `generate(..., custom_generate=\"<repo>\")` (https://huggingface.co/docs/transformers/main/en/generation_strategies#custom-generation-methods). Closing in favor of that route; feel free to share your repo here once it's up!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/39684"}, {"key": "pr#39683", "n": 39683, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "generation", "type": "fix", "title": "Fix issue #39191 respect accelerate config to disable torch.dynamo compilation", "action": "\u2014", "comment": "", "author": "bonpiedlaroute", "idle_days": 266, "models": "", "url": "https://github.com/huggingface/transformers/pull/39683"}, {"key": "issue#39682", "n": 39682, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Accelerate beam search decoding via tree attention", "action": "closed", "comment": "Thanks for BeamSD! As @gante suggested (https://github.com/huggingface/transformers/issues/39682#issuecomment-3151449133), this is a great fit for a Hub `custom_generate` repo loaded with `model.generate(custom_generate=\"<repo>\", trust_remote_code=True)`; custom masks/kernels can be wired via the attention interface (https://huggingface.co/docs/transformers/main/en/attention_interface). Closing in favor of that route.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/39682"}, {"key": "pr#39675", "n": 39675, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[BugFix]: Support dict and config file path for deepspeed", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39675"}, {"key": "pr#39674", "n": 39674, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix loss scaling and token aggregation to use only data parallel group", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39674"}, {"key": "pr#39632", "n": 39632, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix dead NVIDIA link ", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39632"}, {"key": "pr#39631", "n": 39631, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "[serve] Add speech-to-text", "action": "closed", "comment": "since then, `transformers serve` was migrated to Typer in #41487 (`src/transformers/cli/serve.py`) and split into modules under `src/transformers/cli/serving/` in #44796, and the serving docs moved to `docs/source/en/serve-cli/serving.md` in #42853. Transcriptions (`/v1/audio/transcriptions`) are served there already, but the `/v1/audio/speech` endpoint from this draft is not, tho it is coming. Closing this one, thanks!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39631"}, {"key": "pr#39625", "n": 39625, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix: allow Union[str, dict, None] fields like deepspeed to be passed via CLI", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39625"}, {"key": "issue#39618", "n": 39618, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "SageAttention for attention implementation?", "action": "closed", "comment": "As shown in https://github.com/huggingface/transformers/issues/39618#issuecomment-3116875850, no monkey-patching is needed: register a SageAttention wrapper once with `AttentionInterface.register(\"sage\", fn)` and load any model with `attn_implementation=\"sage\"` (https://huggingface.co/docs/transformers/main/en/attention_interface). Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/39618"}, {"key": "pr#39617", "n": 39617, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix FSDP v1 bug: trainer incorrectly uses an unwrapped model", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39617"}, {"key": "pr#39588", "n": 39588, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "WIP, reference modeling", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39588"}, {"key": "pr#39575", "n": 39575, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `vitpose.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39575"}, {"key": "pr#39563", "n": 39563, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `vision-encoder-decoder.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39563"}, {"key": "pr#39560", "n": 39560, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix load_model_end = true work when save_steps < eval_steps", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39560"}, {"key": "pr#39559", "n": 39559, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `main_classes/deepspeed.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39559"}, {"key": "pr#39555", "n": 39555, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[WIP]\u00a0try to relax the tie_weights method", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39555"}, {"key": "pr#39544", "n": 39544, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated feature_extractors.md to Korea", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39544"}, {"key": "pr#39534", "n": 39534, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add Beit3 model", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39534"}, {"key": "pr#39517", "n": 39517, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `compressed_tensor.md`  to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39517"}, {"key": "issue#39495", "n": 39495, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Add Muon Optimiser for 2x faster convergence", "action": "commented", "comment": "Not built in yet, but the plumbing is there: `src/transformers/trainer_optimizer.py` supports optimizer *factories* (see `is_optimizer_factory`) precisely for optimizers like Muon that split parameters between Muon (2D hidden weights) and AdamW (embeddings, norms, head). A PR adding a `muon` value to `OptimizerNames` in `training_args.py` with a handler building `torch.optim.Muon` (PyTorch >= 2.9) plus AdamW would be welcome. Meanwhile you can pass your own via `Trainer(optimizers=(...))`.", "author": "RonanKMcGovern", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/39495"}, {"key": "pr#39493", "n": 39493, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[Voxtral] nit + pin correct mistral common version", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39493"}, {"key": "pr#39480", "n": 39480, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add model arcinstitute state", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39480"}, {"key": "issue#39479", "n": 39479, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Adding Space-Time-MiniLM-v0", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [HaidarJomaa/Space-Time-MiniLM-v0](https://huggingface.co/HaidarJomaa/Space-Time-MiniLM-v0)\n- Downloads (last 30d): 114 \u00b7 Likes: 1 \u00b7 Params: ~36M \u00b7 License: apache-2.0\n- Released: 2025-07 \u00b7 Remote code on the Hub: no\n- Low adoption so far; sharing it with [custom code](https://huggingface.co/docs/transformers/main/en/custom_models) on the Hub is the quickest path until there's broader demand.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "haidarjomaa", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/39479"}, {"key": "issue#39476", "n": 39476, "kind": "issue", "group": "Open issues", "category": "feature", "area": "core / loading", "type": "feature", "title": "Allow `load_best_model_at_end=True` to work when `save_steps < eval_steps` and best model is saved", "action": "commented", "comment": "Still open for a PR (@SunMarc agreed). The restriction is in `TrainingArguments.__post_init__` (`src/transformers/training_args.py`, the \"--load_best_model_at_end requires the saving steps to be a multiple...\" check); the best-checkpoint tracking and rotation logic is in `Trainer._save_checkpoint` / `_rotate_checkpoints` in `src/transformers/trainer.py`. Note `save_strategy=\"best\"` already saves only when the metric improves, which may cover your disk-space use case.", "author": "KrisTHL181", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/39476"}, {"key": "pr#39468", "n": 39468, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix quantized model dispatch with device_map='auto'", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39468"}, {"key": "pr#39466", "n": 39466, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "README: Update Bert Japanese model card", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39466"}, {"key": "pr#39464", "n": 39464, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Skipping `initialize_weights` when model is quantized", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39464"}, {"key": "pr#39456", "n": 39456, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix quantized model initialization for int8 dtypes", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39456"}, {"key": "pr#39449", "n": 39449, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Fix logger warnings in Gemma model test files", "action": "closed", "comment": "Thanks for the PR! TensorFlow and Flax support was removed in v5, so this can't land anymore. Closing, thanks again for the contribution!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39449"}, {"key": "issue#39437", "n": 39437, "kind": "issue", "group": "Open issues", "category": "feature", "area": "generation", "type": "feature", "title": "Support for per-token latency tracking in `generate()` (suggested options: using callback, profiler class, or using a config flag)", "action": "commented", "comment": "There's no built-in per-token latency option yet; as gante said, it was waiting on the generate refactor, part of which has landed (`GenerationMixin._prefill` in `src/transformers/generation/utils.py`). Today you can get per-token timings without library changes by passing a custom `BaseStreamer` whose `put()` records `time.perf_counter()` (it's called once per decoding step, the first call being the prompt). If you'd like to upstream it, a small timing streamer in `src/transformers/generation/streamers.py` would be the least intrusive place.", "author": "spsagar13", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/39437"}, {"key": "issue#39426", "n": 39426, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "object detection : matchin outputs.last_hidden_state with results", "action": "closed", "comment": "As @qubvel suggested (https://github.com/huggingface/transformers/issues/39426#issuecomment-3102259277), a custom post-processing step based on `post_process_object_detection` that keeps the selected query indices is the way to match detections with `outputs.last_hidden_state`. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/39426"}, {"key": "issue#39410", "n": 39410, "kind": "issue", "group": "Open issues", "category": "feature", "area": "quantization", "type": "feature", "title": "FP8 training support for Model Parallel / Tensor Parallel (MP/TP)", "action": "commented", "comment": "Still not supported on main: the FP8 quantizer (`src/transformers/quantizers/quantizer_finegrained_fp8.py`) reports `is_trainable = False`, which triggers this error. A workaround for fine-tuning is loading with `FineGrainedFP8Config(dequantize=True)` to get bf16 weights and then training with TP (https://huggingface.co/docs/transformers/main/en/tensor_parallelism), at the cost of memory. Native FP8 training would need a backward for the FP8 linear (`integrations/finegrained_fp8.py`) compatible with the TP plans; contributions are welcome.", "author": "edgeinfinity1", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/39410"}, {"key": "pr#39403", "n": 39403, "kind": "PR", "group": "Open PRs", "category": "changes requested", "area": "models", "type": "new model", "title": "Add Vocos model", "action": "\u2014", "comment": "", "author": "Manalelaidouni", "idle_days": 267, "models": "vocos, vocos_encodec", "url": "https://github.com/huggingface/transformers/pull/39403"}, {"key": "issue#39367", "n": 39367, "kind": "issue", "group": "Open issues", "category": "feature", "area": "serving / continuous batching", "type": "feature", "title": "Adding api key to `transformers serve`", "action": "\u2014", "comment": "", "author": "xhluca", "idle_days": 53, "models": "", "url": "https://github.com/huggingface/transformers/issues/39367"}, {"key": "pr#39357", "n": 39357, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update docstring for glm4v", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39357"}, {"key": "issue#39350", "n": 39350, "kind": "issue", "group": "Commented (now closed by others)", "category": "new-model details", "area": "", "type": "", "title": "surpport for google/medgemma-27b-it", "action": "commented", "comment": "Good news: this is now supported in transformers as `gemma3` (added in #36658). [google/medgemma-27b-it](https://huggingface.co/google/medgemma-27b-it) loads natively with `AutoModelForImageTextToText` / the `image-text-to-text` pipeline, no remote code needed.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/39350"}, {"key": "issue#39329", "n": 39329, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "Adding support for Gemma 3n GGUFs", "action": "commented", "comment": "Gemma 3n GGUF is still unsupported, but Gemma 2/3/4 are, so there's a template: add a `gemma3n` entry to `GGUF_CONFIG_MAPPING` in `src/transformers/integrations/ggml.py`, map it to `gemma3n_text` in `src/transformers/modeling_gguf_pytorch_utils.py` (like `gemma3` \u2192 `gemma3_text`) and reuse/extend `Gemma2TensorProcessor` for its extra tensors (per-layer embeddings, AltUp/Laurel). Docs: https://huggingface.co/docs/transformers/main/en/quantization/gguf; a PR would be welcome.", "author": "minhnguyenhoang", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/39329"}, {"key": "pr#39309", "n": 39309, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix audio pipeline with torchcodec input", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39309"}, {"key": "pr#39303", "n": 39303, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix critical typos in code example", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39303"}, {"key": "pr#39297", "n": 39297, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix bug with deepspeed  and accelerator args in training_args.py", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39297"}, {"key": "pr#39293", "n": 39293, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add T5LA models", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39293"}, {"key": "pr#39264", "n": 39264, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix: Add version check for timm to support mobilenetv5 models (fixes #39208)", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #40951, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39264"}, {"key": "pr#39257", "n": 39257, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix to tuple conversion with config", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39257"}, {"key": "pr#39251", "n": 39251, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix slow test_moshika_greedy_unconditional_fp16", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39251"}, {"key": "pr#39236", "n": 39236, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "added moment_p sampling", "action": "closed", "comment": "Thanks for the contribution! New decoding/sampling techniques now live on the Hub as custom generation methods rather than in core `generate` (like `dola` and contrastive search). We'd love to see Moment-P shared that way.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39236"}, {"key": "pr#39222", "n": 39222, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Enable granite 4 hybrid integration tests", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39222"}, {"key": "issue#39213", "n": 39213, "kind": "issue", "group": "Open issues", "category": "feature", "area": "attention / kernels", "type": "feature", "title": "Remove device to host sync triggered in _flash_attention_forward", "action": "\u2014", "comment": "", "author": "piyifan123", "idle_days": 83, "models": "", "url": "https://github.com/huggingface/transformers/issues/39213"}, {"key": "pr#39212", "n": 39212, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add Ukrainian translation of README.md", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39212"}, {"key": "pr#39211", "n": 39211, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Add mobilenet_v5 stub implementation to fix \"Unknown Model\" error", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #40951, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39211"}, {"key": "pr#39209", "n": 39209, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Standardize FSMT class naming: PretrainedFSMTModel \u2192 PreTrainedFSMTModel", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39209"}, {"key": "pr#39206", "n": 39206, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix: filter None router logits in Qwen3 MoE and handle empty router logits (#39203)", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39206"}, {"key": "issue#39202", "n": 39202, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Naming incosistencies of `PreTrained*` classes.", "action": "closed", "comment": "Done in #41300: the class is now `PreTrainedConfig` (`src/transformers/configuration_utils.py`), with `PretrainedConfig` kept as a backward-compatible alias. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/39202"}, {"key": "pr#39183", "n": 39183, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add a 'chat' extra", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39183"}, {"key": "pr#39150", "n": 39150, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Efficient Expert Weight Fusion for Moe deepseek v3", "action": "closed", "comment": "Thanks for the contribution! Expert fusion is now handled generically by the experts implementation on main, so we're closing this as adressed and done!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39150"}, {"key": "pr#39140", "n": 39140, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "feat(trainer): emergency checkpointing on crashes & SIGTERM/SIGINT", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39140"}, {"key": "pr#39109", "n": 39109, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix: rename 'eval_strategy' to 'evaluation_strategy' in TrainingArgum\u2026", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39109"}, {"key": "pr#39108", "n": 39108, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Disable static cache on certain MoE models", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39108"}, {"key": "pr#39084", "n": 39084, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Refactor gemma3n", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39084"}, {"key": "pr#39047", "n": 39047, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "RFC: refactor causal lm loss to handle lm_head in loss function", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39047"}, {"key": "pr#39046", "n": 39046, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix deprecated max_size parameter handling in DETR image processors", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #41284, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39046"}, {"key": "pr#39037", "n": 39037, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix `kosmos2` tests", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39037"}, {"key": "pr#39012", "n": 39012, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[WIP] Fix DeepseekV3ModelTest::test_torch_compile_for_training", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39012"}, {"key": "pr#39009", "n": 39009, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add submodels support check function", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/39009"}, {"key": "pr#38999", "n": 38999, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Use deep copies instead of shallow copies for bbox_embed in GroundingDINO decoder (#37333).", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #38238, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38999"}, {"key": "pr#38991", "n": 38991, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Remove `return_dict` kwarg from all the models", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38991"}, {"key": "pr#38988", "n": 38988, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Check docstring inside modular files as well", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38988"}, {"key": "pr#38962", "n": 38962, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update test_candidate_generator.py", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38962"}, {"key": "pr#38959", "n": 38959, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Updated the model card for wav2vec2-phoneme", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38959"}, {"key": "pr#38958", "n": 38958, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Updated model card for wav2vec2-conformer", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38958"}, {"key": "pr#38957", "n": 38957, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update wav2vec2-bert model card", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38957"}, {"key": "pr#38956", "n": 38956, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Updating model card for wav2vec2", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38956"}, {"key": "pr#38955", "n": 38955, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "docs: Musicgen melody model card", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38955"}, {"key": "issue#38950", "n": 38950, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Number Token Loss", "action": "closed", "comment": "As mentioned, we only add losses to `loss_utils.py` when released checkpoints depend on them. NTL can be used today without library changes by passing it as `Trainer(compute_loss_func=...)` (#34198), e.g. CE plus your NTL term computed from `outputs.logits`; packaging it as a small standalone library is the best way to share it.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/38950"}, {"key": "pr#38926", "n": 38926, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Clarify Python and framework version support in installation.md", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38926"}, {"key": "pr#38923", "n": 38923, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Remove deprecated max_size support from YOLOS image processor", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38923"}, {"key": "issue#38917", "n": 38917, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Add past_key_values as inputs to TorchExportableModuleWithStaticCache forward", "action": "commented; closed; comment corrected", "comment": "On `main`, the new `transformers.exporters` (https://huggingface.co/docs/transformers/main/en/exporters) API (#41992) is the recommended way to do this: its exporters take forward kwargs including `past_key_values` (a `Cache`) as example inputs, and there is an ONNX backend in `src/transformers/exporters/exporter_onnx.py`. If something is still missing for your use case, a PR extending the exporters would be very welcome (see https://huggingface.co/docs/transformers/main/en/exporters_extend).", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/38917"}, {"key": "pr#38908", "n": 38908, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add support to use config dtype in HybridChunkedCache", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38908"}, {"key": "pr#38893", "n": 38893, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix/deprecate max size conditional detr", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #41284, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38893"}, {"key": "pr#38888", "n": 38888, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "continue to fix distributed_type from TPU to XLA in LM examples (#38652)", "action": "closed", "comment": "We removed the legacy `examples/` scripts from the repository (#49357) since they were no longer maintained, so this change no longer has a home. Thanks for the fix!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38888"}, {"key": "pr#38886", "n": 38886, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Allow compile with bnb", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38886"}, {"key": "pr#38884", "n": 38884, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Llama 4 conversion fix for moe models", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38884"}, {"key": "pr#38877", "n": 38877, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "DOC: Clarify attention_mask usage in BertModel forward method", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38877"}, {"key": "pr#38861", "n": 38861, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add SamImageProcessorFast with 4x performance improvement", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38861"}, {"key": "pr#38859", "n": 38859, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add MobileViT fast image processor", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38859"}, {"key": "pr#38839", "n": 38839, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[DO NOT MERGE] Testing saftensors 0.6.0", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38839"}, {"key": "pr#38810", "n": 38810, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add kwargs support in WhisperForConditionalGeneration", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38810"}, {"key": "pr#38805", "n": 38805, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add Dust3R", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38805"}, {"key": "pr#38793", "n": 38793, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix Typos in Comments and Improve Clarity", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38793"}, {"key": "pr#38786", "n": 38786, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Provide clearer instructions on how to specify target language.", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38786"}, {"key": "pr#38764", "n": 38764, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Fixes issue #38709 for CLIP and XCLIP", "action": "closed", "comment": "Thanks for the PR! TensorFlow and Flax support was removed in v5, so this can't land anymore. Closing, thanks again for the contribution!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38764"}, {"key": "pr#38759", "n": 38759, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add chatterbox", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38759"}, {"key": "pr#38757", "n": 38757, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add S3 tokenizer", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38757"}, {"key": "pr#38756", "n": 38756, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add HiFTNet", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38756"}, {"key": "pr#38752", "n": 38752, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Fix/gpu memory leak autoprocessor", "action": "closed", "comment": "Thanks for opening this! We asked for a reproducible snippet back in June 2025 and couldn't confirm the leak, and the branch is now far behind `main`. If you still see GPU memory growth with `AutoProcessor` on the latest release, please open an issue with a minimal reproducer and we can go from there. Closing for now.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38752"}, {"key": "issue#38708", "n": 38708, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Bert2D: A 2D-Word Embedding Model for Morphologically Rich Languages", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [yigitbekir/Bert2D-cased-Turkish-128K-WWM-NSW2](https://huggingface.co/yigitbekir/Bert2D-cased-Turkish-128K-WWM-NSW2)\n- Downloads (last 30d): 81 \u00b7 Likes: 6 \u00b7 Params: ~185M \u00b7 License: apache-2.0\n- Released: 2025-05 \u00b7 Remote code on the Hub: yes\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "yigit353", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/38708"}, {"key": "pr#38684", "n": 38684, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix AssertionError when saving CodeT5+ 2B checkpoints (Resolves #38602)", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38684"}, {"key": "issue#38680", "n": 38680, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Add support for Orthogonal Residual Updates", "action": "closed", "comment": "As @Rocketknight1 said in https://github.com/huggingface/transformers/issues/38680#issuecomment-2955949169, we'd need to see wider use of this technique before adding it; in the meantime it can be shared as custom code on the Hub (https://huggingface.co/docs/transformers/main/en/custom_models). Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/38680"}, {"key": "pr#38666", "n": 38666, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "docs: clarify compute_loss parameter num_items_in_batch", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38666"}, {"key": "issue#38657", "n": 38657, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "please develop transformers java/scala  sdk,eagerly to use!", "action": "closed", "comment": "Closing this one: no activity since 2025-06-10, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/38657"}, {"key": "pr#38655", "n": 38655, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "minor docstring fixups", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38655"}, {"key": "pr#38647", "n": 38647, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Reparent all the remaining Causal LM tests", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38647"}, {"key": "pr#38638", "n": 38638, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "5GB to 50GB as a default", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38638"}, {"key": "pr#38626", "n": 38626, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix: Added code to match interpolation of Google's ViT implementation\u2026", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38626"}, {"key": "pr#38604", "n": 38604, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update Mamba configuration to enable bias by default in the mixer blo\u2026", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38604"}, {"key": "pr#38598", "n": 38598, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Deprecated sigopt from hyperparameter search", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38598"}, {"key": "pr#38587", "n": 38587, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Refactor Bamba tests to inherit from CausalLMModelTester base classes", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38587"}, {"key": "pr#38569", "n": 38569, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add Bagel", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38569"}, {"key": "pr#38568", "n": 38568, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update Wav2Vec2 documentation to create model cards", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38568"}, {"key": "pr#38564", "n": 38564, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix ModernBERT tokenizer issue with is_split_into_words flag", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38564"}, {"key": "pr#38550", "n": 38550, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Improve GPTNeoX model card following standardization guidelines", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38550"}, {"key": "pr#38547", "n": 38547, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Fix CTRL model DataParallel compatibility", "action": "closed", "comment": "Thanks for the fix for CTRL with `DataParallel`! This PR targets `amd-hf-ci-branch` rather than `main`, and on `main` CTRL no longer returns stacked 5D `past_key_values` tensors: it was migrated to the `Cache` API (`DynamicCache`) in #38635, which removes the shape mismatch described here. Closing for that reason; if you still hit a `DataParallel` issue on `main`, please open an issue with a reproducer.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38547"}, {"key": "pr#38534", "n": 38534, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fixed a bug which was causing only partial files to be imported", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38534"}, {"key": "issue#38509", "n": 38509, "kind": "issue", "group": "Open issues", "category": "feature", "area": "processors (image/video/audio)", "type": "feature", "title": "SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference", "action": "\u2014", "comment": "", "author": "aryanchauhan31", "idle_days": 504, "models": "", "url": "https://github.com/huggingface/transformers/issues/38509"}, {"key": "pr#38478", "n": 38478, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix meta tensor copy error", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38478"}, {"key": "pr#38473", "n": 38473, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "docs: Add Turkish translation for README", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38473"}, {"key": "pr#38470", "n": 38470, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Add detailed ConvBERT model card with usage, architecture, and refere\u2026", "action": "closed", "comment": "Thanks for your first contribution, @Aesha19, and sorry for the long wait! Model docs are rendered from `docs/source/en/model_doc/convbert.md`, not from a `modelcard.md` next to the code, and that page has been updated several times since (e.g. #45802 and #47748). If you'd still like to improve the ConvBERT docs, please open a new PR against `main` editing `docs/source/en/model_doc/convbert.md`. Closing this one, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38470"}, {"key": "issue#38452", "n": 38452, "kind": "issue", "group": "Open issues", "category": "feature", "area": "generation", "type": "feature", "title": "Memory saving by upcasting logits for only non-ignored positions", "action": "commented", "comment": "Still open: `ForCausalLMLoss` in `src/transformers/loss/loss_utils.py` calls `logits.float()` before any masking. The previous attempt #40065 was closed because it targeted v4; a new PR against main should show memory/speed numbers, since maintainers raised that boolean mask-select adds a device sync (see the discussion there). A version that avoids data-dependent shapes would be the easiest to accept.", "author": "harshit2997", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/38452"}, {"key": "issue#38450", "n": 38450, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "\ud83d\udca1 Proposal: Add temporal-grounding pipeline for video-language tasks", "action": "closed", "comment": "Thanks for the proposal! As @zucchini-nlp (https://github.com/huggingface/transformers/issues/38450#issuecomment-2919852095) and @Rocketknight1 (https://github.com/huggingface/transformers/issues/38450#issuecomment-2919264079) noted, a pipeline first needs pre-trained models in the library that do this task, and none do yet. Closing for now; a custom pipeline on the Hub (https://huggingface.co/docs/transformers/main/en/add_new_pipeline) is a good start.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/38450"}, {"key": "pr#38444", "n": 38444, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add configurable normalization schemes to SigLIP image processors", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38444"}, {"key": "pr#38437", "n": 38437, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Continuous batchin: offer only the next token", "action": "closed", "comment": "Thanks @LysandreJik! `src/transformers/generation/continuous_batching.py` was turned into the `generation/continuous_batching/` package in #40426 and has been reworked a lot since (e.g. #43960), so this diff no longer applies. If a `next_token` field on streamed `GenerationOutput` is still wanted, it would need a new PR against `main` on top of `continuous_batching/requests.py`. Closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38437"}, {"key": "pr#38432", "n": 38432, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "More coverage for LossKwargs + cleaning", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38432"}, {"key": "pr#38421", "n": 38421, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "[Qwen2.5-VL] Fix empty string input crash in processor", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #36555, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38421"}, {"key": "issue#38416", "n": 38416, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Model implmenetation using Liger Kernel layers", "action": "closed", "comment": "This is now covered by the Kernels integration: layers decorated with `use_kernel_forward_from_hub` (RMSNorm, SwiGLU/GeGLU MLPs, activations, ...) are mapped to `kernels-community/liger-kernels` in `src/transformers/integrations/hub_kernels.py` and enabled with `from_pretrained(..., use_kernels=True)`, so no monkey-patching is needed (https://huggingface.co/docs/transformers/main/en/kernels). Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/38416"}, {"key": "pr#38400", "n": 38400, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "[WIP] Tokenizer Refactor", "action": "closed", "comment": "Thanks for the PR! This model is deprecated and no longer maintained in the library, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38400"}, {"key": "pr#38388", "n": 38388, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix Whisper inference regression with backward-compatible logprob calculation", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38388"}, {"key": "pr#38387", "n": 38387, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Expanded `UserWarning` message on parameter edge case", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38387"}, {"key": "issue#38377", "n": 38377, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Why are the model classes in unit tests imported directly from the transformer package instead of directly importing the model classes in the file? Is there any special consideration?", "action": "closed", "comment": "Closing this one: no activity since 2025-05-26, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/38377"}, {"key": "pr#38339", "n": 38339, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Bump minimum version of accelerate", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38339"}, {"key": "pr#38338", "n": 38338, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Support CLIP w/ Registers & Fusion MLP Gate", "action": "closed", "comment": "Thanks for sharing this! We keep CLIP's modeling to the released architecture, so variants like this are best shared as custom code on the Hub (`trust_remote_code`). Leveraging the fusion mapping API is also a great way to get this, conversion mapping as well! (custom register)", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38338"}, {"key": "pr#38272", "n": 38272, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix the shape of ModernBertForMaskedLM's output hidden_states", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38272"}, {"key": "issue#38267", "n": 38267, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add support for BAGEL from ByteDance", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [ByteDance-Seed/BAGEL-7B-MoT](https://huggingface.co/ByteDance-Seed/BAGEL-7B-MoT)\n- Downloads (last 30d): 3,433 \u00b7 Likes: 1,234 \u00b7 Params: ~14.7B total (7B active) \u00b7 License: apache-2.0\n- Released: 2025-05 \u00b7 Remote code on the Hub: no (original-repo code)\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "Shakib-IO", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/38267"}, {"key": "issue#38239", "n": 38239, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "QuantizationMethod.BITNET Support", "action": "commented", "comment": "BitNet inference is supported in transformers as `bitnet` (added in #37742), e.g. [microsoft/bitnet-b1.58-2B-4T](https://huggingface.co/microsoft/bitnet-b1.58-2B-4T) (bf16 master weights: [microsoft/bitnet-b1.58-2B-4T-bf16](https://huggingface.co/microsoft/bitnet-b1.58-2B-4T-bf16)). Training/fine-tuning through `QuantizationMethod.BITNET` is a separate feature; the bf16 checkpoint is the right starting point for anyone who wants to work on it.", "author": "mrjha4050", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/38239"}, {"key": "pr#38237", "n": 38237, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[gptq] Fixing gptq tests after new release of gptqmodel", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38237"}, {"key": "pr#38209", "n": 38209, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[Meta MTIA] Support hoisted rotary embeddings.", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38209"}, {"key": "pr#38196", "n": 38196, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update kwargs for DDP", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38196"}, {"key": "pr#38185", "n": 38185, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Add SVE implementation for Mamba Sequential Scan Algorithm", "action": "closed", "comment": "Thanks for the kernel! We're no longer maintaining custom kernels inside transformers; the way forward is publishing it via the kernels hub (https://github.com/huggingface/kernels) or as an installable module. Closing, but we'd be glad to help integrate it from there.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38185"}, {"key": "pr#38173", "n": 38173, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "[WIP] new BLT", "action": "closed", "comment": "Thanks for the PR! TensorFlow and Flax support was removed in v5, so this can't land anymore. Closing, thanks again for the contribution!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38173"}, {"key": "pr#38141", "n": 38141, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix FSDP + llava-next/llava-onevision", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38141"}, {"key": "pr#38114", "n": 38114, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Force real tensors and clone state_dict in src/transformers/modeling_utils.py", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38114"}, {"key": "pr#38111", "n": 38111, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Support TP for save_pretrained()", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38111"}, {"key": "pr#38064", "n": 38064, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Added scores in the streamer classes based on generation flag", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38064"}, {"key": "pr#38055", "n": 38055, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "SQuat cache implementation", "action": "closed", "comment": "Thanks for SQuat, and for porting it to a `custom_generate` repo! That's exactly the path we recommend for new cache methods, so we'll close this in-library PR.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38055"}, {"key": "pr#38054", "n": 38054, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix-mask2former-overlapping-annotations", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38054"}, {"key": "pr#38030", "n": 38030, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Add `TemplateConstraint` and `OrdredConstraint` features (#27706)", "action": "closed", "comment": "Thanks for the PR! Constrained beam search has been removed from the core library and new decoding strategies are now shared as custom `generate` methods on the Hub, so please consider publishing it there.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/38030"}, {"key": "pr#37998", "n": 37998, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update AMD docker with needed packages", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37998"}, {"key": "pr#37994", "n": 37994, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "proof of concept for using dataset of test cases for tokenizer tests", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37994"}, {"key": "pr#37975", "n": 37975, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[vLLM] GPT2 with transformers backend", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37975"}, {"key": "issue#37971", "n": 37971, "kind": "issue", "group": "Open issues", "category": "feature", "area": "pipelines", "type": "feature", "title": "Add `pruna` integration for loading model through `transformers.from_pretrained` / `pipeline`.", "action": "commented", "comment": "Not on main, and @SunMarc suggested starting with diffusers. If you want `from_pretrained` support here, the route is an `HfQuantizer` + config registered in `src/transformers/quantizers/auto.py`, following https://huggingface.co/docs/transformers/main/en/quantization/contribute; please align on scope with the maintainers before a PR.", "author": "davidberenstein1957", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/37971"}, {"key": "pr#37967", "n": 37967, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "feat: Add encode, patch_token to RagTokenizer (fixes #35532)", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37967"}, {"key": "pr#37963", "n": 37963, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Object detection experiment", "action": "closed", "comment": "We removed the legacy `examples/` scripts from the repository (#49357) since they were no longer maintained, so this change no longer has a home. Thanks for exploring this!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37963"}, {"key": "pr#37944", "n": 37944, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix graph break for torch.compile in attention mask check", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37944"}, {"key": "pr#37933", "n": 37933, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Feat: Add ddp_param_to_hook_all_reduce option", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37933"}, {"key": "pr#37921", "n": 37921, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix wrong example in grounding dino", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37921"}, {"key": "pr#37917", "n": 37917, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "support MiniCPM-o2.6", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37917"}, {"key": "issue#37914", "n": 37914, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Training Qwen2.5 VL with dynamic image size using more balanced Sampler for each GPU mem usage", "action": "closed", "comment": "This should help now: #47340 added `TrainingArguments(train_sampling_strategy=\"batch_rebalance\")`, which uses `BatchRebalanceSampler` to balance token cost across devices and gradient-accumulation steps (data-parallel only for now); `\"group_by_length\"` is the simpler alternative. Closing, please reopen if memory is still unbalanced with it.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/37914"}, {"key": "pr#37904", "n": 37904, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Feat: Add class_proba option to semantic segmentation post-processing", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37904"}, {"key": "pr#37884", "n": 37884, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Get our efficiency back", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37884"}, {"key": "pr#37875", "n": 37875, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add DEIM object detection model", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37875"}, {"key": "pr#37870", "n": 37870, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Bump transformers from 4.48.0 to 4.50.0 in /examples/tensorflow/language-modeling-tpu", "action": "closed", "comment": "TensorFlow support and the TF examples were removed in v5, so this file no longer exists. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37870"}, {"key": "pr#37869", "n": 37869, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Hybrid cache v2", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37869"}, {"key": "issue#37867", "n": 37867, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Option for save_pretrained() to Export Model Source Code Files", "action": "closed", "comment": "As explained above, for an editable snapshot copy the model's files from `src/transformers/models/<name>/` into your repo and register them as custom code (`MyConfig.register_for_auto_class()`, `MyModel.register_for_auto_class(\"AutoModel\")`), after which `save_pretrained` saves the source alongside the weights (https://huggingface.co/docs/transformers/main/en/custom_models). We don't plan to add a flag for library models, so closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/37867"}, {"key": "pr#37846", "n": 37846, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Integrating Kimi-Audio", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37846"}, {"key": "pr#37845", "n": 37845, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Remove redundancies for Qwen3MoeDecoderLayer ", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37845"}, {"key": "pr#37837", "n": 37837, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "feat: Add ConvaiCausalLM model for Hindi Causal Language Modeling", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37837"}, {"key": "issue#37771", "n": 37771, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Force the model to write some tokens mid-generation?", "action": "closed", "comment": "This is doable with existing APIs: stop on `</codetests>` with `stop_strings` (see src/transformers/generation/configuration_utils.py), run your tests, append the forced text to the sequence, and call `generate` again reusing `past_key_values` so only the new tokens are processed (https://huggingface.co/docs/transformers/main/en/kv_cache#iterative-generation). Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/37771"}, {"key": "pr#37769", "n": 37769, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Vectorize deepseek moe", "action": "closed", "comment": "Thanks for this! MoE layers now go through the shared experts implementation with fused weights on main, so this approach is superseded and we're closing it.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37769"}, {"key": "pr#37763", "n": 37763, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "general spm converter", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37763"}, {"key": "pr#37762", "n": 37762, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add update_tp_plan() for Quark", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37762"}, {"key": "pr#37687", "n": 37687, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Handle audio/ video default arguments in processor's apply_chat_template", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37687"}, {"key": "pr#37653", "n": 37653, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Non model inits", "action": "closed", "comment": "Thanks for the PR! TensorFlow and Flax support was removed in v5, so this can't land anymore. Closing, thanks again for the contribution!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37653"}, {"key": "issue#37648", "n": 37648, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Adding Paged Attention to Qwen1.5-MoE-A2.7B-Chat models using PyTorch XLA and Pallas", "action": "closed", "comment": "Custom attention kernels, including a Pallas/XLA paged one, can now be registered with `AttentionInterface.register(...)` and selected via `attn_implementation` without modifying the modules (https://huggingface.co/docs/transformers/main/en/attention_interface); on GPU, paged attention ships with continuous batching (https://huggingface.co/docs/transformers/main/en/paged_attention). Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/37648"}, {"key": "pr#37643", "n": 37643, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add support for manually setting `head_dim` in Qwen2 MoE", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37643"}, {"key": "pr#37642", "n": 37642, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add time-based evaluation strategy to Trainer", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37642"}, {"key": "pr#37635", "n": 37635, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Add resume checkpoint support to ClearML callback", "action": "closed", "comment": "Thanks for the contribution! `ClearMLCallback` in `src/transformers/integrations/integration_utils.py` has changed a lot since this branch was opened ~18 months ago, and resume support still isn't there on `main`. If this is still useful for you, please open a fresh PR against `main` with the test placed under `tests/trainer/`, and we'll gladly take a look. Closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37635"}, {"key": "pr#37634", "n": 37634, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add PLM Model", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37634"}, {"key": "issue#37627", "n": 37627, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "if I want to use my image-text data to finetune the SigLIP2\uff0c where I can get the train code?", "action": "closed", "comment": "Glad the snippet above helped! For fine-tuning, the model's `return_loss=True` argument (see `src/transformers/models/siglip2/modeling_siglip2.py`) returns the contrastive loss directly; for further questions https://discuss.huggingface.co is the best place. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/37627"}, {"key": "pr#37623", "n": 37623, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Make hybrid cache exportable", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37623"}, {"key": "pr#37611", "n": 37611, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add FastImageProcessor for InstructBLIPVideo", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37611"}, {"key": "pr#37596", "n": 37596, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Tests for the new Tensor Parallel integration", "action": "closed", "comment": "Thanks @ailunc for writing tests for the TP integration! Tensor-parallel coverage now lives in `tests/tensor_parallel/test_tensor_parallel.py` (introduced in #36539) and the shared `tests/test_tensor_parallel_mixin.py`, so the top-level `tests/test_tensor_parallel*.py` files here would duplicate it. If you see a gap in the current tests, a PR against `main` extending those files is welcome. Closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37596"}, {"key": "pr#37582", "n": 37582, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add code examples for creating & fine\u2011tuning EncoderDecoderModel (fixes #16135)", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37582"}, {"key": "pr#37539", "n": 37539, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Mllama fast image processor", "action": "closed", "comment": "Thanks a lot @rootonchair for starting this! The Mllama fast image processor was finished and merged in #41391, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37539"}, {"key": "pr#37537", "n": 37537, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fast tokenizer encoding doesn't handle empty string input", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37537"}, {"key": "pr#37535", "n": 37535, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Qwen2.5-VL fix redundant cu_window_seqlens", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37535"}, {"key": "pr#37534", "n": 37534, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Docs: fix docstrings for Gemma3 modeling", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37534"}, {"key": "pr#37528", "n": 37528, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Phi3", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37528"}, {"key": "pr#37522", "n": 37522, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "internalize build_inputs_with_special_tokens and prepare_for_model", "action": "closed", "comment": "Thanks for the PR! This model is deprecated and no longer maintained in the library, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37522"}, {"key": "pr#37512", "n": 37512, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update tokenization_utils_base.py", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37512"}, {"key": "pr#37509", "n": 37509, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[fix] Trainer num_tokens() count", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37509"}, {"key": "issue#37505", "n": 37505, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Tensor parallel support for LLM training.", "action": "closed", "comment": "TP training is supported on main: load with a `tp_plan`/`DistributedConfig` and train with `Trainer`, combined with data parallelism, as shown in https://huggingface.co/docs/transformers/main/en/tensor_parallelism (#44613). `save_pretrained` now gathers TP-sharded weights (or writes a distributed checkpoint with `distributed_checkpoint=True`).", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/37505"}, {"key": "issue#37502", "n": 37502, "kind": "issue", "group": "Open issues", "category": "feature", "area": "core / loading", "type": "feature", "title": "Add resume checkpoint support to ClearML callback", "action": "\u2014", "comment": "", "author": "aarbelle", "idle_days": 64, "models": "", "url": "https://github.com/huggingface/transformers/issues/37502"}, {"key": "pr#37487", "n": 37487, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix: :bug: Support explicitly passing callback", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37487"}, {"key": "pr#37475", "n": 37475, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "trainer.py fix loss aggregation over multiple devices", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37475"}, {"key": "pr#37473", "n": 37473, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Modular m4t speecht5 sew", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37473"}, {"key": "issue#37471", "n": 37471, "kind": "issue", "group": "Commented (now closed by others)", "category": "pointers (how to do it)", "area": "", "type": "", "title": "Assistant Decoding for Llava-Onevision Does Not Work", "action": "commented", "comment": "As noted above, both OneVision checkpoints use the same tokenizer, so passing `tokenizer`/`assistant_tokenizer` routes you into the text-only different-tokenizer path (`AssistedCandidateGeneratorDifferentTokenizers` in `src/transformers/generation/candidate_generator.py`), which doesn't carry `pixel_values`. Try calling `generate(..., assistant_model=draft)` without the tokenizers. If the remaining failure is the padded `vocab_size` mismatch, the fix would be to slice/pad the candidate logits in `AssistedCandidateGenerator` before they're compared with the target's; a PR with a test would be welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/37471"}, {"key": "issue#37465", "n": 37465, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "support flash-attn feature in llama4", "action": "closed", "comment": "Llama 4 has `_supports_flash_attn = False` in `src/transformers/models/llama4/modeling_llama4.py` on main, since, as noted in https://github.com/huggingface/transformers/issues/37465#issuecomment-3000352946, it needs a kernel supporting its chunked attention. Use `attn_implementation=\"sdpa\"` or `\"flex_attention\"` in the meantime. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/37465"}, {"key": "pr#37460", "n": 37460, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix interpolation of convnext image processor", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37460"}, {"key": "pr#37453", "n": 37453, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "feat: Add gradient testing for Flash Attention 2", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37453"}, {"key": "pr#37442", "n": 37442, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Implemented update function in cache_utils.py, with a test file test_cache_utils.py ", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37442"}, {"key": "pr#37432", "n": 37432, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "allow accelerate prepare for torch compiled models", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37432"}, {"key": "pr#37417", "n": 37417, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[bug] fsdp training save checkpoint twice", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37417"}, {"key": "pr#37397", "n": 37397, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add support for Moonlight 16B, add aux loss for Deepseek v3 model finetuning.", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37397"}, {"key": "pr#37396", "n": 37396, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Next batch of models with removed return_dict", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37396"}, {"key": "issue#37390", "n": 37390, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "how to reduce original model's tokenizer vocabulary", "action": "closed", "comment": "Closing this one: no activity since 2025-04-09, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/37390"}, {"key": "pr#37377", "n": 37377, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update fastspeech2 model card", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37377"}, {"key": "pr#37374", "n": 37374, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix typo in Gemma3ForCausalLM doctest", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37374"}, {"key": "issue#37364", "n": 37364, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "How to find a specific func doc when using transformers doc?", "action": "closed", "comment": "Closing this one: no activity since 2025-09-15, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/37364"}, {"key": "pr#37357", "n": 37357, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "support overlapping masks in mask2former image processor", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37357"}, {"key": "pr#37350", "n": 37350, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Haocheng lu2", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37350"}, {"key": "issue#37316", "n": 37316, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Lama4 scout. Any chance it could ever be in the browser?", "action": "closed", "comment": "Closing this one: no activity since 2025-04-06, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/37316"}, {"key": "pr#37297", "n": 37297, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Added fast processor for llava-next-video model", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37297"}, {"key": "issue#37293", "n": 37293, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "[i18n-<Transformers>] Translating docs to <Tibetain>", "action": "closed", "comment": "Closing since the template wasn't filled in (it still has the `<languageName>` placeholders). If you'd like to translate the docs into Tibetan, please open a new issue with the language code and the pages you'll take, following https://github.com/huggingface/transformers/blob/main/docs/TRANSLATING.md", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/37293"}, {"key": "pr#37267", "n": 37267, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add Magma Agentic Model from Microsoft", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37267"}, {"key": "pr#37259", "n": 37259, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Bye bye env vars, keep everything as configs", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37259"}, {"key": "pr#37252", "n": 37252, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[don't merge yet] Revert \"Revert #37031\"", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37252"}, {"key": "pr#37243", "n": 37243, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix Flash Attention guard", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37243"}, {"key": "pr#37231", "n": 37231, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update model-card for Autofomer", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37231"}, {"key": "pr#37226", "n": 37226, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[RFC] Fix Gemma 3 FP16 with activation scaling", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37226"}, {"key": "pr#37225", "n": 37225, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Adding MLPSpeculator support for assisted generation", "action": "closed", "comment": "Thanks for the work on MLPSpeculator! New decoding methods are now added as custom `generate` methods on the Hub rather than in the core library, so please consider publishing it there.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37225"}, {"key": "pr#37210", "n": 37210, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "add fast image processor for pix2struct ", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37210"}, {"key": "pr#37196", "n": 37196, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Added separate limits for saving full checkpoints and model weights", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37196"}, {"key": "issue#37195", "n": 37195, "kind": "issue", "group": "Open issues", "category": "feature", "area": "core / loading", "type": "feature", "title": "Different limits for saving only model weights and saving full checkpoints", "action": "commented", "comment": "Not supported yet: rotation is handled by `rotate_checkpoints` in `src/transformers/trainer_utils.py` (called from `Trainer._save_checkpoint`), driven only by `save_total_limit`. #37196 implemented separate limits but was closed because it targeted v4; a fresh PR on main adding e.g. a separate limit for weights-only checkpoints there would be welcome.", "author": "Tim-Siu", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/37195"}, {"key": "pr#37191", "n": 37191, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add Fast Image Processor for VideoMAE ", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37191"}, {"key": "pr#37177", "n": 37177, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "trying custom tokenizer fix", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37177"}, {"key": "pr#37168", "n": 37168, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Idefics2 Fast Image processor ", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #38157, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37168"}, {"key": "issue#37165", "n": 37165, "kind": "issue", "group": "Commented (now closed by others)", "category": "new-model details", "area": "", "type": "", "title": "Add QLIP into transformers", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [nvidia/QLIP-L-14-392](https://huggingface.co/nvidia/QLIP-L-14-392), [nvidia/QLIP-B-16-256](https://huggingface.co/nvidia/QLIP-B-16-256)\n- Downloads (last 30d): 222 \u00b7 Likes: 14 \u00b7 Params: ~738M \u00b7 License: cc-by-nc-4.0\n- Released: 2025-02 \u00b7 Remote code on the Hub: yes\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/37165"}, {"key": "pr#37155", "n": 37155, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "docs: Update LayoutLMv3 model card with standardized format and impro\u2026", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37155"}, {"key": "pr#37150", "n": 37150, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[New Model Addition] Baichuan M1", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37150"}, {"key": "pr#37127", "n": 37127, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "refactor(audio_processing): replace pipe with temp files for FFmpeg p\u2026", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37127"}, {"key": "pr#37119", "n": 37119, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Add FastImageProcessor for EfficientNet", "action": "closed", "comment": "Thanks @chewyuenrachael for the contribution! A fast EfficientNet image processor was merged in #37055, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37119"}, {"key": "pr#37106", "n": 37106, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Add Sdpa Support: `[Electra]`", "action": "closed", "comment": "Thanks for working on this! Electra now uses the shared attention interface, which already provides SDPA, so a dedicated SDPA attention class isn't needed. Closing, thank you!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37106"}, {"key": "issue#37105", "n": 37105, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Add Sdpa Support for `Electra`", "action": "closed", "comment": "Electra now goes through the attention interface (SDPA, FlashAttention, flex) since #38301. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/37105"}, {"key": "pr#37094", "n": 37094, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Add ImageProcessorFast to Efficientnet processor ", "action": "closed", "comment": "Thanks a lot @Yann-CV for the work here! A fast EfficientNet image processor was merged in #37055, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37094"}, {"key": "pr#37082", "n": 37082, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "[draft] random tests order", "action": "closed", "comment": "Closing this CI experiment draft (random test order) as it has been idle for a long time; feel free to reopen if you want to pursue it.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37082"}, {"key": "issue#37078", "n": 37078, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Do not update cache when use_cache=False and past_key_values are provided?", "action": "closed", "comment": "As discussed above, we don't plan to change the cache update semantics; to reuse a context cache without it growing, use `cache.crop(context_length)` after each forward (https://huggingface.co/docs/transformers/main/en/internal/generation_utils#transformers.Cache.crop) or a custom cache layer in `src/transformers/cache_utils.py`. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/37078"}, {"key": "pr#37072", "n": 37072, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Reverse dependency map shouldn't be created when test_all is set", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37072"}, {"key": "pr#37049", "n": 37049, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Adding a stub for MiniCPM-o to the models", "action": "closed", "comment": "Thanks for getting the ball rolling on MiniCPM-o, @jecrs! As mentioned in the thread, a more complete port is being worked on in #37917, so this stub isn't needed anymore. Contributions to that effort, or a new PR against `main` written with [modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers), are very welcome. Closing this one, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37049"}, {"key": "pr#37045", "n": 37045, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": " Add Fast Image Processor for Idefics3", "action": "closed", "comment": "Thanks for this contribution! The Idefics3 fast image processor landed in #38157 (which credits this PR), so closing as superseded.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37045"}, {"key": "pr#37042", "n": 37042, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Feature universe - \u91cf\u5b50\u7ecf\u5178\u540c\u6784Transformer\u6a21\u578b\u5b9e\u73b0\u4e0e\u4f18\u5316 | Quantum-Classical Isomorphic Transformer Model Implementation and Optimization", "action": "closed", "comment": "Thanks for sharing this work! As suggested in the thread, an experimental architecture like this is a much better fit as a [custom model](https://huggingface.co/docs/transformers/en/custom_models) shared on the Hub with `trust_remote_code`, rather than new top-level and example files in the library. Closing this PR, but we'd be glad to see it on the Hub.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37042"}, {"key": "pr#37025", "n": 37025, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix best_model_checkpoint is None issue when distiributed training", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37025"}, {"key": "pr#37009", "n": 37009, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Export Whisper to ExecuTorch", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/37009"}, {"key": "pr#36999", "n": 36999, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add Fast SamImageProcessor", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36999"}, {"key": "issue#36977", "n": 36977, "kind": "issue", "group": "Open issues", "category": "feature", "area": "models", "type": "feature", "title": "QuestionAnswering for Gemma 3", "action": "commented", "comment": "Not on main yet (it was held back because no official checkpoint is QA-tuned), but it's now small: `GenericForQuestionAnswering` in `src/transformers/modeling_layers.py` provides the head, used e.g. by `ArceeForQuestionAnswering`. Add `Gemma3TextForQuestionAnswering(GenericForQuestionAnswering, Gemma3PreTrainedModel)` in `src/transformers/models/gemma3/modular_gemma3.py` (mirroring the existing `Gemma3TextForSequenceClassification`), register it in `models/auto/modeling_auto.py`, and run `make fix-repo`. Guide: https://huggingface.co/docs/transformers/main/en/modular_transformers", "author": "DavidAdamczyk", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/36977"}, {"key": "pr#36969", "n": 36969, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Make executorch integration more seamless by analyzing model signature", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36969"}, {"key": "pr#36952", "n": 36952, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Added Sapnous Architecture", "action": "closed", "comment": "Thanks for the contribution! As mentioned in review, new architectures without broad adoption are best shared as custom code on the Hub (https://huggingface.co/docs/transformers/custom_models#model). Closing for now; happy to revisit if the model gains wide usage.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36952"}, {"key": "pr#36948", "n": 36948, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update image_processing_qwen2_vl.py\u3002fix bug.", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36948"}, {"key": "pr#36943", "n": 36943, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Static cache should support indexing", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36943"}, {"key": "issue#36941", "n": 36941, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Add param_to_hook_all_reduce parameter in HF Trainer", "action": "commented", "comment": "Still wanted (see SunMarc's comment), but it needs accelerate first: `accelerate.utils.DistributedDataParallelKwargs` doesn't have a `param_to_hook_all_reduce` field yet. Once it does, the Trainer side is small: add a `ddp_param_to_hook_all_reduce` arg in `training_args.py` and forward it in the `ddp_kwargs` dict in `Trainer._build_accelerator_args` (`src/transformers/trainer.py`), like `ddp_static_graph`. #37933 was closed, so a fresh PR (starting in accelerate) would be welcome.", "author": "awsankur", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/36941"}, {"key": "pr#36928", "n": 36928, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Remove the redundant shift during the loss computation in the Moshi m\u2026", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36928"}, {"key": "pr#36919", "n": 36919, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[qwen2-audio] remove default template", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36919"}, {"key": "issue#36904", "n": 36904, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "PixtralVisionModel does not support Flash Attention 2.0 yet", "action": "closed", "comment": "This was fixed by #37576, which is now on main. Closing, but feel free to reopen if you still see it on the latest version!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/36904"}, {"key": "pr#36901", "n": 36901, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add NGen3 ", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36901"}, {"key": "issue#36900", "n": 36900, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "groot n1", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [nvidia/GR00T-N1.7-3B](https://huggingface.co/nvidia/GR00T-N1.7-3B), [nvidia/GR00T-N1.6-3B](https://huggingface.co/nvidia/GR00T-N1.6-3B), [nvidia/GR00T-N1-2B](https://huggingface.co/nvidia/GR00T-N1-2B)\n- Downloads (last 30d): 106,090 \u00b7 Likes: 162 \u00b7 Params: ~3.1B (N1: ~2.2B) \u00b7 License: NVIDIA license (see model card)\n- Released: 2025-03 (N1), 2026-02 (N1.7) \u00b7 Remote code on the Hub: no (Isaac-GR00T repo code)\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "sushmanthreddy", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/36900"}, {"key": "pr#36894", "n": 36894, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Enable tracing for Moshi", "action": "closed", "comment": "Thanks for the work! torch.fx tracing support has been removed from transformers (compile/export are the supported paths), so we won't enable it for Moshi. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36894"}, {"key": "pr#36892", "n": 36892, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "processors (image/video/audio)", "type": "feature / refactor", "title": "[WIP] Computer vision util: vision visualizer", "action": "\u2014", "comment": "", "author": "molbap", "idle_days": 364, "models": "", "url": "https://github.com/huggingface/transformers/pull/36892"}, {"key": "pr#36883", "n": 36883, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix `resume_from_checkpoint` not recognising `\"last-checkpoint\"`", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36883"}, {"key": "issue#36874", "n": 36874, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Support for SpatialLM series model", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [manycore-research/SpatialLM1.1-Qwen-0.5B](https://huggingface.co/manycore-research/SpatialLM1.1-Qwen-0.5B), [manycore-research/SpatialLM1.1-Llama-1B](https://huggingface.co/manycore-research/SpatialLM1.1-Llama-1B), [manycore-research/SpatialLM-Llama-1B](https://huggingface.co/manycore-research/SpatialLM-Llama-1B)\n- Downloads (last 30d): 7,741 \u00b7 Likes: 38 \u00b7 Params: ~604M \u00b7 License: cc-by-nc-4.0\n- Released: 2025-06 \u00b7 Remote code on the Hub: no\n- The checkpoints use a custom `spatiallm_*` model type built on Llama/Qwen2 plus a point-cloud encoder, so a modular file on top of `llama`/`qwen2` should keep this short.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "nekomeowww", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/36874"}, {"key": "pr#36870", "n": 36870, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Improve Model Download Speeds By ~3x For Large Models", "action": "closed", "comment": "Thanks for the great analysis! As discussed with @Wauplin, download logic belongs in `huggingface_hub` rather than transformers, so we'll close this here. Please do open it there!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36870"}, {"key": "pr#36839", "n": 36839, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Haocheng lu", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36839"}, {"key": "issue#36824", "n": 36824, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Support for Ovis2 models", "action": "closed", "comment": "This has been added in #37088 and is available on main. Closing, but feel free to reopen if something is still missing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/36824"}, {"key": "issue#36817", "n": 36817, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add EuroBert Model To Config", "action": "commented", "comment": "Good news: this is now supported in transformers as `eurobert` (added in #39455).", "author": "zynos", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/36817"}, {"key": "pr#36809", "n": 36809, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix unexpected kws of input_ids when setup no speech detection of whisper", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36809"}, {"key": "pr#36808", "n": 36808, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Support loading custom code objects (`trust_remote_code=True`) in offline mode from local", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36808"}, {"key": "pr#36807", "n": 36807, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add long vita", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36807"}, {"key": "pr#36791", "n": 36791, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update configuration_auto.py for NGen3", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36791"}, {"key": "pr#36787", "n": 36787, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Create modeling_ngen3.py for NGen3", "action": "closed", "comment": "Thanks for sharing your model! For new architectures like this the best path is the Hub with custom code (https://huggingface.co/docs/transformers/model_sharing), which is much easier to ship. Closing, and good luck with it!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36787"}, {"key": "pr#36782", "n": 36782, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix attention_mask dimension issue in GPT2Model", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36782"}, {"key": "pr#36760", "n": 36760, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "feat: expose the strict flag to allow catching missing model layers while loading a checkpoint", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36760"}, {"key": "issue#36743", "n": 36743, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Model Card to include key information (e.g. max_sequence_length, etc.)", "action": "closed", "comment": "As @Rocketknight1 explained in https://github.com/huggingface/transformers/issues/36743#issuecomment-2736685144, model cards are controlled by their authors and we can't enforce their content; recommended generation parameters belong in the repo's `generation_config.json`, which `generate()` picks up automatically. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/36743"}, {"key": "pr#36737", "n": 36737, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "[WP] PagedAttention + Prefix Cache for FlashAttention2", "action": "closed", "comment": "Thanks for this! Continuous batching / paged attention landed through the generation stack without touching modeling files, which is the approach we want to keep. Closing, thanks for the inspiration.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36737"}, {"key": "issue#36718", "n": 36718, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add RoMa keypoint matcher", "action": "commented", "comment": "No public weights on the Hub as of today. Official RoMa weights are distributed via the GitHub repo, not the Hub; an HF-format conversion would be part of the port.\n- Open PR: #47082", "author": "tcourat", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/36718"}, {"key": "pr#36711", "n": 36711, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "models", "type": "new model", "title": "[WIP] Add DINO DETR Model to HuggingFace Transformers", "action": "\u2014", "comment": "", "author": "konstantinos-p", "idle_days": 179, "models": "dino_detr", "url": "https://github.com/huggingface/transformers/pull/36711"}, {"key": "pr#36708", "n": 36708, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "generation", "type": "fix", "title": "Fix long lagging when streaming text without spaces and NJK chars", "action": "\u2014", "comment": "", "author": "CNSeniorious000", "idle_days": 2, "models": "", "url": "https://github.com/huggingface/transformers/pull/36708"}, {"key": "pr#36706", "n": 36706, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Limit numpy version to <2.0.0", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36706"}, {"key": "issue#36697", "n": 36697, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Support Flex Attention for encoder only models (XLMRoberta, ModernBERT etc...)", "action": "closed", "comment": "Flex attention is now supported for encoder models: XLM-RoBERTa via #38301 and ModernBERT via #43030. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/36697"}, {"key": "issue#36659", "n": 36659, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Qwen2 MoE manual `head_dim`", "action": "closed", "comment": "Qwen2-MoE supports a custom `head_dim` since #37188. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/36659"}, {"key": "pr#36647", "n": 36647, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix device issue in modeling_qwen2", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36647"}, {"key": "issue#36640", "n": 36640, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "[Feature Request]: refactor _update_causal_mask to  a public utility", "action": "closed", "comment": "This was done in #37866: mask creation now lives in public utilities in `transformers.masking_utils` (e.g. `create_causal_mask`). Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/36640"}, {"key": "issue#36629", "n": 36629, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add Magma from Microsoft to Transformers", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [microsoft/Magma-8B](https://huggingface.co/microsoft/Magma-8B)\n- Downloads (last 30d): 1,373 \u00b7 Likes: 422 \u00b7 Params: ~8.9B \u00b7 License: mit\n- Released: 2025-02 \u00b7 Remote code on the Hub: yes\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "VladOS95-cyber", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/36629"}, {"key": "pr#36621", "n": 36621, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[WIP] Add support to load models with transforms", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36621"}, {"key": "issue#36564", "n": 36564, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Add support for StableAdamW optimizer in Trainer", "action": "closed", "comment": "Good news: StableAdamW was added to the Trainer in #39446; use `optim=\"stable_adamw\"` (requires `torch-optimi`). Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/36564"}, {"key": "pr#36556", "n": 36556, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "add-LongVITAModel", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36556"}, {"key": "pr#36553", "n": 36553, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "add-long-vita", "action": "closed", "comment": "Thanks for the contribution! New models now go through a modular file and our attention interface instead of custom attention code. The fastest way to share Long-VITA is Hub remote code, and #36807 continues the in-library effort. Closing this one. You can add as a model but needs a new POR with latest good practices!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36553"}, {"key": "pr#36514", "n": 36514, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Refactor object-detection models", "action": "closed; comment corrected", "comment": "Thanks for this work! The object-detection models were refactored to the updated standards in #41549 (DETR, Conditional/Deformable DETR, RT-DETR and others), so this PR is superseded. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36514"}, {"key": "issue#36513", "n": 36513, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Allow parameters of the ViTPooler to be configurable, with the default values set to the current hardcoded values", "action": "closed", "comment": "This has been added in #36517 and is available on main. Closing, but feel free to reopen if something is still missing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/36513"}, {"key": "pr#36496", "n": 36496, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add an event related to forward in the TrainerCallback", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36496"}, {"key": "pr#36491", "n": 36491, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Allow OOV Image Token for LLaVa Next Variants", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36491"}, {"key": "issue#36481", "n": 36481, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Add type checking to CI", "action": "closed", "comment": "Type checking now runs in our CI (`make typing`, using `ty`, introduced in #44167), and typing coverage keeps improving. Closing, but feel free to open targeted issues for remaining type errors!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/36481"}, {"key": "pr#36478", "n": 36478, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Sanitize Model Module Names to Follow Python Conventions", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36478"}, {"key": "pr#36432", "n": 36432, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add fetch_paginated_github_data to deduplicate GitHub API pagination \u2026", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36432"}, {"key": "pr#36412", "n": 36412, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[i18n-zh] Translating kv_cache into zh-hans", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36412"}, {"key": "pr#36401", "n": 36401, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "[Model] Optimize BERT memory usage and improve code readability", "action": "closed", "comment": "Thanks for the contribution! As mentioned, the BERT modeling code has since been refactored so these changes are no longer needed; closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36401"}, {"key": "issue#36398", "n": 36398, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Please add support for TransPixar AI", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [wileewang/TransPixar](https://huggingface.co/wileewang/TransPixar)\n- Downloads (last 30d): 84 \u00b7 Likes: 43 \u00b7 Params: n/a \u00b7 License: apache-2.0\n- Released: 2024-12 \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "jozefchutka", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/36398"}, {"key": "issue#36381", "n": 36381, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "WARNING: transformers 4.49.0 does not provide the extra 'emu3'", "action": "closed", "comment": "`setup.py` doesn't define an `emu3` extra, so pip's warning is expected; Emu3 ships in the base package (`Emu3ForConditionalGeneration`). Closing since the reproducer requested in https://github.com/huggingface/transformers/issues/36381#issuecomment-2681133803 wasn't provided.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/36381"}, {"key": "issue#36379", "n": 36379, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add EVEv2 : an Encoder-free VLM", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [BAAI/EVE-7B-HD-v2.0](https://huggingface.co/BAAI/EVE-7B-HD-v2.0)\n- Downloads (last 30d): 127 \u00b7 Likes: 9 \u00b7 Params: ~14.2B (stored), Qwen2-based \u00b7 License: apache-2.0\n- Released: 2025-02 \u00b7 Remote code on the Hub: no (original-repo code)\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "sbucaille", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/36379"}, {"key": "pr#36377", "n": 36377, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix: Use config.use_sliding_window instead of config.sliding_window", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36377"}, {"key": "pr#36356", "n": 36356, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Introduce numpy/numba optimization to `Qwen2VLImageProcessor`", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36356"}, {"key": "issue#36352", "n": 36352, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Implement Titans Architecture with GRPO Fine-Tuning", "action": "commented", "comment": "No public weights on the Hub as of today: there's no official pretrained Titans checkpoint, and we only add architectures once a pretrained model exists.", "author": "rajveer43", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/36352"}, {"key": "issue#36351", "n": 36351, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Support sliding_window for sdpa in qwen2", "action": "closed", "comment": "Qwen2 now builds proper sliding-window masks for every attention backend, SDPA included, after the masking rewrite in #37866. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/36351"}, {"key": "issue#36339", "n": 36339, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Making apply_rotary_pos_emb generic", "action": "closed", "comment": "torch.fx support was removed in #41683, and on main `apply_rotary_pos_emb` is decorated with `@use_kernel_forward_from_hub(\"rotary_pos_emb\")` (e.g. src/transformers/models/llama/modeling_llama.py), which gives a single hook to swap the implementation. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/36339"}, {"key": "issue#36310", "n": 36310, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Include \"time\" as option to save_strategy (and log and eval too!)", "action": "commented", "comment": "Still not available on main. The interval strategies live in `IntervalStrategy` (`src/transformers/trainer_utils.py`) and the save/eval/log decisions are made in `DefaultFlowCallback.on_step_end` (`src/transformers/trainer_callback.py`), so a time-based option would add e.g. `save_minutes`/`eval_minutes` to `TrainingArguments` and a wall-clock check there (synchronized across ranks). The earlier PR #37642 was closed only because it was based on v4; a fresh PR against main is welcome. Meanwhile, a custom `TrainerCallback` setting `control.should_save = True` on a timer works.", "author": "davidhughhenrymack", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/36310"}, {"key": "pr#36303", "n": 36303, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Modify save_steps when resuming from a checkpoint.", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36303"}, {"key": "pr#36300", "n": 36300, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update computation method for training samples", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36300"}, {"key": "pr#36287", "n": 36287, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[modular] allow multiple modular files in the same model folder", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36287"}, {"key": "pr#36221", "n": 36221, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Prevent Reinitialization of Resized LM Head When `tie_word_embeddings` is False #35141", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #45079, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36221"}, {"key": "pr#36216", "n": 36216, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "[WIP] Add a dedicated tokenizer for byte level transformers", "action": "closed", "comment": "Thanks for the proposal! We only add tokenizers alongside a model with released checkpoints, and v5 consolidated tokenizers on the `tokenizers` backend, so a standalone slow tokenizer in utils isn't the direction. Closing, thank you!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36216"}, {"key": "pr#36214", "n": 36214, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix the eval_use_gather_object flag usage", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36214"}, {"key": "pr#36211", "n": 36211, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix: condition bos_token_id and space as token", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36211"}, {"key": "issue#36205", "n": 36205, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Request to add DINO object detector", "action": "commented", "comment": "Some details for anyone who'd like to pick this up: no official DINO (DETR with improved denoising anchors) checkpoints are on the Hugging Face Hub as of today, so adding it would first need converted weights.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "tcourat", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/36205"}, {"key": "issue#36204", "n": 36204, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Request to add DEIM object detector", "action": "commented", "comment": "Good news: the follow-up DEIMv2 is now supported in transformers as `deimv2` (added in #44339), e.g. [Intellindust/DEIMv2_HGNetv2_N_COCO](https://huggingface.co/Intellindust/DEIMv2_HGNetv2_N_COCO). DEIM (v1) itself uses the D-FINE architecture (`d_fine`), and its weights are only on GitHub, not the Hub.", "author": "tcourat", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/36204"}, {"key": "issue#36202", "n": 36202, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Dedicated tokenizer for byte level transformers", "action": "closed", "comment": "Byte-level models are supported on main: the Byte Latent Transformer landed as `blt` (#38579, https://huggingface.co/docs/transformers/main/en/model_doc/blt), and `ByT5Tokenizer` already maps text to raw UTF-8 bytes with no vocabulary. For a custom byte/patch scheme, https://huggingface.co/docs/transformers/main/en/custom_tokenizers explains how to ship your own tokenizer.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/36202"}, {"key": "pr#36201", "n": 36201, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Append best model checkpoint with active adapter when not default", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36201"}, {"key": "pr#36198", "n": 36198, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fixed dynamic module import when there is more than one dot in class \u2026", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36198"}, {"key": "pr#36197", "n": 36197, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "(ugly) Use  `parallelism=4` for `check_repository_consistency`", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36197"}, {"key": "issue#36166", "n": 36166, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add phi3 Vision", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [microsoft/Phi-3-vision-128k-instruct](https://huggingface.co/microsoft/Phi-3-vision-128k-instruct), [microsoft/Phi-3.5-vision-instruct](https://huggingface.co/microsoft/Phi-3.5-vision-instruct)\n- Downloads (last 30d): 44,108 \u00b7 Likes: 976 \u00b7 Params: ~4.1B \u00b7 License: mit\n- Released: 2024-05 \u00b7 Remote code on the Hub: yes\n- Open PR: #41977 (the earlier #36036 was closed)\n- Phi-3.5-vision has far more usage: 615,779 downloads (last 30d), same architecture.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "ArthurZucker", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/36166"}, {"key": "issue#36144", "n": 36144, "kind": "issue", "group": "Open issues", "category": "feature", "area": "gguf", "type": "feature", "title": "Add the support for deepseek architecture .gguf", "action": "commented", "comment": "DeepSeek GGUF is still not supported, but the path is now well trodden: add a `deepseek2` entry (llama.cpp's arch name for V2/V3/R1) to `GGUF_CONFIG_MAPPING` in `src/transformers/integrations/ggml.py`, map it to `deepseek_v3` in `src/transformers/modeling_gguf_pytorch_utils.py`, and handle stacked expert tensors like the existing MoE support (see `_split_moe_expert_tensor` and MiniMax-M2 support in #44526, or `qwen3_moe`). The MLA `kv_b` tensors will likely need a custom `TensorProcessor`. Docs: https://huggingface.co/docs/transformers/main/en/quantization/gguf; contributions welcome.", "author": "z7d1", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/36144"}, {"key": "pr#36133", "n": 36133, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Set evaluation and checkpointing defaults to 'epoch' and reduce loggi\u2026", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36133"}, {"key": "issue#36119", "n": 36119, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "Adapting Whisper to the new loss_function attribute", "action": "\u2014", "comment": "", "author": "yoadsn", "idle_days": 123, "models": "", "url": "https://github.com/huggingface/transformers/issues/36119"}, {"key": "issue#36111", "n": 36111, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add Deepseek-VL2", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [deepseek-ai/deepseek-vl2](https://huggingface.co/deepseek-ai/deepseek-vl2)\n- Downloads (last 30d): 7.5k \u00b7 Likes: 406 \u00b7 Params: ~27.5B \u00b7 License: other\n- Released: 2024-12 \u00b7 Remote code on the Hub: no\n- Note: `deepseek_vl` (v1) is on main; DeepSeek-VL2 (`deepseek_vl_v2`) is not.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "ArthurZucker", "idle_days": 0, "models": "deepseek_vl", "url": "https://github.com/huggingface/transformers/issues/36111"}, {"key": "pr#36095", "n": 36095, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fixup another model + encoder decoder", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36095"}, {"key": "pr#36090", "n": 36090, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix: Llama - adjust the rotary embedding dimensions.", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36090"}, {"key": "pr#36089", "n": 36089, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix: dtype might change during resize", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36089"}, {"key": "pr#36073", "n": 36073, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Speedup modular conversion w/ multiproc", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36073"}, {"key": "issue#36069", "n": 36069, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Question/request: High-Level Model Architecture Visualization Tool", "action": "closed", "comment": "This is something planed, updated in https://arthurz-transformers-block-atlas.static.hf.space/index.html PR is #48271", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/36069"}, {"key": "pr#36038", "n": 36038, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix bug in prepare_inputs_for_generation function: in Qwen2-VL (#36037)", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36038"}, {"key": "pr#36011", "n": 36011, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix ImportError: cannot import name 'GenerationMixin' from 'transformers.generation'", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36011"}, {"key": "pr#36005", "n": 36005, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[`Pegasus(-x)`] Fixup embedding interfaces on Encoder/Decoder", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/36005"}, {"key": "issue#35978", "n": 35978, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "HPD-Transformer: A Hybrid Parsing-Density Transformer for Efficient Structured & Probabilistic Reasoning", "action": "commented", "comment": "No public weights on the Hub as of today. No implementation or checkpoint is linked either.", "author": "infodevlovable", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/35978"}, {"key": "pr#35939", "n": 35939, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[Whisper] Fix num_return_sequences behavior", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35939"}, {"key": "issue#35929", "n": 35929, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add YuE audio model", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [m-a-p/YuE-s1-7B-anneal-en-cot](https://huggingface.co/m-a-p/YuE-s1-7B-anneal-en-cot), [m-a-p/YuE-s2-1B-general](https://huggingface.co/m-a-p/YuE-s2-1B-general)\n- Downloads (last 30d): 3,374 \u00b7 Likes: 462 \u00b7 Params: ~6.2B \u00b7 License: apache-2.0\n- Released: 2025-01 \u00b7 Remote code on the Hub: no\n- Open PR: #41041 (WIP)\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "ArthurZucker", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/35929"}, {"key": "pr#35920", "n": 35920, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": " changed MllamaPreTrainedModel._prepare_4d_causal_attention_mask", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35920"}, {"key": "pr#35904", "n": 35904, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix Mask2Former Weight Initialization Issues #35877", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35904"}, {"key": "pr#35887", "n": 35887, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Adds GGUF support for Gemma models", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35887"}, {"key": "issue#35882", "n": 35882, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Request to add Co-DETR", "action": "commented", "comment": "Some details for anyone who'd like to pick this up: no official Co-DETR checkpoints are on the Hugging Face Hub as of today, so adding it would first need converted weights.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "daniel-bogdoll", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/35882"}, {"key": "issue#35877", "n": 35877, "kind": "issue", "group": "Open issues", "category": "bug", "area": "cache / masks", "type": "bug", "title": "Mask2Former _init_weights", "action": "commented", "comment": "Still valid on main: `Mask2FormerPreTrainedModel._init_weights` in `src/transformers/models/mask2former/modeling_mask2former.py` still normal-inits `nn.Embedding` with `config.init_std` and zeroes the mask-MLP biases. #35904 was closed only because it targeted v4; a fresh PR against main using the `transformers.initialization` (`init.*`) helpers already used in that method, matched against the original Mask2Former init, would be welcome.", "author": "tommiekerssies", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/35877"}, {"key": "issue#35874", "n": 35874, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "ZeroShotClassificationArgumentHandler should be explicit it has a somewhat unsafe internal behaviour.", "action": "\u2014", "comment": "", "author": "nicolasdalsass", "idle_days": 4, "models": "", "url": "https://github.com/huggingface/transformers/issues/35874"}, {"key": "pr#35847", "n": 35847, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix Jitter Noise Passing to Experts in Switch Transformers #33969", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35847"}, {"key": "pr#35841", "n": 35841, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update modeling_attn_mask_utils.py", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35841"}, {"key": "pr#35825", "n": 35825, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add MultipleChoice & QuestionAnswering heads to ModernBERT", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35825"}, {"key": "pr#35819", "n": 35819, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Exploring use of kwargs for timm model and transforms creation", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35819"}, {"key": "pr#35799", "n": 35799, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Fix Noise Computation for NEFTune During Packed Training #34659", "action": "closed", "comment": "Thanks @sambhavnoobcoder for digging into NEFTune with packed sequences! Since then the NEFTune implementation was moved out of the Trainer into standalone functions in `src/transformers/integrations/neftune.py` (#43714), so this branch no longer applies. As far as I can tell packed batches are still not handled per sequence, so if you're still interested, feel free to open a fresh PR against `main` on top of the new module. Closing this one in the meantime.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35799"}, {"key": "pr#35793", "n": 35793, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix docstring for get_candidates return shape", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35793"}, {"key": "pr#35790", "n": 35790, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add StyleTTS 2", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35790"}, {"key": "pr#35769", "n": 35769, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fixed ViTMAE for non-square images", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35769"}, {"key": "pr#35764", "n": 35764, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add timm_wrapper support to AutoFeatureExtractor", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35764"}, {"key": "pr#35732", "n": 35732, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Generate handle per batch idx params", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35732"}, {"key": "pr#35720", "n": 35720, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Added support for GraniteForSequenceClassification", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35720"}, {"key": "pr#35694", "n": 35694, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add internlm3 dense", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35694"}, {"key": "issue#35683", "n": 35683, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "Support LLMs With No Image Placeholder Embedding in LLava-based Models", "action": "commented", "comment": "Still valid: `LlavaModel.forward` embeds `input_ids` directly (`src/transformers/models/llava/modeling_llava.py`) before `get_placeholder_mask` swaps in image features, so an image token outside the LM vocab raises `IndexError`. #36491 implemented this but was closed because it targeted v4; a fresh PR on main could mask placeholder ids (e.g. replace them with 0) before `get_input_embeddings()(input_ids)` in `modeling_llava.py`, then propagate to llava-next and the modular files that inherit from it with `make fix-repo`.", "author": "alex-jw-brooks", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/35683"}, {"key": "pr#35674", "n": 35674, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix Aria CI and testing", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35674"}, {"key": "issue#35645", "n": 35645, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Will Qwen2VL support sequence classification head in the future?", "action": "closed", "comment": "As mentioned above, we add task heads when there are trained checkpoints using them, and none were found for Qwen2-VL sequence classification. In the meantime, a small custom head on top of `Qwen2VLModel` (pooling the last non-padded token, as `GenericForSequenceClassification` in `src/transformers/modeling_layers.py` does) works well. Closing; feel free to reopen with a checkpoint.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/35645"}, {"key": "issue#35634", "n": 35634, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "ValueError: MllamaForConditionalGeneration does not support Flash Attention 2.0 yet", "action": "closed", "comment": "This was fixed by #37576, which is now on main. Closing, but feel free to reopen if you still see it on the latest version!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/35634"}, {"key": "pr#35627", "n": 35627, "kind": "PR", "group": "Open PRs", "category": "awaiting review", "area": "processors (image/video/audio)", "type": "fix", "title": "Fix Batch Size Mismatch When Using `crops_n_layers` in `mask-generation` Pipeline #35530", "action": "\u2014", "comment": "", "author": "sambhavnoobcoder", "idle_days": 3, "models": "sam", "url": "https://github.com/huggingface/transformers/pull/35627"}, {"key": "pr#35585", "n": 35585, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "A new Traditional Chinese version of the README.md for run_on_remote.py", "action": "closed", "comment": "Thank you for the translation! The script this README documents, `examples/run_on_remote.py`, was removed from the repository in #49357, so there's nothing left for this page to describe. Closing this PR, and thanks again for contributing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35585"}, {"key": "pr#35571", "n": 35571, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Exploit symmetry of covariance matrices for faster & more stable diag\u2026", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35571"}, {"key": "pr#35550", "n": 35550, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Add AIMv2 to Transformers", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #36625, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35550"}, {"key": "pr#35535", "n": 35535, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "change README FILE", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35535"}, {"key": "issue#35530", "n": 35530, "kind": "issue", "group": "Open issues", "category": "bug", "area": "generation", "type": "bug", "title": "SAM mask-generation - crops_n_layers", "action": "\u2014", "comment": "", "author": "kusstox", "idle_days": 571, "models": "", "url": "https://github.com/huggingface/transformers/issues/35530"}, {"key": "issue#35525", "n": 35525, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Trainer: Use second last checkpoint if last checkpoint loading fails", "action": "\u2014", "comment": "", "author": "SilverSoldier", "idle_days": 15, "models": "", "url": "https://github.com/huggingface/transformers/issues/35525"}, {"key": "pr#35517", "n": 35517, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add support for 4D custom attention masks in GPT-2", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35517"}, {"key": "pr#35515", "n": 35515, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Modular files", "action": "closed", "comment": "Thanks for the PR! TensorFlow and Flax support was removed in v5, so this can't land anymore. Closing, thanks again for the contribution!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35515"}, {"key": "pr#35511", "n": 35511, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update modeling_utils.py to fix typo", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35511"}, {"key": "pr#35506", "n": 35506, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix Qwen2RotaryEmbedding Device Matching Bug", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35506"}, {"key": "pr#35500", "n": 35500, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update conftest.py", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35500"}, {"key": "issue#35492", "n": 35492, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Instructions to raise PR for addition of shared library files(.so) and .cpp files ", "action": "closed", "comment": "We no longer accept compiled `.so`/C++ code into the repo. Custom kernels are now published on the Hub and loaded with the `kernels` library; see https://huggingface.co/docs/transformers/main/en/kernel_doc/writing_kernels for building one and https://huggingface.co/docs/transformers/main/en/kernels for wiring it into a model (e.g. via `use_kernel_forward_from_hub`). Closing as that covers the question.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/35492"}, {"key": "pr#35488", "n": 35488, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "[tokenizer] fix llama tokenizer (slow)", "action": "closed", "comment": "Thanks for the fix! In v5 the Llama tokenizer is backed by `tokenizers` only and the slow implementation you patched no longer exists, so we're closing this.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35488"}, {"key": "pr#35476", "n": 35476, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Add FAST", "action": "closed", "comment": "Thanks a lot for all the work on FAST, @jadechoghari, and for getting the TextNet backbone merged in #34979! This branch has gone over a year without activity and is now far behind `main`. Models are now added with [modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers) and checked with [mlinter](https://huggingface.github.io/transformers-mlinter/), so a fresh implementation would be much smaller than this diff. If you'd like to pick it up again, a new PR against `main` that builds on `textnet` would be very welcome. Closing this one for now, thanks!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35476"}, {"key": "pr#35475", "n": 35475, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "cover edge case where decoder_input_ids is provided but it is empty", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35475"}, {"key": "pr#35469", "n": 35469, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Use new attention API for LayoutLMv3 (SDPA, Flash Attn v2 support)", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35469"}, {"key": "issue#35467", "n": 35467, "kind": "issue", "group": "Open issues", "category": "feature", "area": "attention / kernels", "type": "feature", "title": "Support SDPA & Flash Attention 2 for LayoutLMv3", "action": "commented", "comment": "Still open: `src/transformers/models/layoutlmv3/modeling_layoutlmv3.py` doesn't use the attention interface yet (earlier attempts #35469, #41801, #42225 were closed). The path is to port its self-attention to `ALL_ATTENTION_FUNCTIONS` as BERT does (https://huggingface.co/docs/transformers/main/en/attention_interface); note LayoutLMv3 adds 1D/2D relative position biases to the scores, so SDPA can take them as an additive mask, while flash-attention would need to be disabled when those biases are used.", "author": "stancld", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/35467"}, {"key": "issue#35464", "n": 35464, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "\"Is it possible for Hugging Face to implement a chat model for quick information retrieval similar to vLLM?\"", "action": "closed", "comment": "Transformers now ships this: `transformers serve` (`src/transformers/cli/serve.py`) starts an OpenAI-compatible server and `transformers chat <model>` (`src/transformers/cli/chat.py`) gives an interactive terminal chat. See https://huggingface.co/docs/transformers/main/en/serve-cli/serving", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/35464"}, {"key": "pr#35455", "n": 35455, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Fix #35447 Tokenizer does not split text according to newly added input tokens", "action": "closed", "comment": "Thanks for digging into the `Trie.split` behaviour! Since then `src/transformers/tokenization_utils.py` was removed in #40936 (the `Trie` now lives in `src/transformers/tokenization_python.py`), and #35447 has been closed, so this branch can't be merged anymore. If you can still reproduce the issue on `main`, please open a fresh PR there with the extra test cases (e.g. `?` / emojis) mentioned in the review. Closing for now.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35455"}, {"key": "issue#35451", "n": 35451, "kind": "issue", "group": "Open issues", "category": "feature", "area": "generation", "type": "feature", "title": "[Feature Request] Add beam search text streaming visualization feature", "action": "\u2014", "comment": "", "author": "MosheOfer1", "idle_days": 655, "models": "", "url": "https://github.com/huggingface/transformers/issues/35451"}, {"key": "pr#35436", "n": 35436, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Multi Beam Text Streamer", "action": "closed", "comment": "Thanks for the idea! It depends on `beam_search.py`, which was removed when beam search was refactored. Something like this could be shared as a Hub custom generation method. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35436"}, {"key": "pr#35420", "n": 35420, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "remove-redundant-code", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35420"}, {"key": "pr#35419", "n": 35419, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[WIP] Add support for flex attention (paged attention)", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35419"}, {"key": "pr#35415", "n": 35415, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "support telechat2", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35415"}, {"key": "pr#35392", "n": 35392, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Pass `hidden_states` to `MllamaVisionEncoderLayer` w.o. keyword", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35392"}, {"key": "issue#35385", "n": 35385, "kind": "issue", "group": "Open issues", "category": "feature", "area": "models", "type": "feature", "title": "Support modernBERT for encoder-decoder models", "action": "commented", "comment": "ModernBERT can already be the encoder in `EncoderDecoderModel`; what's missing is a decoder with cross-attention. `ModernBertDecoderForCausalLM` (#38967, `src/transformers/models/modernbert_decoder/`) has no cross-attention yet, so a PR adding an optional cross-attention block (see `add_cross_attention` in `modeling_bert.py` for the pattern) would make `from_encoder_decoder_pretrained(\"answerdotai/ModernBERT-base\", <modernbert-decoder ckpt>)` work.", "author": "Bachstelze", "idle_days": 0, "models": "encoder_decoder", "url": "https://github.com/huggingface/transformers/issues/35385"}, {"key": "pr#35381", "n": 35381, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Add support for H2O cache eviction with LLaMA", "action": "closed", "comment": "Thanks for the H2O implementation! We no longer add new cache classes or modeling changes to the core library; the right home is a Hub `custom_generate` repo (template: https://huggingface.co/transformers-community/sink_cache). Closing, but please share it there.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35381"}, {"key": "pr#35365", "n": 35365, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "FineTuning AutoModelForSequenceClassification.from_pretrained(meta-llama/Llama-3.2-1B) Bug:RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:1 and cuda:0! (when checking argument for argument target in method wrapper_CUDA_nll_loss_forward) and awq importing", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35365"}, {"key": "issue#35346", "n": 35346, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "[`Mamba2`] Varlen implementation", "action": "closed", "comment": "Varlen/packed inputs are now supported for Mamba2: #48490 forwards kernel kwargs to the mixer so `model(input_ids, seq_idx=seq_idx)` reaches `causal_conv1d_fn` and the chunk-scan kernels, keeping state from bleeding across packed sequences, and #49474 raises if a torch fallback would drop those boundaries. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/35346"}, {"key": "pr#35339", "n": 35339, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix syntax in HfQuantizer docstring", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35339"}, {"key": "pr#35336", "n": 35336, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix hardcoded `float` dtypes in DeBERTa model, which caused multiple RuntimeErrors in `bfloat16`", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35336"}, {"key": "issue#35333", "n": 35333, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "AllAboardBertweetModel", "action": "commented", "comment": "No public weights on the Hub as of today. No implementation or checkpoint is linked either.", "author": "CharC0de", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/35333"}, {"key": "pr#35329", "n": 35329, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Create zero_delay_qkv_benchmark.py", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35329"}, {"key": "pr#35328", "n": 35328, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Create zero-delay-qkv-compression.py", "action": "closed", "comment": "Thanks for sharing! Experimental techniques like this are better published as a Hub repo, blog or cookbook recipe than as files in our docs, so we're closing this.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35328"}, {"key": "pr#35325", "n": 35325, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "avoid un-necessary unwrapping of the model on each train step", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35325"}, {"key": "pr#35324", "n": 35324, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Create Zero-Delay_QKV_Compression.md", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35324"}, {"key": "issue#35317", "n": 35317, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Deepseek v2", "action": "closed", "comment": "This has been added in #36400 and is available on main. Closing, but feel free to reopen if something is still missing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/35317"}, {"key": "issue#35308", "n": 35308, "kind": "issue", "group": "Open issues", "category": "bug", "area": "trainer", "type": "bug", "title": "Multi-GPU training crashes with IterableDataset and different length input (e.g. Next token prediction)", "action": "\u2014", "comment": "", "author": "avishaiElmakies", "idle_days": 96, "models": "", "url": "https://github.com/huggingface/transformers/issues/35308"}, {"key": "pr#35284", "n": 35284, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add config.json explicit change when using custom pipelines.", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35284"}, {"key": "issue#35282", "n": 35282, "kind": "issue", "group": "Open issues", "category": "feature", "area": "gguf", "type": "feature", "title": "Qwen2vl support for GGUF", "action": "commented", "comment": "Qwen2-VL GGUF isn't supported yet, but the text side has a clear template: GGUF loading for `qwen2`/`qwen3` is defined in `GGUF_CONFIG_MAPPING` in `src/transformers/integrations/ggml.py` and the tensor renaming in `src/transformers/modeling_gguf_pytorch_utils.py` (see how `qwen2moe` is remapped there). Adding a `qwen2vl` entry mapped to `Qwen2VLConfig` covers the language model; the vision tower lives in a separate llama.cpp `mmproj` GGUF, so `load_gguf_checkpoint` would also need to accept and map that file. Docs: https://huggingface.co/docs/transformers/main/en/quantization/gguf", "author": "cheald", "idle_days": 0, "models": "qwen2_vl", "url": "https://github.com/huggingface/transformers/issues/35282"}, {"key": "pr#35275", "n": 35275, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "speed up PrefixConstrainedLogitsProcessor", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35275"}, {"key": "issue#35237", "n": 35237, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "[i18n-Chinese] Translating perf_infer_gpu_multi.md to Chinese", "action": "closed", "comment": "This translation was merged in #35271 and lives at `docs/source/zh/perf_infer_gpu_multi.md`. Thanks! Closing; further Chinese translation work can be coordinated in #26803.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/35237"}, {"key": "pr#35209", "n": 35209, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Cleanup: continue the init refactor", "action": "closed", "comment": "Thanks for the PR! TensorFlow and Flax support was removed in v5, so this can't land anymore. Closing, thanks again for the contribution!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35209"}, {"key": "pr#35191", "n": 35191, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Add Flax activation functions and corresponding tests", "action": "closed", "comment": "Thanks for the work here! Flax/JAX support was removed in v5, so we're closing this PR.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35191"}, {"key": "issue#35180", "n": 35180, "kind": "issue", "group": "Open issues", "category": "feature", "area": "models", "type": "feature", "title": "Adding Mamba2ForTokenClassification to Mamba2", "action": "commented", "comment": "Happy to take this: `GenericForTokenClassification` in `src/transformers/modeling_layers.py` makes it a small addition, e.g. `class Mamba2ForTokenClassification(GenericForTokenClassification, Mamba2PreTrainedModel): pass` (see how Llama/Apertus do it), registered in `MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES` in `models/auto/modeling_auto.py`, with docs and the common tests. Guide: https://huggingface.co/docs/transformers/main/en/contributing.", "author": "ryan-minato", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/35180"}, {"key": "pr#35161", "n": 35161, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Use modules_to_not_convert in AQLM", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35161"}, {"key": "issue#35140", "n": 35140, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "FileNotFoundError: Couldn't find a module script at /home/tooko/transformers-course/glue/glue.py.", "action": "commented", "comment": "This isn't a model request: the error comes from `evaluate.load(\"glue\", \"mrpc\")` in the `evaluate` library, not from transformers. Upgrading `evaluate` and `datasets` usually fixes it; otherwise please open an issue on [huggingface/evaluate](https://github.com/huggingface/evaluate).", "author": "AmanoTooko7", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/35140"}, {"key": "issue#35135", "n": 35135, "kind": "issue", "group": "Closed issues", "category": "obsolete", "area": "", "type": "", "title": "[i18n-<languageCode>] Translating docs to Chinese  Translate agents.md into Chinese", "action": "closed", "comment": "`agents.md` was removed from the Transformers docs when agents moved to the `smolagents` library, so there's nothing left to translate here. Chinese translation work is tracked in #26803 if you'd like to pick another page.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/35135"}, {"key": "pr#35128", "n": 35128, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "logic was inverted for passing loss_kwargs to forward pass", "action": "closed", "comment": "Thanks @winglian! The trainer loss_kwargs fix landed in #35121, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35128"}, {"key": "issue#35074", "n": 35074, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "`trainer.evaluate` always creates a new MLFlow run, separate from the one used during `train()`", "action": "\u2014", "comment": "", "author": "nathan-az", "idle_days": 39, "models": "", "url": "https://github.com/huggingface/transformers/issues/35074"}, {"key": "issue#35061", "n": 35061, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "SequenceClassification for all Model types should have the option to add weights in Cross Entropy loss", "action": "closed", "comment": "Closing this one: no activity since 2024-12-03, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/35061"}, {"key": "pr#35040", "n": 35040, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Object detection evaluator", "action": "closed", "comment": "Thanks @qubvel for this refactor of the object-detection evaluation! The `examples/` folder, including `examples/pytorch/object-detection/run_object_detection.py` and `test_pytorch_examples.py`, was removed from the repository in #49357, so there's nothing left for this change to apply to. Closing for that reason, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35040"}, {"key": "pr#35038", "n": 35038, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Fix tokenizer model handling for LLaMA, converting Base64 to SentencePiece (#35037)", "action": "closed", "comment": "Thanks for the fix! In v5 tokenizers were consolidated on the `tokenizers` backend and `LlamaTokenizer` no longer uses the slow SentencePiece path this patches. Closing, but please open an issue if the base64 conversion problem still reproduces on main.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35038"}, {"key": "issue#35033", "n": 35033, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Make it possible to save and evaluate checkpoint on CTRL+C / `KeyboardInterrupt` with Hugging Face Trainer", "action": "closed", "comment": "Trainer now supports just-in-time checkpointing on termination (#41723): set `enable_jit_checkpoint=True` in `TrainingArguments` and a checkpoint is saved when the process receives SIGTERM (e.g. `kill <pid>`, or a scheduler preemption). Closing; if you specifically need Ctrl+C (SIGINT) handled too, a follow-up issue is welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/35033"}, {"key": "pr#35023", "n": 35023, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "adjust beam search early stopping to any criteria as opposed to all", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #35802, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35023"}, {"key": "issue#35022", "n": 35022, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Only Fine-tune the embeddings of the added special tokens", "action": "closed", "comment": "Besides the workaround @zucchini-nlp shared (https://github.com/huggingface/transformers/issues/35022#issuecomment-2511040641), PEFT now supports this directly: `LoraConfig(..., trainable_token_indices=[new_token_ids])` (or `TrainableTokensConfig`) trains only the embeddings of the added tokens. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/35022"}, {"key": "pr#35010", "n": 35010, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "switch from `training_args.bin` `training_args.json`", "action": "closed", "comment": "Thanks for the PR! These examples/templates have since been removed from the repo, so closing this one. Feel free to open a fresh PR against `main` if it's still relevant.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/35010"}, {"key": "issue#35003", "n": 35003, "kind": "issue", "group": "Open issues", "category": "feature", "area": "core / utils", "type": "feature", "title": "Replace all torch.FloatTensor by torch.Tensor", "action": "commented; maintainer pinged", "comment": "@Rocketknight1 could you take a look? Reporter declined to do it (2026-10-20) and asked a maintainer to take it.", "author": "julien-blanchon", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/35003"}, {"key": "issue#34984", "n": 34984, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "No use `no_sync` context manager when using gradient accumulation w/ deepspeed's zero stage 2 or 3 via `accelerate`", "action": "closed", "comment": "Fixed on main in #43919: during gradient accumulation the Trainer no longer enters `accelerator.no_sync` when running with DeepSpeed (it uses a null context), so ZeRO-2/3 with DeepSpeed >= 0.16 works. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/34984"}, {"key": "pr#34980", "n": 34980, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add embedding scaling", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34980"}, {"key": "pr#34970", "n": 34970, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Optimized dola decoding generation function for faster performance", "action": "closed", "comment": "Thanks for the optimization! DoLa decoding was moved out of core into a Hub custom generation method (`transformers-community/dola`), so improvements should go there. Closing here.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34970"}, {"key": "pr#34951", "n": 34951, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Docs updates (russian documentation)", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34951"}, {"key": "pr#34925", "n": 34925, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Allowing dataloader prefetching to work with iterable datasets.", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #47164, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34925"}, {"key": "pr#34918", "n": 34918, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[RWKV] Add RWKV6 model", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34918"}, {"key": "pr#34912", "n": 34912, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "add istft decoder for VITS model", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34912"}, {"key": "issue#34909", "n": 34909, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "[i18n-zh] Translating docs to Chinese Translate bertlogy.md into Chinese", "action": "closed", "comment": "Your translation was merged in #34908 and is live in `docs/source/zh/bertology.md`. Thanks! Further Chinese translation work is coordinated in #26803.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/34909"}, {"key": "pr#34902", "n": 34902, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "ViTImageproc handle pil hw images", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34902"}, {"key": "pr#34900", "n": 34900, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add Relation DETR", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34900"}, {"key": "pr#34899", "n": 34899, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Fix auxiliary_outputs types in loss", "action": "closed", "comment": "Thanks for the PR! This model is deprecated and no longer maintained in the library, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34899"}, {"key": "pr#34844", "n": 34844, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Comments update for better reading", "action": "closed", "comment": "Thanks for the PR! `hubconf.py` has been removed from the repository, so there's nothing left to update. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34844"}, {"key": "issue#34789", "n": 34789, "kind": "issue", "group": "Open issues", "category": "feature", "area": "distributed", "type": "feature", "title": "Add `Tensor Parallel` support for ALL models", "action": "\u2014", "comment": "", "author": "ArthurZucker", "idle_days": 37, "models": "", "url": "https://github.com/huggingface/transformers/issues/34789"}, {"key": "pr#34786", "n": 34786, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Sync table question answering pipeline with Hub spec", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34786"}, {"key": "pr#34772", "n": 34772, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[New Model] X-ALMA", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34772"}, {"key": "pr#34768", "n": 34768, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "skip generation kwargs validation in torch compile", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34768"}, {"key": "pr#34753", "n": 34753, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[DO NOT MERGE] Add pytest-subtests to the dependencies", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34753"}, {"key": "pr#34752", "n": 34752, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Sync DQA pipeline with Hub spec", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34752"}, {"key": "pr#34748", "n": 34748, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[DO NOT MERGE] Testing the new ABI3 tokenizers version.", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34748"}, {"key": "pr#34732", "n": 34732, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Pass callbacks kwarg to study.optimize() in run_hp_search_optuna()", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34732"}, {"key": "pr#34711", "n": 34711, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Add Kannada language documentation to PEGASUS tokenizer class", "action": "closed", "comment": "Thanks for contributing! This file was removed when we consolidated tokenizers in v5, and translations go through the docs (see docs/TRANSLATING.md), so we're closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34711"}, {"key": "pr#34710", "n": 34710, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "SnapKV_Cache support added", "action": "closed", "comment": "Thanks for the SnapKV implementation! We now keep new cache strategies out of the core library and out of modeling files; the recommended path is a Hub `custom_generate` repo (see https://huggingface.co/transformers-community/sink_cache as a template). Closing, but we'd love to see it shared there.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34710"}, {"key": "pr#34708", "n": 34708, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Corrected example code masked_language_modeling.md", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34708"}, {"key": "pr#34707", "n": 34707, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Past Keys Output now working with output router logits", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34707"}, {"key": "pr#34705", "n": 34705, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix: Handle BLIP-2 model output format", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34705"}, {"key": "pr#34691", "n": 34691, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update agents.md ", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34691"}, {"key": "pr#34676", "n": 34676, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "add empty cache before load_best_model to prevent cuda OOM", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34676"}, {"key": "pr#34670", "n": 34670, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[i18n-KO] Translated backbones.md to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34670"}, {"key": "pr#34667", "n": 34667, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Clear unused allocated GPU memory when available GPU memory is low. ", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34667"}, {"key": "pr#34661", "n": 34661, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Fix for TypeError in train_new_from_iterator() in tokenization_utils_fast.py", "action": "closed", "comment": "Thanks @sindhuvahinis for the fix! The `pre_tokenizer is None` crash in `train_new_from_iterator()` was fixed in #35396, and `tokenization_utils_fast.py` itself was later replaced by `tokenization_utils_tokenizers.py` in #40936. Closing as this is covered on `main`.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34661"}, {"key": "issue#34659", "n": 34659, "kind": "issue", "group": "Open issues", "category": "bug", "area": "trainer", "type": "bug", "title": "Neftune computation is probably wrong with packed training", "action": "\u2014", "comment": "", "author": "ccdv-ai", "idle_days": 4, "models": "", "url": "https://github.com/huggingface/transformers/issues/34659"}, {"key": "pr#34655", "n": 34655, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Added explicit error handling to a few modules", "action": "closed", "comment": "Thanks for the PR! This model is deprecated and no longer maintained in the library, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34655"}, {"key": "issue#34654", "n": 34654, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "The support of `Mllama` in AutoModel", "action": "closed", "comment": "`AutoModel` now supports Mllama (`MllamaModel`) since #37033. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/34654"}, {"key": "pr#34652", "n": 34652, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Add EXAONE", "action": "closed", "comment": "Thanks for the PR and your patience! As discussed, the model is architecturally Llama, so converting the checkpoints to Llama format is the path we recommend instead of a new model class. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34652"}, {"key": "issue#34651", "n": 34651, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add EXAONE", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct)\n- Downloads (last 30d): 15.6k \u00b7 Likes: 424 \u00b7 Params: ~7.8B \u00b7 License: other\n- Released: 2024-07 \u00b7 Remote code on the Hub: yes\n- Note: EXAONE 4.0 (`exaone4`) is supported on main; EXAONE 3.0 still runs via remote code.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "lgai-exaone", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/34651"}, {"key": "issue#34649", "n": 34649, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Add functionality for deleting adapter layers in PEFT integration", "action": "closed", "comment": "This has been added in #34650 and is available on main. Closing, but feel free to reopen if something is still missing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/34649"}, {"key": "pr#34636", "n": 34636, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[DO NOT MERGE] Testing dev release of safetensors", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34636"}, {"key": "pr#34624", "n": 34624, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "fix bug in modeling_flax_dinov2 when batch size is more than 1", "action": "closed", "comment": "Thanks for the fix! Flax support was removed in v5, so this file no longer exists. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34624"}, {"key": "pr#34621", "n": 34621, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "T5 Attention fix and tests", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34621"}, {"key": "pr#34620", "n": 34620, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "fix:#34611 by expanding class_pos_embed", "action": "closed", "comment": "Thanks a lot for tackling #34611! The same fix landed in #35138, so closing this as superseded.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34620"}, {"key": "issue#34613", "n": 34613, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "redirect logging output to `stdout` instead of `stderr`", "action": "closed", "comment": "As @Rocketknight1 explained in https://github.com/huggingface/transformers/issues/34613#issuecomment-2602766081, most library logs are warnings/errors and belong on `stderr`. To send them to stdout, use `transformers.utils.logging.disable_default_handler()` then `transformers.utils.logging.add_handler(logging.StreamHandler(sys.stdout))` (both in src/transformers/utils/logging.py). Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/34613"}, {"key": "issue#34612", "n": 34612, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "save `training_args` in another file format ", "action": "commented", "comment": "Still valid: `Trainer` saves `TRAINING_ARGS_NAME = \"training_args.bin\"` with `torch.save` (`src/transformers/trainer.py`). A focused PR that writes `training_args.json` via `TrainingArguments.to_json_string()` (keeping read support for the old `.bin` with a deprecation warning, as @muellerzr suggested in #35010) would be welcome.", "author": "not-lain", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/34612"}, {"key": "pr#34607", "n": 34607, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix spelling of \"your file\"", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34607"}, {"key": "pr#34606", "n": 34606, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix optimum.quanto quantization call in cache_utils", "action": "closed", "comment": "Thanks a lot @w3rew! This fix landed in #34750 (with you as co-author), so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34606"}, {"key": "pr#34598", "n": 34598, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `bloom.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34598"}, {"key": "pr#34576", "n": 34576, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Handle num_items_in_batch in Mistral's forward", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #35875, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34576"}, {"key": "issue#34571", "n": 34571, "kind": "issue", "group": "Open issues", "category": "feature", "area": "generation", "type": "feature", "title": "Feature to configure `stop_strings` in `generation_config.json` or other config files", "action": "\u2014", "comment": "", "author": "suhara", "idle_days": 39, "models": "", "url": "https://github.com/huggingface/transformers/issues/34571"}, {"key": "issue#34555", "n": 34555, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Add more configurations to the class LlavaOnevisionConfig.", "action": "closed", "comment": "As stated in https://github.com/huggingface/transformers/issues/34555#issuecomment-2497576574, different code paths mean a different model, so these switches can't be added as branches in LLaVA-OneVision's forward. Variants can be added as their own model via a modular file (https://huggingface.co/docs/transformers/main/en/modular_transformers). Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/34555"}, {"key": "pr#34547", "n": 34547, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "uniformize kwargs for OneFormer", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34547"}, {"key": "pr#34546", "n": 34546, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "this works right now for rdu", "action": "closed", "comment": "Thanks for the PR! We avoid hardware-specific branches inside the Trainer and model files; device support is best added through accelerate/PyTorch device plumbing, so we're closing this.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34546"}, {"key": "issue#34545", "n": 34545, "kind": "issue", "group": "Open issues", "category": "bug", "area": "processors (image/video/audio)", "type": "bug", "title": "VLMs Processors are not fully consistent in the inputs formats they accept", "action": "\u2014", "comment": "", "author": "yonigozlan", "idle_days": 607, "models": "", "url": "https://github.com/huggingface/transformers/issues/34545"}, {"key": "pr#34542", "n": 34542, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Update Pixtral to use SDPA", "action": "closed", "comment": "Thanks for this! Pixtral now goes through the shared attention interface, which already provides SDPA (and other backends), so hand-wired SDPA calls in the model file aren't needed anymore. Closing, thank you!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34542"}, {"key": "pr#34533", "n": 34533, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[WIP] Clean initialisation of swinlayer class", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34533"}, {"key": "pr#34526", "n": 34526, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[bugfix] incorrect usage of warning_once", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34526"}, {"key": "pr#34519", "n": 34519, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add in free_memory passthrough", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34519"}, {"key": "pr#34508", "n": 34508, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Memory optimization for gpt_bigcode", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34508"}, {"key": "pr#34498", "n": 34498, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "(WIP) cleanup(Mask2Former): Remove unused level_start_index parameter", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34498"}, {"key": "pr#34491", "n": 34491, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix PixtralProcessor to return input IDs for all prompts and images in batch", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34491"}, {"key": "issue#34489", "n": 34489, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Vision (Auto)Processor multiple images finetuning example.", "action": "closed", "comment": "As @qubvel suggested in https://github.com/huggingface/transformers/issues/34489#issuecomment-2450445802, the PaliGemma fine-tuning notebook in https://github.com/merveenoyan/smol-vision can be adapted for multi-image inputs. Examples are no longer hosted in this repo since #49357, so closing here.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/34489"}, {"key": "issue#34488", "n": 34488, "kind": "issue", "group": "Closed issues", "category": "obsolete", "area": "", "type": "", "title": "Please don't kill BetterTransformer \u2014 1.88x faster inference than SDPA", "action": "closed", "comment": "BetterTransformer was removed in v5 (#41367). The supported speedups for encoder models are `attn_implementation=\"sdpa\"`/`\"flash_attention_2\"` with `torch.compile`, padding-free inputs, and Hub kernels (https://huggingface.co/docs/transformers/main/en/kernels, https://huggingface.co/docs/transformers/main/en/attention_interface). If you still see a gap with those, a new issue with numbers on current main is welcome.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/34488"}, {"key": "issue#34463", "n": 34463, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Improve batching speed for Whisper models using VAD based chunking", "action": "closed", "comment": "Closing this one: no activity since 2024-12-19, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/34463"}, {"key": "pr#34451", "n": 34451, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update processing_pixtral.py", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34451"}, {"key": "pr#34439", "n": 34439, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix ValueError by providing labels in the input", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34439"}, {"key": "pr#34431", "n": 34431, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Fix: raise if block_size is too big", "action": "closed", "comment": "We removed the `examples/pytorch` scripts from the repository in #49357, so this change no longer applies. Thanks for the contribution!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34431"}, {"key": "pr#34396", "n": 34396, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Change HybridCache.get_seq_length return type to int", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34396"}, {"key": "pr#34394", "n": 34394, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Fixing strided perplexity calculation for fixed-length models", "action": "closed", "comment": "Thanks for the fix! As discussed, we'll keep the version already on main, so we're closing this.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34394"}, {"key": "pr#34368", "n": 34368, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Added performance.mdx italian translation", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34368"}, {"key": "pr#34364", "n": 34364, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "MsT: chunking the LM-head and MLP to extend sequence length and save memory", "action": "closed", "comment": "Thanks for the proposal and the discussion here, @wdlctc! There wasn't enough community demand to add Mini-Sequence chunking to the library, and the branch is now about two years behind `main`. You can already get chunked/fused LM-head + loss memory savings through the Liger kernel integration (`TrainingArguments(use_liger_kernel=True)`). If you'd still like to propose a generic, opt-in version, please open a new PR against `main` that fits the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets). Closing this one for now, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34364"}, {"key": "pr#34363", "n": 34363, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "#34146 - differential attention implementation for bert", "action": "closed", "comment": "Thanks for implementing this! Since differential attention computes something different and has no official checkpoints, it's best shared as remote code on the Hub (https://huggingface.co/docs/transformers/custom_models) rather than as a branch in BERT.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34363"}, {"key": "pr#34362", "n": 34362, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[MLU] remove deepspeed-mlu dependency", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34362"}, {"key": "pr#34361", "n": 34361, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Modular phi", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34361"}, {"key": "pr#34357", "n": 34357, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Remove coupled type imports from integration_utils.py", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34357"}, {"key": "issue#34344", "n": 34344, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Fast tokenizer for CANINE", "action": "closed", "comment": "Closing this one: no activity since 2024-10-27, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/34344"}, {"key": "issue#34327", "n": 34327, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "LLaMa 3 8B - offloaded_static cache - layer_device_map TypeError", "action": "closed", "comment": "The cache classes were reworked since 4.45: `OffloadedStaticCache` and its `max_batch_size` argument are gone (#43168), and `cache_implementation=\"offloaded_static\"` now builds `StaticCache(offloaded=True)` from `generate` (see `src/transformers/generation/configuration_utils.py`). The `layer_device_map` TypeError shouldn't happen on current versions; please reopen with a repro if it does. Docs: https://huggingface.co/docs/transformers/main/en/kv_cache#cache-offloading", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/34327"}, {"key": "pr#34323", "n": 34323, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Improve EncoderDecoderModel docs", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34323"}, {"key": "pr#34321", "n": 34321, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix PixtralProcessor to return outputs for all examples in a batch", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #34486, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34321"}, {"key": "pr#34318", "n": 34318, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Mask Padded Tokens In Aspect Ratio Mask Preparation", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34318"}, {"key": "pr#34315", "n": 34315, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "fix: unboundlocalerror when trying to retrieve attention feature maps using Qwen2FlashAttention2().forward()", "action": "closed", "comment": "Thanks for the fix! Qwen2 no longer has a dedicated FlashAttention2 class: attention now goes through the shared attention interface, so this code path is gone. Closing, but please open an issue if you still see the problem on main.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34315"}, {"key": "issue#34313", "n": 34313, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "speed up whisper compile time", "action": "closed", "comment": "As @eustlb explained (https://github.com/huggingface/transformers/issues/34313#issuecomment-2434791817), `reduce-overhead`/`max-autotune` capture one CUDA graph per sequence length, hence the warm-up. Closing; see https://huggingface.co/docs/transformers/main/en/perf_torch_compile for ways to reduce compile time.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/34313"}, {"key": "pr#34290", "n": 34290, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Sync VQA pipeline with huggingface_hub spec", "action": "closed", "comment": "Thanks @Rocketknight1 for working through the hub spec sync! Since then, the `visual-question-answering` pipeline and its tests were removed in the v5 pipeline cleanup (#43325), and users are now pointed to the `image-text-to-text` pipeline. There's nothing left on `main` to sync, so I'm closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34290"}, {"key": "pr#34246", "n": 34246, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Fix handling of Sequence post-processors in train_new_from_iterator", "action": "closed", "comment": "Thanks a lot for the fix and the test! Since then, `src/transformers/tokenization_utils_fast.py` was removed in #40936 and `train_new_from_iterator` now lives in `src/transformers/tokenization_utils_tokenizers.py`, so this branch can no longer be merged. As far as we can tell that code still only rewrites top-level `special_tokens`, so if you still run into the `Sequence` post-processor issue, we'd be happy to review a rebased PR against `main`. Closing this one in the meantime.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34246"}, {"key": "issue#34238", "n": 34238, "kind": "issue", "group": "Open issues", "category": "feature", "area": "gguf", "type": "feature", "title": "GGUF support for BERT architecture", "action": "commented", "comment": "GGUF loading for BERT is still not supported, but the path is clear: add a `\"bert\"` entry to `GGUF_CONFIG_MAPPING` (and a WordPiece tokenizer mapping) in `src/transformers/integrations/ggml.py`, plus a `TensorProcessor` in `src/transformers/modeling_gguf_pytorch_utils.py` if tensors need renaming or reshaping (see `T5TensorProcessor`/`GPT2TensorProcessor` as examples), and a test in `tests/quantization/ggml/test_ggml.py`. Docs: https://huggingface.co/docs/transformers/main/en/quantization/gguf. A PR with your mapping would be welcome!", "author": "Dimmension", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/34238"}, {"key": "pr#34233", "n": 34233, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix loss scaling only when compute_loss_func is used", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34233"}, {"key": "pr#34205", "n": 34205, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Support AcceleratorConfig.use_stateful_dataloader in Trainer ", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34205"}, {"key": "pr#34183", "n": 34183, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update stale.py", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34183"}, {"key": "pr#34182", "n": 34182, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Update language_modeling.py", "action": "closed", "comment": "Thanks for the contribution! The legacy `data/datasets` module has been removed from transformers (datasets loading now goes through the `datasets` library), so this change can't apply. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34182"}, {"key": "pr#34179", "n": 34179, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Adds the Molmo vision-language models", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34179"}, {"key": "pr#34151", "n": 34151, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Use safetensors from PR", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34151"}, {"key": "issue#34149", "n": 34149, "kind": "issue", "group": "Open issues", "category": "feature", "area": "models", "type": "feature", "title": "TimeSeriesTransformerModel: add no time lags option", "action": "commented; maintainer pinged", "comment": "@kashif could you take a look? It has been waiting for a maintainer review.", "author": "chiara-fb", "idle_days": -1, "models": "time_series_transformer", "url": "https://github.com/huggingface/transformers/issues/34149"}, {"key": "issue#34146", "n": 34146, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Differential Attention implementation for BERT.", "action": "closed", "comment": "Differential attention is already available via DiffLlama (`src/transformers/models/diffllama/`). Following our criteria for new architectures (https://github.com/huggingface/transformers/pull/36952#pullrequestreview-2724807497), an encoder variant would make sense once pretrained weights exist, ideally as a modular model reusing DiffLlama's attention. Closing for now.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/34146"}, {"key": "pr#34086", "n": 34086, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "add qwen2vl for sequence classification", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34086"}, {"key": "pr#34075", "n": 34075, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add MinGRU", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34075"}, {"key": "pr#34064", "n": 34064, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Sync token-classification pipeline with Hub spec", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34064"}, {"key": "pr#34057", "n": 34057, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Object detection training for Owl-vit/Owlv2", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34057"}, {"key": "pr#34050", "n": 34050, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "lr inconsistency", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34050"}, {"key": "issue#34019", "n": 34019, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Add Loss Functions for QFormer Training in BLIP-2 Model (ITC, ITM, and ITG)", "action": "\u2014", "comment": "", "author": "thisisiron", "idle_days": 64, "models": "blip_2", "url": "https://github.com/huggingface/transformers/issues/34019"}, {"key": "pr#34009", "n": 34009, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Enabled Flash Attention for PaliGemma models", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #37576, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/34009"}, {"key": "issue#33985", "n": 33985, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "../aten/src/ATen/native/cuda/Indexing.cu:1289: indexSelectLargeIndex: block: [267,0,0], thread: [25,0,0] Assertion `srcIndex < srcSelectDimSize` failed.", "action": "closed", "comment": "Closing this one: no activity since 2025-10-11, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/33985"}, {"key": "issue#33983", "n": 33983, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Enhancing RoBERTa Robustness through Adversarial Training", "action": "closed", "comment": "Closing this one: no activity since 2024-10-06, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/33983"}, {"key": "pr#33975", "n": 33975, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `image_text_to_text.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33975"}, {"key": "pr#33971", "n": 33971, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `compressed_tensors.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33971"}, {"key": "pr#33970", "n": 33970, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `quantization/overview.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33970"}, {"key": "pr#33962", "n": 33962, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add Molmo (7B-D, 7B-O)", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33962"}, {"key": "pr#33956", "n": 33956, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add training support to Meta's EnCodec", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33956"}, {"key": "issue#33949", "n": 33949, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Request for Iterative Generation in Pipeline (e.g., LLaMA model)", "action": "closed", "comment": "This is supported: `pipe(messages, num_return_sequences=n, do_sample=True, max_new_tokens=...)` returns n completions from one batched `generate` call (example in the `TextGenerationPipeline` docstring). For an OpenAI-style API with `n`, `transformers serve` exposes one: https://huggingface.co/docs/transformers/main/en/serve-cli/serving", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/33949"}, {"key": "issue#33948", "n": 33948, "kind": "issue", "group": "Open issues", "category": "tracker / RFC", "area": "processors (image/video/audio)", "type": "tracker / RFC", "title": "Major VLM tracker (standardize the API)", "action": "\u2014", "comment": "", "author": "zucchini-nlp", "idle_days": 746, "models": "", "url": "https://github.com/huggingface/transformers/issues/33948"}, {"key": "issue#33947", "n": 33947, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Centralized Page for Domain Whitelisting (Hugging Face Models)", "action": "closed", "comment": "Thanks for raising this! The Hub team listed the current CDN domains above (https://github.com/huggingface/transformers/issues/33947#issuecomment-2402109441). This is Hub infrastructure rather than the Transformers library, so please follow up on the huggingface_hub repo if it changes. Closing here.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/33947"}, {"key": "issue#33916", "n": 33916, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "[Modular Transformers] Request for comments", "action": "closed", "comment": "Modular Transformers is now the standard way of adding models, documented in `modular_transformers.md`. Thanks to everyone for the feedback, closing this RFC!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/33916"}, {"key": "pr#33915", "n": 33915, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `processors.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33915"}, {"key": "pr#33914", "n": 33914, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `image_processor.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33914"}, {"key": "pr#33913", "n": 33913, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Fix tied weigths isuue", "action": "closed", "comment": "Thanks for this fix! An updated version landed in #37031, so closing this one as superseded.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33913"}, {"key": "pr#33908", "n": 33908, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add gguf for granite", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33908"}, {"key": "issue#33905", "n": 33905, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Implement LlamaGen for Image Generation", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [FoundationVision/LlamaGen](https://huggingface.co/FoundationVision/LlamaGen)\n- Downloads (last 30d): 0 \u00b7 Likes: 32 \u00b7 Params: 111M-3.1B (checkpoint collection, raw .pt files) \u00b7 License: mit\n- Released: 2024-06 \u00b7 Remote code on the Hub: no\n- Related models already in transformers: `chameleon`, `janus`, `emu3`.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "ighoshsubho", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/33905"}, {"key": "pr#33892", "n": 33892, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "docs", "type": "docs", "title": "\ud83c\udf10 [i18n-KO] Translated ` deepspeed.md` to Korean", "action": "\u2014", "comment": "", "author": "ahnjj", "idle_days": 365, "models": "", "url": "https://github.com/huggingface/transformers/pull/33892"}, {"key": "pr#33881", "n": 33881, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[feat] add possibility to give kwargs handler to accelerator in trainer", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33881"}, {"key": "pr#33875", "n": 33875, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix update stateful callbacks state to save checkpoint", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33875"}, {"key": "pr#33868", "n": 33868, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Fix: Correct Documentation Errors in Transformers", "action": "closed", "comment": "Thanks @tharun977 for taking the time to look for doc errors! This PR targets `onnx_gpt2_io_definition`, an old 2022 feature branch whose own PR (#16754) was closed, so it can't be merged as is; `.coveragerc` was also folded into `pyproject.toml` in #32650. If you spot specific documentation errors, a small focused PR against `main` is very welcome. Closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33868"}, {"key": "pr#33867", "n": 33867, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `video_text_to_text.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33867"}, {"key": "pr#33854", "n": 33854, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "improved readability, error handling; installed pip and albumentations", "action": "closed", "comment": "Thanks a lot for taking the time to improve this script, @MelissaTanksley! The `examples/pytorch/` scripts, including `run_object_detection.py`, have since been removed from the repo in #49357 because they were no longer maintained. There is nothing left on `main` for this change to apply to, so I'm closing this PR. Thanks again for contributing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33854"}, {"key": "pr#33831", "n": 33831, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Mllama flash attention 2", "action": "closed", "comment": "Thanks for working on this! Models now dispatch attention through the shared attention interface (`ALL_ATTENTION_FUNCTIONS`), and Mllama already supports flash attention that way, so per-model FlashAttention2 classes are no longer needed. Closing, thank you!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33831"}, {"key": "issue#33810", "n": 33810, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Improve image processing time", "action": "closed", "comment": "This is done: image processors now have a torchvision backend that is GPU-accelerated and much faster than PIL (#28847), and since #43514 it is the default backend for all models (see https://huggingface.co/docs/transformers/main/en/image_processors). Passing `device=\"cuda\"` to the processor moves preprocessing to the GPU. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/33810"}, {"key": "pr#33800", "n": 33800, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Idefics3: Error/ Bugs Found when no images token input", "action": "closed", "comment": "Thanks for the PR! TensorFlow and Flax support was removed in v5, so this can't land anymore. Closing, thanks again for the contribution!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33800"}, {"key": "pr#33792", "n": 33792, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `model_doc/bert.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33792"}, {"key": "issue#33755", "n": 33755, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Tuning generation_config in Trainer hyperparameter_search (Optuna backend)", "action": "commented", "comment": "Thanks @jedrzejpolaczek for the analysis, go for it! The fix belongs in `Trainer._hp_search_setup` (`src/transformers/trainer.py`), which should skip the `type(old_attr)(value)` cast for `GenerationConfig` values (or build it via `GenerationConfig.from_dict`), and in `Seq2SeqTrainer` so the trial's `generation_config` is picked up; please add a test in `tests/trainer/`.", "author": "ribesstefano", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/33755"}, {"key": "pr#33751", "n": 33751, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Load a pretrainedfast tokenizer if fast=true and tokenizer.json exists", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33751"}, {"key": "pr#33738", "n": 33738, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix MoE tensor reshape", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33738"}, {"key": "pr#33732", "n": 33732, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[whisper] added dropping of attention weights after DTW calculations related to word timestamps if these weights are not requested in the output", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33732"}, {"key": "pr#33723", "n": 33723, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Added Ukrainian translations", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33723"}, {"key": "pr#33720", "n": 33720, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[i18n-ZH] Sync and Localize Latest English Readme to Simplified Chinese Readme", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33720"}, {"key": "pr#33716", "n": 33716, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "add sdpa and flash_attention2 support to speech2text", "action": "closed", "comment": "Thanks for the PR! TensorFlow and Flax support was removed in v5, so this can't land anymore. Closing, thanks again for the contribution!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33716"}, {"key": "pr#33714", "n": 33714, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Stable dropout show drop prob in model print", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33714"}, {"key": "issue#33710", "n": 33710, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add support for Molmo", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [allenai/Molmo-7B-D-0924](https://huggingface.co/allenai/Molmo-7B-D-0924), [allenai/Molmo-7B-O-0924](https://huggingface.co/allenai/Molmo-7B-O-0924)\n- Downloads (last 30d): 19,176 \u00b7 Likes: 568 \u00b7 Params: ~8.0B \u00b7 License: apache-2.0\n- Released: 2024-09 \u00b7 Remote code on the Hub: yes\n- Open PR: #43448 (also #33962)\n- The newer Molmo2 is already supported as `molmo2` (added in #43451), which should make a modular Molmo (v1) much shorter.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "fakerybakery", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/33710"}, {"key": "pr#33699", "n": 33699, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix paligemma `eager` vs `sdpa` + image transforms test fixup", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33699"}, {"key": "issue#33698", "n": 33698, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Refactor `output_hidden_states` to allow index selection", "action": "closed", "comment": "Closing as not planned: the accompanying PR #33705 was closed, and outputs are now collected through hooks (`_can_record_outputs` / `OutputRecorder` in `src/transformers/utils/output_capturing.py`) rather than per-model flags, so we won't change `output_hidden_states` to accept an index. VLMs select the layer via `vision_feature_layer`.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/33698"}, {"key": "issue#33669", "n": 33669, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add support for OmDet-Turbo multi-gpu inference with DataParallel.", "action": "\u2014", "comment": "", "author": "yonigozlan", "idle_days": 756, "models": "omdet_turbo", "url": "https://github.com/huggingface/transformers/issues/33669"}, {"key": "issue#33666", "n": 33666, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Qwen2-VL: Multi-GPU training", "action": "closed", "comment": "Closing this one: no activity since 2026-02-06. If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/33666"}, {"key": "issue#33664", "n": 33664, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Object detection training/fine-tuning for Owl-vit/Owlv2", "action": "\u2014", "comment": "", "author": "catalys1", "idle_days": 111, "models": "owlv2", "url": "https://github.com/huggingface/transformers/issues/33664"}, {"key": "issue#33659", "n": 33659, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "support schedule sampling", "action": "closed", "comment": "Closing this one: no activity since 2024-09-23, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/33659"}, {"key": "pr#33653", "n": 33653, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Remove attribute that is no longer used", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33653"}, {"key": "issue#33642", "n": 33642, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Enable changing the loss function by making the hard-coded `loss_fct` an attribute of `BertForTokenClassification`.", "action": "closed", "comment": "You can now do this without touching the model: pass `compute_loss_func=` to `Trainer` (a callable receiving `outputs, labels, num_items_in_batch`), e.g. to apply a weighted `CrossEntropyLoss` on `outputs.logits`. See the `compute_loss_func` docs in `src/transformers/trainer.py` and https://huggingface.co/docs/transformers/main/en/trainer", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/33642"}, {"key": "pr#33640", "n": 33640, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix the bug with loading DeepSpeed MoE models", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33640"}, {"key": "pr#33638", "n": 33638, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Updated few changes to the code and updated the files which had few typos in it.", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33638"}, {"key": "pr#33588", "n": 33588, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `quantization/contribute.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33588"}, {"key": "pr#33576", "n": 33576, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Remove repeated test_model_from_pretrained, test_model, test_config", "action": "closed", "comment": "Thanks for the PR! TensorFlow and Flax support was removed in v5, so this can't land anymore. Closing, thanks again for the contribution!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33576"}, {"key": "pr#33568", "n": 33568, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `llava_next.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33568"}, {"key": "pr#33563", "n": 33563, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "[Draft] Add counting tokens", "action": "closed", "comment": "Thanks for this! Agents now live in the smolagents library rather than transformers, so we're closing this; a similar feature would be welcome there.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33563"}, {"key": "pr#33551", "n": 33551, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Adding MSELoss for regression tasks in HubertForSequenceClassification model", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33551"}, {"key": "issue#33548", "n": 33548, "kind": "issue", "group": "Open issues", "category": "feature", "area": "generation", "type": "feature", "title": "Pass input_ids to _update_model_kwargs_for_generation", "action": "commented", "comment": "`GenerationMixin._update_model_kwargs_for_generation` still doesn't receive `input_ids` (`src/transformers/generation/utils.py`). Two options today: write your own decoding loop as a Hub custom generation method (https://huggingface.co/docs/transformers/main/en/generation_strategies#custom-generation-methods), or recompute the index in your model's `prepare_inputs_for_generation`. A small PR passing `input_ids` as an optional kwarg at the call sites would need a generation maintainer's sign-off first.", "author": "alex-hh", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/33548"}, {"key": "pr#33544", "n": 33544, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Correctly support resuming from checkpoint with a dataset without length", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33544"}, {"key": "pr#33537", "n": 33537, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `decision transformer.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33537"}, {"key": "pr#33512", "n": 33512, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[WIP] [Whisper] Fix generate and tokenizer behavior with added tokens", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33512"}, {"key": "pr#33505", "n": 33505, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Run nightly ci test new runner", "action": "closed", "comment": "Thanks for this! Both `.github/workflows/model_jobs.yml` and `.github/workflows/self-scheduled.yml` were removed in #47360, when the daily CI moved to the transformers-ci workflow, so this runner change no longer applies. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33505"}, {"key": "pr#33501", "n": 33501, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix llava next incorrect padding size", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33501"}, {"key": "pr#33496", "n": 33496, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix `inputs` deprecation warning in automatic_speech_recognition.py", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33496"}, {"key": "pr#33495", "n": 33495, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "removed redundant creation of causal mask when attention mask is already 4D", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33495"}, {"key": "issue#33492", "n": 33492, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "add `private` parameter to trainer.push_to_hub", "action": "commented", "comment": "Still open for contributions: `Trainer.push_to_hub` in `src/transformers/trainer.py` has no `private` argument, and visibility is taken from `args.hub_private_repo` when `init_hf_repo` creates the repo. A small PR adding `private: bool | None = None` that overrides `hub_private_repo` at creation (see the closed attempt #33511 for a starting point) would be welcome.", "author": "not-lain", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/33492"}, {"key": "pr#33419", "n": 33419, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Support to use adam_mini from installing directly", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33419"}, {"key": "issue#33396", "n": 33396, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "A Trainer subclass for Decoder-Only LM with generation in evaluate()", "action": "closed", "comment": "We don't plan to add another `Trainer` subclass for this (#32346 was closed). The usual approach is to override `prediction_step` (or add a callback) to call `model.generate` on prompt-only eval inputs, as `Seq2SeqTrainer.prediction_step` in `src/transformers/trainer_seq2seq.py` does; TRL's trainers also support generation-based evaluation for decoder-only models.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/33396"}, {"key": "pr#33394", "n": 33394, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix flash attention comment", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33394"}, {"key": "pr#33377", "n": 33377, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Adding pruning integration to transformer", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33377"}, {"key": "pr#33370", "n": 33370, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix exception AttributeError for missing attribute is_zero3 from HfDeepSpeedConfig object", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33370"}, {"key": "pr#33367", "n": 33367, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Enhanced Model Configuration Testing: An Inclusive Reference Manual for Enhanced Error Management, Reporting and Enhanced Program Code ", "action": "closed", "comment": "Thanks for taking the time to open this! The PR adds a standalone `config_common.py` at the repository root, which isn't part of the library or test suite (config tests live in `tests/test_configuration_common.py` on `main`), and the branch is now ~2 years behind. Closing this one, but if you have a concrete improvement to the config tests, feel free to open a focused PR against `main`.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33367"}, {"key": "pr#33348", "n": 33348, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "AI\u2013Enhanced Benchmarking: New and enhanced features of Open Source and Integration", "action": "closed; comment corrected", "comment": "Closing this one: no activity since 2024-09-10, 1 of the 1 files it changes no longer exist on `main`, it was written against v4 (v5 changed the codebase a lot). Thanks for the contribution, and feel free to open a fresh PR against `main` if it's still relevant!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33348"}, {"key": "pr#33310", "n": 33310, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Enhanced Logging for Seq2SeqTrainingArguments Initialization", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33310"}, {"key": "issue#33289", "n": 33289, "kind": "issue", "group": "Open issues", "category": "feature", "area": "models", "type": "feature", "title": "Qwen2-VL Doesn't Execute on TPUs", "action": "commented", "comment": "As discussed, the Qwen2-VL/Qwen2.5-VL vision path materializes data-dependent shapes (e.g. `grid_thw`-driven `cu_seqlens`, window indices, `repeat_interleave`), which triggers XLA recompilations. Contributions making these static-shape friendly in `modeling_qwen2_vl.py` and `modular_qwen2_5_vl.py` (then `make fix-repo`) are welcome; they would also help `torch.compile`. See https://huggingface.co/docs/transformers/main/en/google_tpu.", "author": "radna0", "idle_days": 0, "models": "qwen2_vl", "url": "https://github.com/huggingface/transformers/issues/33289"}, {"key": "pr#33284", "n": 33284, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add support for JAIS Family of Arabic Language Models", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33284"}, {"key": "pr#33281", "n": 33281, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "scale gradient accumulation steps with train batch size to keep effective batch size about the same", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33281"}, {"key": "issue#33260", "n": 33260, "kind": "issue", "group": "Open issues", "category": "feature", "area": "gguf", "type": "feature", "title": "Community contribution: Adding GGUF support for more architectures", "action": "\u2014", "comment": "", "author": "SunMarc", "idle_days": 39, "models": "", "url": "https://github.com/huggingface/transformers/issues/33260"}, {"key": "pr#33223", "n": 33223, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Agents - Delineate sequences with easier tags", "action": "closed", "comment": "Thanks for the contribution! Agents have moved out of transformers into the smolagents library, so we're closing this here; feel free to propose it on huggingface/smolagents.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33223"}, {"key": "pr#33217", "n": 33217, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add propainter", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33217"}, {"key": "pr#33192", "n": 33192, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Add warning and info  message for beta and gamma parameters", "action": "closed", "comment": "Thanks for the PR! `examples/research_projects` was removed and v5 rewrote weight loading so key renames now go through `conversion_mapping.py`, so this no longer applies. Closing, thank you.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33192"}, {"key": "issue#33184", "n": 33184, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Is it possible to infer the model separately through encoder.onnx and decoder.onnx", "action": "closed; comment corrected", "comment": "As @xenova answered (https://github.com/huggingface/transformers/issues/33184#issuecomment-2318197240), yes: encoder.onnx and decoder.onnx can be run separately, feeding the encoder outputs to the decoder. For exporting, see `transformers.exporters` (https://huggingface.co/docs/transformers/main/en/exporters). Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/33184"}, {"key": "pr#33174", "n": 33174, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[Zero-shot image classification pipeline] Remove tokenizer_kwargs", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33174"}, {"key": "pr#33167", "n": 33167, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add Sapiens model", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33167"}, {"key": "issue#33165", "n": 33165, "kind": "issue", "group": "Open issues", "category": "bug", "area": "models", "type": "bug", "title": "Tensor size mismatch when trying to run RT-DETR on multiple gpus", "action": "\u2014", "comment": "", "author": "mesllo-bc", "idle_days": 533, "models": "rt_detr", "url": "https://github.com/huggingface/transformers/issues/33165"}, {"key": "pr#33111", "n": 33111, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "[Backbone] Remove out_features everywhere", "action": "closed", "comment": "Thanks for the PR! This model is deprecated and no longer maintained in the library, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33111"}, {"key": "pr#33108", "n": 33108, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Enable multi-device for Segformer", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #41693, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33108"}, {"key": "pr#33102", "n": 33102, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "updates_readmedocs", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33102"}, {"key": "issue#33095", "n": 33095, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "llama3 position_ids error with left padding", "action": "closed", "comment": "This is intended: when calling the model directly with left padding, pass `position_ids` (e.g. `attention_mask.long().cumsum(-1) - 1` clamped at 0). `generate()` already does this for you (`_prepare_position_ids_for_generation` in `src/transformers/generation/utils.py`), and for padding-free training `DataCollatorWithFlattening` provides them: https://huggingface.co/docs/transformers/main/en/padding_free", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/33095"}, {"key": "pr#33085", "n": 33085, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Create the FSDP Plugin in the trainer explicitly rather wait for it to be setup by accelerate", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33085"}, {"key": "pr#33035", "n": 33035, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[i18n-ar] Translated file : `docs/source/ar/fsdp.md` into Arabic", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33035"}, {"key": "pr#33033", "n": 33033, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[i18n-ar] Translated file : `docs/source/ar/deepspeed.md` into Arabic", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33033"}, {"key": "pr#33032", "n": 33032, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[i18n-ar] Translated file : `docs/source/ar/debugging.md` into Arabic", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33032"}, {"key": "pr#33028", "n": 33028, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[i18n-ar] Translated file : `docs/source/ar/contributing.md` into Arabic", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33028"}, {"key": "pr#33025", "n": 33025, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[i18n-ar] Translated file : `docs/source/ar/big_models.md` into Arabic", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33025"}, {"key": "pr#33019", "n": 33019, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[i18n-ar] Translated file : `docs/source/ar/add_new_pipeline.md` into Arabic", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33019"}, {"key": "pr#33018", "n": 33018, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[i18n-ar] Translated file : `docs/source/ar/add_new_model.md` into Arabic", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/33018"}, {"key": "pr#32938", "n": 32938, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Convert new SigLIP checkpoints", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32938"}, {"key": "issue#32926", "n": 32926, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Auto model & pipeline for image-text-to-image-text models", "action": "closed", "comment": "Following the discussion here, this landed as the `any-to-any` pipeline plus `AutoModelForMultimodalLM` (#40884), which supports interleaved inputs and a `generation_mode` for models that can emit images (e.g. Janus). See https://huggingface.co/docs/transformers/main/en/tasks/any_to_any. Closing; new models can register there.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/32926"}, {"key": "pr#32925", "n": 32925, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Update convert.py", "action": "closed", "comment": "Thanks for the improvement! The old `convert` command for TensorFlow checkpoints was removed with TF support in v5, so this file no longer exists. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32925"}, {"key": "pr#32920", "n": 32920, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix: Issue #32917, fixed _prune_heads in T5Model ", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32920"}, {"key": "issue#32909", "n": 32909, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Trainer.model.push_to_hub() does not allow a private repository flag", "action": "closed", "comment": "As @asartran confirmed (https://github.com/huggingface/transformers/issues/32909#issuecomment-2315632642), `hub_private_repo` / `private=` only apply when the repo is created; to change the visibility of an existing repo use `huggingface_hub.update_repo_settings(repo_id, private=True)` or the repo settings page. Closing as expected behavior.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/32909"}, {"key": "pr#32906", "n": 32906, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Uniform kwargs for processors of audio-text models", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32906"}, {"key": "issue#32902", "n": 32902, "kind": "issue", "group": "Open issues", "category": "feature", "area": "generation", "type": "feature", "title": "StoppingCriteria for Repetition", "action": "\u2014", "comment": "", "author": "Vipitis", "idle_days": 100, "models": "", "url": "https://github.com/huggingface/transformers/issues/32902"}, {"key": "pr#32893", "n": 32893, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Make qwen2 `attention_qkv_bias` optional", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32893"}, {"key": "issue#32892", "n": 32892, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Optional `bias` for qwen2 model", "action": "closed", "comment": "As answered on #32893, we keep the bias hard-coded since it defines the Qwen2 architecture. A Qwen2-style model without qkv bias is a Llama architecture: use `LlamaConfig(attention_bias=False, ...)`, or define your variant in a modular file (https://huggingface.co/docs/transformers/main/en/modular_transformers).", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/32892"}, {"key": "issue#32881", "n": 32881, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "`truncate_dim` on `BertModel`", "action": "closed", "comment": "As noted in https://github.com/huggingface/transformers/issues/32881#issuecomment-2296869723, this custom use case is best discussed on the forum (the thread is at https://discuss.huggingface.co/). With `BertModel` you can truncate the embedding yourself, e.g. `outputs.last_hidden_state[:, 0, :dim]`. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/32881"}, {"key": "issue#32873", "n": 32873, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "How to use \u3010examples/pytorch/contrastive-image-text\u3011 to inter  inference", "action": "closed", "comment": "The `examples/pytorch` scripts were removed in #49357, so closing this one. Feel free to open a new issue if you hit a problem with the library itself!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/32873"}, {"key": "pr#32869", "n": 32869, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update modeling_seamless_m4t.py SeamlessM4TForSpeechToText needs text_decoder it should not be in _keys_to_ignore_on_load_missing.", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32869"}, {"key": "pr#32859", "n": 32859, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix assertion and value errors in sam-vit-h convertion script", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32859"}, {"key": "issue#32846", "n": 32846, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "split head_dim from hidden_size for llama like gemma or mistral", "action": "closed", "comment": "`LlamaConfig` now accepts a separate `head_dim` (added in #32857). Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/32846"}, {"key": "issue#32839", "n": 32839, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Q-GaLore Support", "action": "closed", "comment": "The Q-GaLore integration PR (#31936) was closed without merging, and there has been no further demand since. You can still use it with `Trainer` by passing the optimizer from the Q-GaLore package via `Trainer(optimizers=(...))` or `optimizer_cls_and_kwargs`. Closing as not planned.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/32839"}, {"key": "pr#32831", "n": 32831, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[Docs] Update resources", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32831"}, {"key": "pr#32666", "n": 32666, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Use Self type for PreTrainedModel.from_pretrained", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32666"}, {"key": "pr#32658", "n": 32658, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix RT-DETR optional losses", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32658"}, {"key": "pr#32621", "n": 32621, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Device selection for Qwen2Audio", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32621"}, {"key": "pr#32620", "n": 32620, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Fix of type mis-match error in make_log_bucket_position", "action": "closed", "comment": "Thanks for the fix! TensorFlow models were removed in v5, so these files no longer exist and we're closing this.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32620"}, {"key": "pr#32619", "n": 32619, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add warm up in evaluation_loop to get more accurate performance data", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32619"}, {"key": "pr#32602", "n": 32602, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "bump tokenizers to support latest release", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32602"}, {"key": "pr#32579", "n": 32579, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "RT-DETR postprocessing bug", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32579"}, {"key": "pr#32574", "n": 32574, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "cleaned get_test_pipeline() unused variables from tests/pipelines files", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32574"}, {"key": "pr#32567", "n": 32567, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "use new home for AnyPrecisionAdamW", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32567"}, {"key": "pr#32556", "n": 32556, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add reorder_cache class attribute For Starcoder", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32556"}, {"key": "pr#32546", "n": 32546, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Quickfix generate tests", "action": "closed", "comment": "Thanks! Flax support was removed in v5, so most of the files touched here no longer exist. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32546"}, {"key": "pr#32517", "n": 32517, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Phi3FlashAttention2 call to rotary embeddings bugfix", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32517"}, {"key": "pr#32490", "n": 32490, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix zero shot detection pipeline", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32490"}, {"key": "issue#32488", "n": 32488, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "SAM can't process batches of nonhomogenous-count of bounding-boxes per image", "action": "commented", "comment": "Still unsupported on main: `SamProcessor._check_and_preprocess_points` in `src/transformers/models/sam/processing_sam.py` converts `input_boxes` with `np.array`, which fails on ragged lists (SAM2's processor raises explicitly). The proposal discussed above is the way to go: pad boxes per image and return a box mask, mirroring how points are padded with `point_pad_value`, then ignore padded boxes in `SamPromptEncoder` (`modeling_sam.py`) and in the output masks. A PR with tests in `tests/models/sam/` would be welcome.", "author": "royvelich", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/32488"}, {"key": "pr#32483", "n": 32483, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Adding new zero-shot examples", "action": "closed", "comment": "We removed the `examples/pytorch` scripts from the repository in #49357, so we can't accept new example scripts there. Thanks for the work; publishing them as a standalone repo or Hub Space would be a great home.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32483"}, {"key": "issue#32478", "n": 32478, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "[ I built it! ] Server application with on-the-fly quantization to serve up HF-Hub models for back-and-forth request-response inferencing   ", "action": "closed; comment corrected", "comment": "Thanks for sharing your project! The library now ships `transformers serve` (`src/transformers/cli/serve.py`), a local OpenAI-compatible server for Hub models, see https://huggingface.co/docs/transformers/main/en/serve-cli/serving.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/32478"}, {"key": "issue#32459", "n": 32459, "kind": "issue", "group": "Closed issues", "category": "obsolete", "area": "", "type": "", "title": "Zero-shot finetuning examples", "action": "closed", "comment": "The prerequisites landed (Grounding DINO loss fix #31828 and processor uniformization #31964), but example scripts are no longer hosted in this repo since #49357, so a new `examples/pytorch/zero-shot` script can't be added here. A blog post or notebook on the Hub would be the best home for it. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/32459"}, {"key": "pr#32439", "n": 32439, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[WiP] Add buf_init_callbacks for dist inference = auto generate rope embeddings after meta/fake tracing", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32439"}, {"key": "issue#32435", "n": 32435, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "[i18n-ar] Translating docs to Arabic (\u0627\u0644\u0639\u0631\u0628\u064a\u0629)", "action": "closed", "comment": "As @stevhliu pointed out in https://github.com/huggingface/transformers/issues/32435#issuecomment-2270013335, Arabic translation progress is tracked in #29045, so closing this one as a duplicate. New Arabic translation PRs are still welcome!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/32435"}, {"key": "pr#32428", "n": 32428, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Register buffer init callbacks in llama", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32428"}, {"key": "issue#32411", "n": 32411, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "[i18n-<languageCode>] Translating docs to <languageName>\u4e2d\u6587", "action": "closed", "comment": "Closing as a duplicate: Chinese translation work is coordinated in #26803, please continue there.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/32411"}, {"key": "pr#32396", "n": 32396, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Avoid decoding the entire token cache for every new token during text streaming", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32396"}, {"key": "pr#32363", "n": 32363, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Remove irrelevant comment in `_update_causal_mask`", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32363"}, {"key": "pr#32358", "n": 32358, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-KO] Translated `<hqq>.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32358"}, {"key": "pr#32355", "n": 32355, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix: MPS device is not being set on current implementation ", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32355"}, {"key": "pr#32319", "n": 32319, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Filtering logits of non tensors", "action": "closed", "comment": "Thanks for the PR! TensorFlow and Flax support was removed in v5, so this can't land anymore. Closing, thanks again for the contribution!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32319"}, {"key": "pr#32305", "n": 32305, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "DeepSpeed sequence parallelism (aka Ulysses) integration with HF transformer", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32305"}, {"key": "pr#32291", "n": 32291, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix chinese clip", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32291"}, {"key": "pr#32284", "n": 32284, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[i18n-KO] Translated `aqlm.md` to Korean", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32284"}, {"key": "issue#32266", "n": 32266, "kind": "issue", "group": "Open issues", "category": "feature", "area": "gguf", "type": "feature", "title": "Support loading shard GGUF models", "action": "commented", "comment": "Still single-file only: `gguf_file` is resolved to one file in `src/transformers/modeling_utils.py` and parsed by `load_gguf_checkpoint` in `src/transformers/modeling_gguf_pytorch_utils.py`. Supporting `*-00001-of-0000N.gguf` would mean detecting the split pattern, resolving all shards, reading metadata from shard 1 and merging tensors across shards in `load_gguf_checkpoint`. Docs: https://huggingface.co/docs/transformers/main/en/quantization/gguf; a PR would be welcome (no open one as far as we can tell).", "author": "Isotr0py", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/32266"}, {"key": "issue#32256", "n": 32256, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "JudgeXL-LLM", "action": "commented", "comment": "Some details:\n- Hub: [Wonder-Griffin/XL-Judge-LLM](https://huggingface.co/Wonder-Griffin/XL-Judge-LLM)\n- Downloads (last 30d): 72 \u00b7 Likes: 0 \u00b7 Params: ~183M \u00b7 License: wtfpl\n- Released: 2024-07 \u00b7 Remote code on the Hub: no\n\nThis is a personal experimental checkpoint rather than a published architecture, so the recommended path is sharing it on the Hub with [custom code](https://huggingface.co/docs/transformers/main/en/custom_models) instead of adding it to the library.", "author": "GriffinDevelopments", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/32256"}, {"key": "issue#32255", "n": 32255, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "JudgeXL-LLM", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [Wonder-Griffin/XL-Judge-LLM](https://huggingface.co/Wonder-Griffin/XL-Judge-LLM)\n- Downloads (last 30d): 72 \u00b7 Likes: 0 \u00b7 Params: ~183M \u00b7 License: wtfpl\n- Released: 2024-07 \u00b7 Remote code on the Hub: no\n- This is an experimental personal model with little usage so far. The best route is to share it with remote code on the Hub ([custom models guide](https://huggingface.co/docs/transformers/main/en/custom_models)) instead of adding it to the library.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "GriffinDevelopments", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/32255"}, {"key": "pr#32231", "n": 32231, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[WIP] Add support for XTR", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32231"}, {"key": "issue#32225", "n": 32225, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Add New Optimizer", "action": "closed", "comment": "As mentioned above, you can use a custom optimizer with `Trainer` without changing the library via `Trainer(optimizers=(optimizer, scheduler))` or `optimizer_cls_and_kwargs` (https://huggingface.co/docs/transformers/main/en/main_classes/trainer#transformers.Trainer.optimizers). If it becomes a widely used, published optimizer, a new issue with paper and benchmarks is welcome. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/32225"}, {"key": "issue#32200", "n": 32200, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "Support `from_pretrained` of `FlaxPretrainedModel` from sharded `.safetensors` weights", "action": "closed", "comment": "TensorFlow and Flax/JAX support was fully removed in v5 (#40760), so this no longer applies. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/32200"}, {"key": "pr#32189", "n": 32189, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[WIP] - Enable speculative decoding with batch size >1", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32189"}, {"key": "issue#32165", "n": 32165, "kind": "issue", "group": "Open issues", "category": "feature", "area": "generation", "type": "feature", "title": "Enable speculative decoding with batch size >1", "action": "commented", "comment": "Still unsupported on main: `_assisted_decoding` in `src/transformers/generation/utils.py` raises for `batch_size > 1`. The candidate generation side lives in `src/transformers/generation/candidate_generator.py`; the second approach you describe (per-sequence rollback with padding and `Cache.crop` per row) is the one that would be accepted, and #26875 is a useful starting point. If you need batched speculative decoding now, it can also be prototyped as a Hub `custom_generate` repo: https://huggingface.co/docs/transformers/main/en/generation_strategies#custom-generation-methods", "author": "kamilakesbi", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/32165"}, {"key": "issue#32164", "n": 32164, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add Matching Anything by Segmenting Anything (MASA) MOT tracking model", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [dereksiyuanli/masa](https://huggingface.co/dereksiyuanli/masa)\n- Downloads (last 30d): 2 \u00b7 Likes: 7 \u00b7 Params: n/a \u00b7 License: apache-2.0\n- Released: 2024-06 \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "rolson24", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/32164"}, {"key": "issue#32145", "n": 32145, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "callback to implement how the predictions should be stored.", "action": "closed", "comment": "As @asartran answered (https://github.com/huggingface/transformers/issues/32145#issuecomment-2244824544), you can write your own callback by inheriting from `TrainerCallback` to control how predictions are stored. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/32145"}, {"key": "pr#32141", "n": 32141, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "add scaling_factor to GemmaRotaryEmbedding for fix error in GemmaLine\u2026", "action": "closed", "comment": "Thanks for the fix! `diff_gemma.py` was replaced by the modular system (`modular_gemma.py`) and the scaling rotary classes were removed, so this no longer applies. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32141"}, {"key": "pr#32116", "n": 32116, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update benchmark.py-- Enhance Benchmarking with Multi-Commit Support \u2026", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32116"}, {"key": "pr#32114", "n": 32114, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix: resolve bug with `use_mps_device` setting not taking effect", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32114"}, {"key": "issue#32113", "n": 32113, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "_prepare_4d_causal_attention_mask mask inversion should work boolean masks", "action": "closed", "comment": "This code path is no longer used by models: the attention masking was rewritten in #37866 (`src/transformers/masking_utils.py`), which works with boolean masks internally (`attention_mask.to(dtype=torch.bool)`). Closing as superseded.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/32113"}, {"key": "pr#32097", "n": 32097, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Custom beam search scorer argument in generate function", "action": "closed", "comment": "Thanks for the PR! Beam search was refactored and `BeamScorer` / `beam_search.py` no longer exist, so this API no longer applies. Custom beam logic can be shared as a Hub custom generation method. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/32097"}, {"key": "issue#32067", "n": 32067, "kind": "issue", "group": "Open issues", "category": "feature", "area": "core / loading", "type": "feature", "title": "Checkpoint validation as an option", "action": "\u2014", "comment": "", "author": "ChenchaoZhao", "idle_days": 30, "models": "", "url": "https://github.com/huggingface/transformers/issues/32067"}, {"key": "issue#32055", "n": 32055, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Using `numpy==2.0.0`", "action": "closed", "comment": "numpy 2 is fully supported now (no upper pin), and the TF/Flax-related incompatibilities went away with their removal in v5. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/32055"}, {"key": "issue#31992", "n": 31992, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Plans to Integrate LongRoPE into LLaMA?", "action": "closed", "comment": "This has been added in #32135 and is available on main. Closing, but feel free to reopen if something is still missing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/31992"}, {"key": "pr#31976", "n": 31976, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Deepseek v2 support", "action": "closed", "comment": "Thanks @itazap for the early work on DeepSeek-V2! The model has since been added to Transformers in #36400 (`src/transformers/models/deepseek_v2`), so this draft is superseded. Closing it, thanks to everyone who helped here!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31976"}, {"key": "issue#31962", "n": 31962, "kind": "issue", "group": "Closed issues", "category": "obsolete", "area": "", "type": "", "title": "Keep Tuple of past key values as an option", "action": "closed; comment corrected", "comment": "The tuple cache format was removed in v5 (#41378); models now always return `Cache` objects, so we won't keep a tuple option. For export, use `transformers.exporters` (https://huggingface.co/docs/transformers/main/en/exporters) with the static cache (see https://huggingface.co/docs/transformers/main/en/kv_cache#fixed-size-cache).", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/31962"}, {"key": "pr#31936", "n": 31936, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "FEAT / Optim: Add QGaLore optimizer", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31936"}, {"key": "pr#31933", "n": 31933, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Added optimizer adam mini", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31933"}, {"key": "pr#31929", "n": 31929, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add Microsoft CLAP model", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31929"}, {"key": "pr#31906", "n": 31906, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Added french version of preprocessing data and corrected English version", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31906"}, {"key": "pr#31899", "n": 31899, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Set position_embeddings in BertEmbeddings for absolute position type only, to avoid unused parameters", "action": "closed", "comment": "Thanks for the PR! TensorFlow and Flax support was removed in v5, so this can't land anymore. Closing, thanks again for the contribution!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31899"}, {"key": "pr#31895", "n": 31895, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Initializes generators in trainer.py with the device specified in sel\u2026", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31895"}, {"key": "pr#31849", "n": 31849, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Generate: XGLM can generate with inputs_embeds", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31849"}, {"key": "issue#31847", "n": 31847, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "NotImplementedError: ggml_type 3 not implemented", "action": "closed", "comment": "Loading GGUF files with Q4_1 (ggml_type 3) and most other GGML quant types has been supported since #32625. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/31847"}, {"key": "pr#31842", "n": 31842, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix some FastSpeechConformer2 failing tests", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31842"}, {"key": "issue#31836", "n": 31836, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add support for MiniCPM-V-2 and MiniCPM-Llama3-V-2_5", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [openbmb/MiniCPM-Llama3-V-2_5](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5), [openbmb/MiniCPM-V-2](https://huggingface.co/openbmb/MiniCPM-V-2)\n- Downloads (last 30d): 19,248 \u00b7 Likes: 1,419 \u00b7 Params: ~8.5B (V-2: ~3.4B) \u00b7 License: MiniCPM model license (see model card)\n- Released: 2024-04 / 2024-05 \u00b7 Remote code on the Hub: yes\n- A previous attempt (#40448) was closed as stale; feel free to build on it with a fresh PR against `main`.\n- Newer MiniCPM-V generations are already supported (`minicpmv4_6`, `minicpmv4_7`); open PR #40448 targets MiniCPM-V 4.5.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "HwwwwwwwH", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/31836"}, {"key": "pr#31834", "n": 31834, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add NSP Labels Handling to DataCollatorForWholeWordMask for Simultaneous WWM and NSP Pre-training", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31834"}, {"key": "issue#31831", "n": 31831, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Add MultiStepLR with Warmup Scheduler", "action": "closed", "comment": "Closing this one: no activity since 2024-07-08, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/31831"}, {"key": "issue#31823", "n": 31823, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "DirectML storage compatibility issues", "action": "closed", "comment": "Closing since the reproducer requested in https://github.com/huggingface/transformers/issues/31823#issuecomment-2213488977 wasn't provided. If you still hit this with `torch_directml`, please open a new issue with a minimal script and the full error.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/31823"}, {"key": "issue#31814", "n": 31814, "kind": "issue", "group": "Open issues", "category": "feature", "area": "models", "type": "feature", "title": "Why there is no LlavaForSequenceClassification ?", "action": "commented", "comment": "Not on main yet, but it's now straightforward: `GenericForSequenceClassification` in `src/transformers/modeling_layers.py` provides the head, and `Gemma3ForSequenceClassification` in `src/transformers/models/gemma3/modular_gemma3.py` shows how a VLM does it (forwarding `pixel_values`). Adding `LlavaForSequenceClassification` the same way in `modular_llava.py`, registering it in `models/auto/modeling_auto.py` and adding it to the tests would close this; see https://huggingface.co/docs/transformers/main/en/modular_transformers", "author": "Zuhashaik", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/31814"}, {"key": "pr#31809", "n": 31809, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix pipeline tests - don't set torch_dtype on non-torch pipelines", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31809"}, {"key": "issue#31804", "n": 31804, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "TinyModel addition", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [noanabeshima/tiny_model](https://huggingface.co/noanabeshima/tiny_model) (original format, not a transformers checkpoint)\n- Downloads (last 30d): 0 \u00b7 Likes: 0 \u00b7 License: apache-2.0\n- Released: 2024-06 \u00b7 Remote code on the Hub: no\n- Note: given the very low usage, sharing it with remote code on the Hub ([custom models guide](https://huggingface.co/docs/transformers/main/en/custom_models)) is probably the better route.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "noanabeshima", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/31804"}, {"key": "pr#31792", "n": 31792, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Adding mplugdocowl", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31792"}, {"key": "issue#31768", "n": 31768, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Train on logits instead of one hot vectors", "action": "closed", "comment": "Training on teacher logits (knowledge distillation) is available in TRL, e.g. its `GKDTrainer` and distillation trainers (https://huggingface.co/docs/transformers/main/en/community_integrations/trl); with plain `Trainer` you can also pass a custom `compute_loss_func` that takes a KL term against teacher logits. Closing in favor of those.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/31768"}, {"key": "issue#31752", "n": 31752, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Suport sdpa for RoBERTa and XLM-RoBERTa models", "action": "closed", "comment": "SDPA (and the other attention backends) is supported for RoBERTa-based models since #30510. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/31752"}, {"key": "issue#31744", "n": 31744, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Fix 'Can't infer missing attention mask on `mps` device'", "action": "closed", "comment": "The mps limitation was lifted in #33099 (`isin_mps_friendly`), so the attention mask can now be inferred on `mps`. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/31744"}, {"key": "issue#31647", "n": 31647, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Data Map Trainer Callback", "action": "closed", "comment": "Thanks for sharing the DataMap callback! Since no PR materialized, closing for now; the callback works as-is by passing it to `Trainer(callbacks=[DataMapCallback(...)])`, and sharing it as a gist or small package is a great way for others to use it.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/31647"}, {"key": "pr#31623", "n": 31623, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Added HHCache class implementing H2O Cache", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31623"}, {"key": "pr#31579", "n": 31579, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[WIP] Add implementation of `_extract_fbank_features_batch` ", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31579"}, {"key": "pr#31541", "n": 31541, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "handle when from_pretrained_id is a list", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31541"}, {"key": "pr#31538", "n": 31538, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix wav2vec2 with torch.compile", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31538"}, {"key": "issue#31497", "n": 31497, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "will this be a model or just a tokenizer/embeddings ? technically a sentence trancformer ?", "action": "closed", "comment": "Closing this one: no activity since 2024-06-19, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/31497"}, {"key": "pr#31482", "n": 31482, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Update beam_constraints with KMP", "action": "closed", "comment": "Thanks for the optimization! Constrained beam search was removed from the core library in v5 (it now lives as a Hub `custom_generate` recipe), so this file no longer exists. Closing, thanks for the effort.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31482"}, {"key": "issue#31475", "n": 31475, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "How do I replace a spare tokens?", "action": "closed", "comment": "As @itazap explained in https://github.com/huggingface/transformers/issues/31475#issuecomment-2313336006, editing the `added_tokens` entries in the tokenizer configs and re-saving works (confirmed by the reporter), and programmatic renaming is tracked in tokenizers (https://github.com/huggingface/tokenizers/pull/1570). Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/31475"}, {"key": "pr#31463", "n": 31463, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "add changes in mistral model to avoid problems in pytorch hooks", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31463"}, {"key": "issue#31441", "n": 31441, "kind": "issue", "group": "Open issues", "category": "feature", "area": "core / loading", "type": "feature", "title": "Add `StatefulDataLoader` support", "action": "\u2014", "comment": "", "author": "yzhangcs", "idle_days": 76, "models": "", "url": "https://github.com/huggingface/transformers/issues/31441"}, {"key": "pr#31426", "n": 31426, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Make CI failure summary more useful", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31426"}, {"key": "pr#31393", "n": 31393, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Set ffmpeg_microphone input to system default on Darwin platform", "action": "closed", "comment": "Thanks @ruisilvestre! These changes were incorporated in #31846, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31393"}, {"key": "pr#31388", "n": 31388, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Preserve `ModelOutput` object for compute_metrics evaluation loop in trainer", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31388"}, {"key": "issue#31338", "n": 31338, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Support saving models trained with DeepSpeed in Trainer callbacks", "action": "commented", "comment": "This is still not possible: `CallbackHandler.call_event` in `src/transformers/trainer_callback.py` passes `model`, `optimizer`, `lr_scheduler` and the dataloaders, but not `accelerator`/`model_wrapped`, which ZeRO-3 saving needs. A PR along the lines you proposed (set them on the handler in `Trainer.__init__`/`_inner_training_loop` and forward them as kwargs) would be welcome; meanwhile, overriding `Trainer._save` or calling `trainer.save_model()` from an `on_train_end` hook that holds a reference to the trainer works.", "author": "dwyatte", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/31338"}, {"key": "pr#31330", "n": 31330, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Rest of model init refactors", "action": "closed", "comment": "Thanks for the PR! TensorFlow and Flax support was removed in v5, so this can't land anymore. Closing, thanks again for the contribution!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31330"}, {"key": "pr#31316", "n": 31316, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "SPLIT PR: eos bos tokens", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31316"}, {"key": "pr#31315", "n": 31315, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "SPLIT PR: add_prefix_space fix", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31315"}, {"key": "pr#31297", "n": 31297, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[torch.compile] Graph break+warning instead of error in _prepare_generation_config", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31297"}, {"key": "pr#31268", "n": 31268, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Potential fix for #30819. Check for best metric only on gpu 0", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31268"}, {"key": "issue#31250", "n": 31250, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Adaptive Decoding Support", "action": "closed", "comment": "As @gante explained (https://github.com/huggingface/transformers/pull/38055#issuecomment-2900553588), new decoding methods are first shared as Hub `custom_generate` repos and added to transformers if they become popular. Adaptive decoding can be shipped that way: https://huggingface.co/docs/transformers/main/en/generation_strategies#custom-generation-methods. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/31250"}, {"key": "issue#31248", "n": 31248, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Add support for non-CUDA architectures at the same time Bitsandbytes is doing it", "action": "closed", "comment": "This has been added in #31098 and is available on main. Closing, but feel free to reopen if something is still missing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/31248"}, {"key": "issue#31222", "n": 31222, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Batch size schedulers", "action": "\u2014", "comment": "", "author": "BramVanroy", "idle_days": 85, "models": "", "url": "https://github.com/huggingface/transformers/issues/31222"}, {"key": "pr#31167", "n": 31167, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "SDPA for T5 Attention", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31167"}, {"key": "issue#31126", "n": 31126, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Request for Static Cache Support for XLA Compiler in Transformers", "action": "closed", "comment": "This landed: `StaticCache` now updates in place with `index_copy_` (#31857, see `src/transformers/cache_utils.py`), which is the approach proposed in #31129 and avoids recompilation from slicing under XLA. See https://huggingface.co/docs/transformers/main/en/google_tpu for TPU/XLA usage. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/31126"}, {"key": "issue#31105", "n": 31105, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "need better token_type_id processing on transformer GPT2Model ", "action": "closed", "comment": "Closing this one: no activity since 2024-05-30, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/31105"}, {"key": "issue#31096", "n": 31096, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Caching Past Key values of any length for Vision LLM's", "action": "closed", "comment": "This works on main: since processors expand image tokens (#30962) and the VLM cleanup (#34502), LLaVA-style models no longer have a special one-new-token cache path; they scatter image features into the embeddings and forward with `past_key_values` like text LLMs, so you can cache a prompt and score several-token continuations. See https://huggingface.co/docs/transformers/main/en/kv_cache#prefill-a-cache-prefix-caching", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/31096"}, {"key": "pr#31059", "n": 31059, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "feat: adding mplugdocowl", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31059"}, {"key": "pr#31050", "n": 31050, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add basic eval table logging for WandbCallback", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/31050"}, {"key": "issue#31049", "n": 31049, "kind": "issue", "group": "Open issues", "category": "feature", "area": "core / loading", "type": "feature", "title": "Checkpoint saving by different evaluation criterias", "action": "\u2014", "comment": "", "author": "daehuikim", "idle_days": 875, "models": "", "url": "https://github.com/huggingface/transformers/issues/31049"}, {"key": "issue#30998", "n": 30998, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Trainer should throw a warning if max_sequence_length < number of tokens in dataset sample record.", "action": "closed", "comment": "Since `Trainer` receives already-tokenized data, it can't tell whether samples were truncated; the tokenizer is where that happens (`truncation=True, max_length=...`), and you can check lengths with `dataset.map(lambda x: {\"n\": len(x[\"input_ids\"])})` beforehand. Closing as not planned.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/30998"}, {"key": "issue#30995", "n": 30995, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add Nomic Embed Code to Transformers", "action": "commented", "comment": "Good news: this is now supported in transformers as `nomic_bert` (added in #43067).", "author": "zanussbaum", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/30995"}, {"key": "issue#30969", "n": 30969, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "Training GPT2 with run_clm.py exceeds the described memory amount .", "action": "closed", "comment": "The `examples/pytorch` scripts were removed in #49357, so closing this one. Feel free to open a new issue if you hit a problem with the library itself!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/30969"}, {"key": "pr#30883", "n": 30883, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add IRIS", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/30883"}, {"key": "issue#30882", "n": 30882, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add IRIS", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [eloialonso/iris](https://huggingface.co/eloialonso/iris) (Atari checkpoints, original-repo format)\n- Downloads (last 30d): 0 \u00b7 Likes: 1 \u00b7 License: not specified\n- Released: 2024-05 \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "RUFFY-369", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/30882"}, {"key": "issue#30877", "n": 30877, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Kosmos-2.5 implementation in transformers", "action": "closed", "comment": "This has been added in #31711 and is available on main. Closing, but feel free to reopen if something is still missing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/30877"}, {"key": "pr#30862", "n": 30862, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[WIP] Dynamic length in static cache", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/30862"}, {"key": "pr#30852", "n": 30852, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add bounding box converter", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/30852"}, {"key": "issue#30818", "n": 30818, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Integrate IndicTrans2 models and tokenizer into HF Transformers", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [ai4bharat/indictrans2-en-indic-1B](https://huggingface.co/ai4bharat/indictrans2-en-indic-1B), [ai4bharat/indictrans2-indic-en-1B](https://huggingface.co/ai4bharat/indictrans2-indic-en-1B), [ai4bharat/indictrans2-en-indic-dist-200M](https://huggingface.co/ai4bharat/indictrans2-en-indic-dist-200M)\n- Downloads (last 30d): 67,405 \u00b7 Likes: 82 \u00b7 Params: ~1.1B \u00b7 License: mit\n- Released: 2023-09 \u00b7 Remote code on the Hub: yes\n- The distilled en-indic 200M is the most used variant (239,600 downloads, last 30d). The architecture is close to `m2m_100`/`nllb`, so a modular file should be short.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "VarunGumma", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/30818"}, {"key": "issue#30791", "n": 30791, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "new model request: DeepSeek-V2", "action": "commented", "comment": "Good news: this is now supported in transformers as `deepseek_v2` (added in #36400).", "author": "Atry", "idle_days": 0, "models": "deepseek_v2", "url": "https://github.com/huggingface/transformers/issues/30791"}, {"key": "issue#30758", "n": 30758, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Implement kv cache sparsity like H2O with attention score", "action": "closed", "comment": "We decided not to add eviction caches like H2O to core (#35381 was closed). They can be built on the `Cache`/cache-layer classes in `src/transformers/cache_utils.py` and passed as `past_key_values=...` to `generate` (see https://huggingface.co/docs/transformers/main/en/cache_explanation); NVIDIA's kvpress library already implements H2O-style presses on top of transformers.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/30758"}, {"key": "issue#30731", "n": 30731, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Mixtral past_key_values and output_router_logits incompatible", "action": "closed", "comment": "This works on main: running the snippet above with a tiny `MixtralConfig`, the second forward with `past_key_values`, an extended `attention_mask` and `output_router_logits=True` returns `aux_loss` without errors. Closing, please reopen with a reproducer if you still hit it on a recent version.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/30731"}, {"key": "issue#30705", "n": 30705, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Refusal rejection removal as a feature", "action": "closed", "comment": "Closing this one: no activity since 2024-05-08, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/30705"}, {"key": "issue#30638", "n": 30638, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add Prismatic VLMs to Transformers", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [TRI-ML/prismatic-vlms](https://huggingface.co/TRI-ML/prismatic-vlms)\n- Downloads (last 30d): 0 \u00b7 Likes: 30 \u00b7 Params: 3B-13B (~50 checkpoints, original-repo format) \u00b7 License: mit\n- Released: 2024-02 \u00b7 Remote code on the Hub: no\n- Related models already in transformers: `llava`, `llava_next`.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "siddk", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/30638"}, {"key": "issue#30529", "n": 30529, "kind": "issue", "group": "Closed issues", "category": "obsolete", "area": "", "type": "", "title": "Use models as Seq2Seq model", "action": "closed", "comment": "MEGA was deprecated and has since been removed from the library (#42235), so we won't be able to fix this. For seq2seq with arbitrary encoder/decoder pairs, `EncoderDecoderModel` (https://huggingface.co/docs/transformers/main/en/model_doc/encoder-decoder) is the supported route. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/30529"}, {"key": "pr#30511", "n": 30511, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Image + text + audio uniform processors", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/30511"}, {"key": "pr#30489", "n": 30489, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add trainer integration test for llava to ensure accelerate autocasting works correctly", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/30489"}, {"key": "pr#30486", "n": 30486, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Move out input validation into base image processor", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/30486"}, {"key": "pr#30479", "n": 30479, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[WIP] adds AudioMAE model", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/30479"}, {"key": "issue#30447", "n": 30447, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "(Willing to PR) Make `tokenizer.padding_side` an argument instead of only being a field", "action": "closed", "comment": "This has been added in #33385 and is available on main. Closing, but feel free to reopen if something is still missing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/30447"}, {"key": "issue#30384", "n": 30384, "kind": "issue", "group": "Open issues", "category": "feature", "area": "processors (image/video/audio)", "type": "feature", "title": "Update `make-fixup` to make sure image processor tests present", "action": "commented", "comment": "Still not enforced: `utils/check_repo.py` (run by `make check-repo`) has access to `IMAGE_PROCESSOR_MAPPING_NAMES` but doesn't verify that each model with an image processor has `tests/models/<model>/test_image_processing_<model>.py`. A small check there (with an allowlist for exceptions) would close this.", "author": "ydshieh", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/30384"}, {"key": "issue#30369", "n": 30369, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "[DynamicHead - Microsoft] Implementation request", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: no DyHead checkpoint on the Hub; weights are only on the [original GitHub repo](https://github.com/microsoft/DynamicHead)\n- Released: 2021-06 (paper) \u00b7 Remote code on the Hub: no\n- Note: DyHead is a detection head on top of ATSS / Faster R-CNN, so it also needs one of those detectors.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "lsch0lz", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/30369"}, {"key": "issue#30360", "n": 30360, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Flash Attention Support for Blip2ForConditionalGeneration ", "action": "closed", "comment": "Blip2 supports FlashAttention since #37576; pass `attn_implementation=\"flash_attention_2\"` when loading. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/30360"}, {"key": "issue#30292", "n": 30292, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Token merging for LLM Inference", "action": "closed", "comment": "Thanks for the experiments! As @gante outlined (https://github.com/huggingface/transformers/issues/30292#issuecomment-2083003832), adding this to core would need reproducible benchmarks across architectures; otherwise the best home is your own package or a Hub `custom_generate` repo (https://huggingface.co/docs/transformers/main/en/generation_strategies#custom-generation-methods). Closing for now; happy to revisit with those benchmarks.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/30292"}, {"key": "issue#30280", "n": 30280, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "modeling_t5 incompatible with multiprocessing", "action": "closed", "comment": "As @ArthurZucker explained in https://github.com/huggingface/transformers/issues/30280#issuecomment-2085045318, custom multiprocessing is outside the scope of transformers and `accelerate` (or FSDP/DeepSpeed) is the recommended path for multi-GPU inference. Closing; if T5 still fails that way on current main, please open a new issue with a minimal script.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/30280"}, {"key": "issue#30214", "n": 30214, "kind": "issue", "group": "Open issues", "category": "feature", "area": "processors (image/video/audio)", "type": "feature", "title": "Create panoptic segmentation task guide", "action": "commented", "comment": "Still open for contributors: `docs/source/en/tasks/` has `semantic_segmentation.md` and `instance_segmentation.md` but no panoptic guide. A `panoptic_segmentation.md` modeled on the instance-segmentation guide (Mask2Former/OneFormer fine-tuning with `Trainer`, PQ evaluation, and `post_process_panoptic_segmentation`), registered in `docs/source/en/_toctree.yml`, would close this.", "author": "NielsRogge", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/30214"}, {"key": "issue#30172", "n": 30172, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "4bit Adam", "action": "closed", "comment": "Good news: a 4-bit AdamW is available in the Trainer via `optim=\"adamw_torch_4bit\"` (added in #31865). Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/30172"}, {"key": "issue#30109", "n": 30109, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Can Transformers AutoModel Support to(dml)", "action": "closed", "comment": "As @asartran noted (https://github.com/huggingface/transformers/issues/30109#issuecomment-2042072416), device support like DirectML belongs with `torch-directml` / `accelerate` rather than Transformers (models are plain `nn.Module`s, so `model.to(torch_directml.device())` is the path). Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/30109"}, {"key": "issue#30045", "n": 30045, "kind": "issue", "group": "Closed issues", "category": "obsolete", "area": "", "type": "", "title": "`test_encode_decode_fast_slow_all_tokens` is failing", "action": "closed", "comment": "This test compared slow and fast tokenizers; slow tokenizers were removed in v5 (#40936) and the test no longer exists, so closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/30045"}, {"key": "issue#30019", "n": 30019, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Model trained with Flash Attention 2.0 raises \"RuntimeError: query and key must have the same dtype\" when generating", "action": "closed", "comment": "On main the flash-attention path casts query/key/value to a common dtype (`get_target_dtype` in `src/transformers/integrations/flash_attention.py`, #41605), so the 'query and key must have the same dtype' error after training with fp32-upcast layers should be gone. Closing; please open a new issue with a reproducer if you still see it on v5.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/30019"}, {"key": "issue#29994", "n": 29994, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Flash Attention 2 support for Wav2Vec2ForCTC", "action": "closed", "comment": "Flash Attention 2 (and SDPA) support for the Wav2Vec2 family was added in #30121, so `Wav2Vec2ForCTC` can use it. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29994"}, {"key": "issue#29983", "n": 29983, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Nested Tensor support", "action": "closed", "comment": "The goal (no compute wasted on padding) is covered by padding-free batching: `DataCollatorWithFlattening` (#31629) packs sequences and passes `position_ids`, and the flash/sdpa attention paths treat them as variable-length sequences. See https://huggingface.co/docs/transformers/main/en/padding_free. We don't plan NestedTensor inputs on top of that.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29983"}, {"key": "issue#29972", "n": 29972, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Enhance Generality of Trainer Class", "action": "closed", "comment": "As @ArthurZucker suggested (https://github.com/huggingface/transformers/issues/29972#issuecomment-2029584895), you can override `compute_loss` and call `self.log({...})` with your extra loss terms, or use callbacks (https://huggingface.co/docs/transformers/main/en/trainer_customize). Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29972"}, {"key": "pr#29969", "n": 29969, "kind": "PR", "group": "Open PRs", "category": "approved", "area": "models", "type": "feature / refactor", "title": "[SigLIP] Add fast tokenizer", "action": "\u2014", "comment": "", "author": "NielsRogge", "idle_days": 721, "models": "glm, siglip", "url": "https://github.com/huggingface/transformers/pull/29969"}, {"key": "pr#29963", "n": 29963, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add serialization to hf_argparser.py", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/29963"}, {"key": "pr#29950", "n": 29950, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "[WIP] Add musecoco", "action": "closed", "comment": "Thanks for the work! New models now go through a `modular_*.py` file, and we recommend sharing MuseCoco on the Hub with remote code first, as suggested in review. Closing for now.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/29950"}, {"key": "issue#29925", "n": 29925, "kind": "issue", "group": "Open issues", "category": "feature", "area": "tokenization", "type": "feature", "title": "SigLIP tokenizer not enforcing use_fast=True", "action": "\u2014", "comment": "", "author": "yxchng", "idle_days": 4, "models": "", "url": "https://github.com/huggingface/transformers/issues/29925"}, {"key": "issue#29908", "n": 29908, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Change Device Allocation and Improve Shape Handling in Trainer evaluation loop", "action": "closed", "comment": "Closing this one: no activity since 2024-03-27, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29908"}, {"key": "issue#29827", "n": 29827, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "About new ClearML Intergrations", "action": "commented", "comment": "Still the case on main: `ClearMLCallback.on_log` in `src/transformers/integrations/integration_utils.py` reports every train metric as a series of a single `train` plot (same for `eval`/`test`), with only a few summary values sent via `report_single_value`. A PR making `title` per metric (e.g. `title=k, series=\"train\"`), or configurable via an env var like the existing `CLEARML_*` ones, would be welcome; until then subclassing `ClearMLCallback` and overriding `on_log` works.", "author": "dimka11", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/29827"}, {"key": "issue#29823", "n": 29823, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Mixture of All Intelligence (MoAI)", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [BK-Lee/MoAI-7B](https://huggingface.co/BK-Lee/MoAI-7B)\n- Downloads (last 30d): 129 \u00b7 Likes: 44 \u00b7 Params: ~8.9B \u00b7 License: mit\n- Released: 2024-03 \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "Dev-Khant", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/29823"}, {"key": "issue#29740", "n": 29740, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Add option to not re-load the model when resuming from checkpoint.", "action": "commented", "comment": "There's still no opt-out: `Trainer.train` calls `self._load_from_checkpoint(resume_from_checkpoint)` whenever resuming (`src/transformers/trainer.py`, plus the FSDP/SageMaker branch in `_prepare_for_training`). As a workaround you can subclass `Trainer` and override `_load_from_checkpoint` to a no-op; a PR adding an explicit flag around those two calls would be reviewed by the trainer maintainers.", "author": "stevemadere", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/29740"}, {"key": "issue#29732", "n": 29732, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "[Feature Request] add layer-wise optimizers", "action": "commented", "comment": "Layer-wise stepping is currently only wired for GaLore and APOLLO (`galore_adamw_layerwise`, `apollo_adamw_layerwise` in `src/transformers/training_args.py`), via the per-parameter `register_post_accumulate_grad_hook` logic in `src/transformers/trainer_optimizer.py`. Generalizing that wrapper to any optimizer (e.g. an `adamw_8bit_layerwise` option) would be the path; note it still can't support DDP or gradient accumulation, as discussed above.", "author": "peterjc123", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/29732"}, {"key": "issue#29730", "n": 29730, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "[Mamba] Possible Issue in beam search with Mamba HF models", "action": "closed", "comment": "The missing beam reordering of Mamba's conv/SSM states is handled on main: Mamba now uses the native cache (#44950), whose `reorder_cache` reorders both `conv_states` and `recurrent_states` (`LinearAttentionCacheLayerMixin` in `src/transformers/cache_utils.py`). Beam search with `state-spaces/mamba-130m-hf` should no longer degrade; please reopen with a repro if it does.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29730"}, {"key": "pr#29723", "n": 29723, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "fix model names", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/29723"}, {"key": "issue#29704", "n": 29704, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Grok-1 MoE support", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [xai-org/grok-1](https://huggingface.co/xai-org/grok-1), [xai-org/grok-2](https://huggingface.co/xai-org/grok-2)\n- Downloads (last 30d): 1,367 \u00b7 Likes: 2,476 \u00b7 Params: ~314B (MoE, 8 experts / 2 active) \u00b7 License: apache-2.0 (grok-2: Grok 2 Community License)\n- Released: 2024-03 (grok-2: 2025-08) \u00b7 Remote code on the Hub: no (JAX checkpoint)\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "AlpinDale", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/29704"}, {"key": "pr#29693", "n": 29693, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add distribution params to time series output", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/29693"}, {"key": "pr#29646", "n": 29646, "kind": "PR", "group": "Closed PRs", "category": "stale", "area": "", "type": "", "title": "Natively support SONAR text models as M2M100 encoder and decoder models", "action": "closed", "comment": "Thanks @avidale for the original work! This has not been active and would need to be rebased on v5!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/29646"}, {"key": "issue#29556", "n": 29556, "kind": "issue", "group": "Open issues", "category": "feature", "area": "models", "type": "feature", "title": "Add predict_proba method for Autoformer/Informer", "action": "commented", "comment": "Still open: `generate()` for Informer/Autoformer returns `SampleTSPredictionOutput`, which only carries `sequences` (`src/transformers/modeling_outputs.py`), while the per-step distribution parameters are only exposed as `params` on the forward output (`Seq2SeqTSPredictionOutput`). Following kashif's suggestion, a PR could add an optional `params` field to `SampleTSPredictionOutput` and collect the `self.parameter_projection(...)` outputs in the generation loop of `modeling_informer.py` / `modeling_autoformer.py` (and `time_series_transformer`).", "author": "benHeid", "idle_days": 0, "models": "informer", "url": "https://github.com/huggingface/transformers/issues/29556"}, {"key": "issue#29547", "n": 29547, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Adherence to semantic versioning with APIs", "action": "closed", "comment": "Thanks for raising this. As @asartran and @gante explained (https://github.com/huggingface/transformers/issues/29547#issuecomment-1988931198, https://github.com/huggingface/transformers/issues/29547#issuecomment-1994590934), we aim for backward compatibility with deprecation cycles, and the larger breaking changes were batched into v5 with `MIGRATION_GUIDE_V5.md`. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29547"}, {"key": "issue#29546", "n": 29546, "kind": "issue", "group": "Open issues", "category": "feature", "area": "processors (image/video/audio)", "type": "feature", "title": "Support more memory efficient processing for segmentation models.", "action": "\u2014", "comment": "", "author": "jveitchmichaelis", "idle_days": 348, "models": "", "url": "https://github.com/huggingface/transformers/issues/29546"}, {"key": "issue#29516", "n": 29516, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Track for flaky batching tests", "action": "closed", "comment": "Closing this tracker: ViT-hybrid was removed with the deprecated models (#42235), and DPT-hybrid's `test_batching_equivalence` now uses a relaxed tolerance in `tests/models/dpt/test_modeling_dpt_hybrid.py`. If MobileNetV2 flakes again, please open a focused issue with the failing seed.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29516"}, {"key": "issue#29490", "n": 29490, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add Microsoft's Code Reviewer model as a dedicated model", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [microsoft/codereviewer](https://huggingface.co/microsoft/codereviewer)\n- Downloads (last 30d): 624 \u00b7 Likes: 147 \u00b7 License: apache-2.0\n- Released: 2022-07 \u00b7 Remote code on the Hub: no\n- Note: the checkpoint is a T5 model and already loads with `T5ForConditionalGeneration`; only the extra heads in the original repo are missing.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "fasterinnerlooper", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/29490"}, {"key": "issue#29482", "n": 29482, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Large World Model and Ring Attention", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [LargeWorldModel/LWM-Text-1M](https://huggingface.co/LargeWorldModel/LWM-Text-1M), [LargeWorldModel/LWM-Text-Chat-1M](https://huggingface.co/LargeWorldModel/LWM-Text-Chat-1M)\n- Downloads (last 30d): 251 \u00b7 Likes: 29 \u00b7 Params: n/a \u00b7 License: n/a\n- Released: 2024-01 \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "kabachuha", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/29482"}, {"key": "issue#29466", "n": 29466, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "BASED ", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [hazyresearch/based-1b](https://huggingface.co/hazyresearch/based-1b), [hazyresearch/based-360m](https://huggingface.co/hazyresearch/based-360m)\n- Downloads (last 30d): 62 \u00b7 Likes: 8 \u00b7 Params: ~1.3B / 360M \u00b7 License: mit\n- Released: 2024-02 \u00b7 Remote code on the Hub: no (original-repo code)\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "axelmagn", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/29466"}, {"key": "issue#29464", "n": 29464, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Implementing the features of the TextStreamer into the pipeline", "action": "closed", "comment": "This is supported now: the text-generation pipeline accepts `return_full_text=False` (drops the prompt) and, since #40356, `skip_special_tokens=True`, so `pipe(prompt, return_full_text=False, skip_special_tokens=True)` returns just the clean completion. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29464"}, {"key": "issue#29458", "n": 29458, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Add a warning when the dataset has no matching columns", "action": "closed", "comment": "This was added in #31466: `Trainer._remove_unused_columns` now raises a clear error ('No columns in the dataset match the model's forward method signature ...') instead of an opaque `IndexError`. Closing, thanks for the suggestion!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29458"}, {"key": "issue#29416", "n": 29416, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "The results of run_mae.py pre-training were abnormal", "action": "closed", "comment": "The `examples/pytorch` scripts were removed in #49357, so closing this one. Feel free to open a new issue if you hit a problem with the library itself!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29416"}, {"key": "issue#29411", "n": 29411, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "Transformers Agents Collab Notebook - OpenAI Run Mode Issues", "action": "closed", "comment": "Agents were removed from Transformers in #37368 and now live in smolagents, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29411"}, {"key": "issue#29410", "n": 29410, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "Mistral in Flax: generation is slow, JIT fails", "action": "closed", "comment": "Flax/JAX support was removed from Transformers in v5 (#40760), so this Flax Mistral issue no longer applies. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29410"}, {"key": "issue#29391", "n": 29391, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "hyperparameter_serch() does not consider LoRA parameters like r  to be finetuned. ", "action": "closed", "comment": "This is already possible: give `model_init` one argument and it receives the trial (`Trainer.call_model_init` in `src/transformers/trainer.py`). Sample LoRA params there, e.g. `def model_init(trial): r = trial.suggest_int(\"r\", 4, 64) if trial else 8; return get_peft_model(base, LoraConfig(r=r, ...))`, and include `r` in your `hp_space`. See https://huggingface.co/docs/transformers/main/en/hpo_train", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29391"}, {"key": "issue#29385", "n": 29385, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "MPNet doesn't have an implemented LMHead subclass", "action": "closed", "comment": "As @ArthurZucker explained in https://github.com/huggingface/transformers/issues/29385#issuecomment-2029572961, `MPNetForMaskedLM` is already the LM-head model for MPNet and `*WithLMHead` would be equivalent. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29385"}, {"key": "pr#29371", "n": 29371, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add Mixture of Tokens model", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/29371"}, {"key": "issue#29319", "n": 29319, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "Add Mixtral Model to Flax", "action": "closed", "comment": "Flax/JAX support was removed from Transformers in v5 (#40760), so a Flax Mixtral won't be added. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29319"}, {"key": "pr#29315", "n": 29315, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Resolved device mismatch error when exporting Hugging Face's BERTModel to TorchScript", "action": "closed; comment corrected", "comment": "Thanks for looking into this! TorchScript support has been removed in v5, and export now goes through `transformers.exporters` (https://huggingface.co/docs/transformers/main/en/exporters), so we won't be changing models for jit tracing. Closing, thanks for the contribution.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/29315"}, {"key": "issue#29293", "n": 29293, "kind": "issue", "group": "Open issues", "category": "feature", "area": "processors (image/video/audio)", "type": "feature", "title": "Whisper - get probability of detected language", "action": "commented", "comment": "This is still not exposed: `WhisperGenerationMixin.detect_language` in `src/transformers/models/whisper/generation_whisper.py` computes the language-token logits but only returns the argmax ids. A contributor could add an opt-in flag (e.g. `return_probs=True`) that returns the softmax over the language tokens alongside the ids, plus a test in `tests/models/whisper/test_modeling_whisper.py`; see https://huggingface.co/docs/transformers/main/en/contributing.", "author": "antoinethl", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/29293"}, {"key": "issue#29288", "n": 29288, "kind": "issue", "group": "Closed issues", "category": "obsolete", "area": "", "type": "", "title": "Unable to trace DebertaForMaskedLM", "action": "closed; comment corrected", "comment": "`transformers.utils.fx` and HF's `symbolic_trace` support were removed in #41683, so this won't be addressed. For export, use `transformers.exporters` (https://huggingface.co/docs/transformers/main/en/exporters); `torch.compile` covers graph capture for speed.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29288"}, {"key": "pr#29275", "n": 29275, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Add support for Phi in Flax", "action": "closed", "comment": "Thanks for putting this together! Transformers dropped Flax/JAX in v5 and is PyTorch-only now, so we can't add Flax ports. Closing, but thank you.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/29275"}, {"key": "issue#29259", "n": 29259, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "AutoModel.from_config() should attempt to load model relative to where the config was loaded.", "action": "closed", "comment": "Closing this one: no activity since 2024-02-23, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29259"}, {"key": "issue#29227", "n": 29227, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Fine-tuning a pruned model", "action": "closed", "comment": "Thanks for the request! Keeping pruned weights frozen can be done in plain PyTorch, e.g. by registering a gradient hook on each parameter that multiplies by the pruning mask (`param.register_hook(lambda g: g * mask)`) or with `torch.nn.utils.prune`, which works with DDP and Trainer. As there is nothing transformers-specific to add, closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29227"}, {"key": "issue#29204", "n": 29204, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "prophetnet_multi", "action": "commented", "comment": "The multilingual ProphetNet (`xlm_prophetnet`) used to be in transformers but was deprecated for low usage in #30781 and has since been removed.\n\n- Hub: [microsoft/xprophetnet-large-wiki100-cased](https://huggingface.co/microsoft/xprophetnet-large-wiki100-cased)\n- Downloads (last 30d): 588 \u00b7 Likes: 5\n- Released: 2020 (Hub repo 2022-03) \u00b7 Remote code on the Hub: no\n\nYou can still use it with an older transformers release (`<4.45`). Given the usage, re-adding it is unlikely.", "author": "wyl-willing", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/29204"}, {"key": "pr#29194", "n": 29194, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "add Fusion-In-Decoder(FiD) Models", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/29194"}, {"key": "issue#29129", "n": 29129, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Flash attention implementation with BERT base model", "action": "\u2014", "comment": "", "author": "", "idle_days": 866, "models": "", "url": "https://github.com/huggingface/transformers/issues/29129"}, {"key": "issue#29068", "n": 29068, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Add Support for Dataclasses to Trainer", "action": "commented", "comment": "Still open for a contribution (@muellerzr was open to it). `Trainer._prepare_input` in `src/transformers/trainer.py` recurses only into `Mapping` and `tuple`/`list`; adding a `dataclasses.is_dataclass(data)` branch that rebuilds it with `dataclasses.replace(data, **{f.name: self._prepare_input(getattr(data, f.name)) for f in dataclasses.fields(data)})`, plus a test in `tests/trainer/`, would cover it.", "author": "ntenenz", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/29068"}, {"key": "issue#29060", "n": 29060, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Request for Flash Attention 2.0 Support in GPNRoFormerForMaskedLM", "action": "closed", "comment": "`GPNRoFormerForMaskedLM` isn't part of Transformers (it lives in the external `gpn` repo), so this needs to happen there; its authors can add Flash Attention through the attention interface (https://huggingface.co/docs/transformers/main/en/attention_interface). Closing here.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29060"}, {"key": "issue#29006", "n": 29006, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "load_state_dict doesnt support torch._subclasses.fake_tensor.FakeTensorMode", "action": "closed", "comment": "As @Narsil explained in https://github.com/huggingface/transformers/issues/29006#issuecomment-2011906660, `torch._subclasses.fake_tensor` is a private PyTorch API, so we won't special-case it while it isn't officially supported. To build a model without allocating memory, use `with torch.device(\"meta\"): model = AutoModel.from_config(config)`. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/29006"}, {"key": "issue#28936", "n": 28936, "kind": "issue", "group": "Open issues", "category": "docs", "area": "docs", "type": "docs", "title": "[i18n-es] Translating docs to Spanish", "action": "\u2014", "comment": "", "author": "stevhliu", "idle_days": 48, "models": "", "url": "https://github.com/huggingface/transformers/issues/28936"}, {"key": "issue#28935", "n": 28935, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Add support for prefix_allowed_tokens_fn to maintain a state throughout decoding", "action": "closed", "comment": "As @gante explained (https://github.com/huggingface/transformers/issues/28935#issuecomment-1939109113), no library change is needed: the state can live in a mutable input bound to your `prefix_allowed_tokens_fn` (closure / `functools.partial`) or in a custom `LogitsProcessor`. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/28935"}, {"key": "issue#28865", "n": 28865, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Detr models crashes when changing the num_queries parameter in the config", "action": "closed", "comment": "As found in the thread, the error comes from NaN/degenerate predicted boxes when the query embeddings are re-initialized (`num_queries` change with `ignore_mismatched_sizes=True`) and the learning rate is too high; `generalized_box_iou` in `src/transformers/loss/loss_for_object_detection.py` then raises. Keeping the pretrained 100 queries or using a lower LR (e.g. 1e-5 with a frozen backbone) avoids it. Closing as this is training dynamics, not a modeling bug.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/28865"}, {"key": "issue#28849", "n": 28849, "kind": "issue", "group": "Open issues", "category": "feature", "area": "generation", "type": "feature", "title": "TextStreamer: An option to print / put every single token instead of whole words", "action": "\u2014", "comment": "", "author": "drummerv", "idle_days": 6, "models": "", "url": "https://github.com/huggingface/transformers/issues/28849"}, {"key": "issue#28813", "n": 28813, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Adding Gradient Checkpointing and Flash Attention 2 implementation to VisionTextDualEncoderModel", "action": "\u2014", "comment": "", "author": "Syrinechen", "idle_days": 37, "models": "vision_text_dual_encoder", "url": "https://github.com/huggingface/transformers/issues/28813"}, {"key": "issue#28781", "n": 28781, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "Unable to use torch scripting to export Mask2Former model", "action": "closed; comment corrected", "comment": "TorchScript support was removed from Transformers in v5 (#41688), so we're closing this. For export, use `transformers.exporters` (https://huggingface.co/docs/transformers/main/en/exporters).", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/28781"}, {"key": "issue#28763", "n": 28763, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Allow setting different decoder_start_token_ids for each item in a batch in the generate function.", "action": "closed", "comment": "This has been added in #28887 and is available on main. Closing, but feel free to reopen if something is still missing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/28763"}, {"key": "pr#28750", "n": 28750, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Fix the StarCoder agent max_new_tokens input validation error", "action": "closed", "comment": "Thanks for the fix! The agents/tools code has been removed from transformers and now lives in smolagents, so this file no longer exists. Closing, but feel free to open the fix there if it still applies.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/28750"}, {"key": "issue#28740", "n": 28740, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "DETR: IndexError: Caught IndexError in replica 0 on device 0. IndexError: index 8 is out of bounds for dimension 0 with size 8", "action": "closed", "comment": "When several GPUs are visible without a distributed launcher, `Trainer` wraps the model in `nn.DataParallel` (`src/transformers/trainer.py`), which can't split DETR's list-of-dict `labels`. Launching with `torchrun --nproc_per_node N train.py` or `accelerate launch train.py` (DDP) avoids this. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/28740"}, {"key": "issue#28721", "n": 28721, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Load an EncoderDecoderModel as AutoModel", "action": "closed", "comment": "As discussed above, `EncoderDecoderModel` is registered under `AutoModelForSeq2SeqLM` (`\"encoder-decoder\"` in `src/transformers/models/auto/modeling_auto.py`), so `AutoModelForSeq2SeqLM.from_pretrained(\"Bachstelze/instructionRoberta-base\")` is the way to load it; `AutoModel` isn't meant to cover composite encoder-decoder checkpoints. Closing as the loading path exists and the leaderboard question is outside this repo.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/28721"}, {"key": "pr#28711", "n": 28711, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[WIP] Improve multimodal processors - rely less on kwargs", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/28711"}, {"key": "issue#28635", "n": 28635, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Tokenizer `encode/decode` methods are inconsistent, TypeError: argument 'ids': 'list' object cannot be interpreted as an integer", "action": "closed", "comment": "This was fixed by #40936, which is now on main. Closing, but feel free to reopen if you still see it on the latest version!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/28635"}, {"key": "issue#28611", "n": 28611, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "PatchTST and PatchTSMixer categorical features and exogenous variables", "action": "closed", "comment": "As @kashif explained in https://github.com/huggingface/transformers/issues/28611#issuecomment-2080400501, PatchTST/PatchTSMixer don't have the inductive bias to incorporate temporal covariates straightforwardly; lag-tst in GluonTS is a better fit. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/28611"}, {"key": "pr#28556", "n": 28556, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Feature Update [added `initial_prompt` support for automatic-speech-recognition whisper pipeline]", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/28556"}, {"key": "pr#28473", "n": 28473, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "feat: support indicating prefix token of chat template", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/28473"}, {"key": "issue#28462", "n": 28462, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Move layer_idx from a layer property to function argument.", "action": "closed", "comment": "As @gante explained in https://github.com/huggingface/transformers/issues/28462#issuecomment-2293787664, the layer index is tied to the layer (it maps to specific weights and the KV cache), so it stays a module attribute; for layer-stacking experiments, update `layer.self_attn.layer_idx` after reordering. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/28462"}, {"key": "issue#28441", "n": 28441, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Proposal for Adding a New Scheduler Strategy for Language Model Pretraining", "action": "closed", "comment": "A warmup + cosine decay with a minimum learning rate is available as the `cosine_with_min_lr` scheduler (#29341), and warmup-stable-decay was added in #30231. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/28441"}, {"key": "issue#28388", "n": 28388, "kind": "issue", "group": "Open issues", "category": "feature", "area": "models", "type": "feature", "title": "How to use an efficient encoder as shared EncoderDecoderModel?", "action": "commented", "comment": "ELECTRA can already be used as an `EncoderDecoderModel` decoder via `ElectraForCausalLM` (it's in `MODEL_FOR_CAUSAL_LM_MAPPING_NAMES`), so `EncoderDecoderModel.from_encoder_decoder_pretrained(\"google/electra-base-discriminator\", \"google/electra-base-discriminator\", tie_encoder_decoder=True)` should work. ALBERT, DistilBERT and DeBERTa still lack a `*ForCausalLM` with cross-attention; adding one would mean following `BertLMHeadModel`/`ElectraForCausalLM` (cross-attention in the layer, `is_decoder`/`add_cross_attention` config flags) and registering it in `src/transformers/models/auto/modeling_auto.py`.", "author": "Bachstelze", "idle_days": 0, "models": "encoder_decoder", "url": "https://github.com/huggingface/transformers/issues/28388"}, {"key": "issue#28372", "n": 28372, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Support setting multiple adapters", "action": "closed", "comment": "Multiple active adapters are supported at the model level: load them with `model.load_adapter(...)`, call `model.set_adapter([\"adapter_a\", \"adapter_b\"])` (see `src/transformers/integrations/peft.py`, #26407), then pass that model to `pipeline(...)`. Closing, feel free to reopen if this doesn't work for your case.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/28372"}, {"key": "issue#28368", "n": 28368, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "[Flax] Migration from frozen to regular dicts with v0.7.1+", "action": "closed", "comment": "TensorFlow and Flax/JAX support was removed in v5 (#40760), so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/28368"}, {"key": "issue#28356", "n": 28356, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "[generation] Exact Search Decoding", "action": "closed", "comment": "As @gante said earlier, there wasn't enough demand to add exact search to core. It can be shared as a Hub `custom_generate` repo that anyone can load with `model.generate(custom_generate=\"<repo>\", trust_remote_code=True)`: https://huggingface.co/docs/transformers/main/en/generation_strategies#custom-generation-methods. Closing as not planned.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/28356"}, {"key": "issue#28319", "n": 28319, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Allow gradient for generate()", "action": "closed", "comment": "As explained in https://github.com/huggingface/transformers/issues/28319#issuecomment-1875030213, `generate()` is for inference only; training needs custom sampling logic (e.g. TRL: https://huggingface.co/docs/trl). Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/28319"}, {"key": "issue#28301", "n": 28301, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "[i18n-en] Translating docs to Japanese", "action": "closed", "comment": "Thanks to everyone who contributed: `docs/source/ja` on main now has 150+ translated pages! This tracker's list is outdated, so closing it; new Japanese translation PRs are still welcome following https://github.com/huggingface/transformers/blob/main/docs/TRANSLATING.md.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/28301"}, {"key": "issue#28285", "n": 28285, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add Mixture of Tokens model", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [jaszczur/mixture_of_tokens](https://huggingface.co/jaszczur/mixture_of_tokens)\n- Downloads (last 30d): 27 \u00b7 Likes: 1 \u00b7 Params: ~573M \u00b7 License: apache-2.0\n- Released: 2024-05 \u00b7 Remote code on the Hub: yes\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "chedatomasz", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/28285"}, {"key": "pr#28250", "n": 28250, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Fix model code to accurately convert fairseq wav2vec2 model", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/28250"}, {"key": "issue#28218", "n": 28218, "kind": "issue", "group": "Open issues", "category": "bug", "area": "tokenization", "type": "bug", "title": "Tokenizer adds an additional space after the added token", "action": "commented; labelled: bug", "comment": "This still reproduces on main (`'abcdgym'` decodes as `'abcd gym'`). The cause is the `Metaspace(prepend_scheme=\"always\")` pre-tokenizer/decoder in `src/transformers/models/nllb/tokenization_nllb.py`, which prepends `\u2581` to every split, including the text right after an added token. SeamlessM4T in `tokenization_seamless_m4t.py` already uses `prepend_scheme=\"first\"`; making NLLB do the same (behind `legacy`-style handling for backward compatibility, see `_get_prepend_scheme` in `convert_slow_tokenizer.py`) with a test would fix it.", "author": "kitkhai", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/28218"}, {"key": "issue#28160", "n": 28160, "kind": "issue", "group": "Open issues", "category": "feature", "area": "attention / kernels", "type": "feature", "title": "[Flash Attention 2] Performance improvement", "action": "labelled: Feature request", "comment": "", "author": "li-plus", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/28160"}, {"key": "issue#28121", "n": 28121, "kind": "issue", "group": "Open issues", "category": "feature", "area": "pipelines", "type": "feature", "title": "Add StyleTTS 2 to HF Transformers Pipeline", "action": "commented", "comment": "StyleTTS 2 (and Kokoro, which shares its architecture) isn't in the library yet. There was a PR, #35790, that is now closed; anyone wanting to pick it up can start from it and follow https://huggingface.co/docs/transformers/main/en/add_new_model with a modular file (https://huggingface.co/docs/transformers/main/en/modular_transformers). Recent TTS additions under `src/transformers/models/` are good templates for the `text-to-audio` pipeline integration.", "author": "fakerybakery", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/28121"}, {"key": "issue#28079", "n": 28079, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Expose `gradient_as_bucket_view` as training argument for `DDP`", "action": "commented", "comment": "Still not exposed: `Trainer` builds `DistributedDataParallelKwargs` from `ddp_find_unused_parameters`, `ddp_bucket_cap_mb`, `ddp_broadcast_buffers` and `ddp_static_graph` in `Trainer._build_accelerator_args` (`src/transformers/trainer.py`). Adding a `ddp_gradient_as_bucket_view` field next to `ddp_static_graph` in `training_args.py` and forwarding it there is a small, self-contained PR, welcome if you'd like to open it.", "author": "chiragjn", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/28079"}, {"key": "issue#28072", "n": 28072, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Can i convert open-clip trained models (.pt) using code \u201csrc/transformers/models/clip/convert_clip_original_pytorch_to_hf.py\u201d \uff1f", "action": "commented", "comment": "This isn't a model request: `convert_clip_original_pytorch_to_hf.py` targets the OpenAI CLIP format. For open_clip checkpoints, the easiest path is to push them to the Hub with `open_clip` (`push_to_hf_hub`) and load them with `open_clip`, or convert the state dict to CLIP key names manually.", "author": "jzssz", "idle_days": 0, "models": "clip", "url": "https://github.com/huggingface/transformers/issues/28072"}, {"key": "issue#27975", "n": 27975, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "ImageToTextPipeline does not support InstructBlip Models", "action": "closed", "comment": "The `image-to-text` pipeline was removed in v5 (#43306) and now redirects to `image-text-to-text`, which supports InstructBLIP. Closing, but please open a new issue if something still fails there!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/27975"}, {"key": "issue#27961", "n": 27961, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "CLIPTokenizer (and others based on the same telephoned OpenAI code) incorrect tokenize 1138 out of 34483 words that have an exact match in vocab", "action": "closed", "comment": "As @ArthurZucker explained (https://github.com/huggingface/transformers/issues/27961#issuecomment-1856131475), BPE doesn't guarantee one token per vocab word, and the vocab entries in the repro are byte-level encoded, so they need decoding (e.g. with `tokenizers.decoders.ByteLevel()`) before re-tokenizing. Closing as this matches the original tokenizer's behavior.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/27961"}, {"key": "issue#27869", "n": 27869, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Adding truncation to text-generation pipeline", "action": "closed", "comment": "This has been added in #28362 and is available on main. Closing, but feel free to reopen if something is still missing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/27869"}, {"key": "issue#27826", "n": 27826, "kind": "issue", "group": "Open issues", "category": "feature", "area": "processors (image/video/audio)", "type": "feature", "title": "Add LayoutLMProcessor", "action": "commented", "comment": "Still not on main: `src/transformers/models/layoutlm/` has no processor, and the earlier PR #27921 was closed for inactivity. A contributor can revive it by adding `processing_layoutlm.py` modeled on `src/transformers/models/layoutlmv2/processing_layoutlmv2.py` (OCR via the v2 image processor + tokenizer, without returning `pixel_values`), registering it in `models/auto/processing_auto.py`, and then simplifying the LayoutLM branch in `src/transformers/pipelines/document_question_answering.py`.", "author": "gau-nernst", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/27826"}, {"key": "issue#27824", "n": 27824, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Could we Add Linear Projection Layer in pre-trained model?", "action": "\u2014", "comment": "", "author": "DopeorNope-Lee", "idle_days": 1051, "models": "", "url": "https://github.com/huggingface/transformers/issues/27824"}, {"key": "pr#27768", "n": 27768, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[WIP] Uniformize processors in text+image multimodal models.", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/27768"}, {"key": "pr#27759", "n": 27759, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Add the CRATE (Coding RATE) backbone model", "action": "closed", "comment": "Thanks for the contribution! For architectures like this we recommend sharing the model on the Hub with custom code (https://huggingface.co/docs/transformers/custom_models), which lets you ship it without waiting on library review. Closing for now, but feel free to reopen if the model gains wide adoption.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/27759"}, {"key": "issue#27752", "n": 27752, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "TECO - Temporally Consistent Transformers for Video Generation", "action": "commented", "comment": "No public weights on the Hub as of today. TECO weights are released on GitHub/Google Drive only.", "author": "asartran", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/27752"}, {"key": "pr#27734", "n": 27734, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "Deberta can now be exported to TorchScript ", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #22105, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/27734"}, {"key": "issue#27730", "n": 27730, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Add flag for easily finetuning heads / linear probing to AutoModelforSequenceClassification", "action": "closed", "comment": "Closing this one: no activity since 2023-11-27, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/27730"}, {"key": "pr#27720", "n": 27720, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add common processor tests", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/27720"}, {"key": "issue#27712", "n": 27712, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Add support for llama.cpp", "action": "closed", "comment": "GGUF checkpoints load directly with `from_pretrained(..., gguf_file=...)`, and since #47779 supported architectures can keep weights packed and run llama.cpp's own kernels (`ggml-org/ggml-quantization`, `ggml-attn`) via the `kernels` package, see https://huggingface.co/docs/transformers/main/en/quantization/gguf. Closing as this covers using llama.cpp's quantization and kernels from transformers.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/27712"}, {"key": "pr#27671", "n": 27671, "kind": "PR", "group": "Closed PRs", "category": "superseded", "area": "", "type": "", "title": "[WIP][Splinter] Fixes #16627 by implementing the test cases for splinter", "action": "closed", "comment": "Thanks a lot for the PR! This was superseded by #32652, which landed on main and covers the same change, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/27671"}, {"key": "issue#27649", "n": 27649, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Adding support for lookahead decoding for autoregressive (decoder + encoder-decoder) models", "action": "closed", "comment": "As @gante explained after benchmarking in https://github.com/huggingface/transformers/issues/27649#issuecomment-1824621466, the gains didn't justify the model-level changes lookahead decoding needs in core. It can be shared as a Hub `custom_generate` repo loadable with `model.generate(custom_generate=\"<repo>\", trust_remote_code=True)` (https://huggingface.co/docs/transformers/main/en/generation_strategies). Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/27649"}, {"key": "issue#27627", "n": 27627, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Unable to register own custom tokenizer", "action": "commented", "comment": "This isn't a model request. The linked repo [RANITBAG/CustomAItokenizer](https://huggingface.co/RANITBAG/CustomAItokenizer) has no downloads. Custom tokenizers can be shared with remote code following the [custom models guide](https://huggingface.co/docs/transformers/main/en/custom_models), and questions are best asked on the [forum](https://discuss.huggingface.co/).", "author": "Ranitbag007", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/27627"}, {"key": "issue#27609", "n": 27609, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Extend Chat Template Tokenization for Training/Finetuning", "action": "closed", "comment": "This has been added in #30650 and is available on main. Closing, but feel free to reopen if something is still missing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/27609"}, {"key": "pr#27597", "n": 27597, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Paged Attention Based the Latest Cache Design", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/27597"}, {"key": "issue#27577", "n": 27577, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Co-Pilot Store in Huggingface", "action": "closed", "comment": "Thanks for the idea! This is a Hub/product request rather than something for the transformers library (HuggingChat already offers community assistants), so closing here.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/27577"}, {"key": "issue#27453", "n": 27453, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Audio-MAE - ViTMAE for audio", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [gaunernst/vit_base_patch16_1024_128.audiomae_as2m](https://huggingface.co/gaunernst/vit_base_patch16_1024_128.audiomae_as2m)\n- Downloads (last 30d): 2.0k \u00b7 Likes: 1 \u00b7 Params: ~86M \u00b7 License: cc-by-4.0\n- Released: 2023-12 \u00b7 Remote code on the Hub: no\n- Note: this is a timm-format port, loadable today via `TimmWrapperModel`.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "justinluong", "idle_days": 0, "models": "vit_mae", "url": "https://github.com/huggingface/transformers/issues/27453"}, {"key": "issue#27427", "n": 27427, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "cannot recall the model after registering it using AutoModelForCausalLM.register", "action": "commented", "comment": "This is a usage question rather than a new model request. Registering with `AutoModelForCausalLM.register` doesn't add the class to the `transformers` namespace: import `CustomAI` from your own module, or share it via [custom code on the Hub](https://huggingface.co/docs/transformers/main/en/custom_models). No public weights on the Hub as of today.", "author": "Ranitbag007", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/27427"}, {"key": "issue#27308", "n": 27308, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add ReCag model", "action": "commented", "comment": "No public weights to evaluate yet: [Recag/RECag](https://huggingface.co/Recag/RECag) has 21 downloads (last 30d) and no model card or paper. For a personal model, share it with remote code on the Hub ([custom models guide](https://huggingface.co/docs/transformers/main/en/custom_models)) instead of adding it to the library.", "author": "Ranitbag007", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/27308"}, {"key": "issue#27289", "n": 27289, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "[Efficiency] Decoding can be made faster by not converting special tokens to ids for each token.", "action": "closed", "comment": "Fixed on main: the slow-tokenizer decode path now builds `ids_to_skip = set(self.all_special_ids)` once before the loop instead of per token (see #37773 and `src/transformers/tokenization_python.py`). Thanks for the report!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/27289"}, {"key": "pr#27259", "n": 27259, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add NucleusX Model", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/27259"}, {"key": "issue#27238", "n": 27238, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Infomer with different output_size", "action": "closed", "comment": "As @kashif suggested (https://github.com/huggingface/transformers/issues/27238#issuecomment-1889482745), PatchTSMixer handles multivariate forecasting (https://huggingface.co/docs/transformers/main/en/model_doc/patchtsmixer). Closing this.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/27238"}, {"key": "issue#27088", "n": 27088, "kind": "issue", "group": "Open issues", "category": "docs", "area": "docs", "type": "docs", "title": "[i18n-TR] Translating docs to Turkish", "action": "\u2014", "comment": "", "author": "mertyyanik", "idle_days": 200, "models": "", "url": "https://github.com/huggingface/transformers/issues/27088"}, {"key": "issue#27031", "n": 27031, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Microsoft's GLIP Grounding Language Image Pretraining", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [GLIPModel/GLIP](https://huggingface.co/GLIPModel/GLIP)\n- Downloads (last 30d): 0 \u00b7 Likes: 26 \u00b7 Params: n/a \u00b7 License: n/a\n- Released: 2023-06 \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "ethansmith2000", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/27031"}, {"key": "issue#26994", "n": 26994, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "The current architecture does not support Flash Attention 2.0 - distilgpt2, gpt2-medium ", "action": "closed", "comment": "Flash Attention 2 support for GPT-2 (incl. distilgpt2/gpt2-medium) landed in #29226. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/26994"}, {"key": "issue#26862", "n": 26862, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Support reverse_prompt", "action": "closed", "comment": "This exists now: pass `stop_strings=[\"USER:\"]` (and `tokenizer=tokenizer`) to `generate()`, added in #28932. See https://huggingface.co/docs/transformers/main/en/generation_strategies", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/26862"}, {"key": "issue#26840", "n": 26840, "kind": "issue", "group": "Closed issues", "category": "obsolete", "area": "", "type": "", "title": "log interval", "action": "closed", "comment": "The `*_no_trainer` example scripts this targets were removed from the repo in #49357. With `Trainer`, step-based logging/evaluation is already available via `logging_strategy=\"steps\"`, `logging_steps`, `eval_strategy=\"steps\"` and `eval_steps`. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/26840"}, {"key": "issue#26803", "n": 26803, "kind": "issue", "group": "Open issues", "category": "docs", "area": "docs", "type": "docs", "title": "[i18n-zh] Translating docs to Chinese (Simplified)", "action": "\u2014", "comment": "", "author": "jiaqiw09", "idle_days": 33, "models": "", "url": "https://github.com/huggingface/transformers/issues/26803"}, {"key": "issue#26802", "n": 26802, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "[Feature Request] We might need a function to change the sampler used in trainer dataloader", "action": "closed", "comment": "Custom samplers are supported by subclassing `Trainer` and overriding `_get_train_sampler` (`src/transformers/trainer.py`); for multipack-style efficiency, `DataCollatorWithFlattening` provides padding-free packing out of the box. Closing as these cover the use case.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/26802"}, {"key": "issue#26787", "n": 26787, "kind": "issue", "group": "Open issues", "category": "docs", "area": "docs", "type": "docs", "title": "[i18n-<hi>] Translating docs to <Hindi>", "action": "\u2014", "comment": "", "author": "hakunamatata1997", "idle_days": 721, "models": "", "url": "https://github.com/huggingface/transformers/issues/26787"}, {"key": "issue#26786", "n": 26786, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "[i18n-<te>] Translating docs to <Telugu>", "action": "closed", "comment": "Thanks for the contributions! The Telugu docs pages (#26828) were removed during the v5 docs cleanup (#40999); `i18n/README_te.md` remains. Since the English docs have been reorganized, this checklist is outdated, so closing; to restart, please open a new issue with the translation template and follow https://github.com/huggingface/transformers/blob/main/docs/TRANSLATING.md.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/26786"}, {"key": "issue#26781", "n": 26781, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "MaskCLIP", "action": "commented", "comment": "No public weights on the Hub as of today. Only an empty placeholder repo exists (`nebula/MaskCLIP-weights`, 0 downloads); official weights are on GitHub.", "author": "sushmanthreddy", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/26781"}, {"key": "issue#26726", "n": 26726, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Tutorial on implementing tree of thoughts(ToT) framework using a model ", "action": "closed", "comment": "As suggested by @gante in https://github.com/huggingface/transformers/issues/26726#issuecomment-1776906537, a Tree of Thoughts tutorial fits best as a community blog post, and as noted in https://github.com/huggingface/transformers/issues/26726#issuecomment-2004635543 this was to close if work stalled. Closing, thanks to everyone who contributed!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/26726"}, {"key": "issue#26706", "n": 26706, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Add an option to decide whether to store the checkpoint and rng_state.", "action": "closed", "comment": "This has been added in #27652 and is available on main. Closing, but feel free to reopen if something is still missing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/26706"}, {"key": "issue#26690", "n": 26690, "kind": "issue", "group": "Open issues", "category": "feature", "area": "pipelines", "type": "feature", "title": "[RFC] Updating pipeline models", "action": "\u2014", "comment": "", "author": "Rocketknight1", "idle_days": 439, "models": "", "url": "https://github.com/huggingface/transformers/issues/26690"}, {"key": "pr#26657", "n": 26657, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add Fast model ", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/26657"}, {"key": "issue#26647", "n": 26647, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "add T5 as decoder only", "action": "commented", "comment": "This is a feature request rather than a new model: there are no public T5-decoder-only checkpoints on the Hub. You can keep the subclass in your own code or share it with remote code on the Hub ([custom models guide](https://huggingface.co/docs/transformers/main/en/custom_models)).", "author": "Biyani404198", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/26647"}, {"key": "issue#26628", "n": 26628, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "[New Model] Retrieval-based Voice Conversion", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [lj1995/VoiceConversionWebUI](https://huggingface.co/lj1995/VoiceConversionWebUI)\n- Downloads (last 30d): 0 \u00b7 Likes: 1231 \u00b7 Params: n/a \u00b7 License: mit\n- Released: 2023-01 \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "wfjsw", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/26628"}, {"key": "pr#26520", "n": 26520, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "[WIP] Add Support for masking output using a Context-Free-Grammar", "action": "closed", "comment": "Thanks for this work and your patience! Constrained decoding was moved out of core `generate` in v5; this fits much better as a Hub custom generation method or with dedicated libraries like outlines. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/26520"}, {"key": "issue#26504", "n": 26504, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add LISA Model", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [xinlai/LISA-7B-v1](https://huggingface.co/xinlai/LISA-7B-v1), [xinlai/LISA-7B-v1-explanatory](https://huggingface.co/xinlai/LISA-7B-v1-explanatory)\n- Downloads (last 30d): 877 \u00b7 Likes: 8 \u00b7 Params: ~7B (LLaVA + SAM) \u00b7 License: not specified\n- Released: 2023-08 \u00b7 Remote code on the Hub: no (original-repo code)\n- Building blocks already in transformers: `llava`, `sam`.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "Dev-Khant", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/26504"}, {"key": "issue#26501", "n": 26501, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add FAST model", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: no FAST checkpoint on the Hub; weights are only on the [original GitHub repo](https://github.com/czczup/FAST)\n- Released: 2021-11 (paper) \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "raghavanone", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/26501"}, {"key": "issue#26474", "n": 26474, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "compute_metrics with causal LM training", "action": "labelled: Feature request", "comment": "", "author": "steremma", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/26474"}, {"key": "pr#26441", "n": 26441, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[T5] lm_head weights initialization: set variance to reciprocal of hidden dim", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/26441"}, {"key": "issue#26412", "n": 26412, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "How to run Trainer + DeepSpeed + Zero3 + PEFT ", "action": "closed", "comment": "Trainer + DeepSpeed ZeRO-3 + PEFT, including QLoRA, is supported since #29587; see the PEFT guide https://huggingface.co/docs/peft/accelerate/deepspeed for configs. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/26412"}, {"key": "issue#26388", "n": 26388, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Implementation of Git-Rebasin to Huggingface models", "action": "\u2014", "comment": "", "author": "NamburiSrinath", "idle_days": 1090, "models": "", "url": "https://github.com/huggingface/transformers/issues/26388"}, {"key": "issue#26360", "n": 26360, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add ProPainter to transformers", "action": "commented", "comment": "No public weights on the Hub as of today. Official weights are on GitHub releases only; community copies exist (e.g. `ruffy369/propainter`, 42 downloads/30d).", "author": "shauray8", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/26360"}, {"key": "issue#26350", "n": 26350, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Community contribution: Adding Flash Attention 2 support for more architectures", "action": "closed", "comment": "Thanks everyone who contributed! Flash Attention is now provided through the unified attention interface, and the remaining items on this list are done (e.g. CLIP in #31940, BERT-based models in #38301). New models get it automatically by using `ALL_ATTENTION_FUNCTIONS`; see https://huggingface.co/docs/transformers/main/en/attention_interface. If a specific architecture is still missing support, please open a dedicated issue.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/26350"}, {"key": "issue#26298", "n": 26298, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Synk_Dive", "action": "commented", "comment": "No public weights on the Hub as of today (no paper, code or checkpoint for Synk Dive could be found).", "author": "yasmws", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/26298"}, {"key": "issue#26289", "n": 26289, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Make the skip_batches_dataloader used in the Trainer customizable", "action": "commented", "comment": "Still open for contributions, and @SunMarc agreed with the approach. On main, `Trainer` calls accelerate's `skip_first_batches(train_dataloader, steps_trained_in_current_epoch)` inline in the training loop (`src/transformers/trainer.py`); moving that into an overridable method (e.g. `Trainer.get_resumed_dataloader(dataloader, num_batches)`) with a small test in `tests/trainer/` would solve it. See https://huggingface.co/docs/transformers/main/en/trainer_customize for the subclassing pattern.", "author": "bruno-hays", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/26289"}, {"key": "issue#26269", "n": 26269, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "AugViT TensorFlow implementation ", "action": "\u2014", "comment": "", "author": "ushareng", "idle_days": 1123, "models": "", "url": "https://github.com/huggingface/transformers/issues/26269"}, {"key": "issue#26265", "n": 26265, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "\ud83c\udf10 [i18n-zh-hant] Translating docs to Traditional Chinese (zh-hant)", "action": "closed", "comment": "Thanks to everyone who volunteered! No `docs/source/zh-hant` translation landed on main (only `i18n/README_zh-hant.md`), so closing for now; anyone can restart it with a new issue following https://github.com/huggingface/transformers/blob/main/docs/TRANSLATING.md.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/26265"}, {"key": "issue#26256", "n": 26256, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "A new category for recsys", "action": "closed", "comment": "As @LysandreJik said (https://github.com/huggingface/transformers/issues/26256#issuecomment-1746785376), we don't plan to add recommendation-system capabilities to transformers for now, but recsys models and utilities can still be shared on the Hub. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/26256"}, {"key": "issue#26144", "n": 26144, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "UMT5 incredibly slow in generating", "action": "closed; labelled: bug", "comment": "The relative position bias in UMT5 is now computed only for the current query positions during generation (#44759), rather than for the full sequence on every step, which was the main source of the slowdown reported here. Closing, but if you still see a big gap against mT5 on current `main`, please share an updated benchmark and we'll reopen.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/26144"}, {"key": "issue#26062", "n": 26062, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "More explicit bounding box formats", "action": "closed", "comment": "The formats are now documented in the image processors: inputs are COCO-style annotations (selected by the `format` argument, converted internally to normalized `(center_x, center_y, width, height)`), and `post_process_object_detection` returns `(top_left_x, top_left_y, bottom_right_x, bottom_right_y)` in absolute pixels (see `src/transformers/models/detr/image_processing_detr.py`). `center_to_corners_format`/`corners_to_center_format` in `image_transforms.py` convert between them.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/26062"}, {"key": "issue#25925", "n": 25925, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Edit train_dataloader in callbacks", "action": "closed", "comment": "Swapping the dataloader from a callback isn't supported, but `Trainer` calls `set_epoch(epoch)` on the train dataloader at the start of each epoch (`src/transformers/trainer.py`), so a dataset/sampler with a `set_epoch` method, or an overridden `get_train_dataloader`, can vary data per epoch. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/25925"}, {"key": "issue#25904", "n": 25904, "kind": "issue", "group": "Open issues", "category": "feature", "area": "cache / masks", "type": "feature", "title": "Support for the custom kernel for Mask2Former and OneFormer", "action": "commented", "comment": "Custom kernels are now loaded from the Hub rather than compiled in-repo: Deformable DETR does it via `@use_kernel_forward_from_hub(\"MultiScaleDeformableAttention\")` in `src/transformers/models/deformable_detr/modeling_deformable_detr.py`. Mask2Former (`Mask2FormerPixelDecoderEncoderMultiscaleDeformableAttention`) and OneFormer still use the pure-PyTorch `multi_scale_deformable_attention`; a PR refactoring them to reuse the Deformable DETR module (ideally via modular, see https://huggingface.co/docs/transformers/main/en/modular_transformers and https://huggingface.co/docs/transformers/main/en/kernels) would be welcome.", "author": "function2-llx", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/25904"}, {"key": "issue#25778", "n": 25778, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Support for context-free-grammars (CFG) to constrain model output", "action": "closed", "comment": "As @gante stated (https://github.com/huggingface/transformers/issues/25778#issuecomment-3179305864), we're intentionally not adding this to core since outlines, transformers-CFG and syncode already provide it on top of `generate`. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/25778"}, {"key": "issue#25723", "n": 25723, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "NEW Model GPT-JX", "action": "commented", "comment": "No public weights on the Hub as of today. `alien-ai/gpt-jx-3b` is no longer publicly accessible.", "author": "alignment-ai", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/25723"}, {"key": "issue#25677", "n": 25677, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Transformers documentation translation to Macedonian", "action": "closed", "comment": "Closing this one: no activity since 2023-08-23, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/25677"}, {"key": "issue#25654", "n": 25654, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Make \ud83e\udd17Transformers tests device agnostic", "action": "closed", "comment": "This landed in #25870: `testing_utils.py` now has `@require_torch_accelerator`, device-agnostic backend helpers (`backend_empty_cache`, `BACKEND_DEVICE_COUNT`, ...) and a `TRANSFORMERS_TEST_DEVICE_SPEC` file to plug in custom devices. Closing as completed, thanks for driving it!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/25654"}, {"key": "issue#25620", "n": 25620, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "TimeSeriesTransformerForPrediction model unused parameters Runtime error in Distributed environment", "action": "closed", "comment": "Closing this one: no activity since 2024-05-01, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/25620"}, {"key": "issue#25526", "n": 25526, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add FastViT model", "action": "commented", "comment": "FastViT is usable from transformers today via `timm_wrapper`: `AutoModelForImageClassification.from_pretrained(\"timm/fastvit_t8.apple_in1k\")`.\n\nSome details for anyone who'd like to pick this up:\n- Hub: [timm/fastvit_t8.apple_in1k](https://huggingface.co/timm/fastvit_t8.apple_in1k), [timm/fastvit_sa36.apple_in1k](https://huggingface.co/timm/fastvit_sa36.apple_in1k)\n- Downloads (last 30d): 54,203 \u00b7 Likes: 2 \u00b7 Params: ~4M \u00b7 License: other\n- Released: 2023-08 \u00b7 Remote code on the Hub: no\n- These are the `timm` ports of Apple's official weights; there are no Apple-org checkpoints on the Hub.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "atturaioe", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/25526"}, {"key": "pr#25474", "n": 25474, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Add support for BLIP-2 multimodal feature extraction", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/25474"}, {"key": "issue#25455", "n": 25455, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "The length used by length_penalty during beam_search is not correct when input batch size > 1", "action": "closed", "comment": "Fixed by the beam score rework in #27351 and #27808: length penalty now uses the number of generated tokens (`cur_len + 1 - decoder_prompt_len` in `generation/utils.py`), so left-padding in the prompt no longer changes the penalty across a batch.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/25455"}, {"key": "issue#25355", "n": 25355, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "Add Flax diverse group search", "action": "closed", "comment": "TensorFlow and Flax/JAX support was fully removed in v5 (#40760), so this no longer applies. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/25355"}, {"key": "issue#25307", "n": 25307, "kind": "issue", "group": "Open issues", "category": "feature", "area": "cache / masks", "type": "feature", "title": "Return updated attention mask from Wav2Vec 2.0", "action": "labelled: Feature request", "comment": "", "author": "gau-nernst", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/25307"}, {"key": "issue#25300", "n": 25300, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Add zero-shot classification task for BLIP-2", "action": "closed", "comment": "This has been added in #29261 and is available on main. Closing, but feel free to reopen if something is still missing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/25300"}, {"key": "issue#25293", "n": 25293, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "MassFormer", "action": "commented", "comment": "Some details for anyone who'd like to pick this up: no official MassFormer checkpoints are on the Hugging Face Hub as of today, so adding it would first need converted weights.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "yunyicheng", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/25293"}, {"key": "issue#25253", "n": 25253, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "RWKV-WORLD-4", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [BlinkDL/rwkv-4-world](https://huggingface.co/BlinkDL/rwkv-4-world) (raw .pth), [RWKV/rwkv-4-world-3b](https://huggingface.co/RWKV/rwkv-4-world-3b) (HF format)\n- Downloads (last 30d): 25,035 \u00b7 Likes: 213 \u00b7 License: apache-2.0\n- Released: 2023-05 \u00b7 Remote code on the Hub: no\n\nThe RWKV-4 architecture is already supported as `rwkv` (#22797); what's missing is the World tokenizer. Separately, RWKV-7 has an open PR (#47780).\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "CosmoLM", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/25253"}, {"key": "issue#25247", "n": 25247, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Enable use of best epoch in Trial, with early stopping, during hyperparameter search", "action": "commented", "comment": "This is still only supported for the Ray backend, via the `RAY_SCOPE` env var (e.g. `RAY_SCOPE=all`, read in `src/transformers/integrations/integration_utils.py`). For other backends, the objective comes from `Trainer._report_to_hp_search` / the `run_hp_search_*` functions in the same file; a PR adding a 'best' scope there (e.g. using `state.best_metric` when `load_best_model_at_end=True`) would be welcome, see https://huggingface.co/docs/transformers/main/en/hpo_train.", "author": "antonioalegria", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/25247"}, {"key": "issue#25243", "n": 25243, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "RetNet model support", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [fla-hub/retnet-1.3B-100B](https://huggingface.co/fla-hub/retnet-1.3B-100B), [parsee-mizuhashi/retnet-tiny-wikitext-undertrained](https://huggingface.co/parsee-mizuhashi/retnet-tiny-wikitext-undertrained)\n- Downloads (last 30d): 1,365 \u00b7 Likes: 3 \u00b7 Params: ~1.4B \u00b7 License: mit\n- Released: 2024-04 \u00b7 Remote code on the Hub: no\n- There are still no official Microsoft RetNet checkpoints; the `fla-hub` models are community-trained and run with the `fla` library.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "yoinked-h", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/25243"}, {"key": "issue#25178", "n": 25178, "kind": "issue", "group": "Open issues", "category": "feature", "area": "models", "type": "feature", "title": "BERT: TensorFlow Model Garden Conversion scripts", "action": "labelled: Feature request", "comment": "", "author": "stefan-it", "idle_days": -1, "models": "bert", "url": "https://github.com/huggingface/transformers/issues/25178"}, {"key": "issue#25147", "n": 25147, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Add PromptTemplate and allow for default PromptTemplate in model configuration", "action": "closed", "comment": "This has been added in #25323 and is available on main. Closing, but feel free to reopen if something is still missing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/25147"}, {"key": "issue#25130", "n": 25130, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "an inplace operation preventing TorchDistributor training", "action": "closed", "comment": "The in-place add in `BertEmbeddings` is gone on main: it now does `embeddings = embeddings + position_embeddings` (`src/transformers/models/bert/modeling_bert.py`, changed in #41170), so this should work with TorchDistributor on v5. Closing; please open a new issue if you hit another in-place op.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/25130"}, {"key": "issue#25123", "n": 25123, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add Vocos model", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [charactr/vocos-mel-24khz](https://huggingface.co/charactr/vocos-mel-24khz), [charactr/vocos-encodec-24khz](https://huggingface.co/charactr/vocos-encodec-24khz)\n- Downloads (last 30d): 3.0M \u00b7 Likes: 51 \u00b7 Params: n/a \u00b7 License: mit\n- Released: 2023-06 \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "ylacombe", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/25123"}, {"key": "issue#24836", "n": 24836, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Pipeline feature request for min_new_tokens", "action": "closed", "comment": "The `image-to-text` pipeline was replaced by `image-text-to-text` in #43306, which forwards generation parameters, e.g. `pipe(images=..., text=..., generate_kwargs={\"min_new_tokens\": 20})`; `text-generation` also accepts `min_new_tokens` (`src/transformers/pipelines/text_generation.py`). Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/24836"}, {"key": "issue#24711", "n": 24711, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "Initialize Flax model params on CPU ", "action": "closed", "comment": "TensorFlow and Flax/JAX support was removed in v5 (#40760), so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/24711"}, {"key": "issue#24659", "n": 24659, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add HyenaDNA model", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [LongSafari/hyenadna-large-1m-seqlen-hf](https://huggingface.co/LongSafari/hyenadna-large-1m-seqlen-hf), [LongSafari/hyenadna-tiny-1k-seqlen-hf](https://huggingface.co/LongSafari/hyenadna-tiny-1k-seqlen-hf)\n- Downloads (last 30d): 7,871 \u00b7 Likes: 26 \u00b7 Params: ~55M (tiny: ~0.45M) \u00b7 License: bsd-3-clause\n- Released: 2023-11 \u00b7 Remote code on the Hub: yes\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "tanaymeh", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/24659"}, {"key": "issue#24295", "n": 24295, "kind": "issue", "group": "Open issues", "category": "feature", "area": "trainer", "type": "feature", "title": "Add training support for EnCodec", "action": "commented", "comment": "Still open for contributions: the last attempt (#33956) was closed because it targeted v4, with an invite to reopen on main. A new PR would add the losses (reconstruction, commitment, and optionally the adversarial/feature-matching terms) to `EncodecModel` in `src/transformers/models/encodec/modeling_encodec.py`, ideally with a small convergence run (wandb logs) as ArthurZucker asked above. https://huggingface.co/docs/transformers/main/en/contributing has the PR checklist.", "author": "ArthurZucker", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/24295"}, {"key": "issue#24235", "n": 24235, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add SPTSv2", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: no SPTSv2 checkpoint on the Hub; weights are only on the [original GitHub repo](https://github.com/bytedance/SPTSv2)\n- Released: 2023-01 (paper) \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "NielsRogge", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/24235"}, {"key": "issue#23880", "n": 23880, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Whisper with Elastic Weight Consolidation", "action": "closed", "comment": "Closing this one: no activity since 2023-10-12, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/23880"}, {"key": "issue#23846", "n": 23846, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add LaVIN model", "action": "commented", "comment": "Some details for anyone who'd like to pick this up: no official LaVIN checkpoints are on the Hugging Face Hub as of today, so adding it would first need converted weights.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "tensorpro", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/23846"}, {"key": "issue#23814", "n": 23814, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Adding GPTNeoX (Tensorflow version)", "action": "commented", "comment": "TensorFlow support has been removed from transformers (PyTorch-only since v5), so a TF port of GPT-NeoX won't be added. The PyTorch model is supported as `gpt_neox`.", "author": "athleticcoder21", "idle_days": 0, "models": "gpt_neo, gpt_neox", "url": "https://github.com/huggingface/transformers/issues/23814"}, {"key": "issue#23675", "n": 23675, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "[i18n-ms, ISO 639-1] Translating docs to Malay", "action": "closed", "comment": "Thanks for kicking this off! No Malay translation landed (there is no `docs/source/ms` on main), so closing this tracker for now. A PR adding `docs/source/ms` following https://github.com/huggingface/transformers/blob/main/docs/TRANSLATING.md is welcome to restart it.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/23675"}, {"key": "issue#23648", "n": 23648, "kind": "issue", "group": "Open issues", "category": "bug", "area": "other", "type": "bug", "title": "Unexpected padding behaviour of `ClapFeatureExtractor`", "action": "commented; labelled: bug", "comment": "Still reproducible on main: in `src/transformers/models/clap/feature_extraction_clap.py`, `padding = padding if padding else self.padding` lets `padding=True` replace the default `\"repeatpad\"` with constant padding. A small PR mapping boolean `padding=True` to `self.padding` (plus a test in `tests/models/clap/test_feature_extraction_clap.py`) would fix it; with v5 processors you can also route padding to the tokenizer only via `processor(..., text_kwargs={\"padding\": True})`.", "author": "anmol-wiai", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/23648"}, {"key": "issue#23258", "n": 23258, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "Flaky Whisper PT-TF & PT-Flax Equivalence Test", "action": "closed", "comment": "TensorFlow and Flax/JAX support was removed in v5 (#40760), so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/23258"}, {"key": "issue#23226", "n": 23226, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "NSP Support for Zero-shot Text Classification Pipeline", "action": "closed", "comment": "Thanks for sharing the approach. We won't add an NSP mode to the core zero-shot pipeline, but this is a great fit for a custom pipeline shared on the Hub (subclass `Pipeline` and register it), see https://huggingface.co/docs/transformers/main/en/add_new_pipeline", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/23226"}, {"key": "issue#23145", "n": 23145, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Detr Models cannot be loaded with `device_map=\"auto\"`", "action": "closed", "comment": "This was fixed by #26849, which is now on main. Closing, but feel free to reopen if you still see it on the latest version!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/23145"}, {"key": "issue#22994", "n": 22994, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation ", "action": "closed", "comment": "As found in https://github.com/huggingface/transformers/issues/22994#issuecomment-1956924166, this in-place error comes from DDP buffer broadcasting; with `Trainer` you can set `TrainingArguments(ddp_broadcast_buffers=False)`, which is forwarded to DDP in `src/transformers/trainer.py`. Closing; please open a new issue with a reproducer on a current version if it still happens.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/22994"}, {"key": "issue#22923", "n": 22923, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Need support for Sentence Similarity Pipeline", "action": "closed", "comment": "As discussed above, sentence similarity is served by `sentence-transformers` (https://huggingface.co/sentence-transformers), and raw embeddings are available via `pipeline(\"feature-extraction\")`. We don't plan a dedicated pipeline in transformers, so closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/22923"}, {"key": "issue#22756", "n": 22756, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "TypeError: export() got an unexpected keyword argument 'preprocessor'", "action": "commented; comment corrected", "comment": "This isn't a model request: `transformers.onnx` has been removed. Export (ONNX included) now goes through [`transformers.exporters`](https://huggingface.co/docs/transformers/main/en/exporters).", "author": "susht3", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/22756"}, {"key": "issue#22685", "n": 22685, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Add SwiftFormer", "action": "closed", "comment": "This has been added in #22686 and is available on main. Closing, but feel free to reopen if something is still missing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/22685"}, {"key": "issue#22612", "n": 22612, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "Add `output_hidden_state` and `output_scores` to Flax generate", "action": "closed", "comment": "TensorFlow and Flax/JAX support was fully removed in v5 (#40760), so this no longer applies. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/22612"}, {"key": "pr#22608", "n": 22608, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[DO NOT MERGE] Add Crop Transformation", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/22608"}, {"key": "issue#22363", "n": 22363, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Multi-node training with Deepspeed hangs when `full_determinism = True`", "action": "closed", "comment": "As found in the thread, the hang came from a PyTorch bug fixed upstream in pytorch/pytorch#97843 (in releases after 2.0.1). With a current PyTorch, `full_determinism=True` (`enable_full_determinism` in `src/transformers/trainer_utils.py`) should no longer hang multi-node DeepSpeed runs; please reopen with a fresh repro if it does.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/22363"}, {"key": "issue#22315", "n": 22315, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add MegatronT5", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [nvidia/nemo-megatron-t5-3B](https://huggingface.co/nvidia/nemo-megatron-t5-3B)\n- Downloads (last 30d): 116 \u00b7 Likes: 9 \u00b7 Params: n/a \u00b7 License: cc-by-4.0\n- Released: 2022-09 \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "eagle705", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/22315"}, {"key": "issue#22290", "n": 22290, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Native support of ChatGLM-6b", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [zai-org/chatglm-6b](https://huggingface.co/zai-org/chatglm-6b), [zai-org/chatglm2-6b](https://huggingface.co/zai-org/chatglm2-6b)\n- Downloads (last 30d): 2,791 \u00b7 Likes: 2,926 \u00b7 Params: ~6.2B \u00b7 License: ChatGLM-6B license (see model card)\n- Released: 2023-03 \u00b7 Remote code on the Hub: yes\n- `THUDM/*` now redirects to `zai-org/*`. chatglm2-6b is far more used (459,952 downloads/30d). Newer GLM generations are already supported natively (`glm`, `glm4`, ...), but ChatGLM-6B/2/3 still rely on remote code.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "xianbaoqian", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/22290"}, {"key": "issue#22221", "n": 22221, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "export clip to text encoder and image encoder two onnxs", "action": "commented; comment corrected", "comment": "This isn't a model request: export (ONNX included) now goes through [`transformers.exporters`](https://huggingface.co/docs/transformers/main/en/exporters). To get separate encoders, export `CLIPTextModelWithProjection` and `CLIPVisionModelWithProjection` on their own.", "author": "susht3", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/22221"}, {"key": "issue#22220", "n": 22220, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Positinal Encoding for T5 family of models", "action": "closed", "comment": "T5 uses relative position bias computed inside `T5Attention` (`compute_bias` in `src/transformers/models/t5/modeling_t5.py`), so to change it you can subclass/monkey-patch that module (https://huggingface.co/docs/transformers/main/en/how_to_hack_models) or pass `inputs_embeds` with your own embeddings. We don't plan a dedicated hook, so closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/22220"}, {"key": "issue#22178", "n": 22178, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add BEiTv3", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [timm/beit3_base_patch16_224.in22k_ft_in1k](https://huggingface.co/timm/beit3_base_patch16_224.in22k_ft_in1k)\n- Downloads (last 30d): 555 \u00b7 Likes: 0 \u00b7 Params: ~87M \u00b7 License: apache-2.0\n- Released: 2025-05 \u00b7 Remote code on the Hub: no\n- Note: the timm BEiT-3 vision checkpoints can already be loaded through `TimmWrapperModel`; the multimodal BEiT-3 is not supported.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "NielsRogge", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/22178"}, {"key": "issue#21777", "n": 21777, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "Add TensorFlow Whisper model for audio classification", "action": "closed", "comment": "TensorFlow and Flax/JAX support was removed in v5 (#40760), so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/21777"}, {"key": "issue#21682", "n": 21682, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "kros_test", "action": "commented", "comment": "No public weights on the Hub as of today. No implementation or checkpoint is linked either.", "author": "brigs1", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/21682"}, {"key": "issue#21668", "n": 21668, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add SeaFormer model", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: no official SeaFormer checkpoint on the Hub; weights are only on the [original GitHub repo](https://github.com/fudan-zvg/SeaFormer)\n- Released: 2023-01 (paper) \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "inderpreetsingh01", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/21668"}, {"key": "issue#21590", "n": 21590, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Project_Test01", "action": "\u2014", "comment": "", "author": "RutujaTalekar", "idle_days": 1345, "models": "", "url": "https://github.com/huggingface/transformers/issues/21590"}, {"key": "issue#21515", "n": 21515, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "add MaxViT [TF]", "action": "commented", "comment": "TensorFlow support has been removed from transformers (PyTorch-only since v5), so a TF MaxViT won't be added. MaxViT checkpoints on the Hub, e.g. [timm/maxvit_tiny_tf_224.in1k](https://huggingface.co/timm/maxvit_tiny_tf_224.in1k) (21,667 downloads/30d), already load in transformers via `TimmWrapperModel` / `AutoModelForImageClassification`.", "author": "awsaf49", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/21515"}, {"key": "issue#21456", "n": 21456, "kind": "issue", "group": "Open issues", "category": "docs", "area": "docs", "type": "docs", "title": "[i18n-fr] Translating docs to fr", "action": "\u2014", "comment": "", "author": "NoB0", "idle_days": 5, "models": "", "url": "https://github.com/huggingface/transformers/issues/21456"}, {"key": "issue#21079", "n": 21079, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "TokenGT", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [raman-ai/tokengt-base-lap-pcqm4mv2](https://huggingface.co/raman-ai/tokengt-base-lap-pcqm4mv2), [raman-ai/tokengt-base-orf-pcqm4mv2](https://huggingface.co/raman-ai/tokengt-base-orf-pcqm4mv2)\n- Downloads (last 30d): 19 \u00b7 Likes: 0 \u00b7 License: not specified\n- Released: 2023-03 \u00b7 Remote code on the Hub: no\n- These are contributor conversions, not official uploads. Note that the related `graphormer` model was deprecated in transformers for low usage.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "clefourrier", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/21079"}, {"key": "issue#20822", "n": 20822, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Train mobileBERT from scratch for other languages", "action": "\u2014", "comment": "", "author": "fabianbrandscheid", "idle_days": 1250, "models": "", "url": "https://github.com/huggingface/transformers/issues/20822"}, {"key": "issue#20545", "n": 20545, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "add MeMViT model", "action": "commented", "comment": "No public weights on the Hub as of today. MeMViT weights are on GitHub only.", "author": "fcakyon", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/20545"}, {"key": "issue#20392", "n": 20392, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add BART-LS", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: no BART-LS checkpoint on the Hub; weights are only on the [original GitHub repo](https://github.com/facebookresearch/bart_ls)\n- Released: 2022-09 (paper) \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "KMFODA", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/20392"}, {"key": "issue#20133", "n": 20133, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Compact Transformer", "action": "commented", "comment": "Some details for anyone who'd like to pick this up: no official Compact Convolutional Transformer checkpoints are on the Hugging Face Hub as of today, so adding it would first need converted weights.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "astariul", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/20133"}, {"key": "issue#20091", "n": 20091, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "OWL-ViT training / fine-tuning code", "action": "closed", "comment": "Closing in favour of #33664, which tracks the same request (OWL-ViT/OWLv2 detection training) and has a contributor assigned; please follow along there.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/20091"}, {"key": "issue#19730", "n": 19730, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Lite Transformer with Long-Short Range Attention", "action": "commented", "comment": "No public weights on the Hub as of today. Weights are on GitHub only.", "author": "astariul", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/19730"}, {"key": "issue#19511", "n": 19511, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "ERNIE and tensorflow2", "action": "commented", "comment": "ERNIE is supported in transformers as `ernie` (PyTorch). TensorFlow support has been removed from transformers as of v5, so a `TFErnieModel` won't be added.", "author": "Smile-L-up", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/19511"}, {"key": "issue#19123", "n": 19123, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "Adding TensorFlow port of LeViT", "action": "closed", "comment": "TensorFlow and Flax/JAX support was removed in v5 (#40760), so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/19123"}, {"key": "issue#19081", "n": 19081, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "add Unified-IO", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [allenai/uio2-large](https://huggingface.co/allenai/uio2-large), [allenai/uio2-xxl-bfloat16](https://huggingface.co/allenai/uio2-xxl-bfloat16)\n- Downloads (last 30d): 166 \u00b7 Likes: 2 \u00b7 Params: n/a \u00b7 License: n/a\n- Released: 2024-02 \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "thedarkzeno", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/19081"}, {"key": "issue#18817", "n": 18817, "kind": "issue", "group": "Closed issues", "category": "obsolete", "area": "", "type": "", "title": "Identifying backend compatibility versions", "action": "closed", "comment": "This survey is no longer relevant: TensorFlow and Flax support were removed in v5, and the supported PyTorch range is now defined directly in `setup.py` (currently `torch>=2.6`). Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/18817"}, {"key": "issue#18661", "n": 18661, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Refactor Pytorch `model.generate` method to work on TPU", "action": "closed", "comment": "The blocker mentioned here (a fixed-size KV cache) landed in #27931: `generate(..., cache_implementation=\"static\")` keeps shapes fixed, which is what XLA/torch.compile need. See https://huggingface.co/docs/transformers/main/en/google_tpu and https://huggingface.co/docs/transformers/main/en/perf_torch_compile. Closing; if you hit a specific XLA graph break with a static cache, please open a new issue with a repro.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/18661"}, {"key": "issue#18641", "n": 18641, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "Add TF VideoMAE", "action": "closed", "comment": "TensorFlow support was removed from Transformers in v5 (#40760), so we're closing this TF VideoMAE port request.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/18641"}, {"key": "issue#18636", "n": 18636, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "Support output_scores in XLA TF generate", "action": "closed", "comment": "TensorFlow and Flax/JAX support was fully removed in v5 (#40760), so this no longer applies. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/18636"}, {"key": "pr#18605", "n": 18605, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "[WIP] Introduce NestLayer", "action": "closed", "comment": "Thanks for exploring this! We removed TensorFlow support in v5, so modeling_tf_utils.py no longer exists and we're closing this PR.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/18605"}, {"key": "issue#18395", "n": 18395, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "GFT: Generative Fundamental Training", "action": "commented", "comment": "No public weights on the Hub as of today.", "author": "wangyi-fudan", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/18395"}, {"key": "issue#18322", "n": 18322, "kind": "issue", "group": "Open issues", "category": "docs", "area": "docs", "type": "docs", "title": "Transformers documentation translation to French", "action": "\u2014", "comment": "", "author": "omarespejel", "idle_days": 38, "models": "", "url": "https://github.com/huggingface/transformers/issues/18322"}, {"key": "issue#18243", "n": 18243, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Onnx Runtime Errors With LongT5", "action": "closed; comment corrected", "comment": "The `transformers.onnx` exporter was removed in #41700; export (ONNX included) now goes through `transformers.exporters` (https://huggingface.co/docs/transformers/main/en/exporters). Closing; if LongT5 export still fails with `transformers.exporters` on current `main`, please open a new issue with a reproducer.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/18243"}, {"key": "issue#18181", "n": 18181, "kind": "issue", "group": "Closed issues", "category": "obsolete", "area": "", "type": "", "title": "Test summary with previous PyTorch/TensorFlow versions", "action": "closed", "comment": "This 2022 summary about old PyTorch/TensorFlow versions no longer applies: TensorFlow support was removed in v5 and the minimum supported PyTorch version has moved well past those tested here. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/18181"}, {"key": "pr#18152", "n": 18152, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "dalle mega", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/18152"}, {"key": "issue#18118", "n": 18118, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Model parallelism for m2m100", "action": "commented", "comment": "M2M100 is supported in transformers as `m2m_100` (added in #10236). For large variants, use `device_map=\"auto\"` for inference, or Accelerate / FSDP / DeepSpeed for training.", "author": "elricwan", "idle_days": 0, "models": "m2m_100", "url": "https://github.com/huggingface/transformers/issues/18118"}, {"key": "issue#18085", "n": 18085, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "Adding TF Implementation of BEiT", "action": "closed", "comment": "TensorFlow and Flax/JAX support was fully removed in v5 (#40760), so this no longer applies. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/18085"}, {"key": "issue#18066", "n": 18066, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "An example for finetuning FLAVA or any VLP multimodel using trainer (for example for classification)", "action": "closed", "comment": "As @NielsRogge pointed out (https://github.com/huggingface/transformers/issues/18066#issuecomment-1178661272), FLAVA fine-tuning tutorials are available at https://github.com/apsdehal/flava-tutorials and https://github.com/NielsRogge/Transformers-Tutorials. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/18066"}, {"key": "issue#18063", "n": 18063, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add TF implementation of LongT5 model", "action": "\u2014", "comment": "", "author": "stancld", "idle_days": 1515, "models": "", "url": "https://github.com/huggingface/transformers/issues/18063"}, {"key": "issue#18004", "n": 18004, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add DFFT", "action": "commented", "comment": "No public weights on the Hub as of today. Weights are on GitHub only.", "author": "NielsRogge", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/18004"}, {"key": "pr#17739", "n": 17739, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "[WIP] DETR TF implementation", "action": "closed", "comment": "Thanks a lot for this work! Transformers removed TensorFlow support in v5 and is now PyTorch-only, so we won't be adding new TF models. Closing, with thanks.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/17739"}, {"key": "issue#17703", "n": 17703, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add Flax implementation for BLOOM", "action": "commented", "comment": "BLOOM is supported in transformers as `bloom` (PyTorch). Flax support has been removed from transformers as of v5, so a Flax BLOOM won't be added.", "author": "haileyschoelkopf", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/17703"}, {"key": "issue#17582", "n": 17582, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Loading repository after rename does not work (with old name)", "action": "closed", "comment": "This works now: loading goes through `huggingface_hub`, which follows renamed-repo redirects, and `hf_hub_download(\"KB/bert-base-swedish-cased\", \"config.json\")` resolves to the `KBLab` repo. Closing, please reopen if you still see the error with an up-to-date `huggingface_hub`.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/17582"}, {"key": "issue#17459", "n": 17459, "kind": "issue", "group": "Closed issues", "category": "obsolete", "area": "", "type": "", "title": "Tranformers documentation translation to Italian", "action": "closed", "comment": "Thanks to everyone who contributed Italian translations, they live in `docs/source/it` on main! The `.mdx` files listed here no longer exist (docs are `.md` now), so closing this tracker; new translation PRs are welcome following https://github.com/huggingface/transformers/blob/main/docs/TRANSLATING.md.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/17459"}, {"key": "pr#17387", "n": 17387, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "Add Google's Trillson Audio Classification Model", "action": "closed", "comment": "Thanks for the PR! TensorFlow and Flax support was removed in v5, so this can't land anymore. Closing, thanks again for the contribution!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/17387"}, {"key": "issue#17339", "n": 17339, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Google's Trillson Audio Classification", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [vumichien/nonsemantic-speech-trillsson3](https://huggingface.co/vumichien/nonsemantic-speech-trillsson3)\n- Downloads (last 30d): 102 \u00b7 Likes: 1 \u00b7 Params: n/a \u00b7 License: n/a\n- Released: 2022-09 \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "patrickvonplaten", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/17339"}, {"key": "issue#17309", "n": 17309, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "UNETR: Transformers for 3D Medical Image Segmentation", "action": "commented", "comment": "No public weights on the Hub as of today. UNETR weights are distributed through MONAI, not the Hub.", "author": "pri1311", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/17309"}, {"key": "issue#17283", "n": 17283, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "GPT-neo generate is ignoring passed position ids", "action": "closed", "comment": "Fixed on main by #43734: `generate()` now keeps the `position_ids` you pass and extends them step by step instead of recomputing them, so GPT-Neo (and other models) respect custom positions. Please reopen if you still see it on the latest release.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/17283"}, {"key": "issue#17234", "n": 17234, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "add FAN model (vision)", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: no FAN checkpoint on the Hub; weights are only on the [original GitHub repo](https://github.com/NVlabs/FAN)\n- Released: 2022-04 (paper) \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "johnnv1", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/17234"}, {"key": "issue#17224", "n": 17224, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "ALBEF: Align Before Fuse", "action": "commented", "comment": "Some details for anyone who'd like to pick this up: no official ALBEF checkpoints are on the Hugging Face Hub as of today, so adding it would first need converted weights.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "ggoggam", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/17224"}, {"key": "pr#17204", "n": 17204, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "[Kernel Fusion] Training benchmarks of Torchdynamo + AOTAutograd + NVFuser(many models)", "action": "closed", "comment": "Thanks a lot for this exploration! torch.compile is now native in PyTorch and wired into Trainer through `torch_compile`, so we don't plan to add standalone benchmark scripts to the repo. Closing, but thank you for pushing on this early on.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/17204"}, {"key": "issue#17165", "n": 17165, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Set Transformer", "action": "commented", "comment": "No public weights on the Hub as of today. No official pretrained Set Transformer checkpoint exists.", "author": "leondz", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/17165"}, {"key": "issue#16747", "n": 16747, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "pointer to transformer (big) model", "action": "commented", "comment": "No public weights on the Hub as of today: the original Transformer-big weights were never released as a checkpoint. For a vanilla encoder-decoder, `MarianMTModel` / `FSMTForConditionalGeneration` are the closest.", "author": "anirudt", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/16747"}, {"key": "pr#16712", "n": 16712, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "Run the scheduled tests", "action": "closed", "comment": "Thanks @LysandreJik! This was a temporary draft (\"do not merge\") for running the scheduled suite against `test-tokenizers-main` in 2022; its companion #16708 was closed long ago, and `self-scheduled.yml` was removed when daily CI moved to the transformers-ci workflow in #47360. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/16712"}, {"key": "issue#16497", "n": 16497, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "[TODO] Investigate equivalence tests", "action": "closed", "comment": "The PT/TF/Flax equivalence tests are gone since TensorFlow and Flax support were removed in v5 (#40760). Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/16497"}, {"key": "issue#16466", "n": 16466, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "TEAMS: Training ELECTRA Augmented with Multi-word Selection", "action": "commented", "comment": "Some details for anyone who'd like to pick this up: no official TEAMS checkpoints are on the Hugging Face Hub as of today, so adding it would first need converted weights.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "stefan-it", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/16466"}, {"key": "issue#16403", "n": 16403, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "R3M: A Universal Visual Representation for Robot Manipulation", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [surajnair/r3m-50](https://huggingface.co/surajnair/r3m-50), [surajnair/r3m-18](https://huggingface.co/surajnair/r3m-18) (original-repo format)\n- Downloads (last 30d): 13 \u00b7 Likes: 0 \u00b7 License: not specified\n- Released: 2022-04 \u00b7 Remote code on the Hub: no\n\nR3M is a ResNet backbone, so it may map onto the existing `resnet` model with a weight conversion.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "edbeeching", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/16403"}, {"key": "issue#16349", "n": 16349, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": " [WIP] New Model Add FastPitch 1.1", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [nvidia/tts_en_fastpitch](https://huggingface.co/nvidia/tts_en_fastpitch) (NeMo `.nemo` format)\n- Downloads (last 30d): 85 \u00b7 Likes: 43 \u00b7 License: cc-by-4.0\n- Released: 2022-06 \u00b7 Remote code on the Hub: no\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "ArEnSc", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/16349"}, {"key": "issue#16157", "n": 16157, "kind": "issue", "group": "Open issues", "category": "feature", "area": "other", "type": "feature", "title": "Implement Maximal Update Parametrization (muP)", "action": "labelled: Feature request", "comment": "", "author": "thegregyang", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/16157"}, {"key": "pr#15902", "n": 15902, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[WIP] Add Fusion-in-Decoder", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/15902"}, {"key": "pr#15773", "n": 15773, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[Community] Add fairseq FastSpeech2", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/15773"}, {"key": "issue#15763", "n": 15763, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add TUNet", "action": "commented", "comment": "Some details for anyone who'd like to pick this up: no official TUNet checkpoints are on the Hugging Face Hub as of today, so adding it would first need converted weights.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "OllieBroadhurst", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/15763"}, {"key": "issue#15550", "n": 15550, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "DeepMind Retro", "action": "commented", "comment": "No public weights on the Hub as of today. DeepMind never released RETRO weights.", "author": "shamanez", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/15550"}, {"key": "issue#15535", "n": 15535, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "vits support? ", "action": "commented", "comment": "Good news: this is now supported in transformers as `vits` (added in #24085), e.g. [facebook/mms-tts-eng](https://huggingface.co/facebook/mms-tts-eng).", "author": "lucasjinreal", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/15535"}, {"key": "issue#15451", "n": 15451, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Adding RelationExtraction head to layoutLMv2 and layoutXLM models", "action": "\u2014", "comment": "", "author": "R0bk", "idle_days": 796, "models": "", "url": "https://github.com/huggingface/transformers/issues/15451"}, {"key": "issue#15387", "n": 15387, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Evidentiality-guided Generator - Retrieval model", "action": "commented", "comment": "No public weights on the Hub as of today.", "author": "patrickvonplaten", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/15387"}, {"key": "pr#15295", "n": 15295, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "feat(flax): leave restored weights on CPU", "action": "closed", "comment": "Thanks for the contribution! Flax support (including `modeling_flax_utils.py`) was removed in v5, so this change no longer applies. Closing for that reason.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/15295"}, {"key": "pr#15264", "n": 15264, "kind": "PR", "group": "Closed PRs", "category": "old branch / code moved", "area": "", "type": "", "title": "[Kernel Fusion] training benchmarks of AOTAutograd (multiple models)", "action": "closed", "comment": "Thanks a lot @Chillee for these experiments and the collaboration here! This draft was explicitly not meant to be merged, and AOTAutograd has since landed in PyTorch as the backbone of `torch.compile`, which Transformers models now support out of the box. The `scripts/aot_*.py` benchmarks were never added to `main`, so I'm closing this to keep the PR list tidy.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/15264"}, {"key": "issue#15166", "n": 15166, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add FastSpeech2", "action": "commented", "comment": "Good news: this is now supported in transformers as `fastspeech2_conformer` (added in #23439), e.g. [espnet/fastspeech2_conformer](https://huggingface.co/espnet/fastspeech2_conformer).", "author": "jaketae", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/15166"}, {"key": "issue#15072", "n": 15072, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "[JAX/FLAX]: CLM Tokenizer Training confusion", "action": "closed", "comment": "TensorFlow and Flax/JAX support was removed in v5 (#40760), so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/15072"}, {"key": "issue#15026", "n": 15026, "kind": "issue", "group": "Closed issues", "category": "obsolete", "area": "", "type": "", "title": "[Benchmark] HF Trainer on A100", "action": "closed", "comment": "Closing this one: no activity since 2022-03-23, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/15026"}, {"key": "issue#14839", "n": 14839, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Fine-tuning GPT-J-6B in colab: 8-bit weights with low-rank adaptors", "action": "\u2014", "comment": "", "author": "dvmazur", "idle_days": 1272, "models": "", "url": "https://github.com/huggingface/transformers/issues/14839"}, {"key": "issue#14814", "n": 14814, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Support on Mixture of expert models", "action": "commented", "comment": "Good news: mixture-of-experts models are now widely supported in transformers, e.g. `switch_transformers` (#19323), `mixtral`, `qwen3_moe`, `deepseek_v3` and many more.", "author": "fym0503", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/14814"}, {"key": "pr#14708", "n": 14708, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "[WIP] [performance doc] faster/leaner optimizers", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/14708"}, {"key": "issue#14608", "n": 14608, "kind": "issue", "group": "Closed issues", "category": "obsolete", "area": "", "type": "", "title": "[Benchmark] HF Trainer on RTX-3090", "action": "closed", "comment": "Closing this 2022 benchmark index: the benchmark tool (#14934) and the changes it motivated (tf32 control in #14606, `--optim` in #14744) are merged, and the findings are reflected in docs/source/en/optimization_overview.md. Thanks @stas00!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/14608"}, {"key": "issue#14494", "n": 14494, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add TAPAS trained on NQ", "action": "commented", "comment": "`tapas` is already in transformers, and the NQ weights have been converted:\n\n- Hub: [xhluca/tapas-nq-hn-retriever-medium-0](https://huggingface.co/xhluca/tapas-nq-hn-retriever-medium-0), [deepset/tapas-large-nq-hn-reader](https://huggingface.co/deepset/tapas-large-nq-hn-reader)\n- Downloads (last 30d): 94 / 644\n\nThe retriever loads as a `TapasModel` encoder; only a dual-encoder retrieval head is missing.", "author": "xhluca", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/14494"}, {"key": "pr#14356", "n": 14356, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "[WIP] Adding support for `flax` for `pipelines`.", "action": "closed", "comment": "Thanks a lot for working on this! Transformers dropped TensorFlow/Flax/JAX support in v5 and is now PyTorch-only, so we won't be adding Flax pipelines. Closing, but we appreciate the effort.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/14356"}, {"key": "pr#14129", "n": 14129, "kind": "PR", "group": "Closed PRs", "category": "not aligned with v5", "area": "", "type": "", "title": "[Flax] Fix eval and data_args usage in streaming example", "action": "closed", "comment": "Thanks for the fix! The research projects and Flax/JAX examples have been removed from the repo in v5, so we're closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/14129"}, {"key": "issue#13997", "n": 13997, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "[performance/precision] adding `jit.script` to activation functions", "action": "closed", "comment": "Fused activations are now handled differently: `src/transformers/activations.py` decorates activations with `@use_kernel_forward_from_hub` so fused kernels can be loaded from the Hub (https://huggingface.co/docs/transformers/main/en/kernels), and `torch.compile` fuses the rest (https://huggingface.co/docs/transformers/main/en/perf_torch_compile). `torch.jit.script` is deprecated upstream, so closing this.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/13997"}, {"key": "issue#13844", "n": 13844, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "How to tokenize big dataset ", "action": "closed", "comment": "As @sgugger explained (https://github.com/huggingface/transformers/issues/13844#issuecomment-934479525), `datasets` memory-maps data via Arrow, so training with a batch iterator never loads the full dataset in memory. For further questions please use https://discuss.huggingface.co. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/13844"}, {"key": "issue#13784", "n": 13784, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Feature: Tail Free Sampling", "action": "closed", "comment": "As @gante explained (https://github.com/huggingface/transformers/pull/38055#issuecomment-2900553588), new generation features go to the Hub as `custom_generate` repos first and are added to transformers if they become popular. Tail-free sampling can be shipped that way (https://huggingface.co/docs/transformers/main/en/generation_strategies#custom-generation-methods) or as a custom `LogitsProcessor`. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/13784"}, {"key": "pr#13752", "n": 13752, "kind": "PR", "group": "Closed PRs", "category": "outdated (v4-era, inactive)", "area": "", "type": "", "title": "Update FNet Fourier Transform", "action": "closed", "comment": "Thanks for the PR, and sorry it didn't make it in! This branch is based on v4, and v5 is very different, so rather than rebasing we'd love to invite you to open a new PR against `main` using our latest tools and good practices:\n- [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets)\n- [mlinter](https://huggingface.github.io/transformers-mlinter/) and the [modeling rules](https://huggingface.co/docs/transformers/main/en/modeling_rules) it enforces (`make typing`)\n- [Modular transformers](https://huggingface.co/docs/transformers/main/en/modular_transformers)\n- [Contributing guide](https://huggingface.co/docs/transformers/main/en/contributing)\n\nClosing this one in the meantime, thanks again!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/pull/13752"}, {"key": "issue#13355", "n": 13355, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Dependency parsing head for pretrained models", "action": "closed", "comment": "Closing this one: no activity since 2022-01-25, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/13355"}, {"key": "issue#13246", "n": 13246, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "[model loading] framework-agnostic dtype parameter", "action": "closed", "comment": "This has been added in #39782 and is available on main. Closing, but feel free to reopen if something is still missing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/13246"}, {"key": "issue#13116", "n": 13116, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Problem about using mBART50 for Russian to Chinese translation", "action": "closed", "comment": "This is a known limitation of the mBART-50 checkpoints on some non-English pairs rather than a bug in transformers (see #25425); make sure `forced_bos_token_id=tokenizer.lang_code_to_id[tgt_lang]` is passed to `generate`, and for ru->zh a model like NLLB or M2M100 may work better. Closing as there is nothing to fix on the library side.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/13116"}, {"key": "issue#12909", "n": 12909, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "Truncating the prefix of a sequence rather than the suffix", "action": "closed", "comment": "This is supported: set `truncation_side=\"left\"` on the tokenizer (added in #14947) to truncate the beginning of the sequence instead of the end. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/12909"}, {"key": "issue#12851", "n": 12851, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "Got `ONNXRuntimeError` when try to run BART in ONNX format", "action": "closed; comment corrected", "comment": "The `transformers.onnx` export module was removed in v5 (#41700); export now goes through `transformers.exporters` (https://huggingface.co/docs/transformers/main/en/exporters). Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/12851"}, {"key": "issue#12807", "n": 12807, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Model Request: Blenderbot 2.0", "action": "commented", "comment": "Some details for anyone who'd like to pick this up: no official BlenderBot 2.0 checkpoints are on the Hugging Face Hub as of today, so adding it would first need converted weights.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "Xirider", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/12807"}, {"key": "issue#12768", "n": 12768, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Alphafold 2.0", "action": "commented", "comment": "No public weights on the Hub as of today. DeepMind distributes AlphaFold weights outside the Hub. A related protein structure model, ESMFold, is supported (`esm`, `esmfold2`).", "author": "elk-cloner", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/12768"}, {"key": "issue#12647", "n": 12647, "kind": "issue", "group": "Closed issues", "category": "obsolete", "area": "", "type": "", "title": "`TestMarian_MT_EN::test_batch_generation_mt_en` Failing due to randomly generated tokens", "action": "closed", "comment": "This test was for the TensorFlow Marian model, and `test_modeling_tf_marian.py` was removed together with TensorFlow support in #40760. Closing as no longer applicable.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/12647"}, {"key": "issue#12460", "n": 12460, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Extractive summarization pipeline", "action": "closed", "comment": "Thanks for the request! There's no standard model head or checkpoint family for extractive summarization to build a pipeline around, so we don't plan to add one; sentence scoring with `pipeline(\"text-classification\")` or embeddings via sentence-transformers covers most extractive setups. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/12460"}, {"key": "issue#12443", "n": 12443, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "[Wav2Vec2] Better names for internal classes", "action": "closed", "comment": "Both points are resolved on main: the conv module in `modeling_wav2vec2.py` is now `Wav2Vec2FeatureEncoder` (no longer clashing with the feature extractor), and the Flax classes such as `FlaxWav2Vec2EncoderLayerStableLayerNormCollection` were removed along with Flax support in v5.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/12443"}, {"key": "issue#12411", "n": 12411, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "\ud83c\udf1f New model addition: FNet", "action": "closed", "comment": "This has been added in #13045 and is available on main. Closing, but feel free to reopen if something is still missing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/12411"}, {"key": "issue#12381", "n": 12381, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "A fast tokenizer for BertJapaneseTokenizer", "action": "closed", "comment": "Closing this one: no activity since 2023-02-19, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/12381"}, {"key": "issue#12272", "n": 12272, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "[Deepspeed zero3] lazy weights init ", "action": "closed", "comment": "Lazy weight init under ZeRO-3 is handled on main: missing parameters are initialized inside `deepspeed.zero.GatheredParameters` (#28245), and the init path was further reworked in #44236 (`initialize_weights_zero3` in `src/transformers/integrations/deepspeed.py`). Closing as completed.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/12272"}, {"key": "issue#12269", "n": 12269, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "Add TFSpeech2Text", "action": "closed", "comment": "TensorFlow and Flax/JAX support was fully removed in v5 (#40760), so this no longer applies. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/12269"}, {"key": "issue#12241", "n": 12241, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Modify BERT encoder layers?", "action": "commented", "comment": "This isn't a model request: you can wrap `BertModel` in your own `nn.Module` or subclass `BertLayer`. For questions like this the [forum](https://discuss.huggingface.co/) is the best place.", "author": "calusbr", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/12241"}, {"key": "issue#12210", "n": 12210, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Better documentation for generation parameter defaults", "action": "closed", "comment": "This was addressed by `GenerationConfig` (#20218): defaults are now an explicit object (`model.generation_config`), and the precedence between it and kwargs passed to `generate()` is documented at https://huggingface.co/docs/transformers/main/en/main_classes/text_generation", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/12210"}, {"key": "issue#12182", "n": 12182, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "KeyError: 'labels' during Distilling Zero Shot Classification", "action": "closed", "comment": "The zero-shot distillation script lived in `examples/research_projects`, which was removed in #36645, so closing this one.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/12182"}, {"key": "issue#12144", "n": 12144, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "How to train the new wav2vec unsupervised model using hugging face ?", "action": "closed", "comment": "We don't plan to add wav2vec-U's GAN training pipeline to Transformers; the reference implementation remains in fairseq. Pretrained Wav2Vec2/Data2Vec checkpoints can still be fine-tuned with CTC as in https://huggingface.co/docs/transformers/main/en/tasks/asr. Closing, thanks for the discussion!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/12144"}, {"key": "issue#12103", "n": 12103, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "ViT tensorflow Implementation", "action": "closed", "comment": "TensorFlow support was removed from Transformers in v5 (#40760), so a TF ViT implementation won't be added. Closing this one!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/12103"}, {"key": "issue#11998", "n": 11998, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Add SENet Blocks in Encoding Layers", "action": "closed", "comment": "Closing this one: no activity since 2021-08-26, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/11998"}, {"key": "issue#11847", "n": 11847, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Request addition of 'GPT2ForwardBackward' models", "action": "\u2014", "comment": "", "author": "GenTxt", "idle_days": 1975, "models": "", "url": "https://github.com/huggingface/transformers/issues/11847"}, {"key": "issue#11732", "n": 11732, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Import `SPIECE_UNDERLINE` from `file_utils` instead of WET definition", "action": "closed", "comment": "Closing this one: no activity since 2021-06-15, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/11732"}, {"key": "issue#11647", "n": 11647, "kind": "issue", "group": "Open issues", "category": "bug", "area": "other", "type": "bug", "title": "Key Error: 'pre-processing' during conversion from tatoeba to Marian model", "action": "commented; labelled: bug", "comment": "The converter still lives at `src/transformers/models/marian/convert_marian_tatoeba_to_pytorch.py` and still reads `model[\"pre-processing\"]` (around line 80 and 244), which newer Tatoeba-Challenge YAML metadata doesn't always contain. A small PR making that lookup tolerant (e.g. `model.get(\"pre-processing\", \"\")` and checking the SentencePiece files directly) would fix it; many converted Opus-MT checkpoints are also already on the Hub under `Helsinki-NLP/`.", "author": "velocityCavalry", "idle_days": -1, "models": "", "url": "https://github.com/huggingface/transformers/issues/11647"}, {"key": "issue#11632", "n": 11632, "kind": "issue", "group": "Commented (now closed by others)", "category": "new-model details", "area": "", "type": "", "title": "Felix", "action": "commented", "comment": "No public weights on the Hub as of today. Felix weights are on GitHub/GCS only.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/11632"}, {"key": "issue#11129", "n": 11129, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "denoising with sentence permutation, and language sampling", "action": "closed", "comment": "Closing this one: no activity since 2025-03-15, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/11129"}, {"key": "issue#10978", "n": 10978, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "Add GPT Neo models to Write With Transformer", "action": "closed", "comment": "Closing this one: no activity since 2021-06-04, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/10978"}, {"key": "issue#10958", "n": 10958, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Returning Confidence Score For Extractive QA Task When Using Non-Pipeline Approach ", "action": "closed", "comment": "Scores are a post-processing step rather than a model output, so we don't plan to add them to `ForQuestionAnswering` outputs. You can reproduce what the pipeline does: softmax the `start_logits`/`end_logits` and multiply the start and end probabilities of the chosen span (see `QuestionAnsweringPipeline.postprocess` in `src/transformers/pipelines/question_answering.py`), or just call the `question-answering` pipeline.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/10958"}, {"key": "issue#10935", "n": 10935, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add DALL-E: Zero-Shot Text-to-Image Generation", "action": "commented", "comment": "No public weights on the Hub as of today: OpenAI only released the dVAE of DALL-E, not the 12B transformer. Community replicas exist (e.g. [dalle-mini/dalle-mini](https://huggingface.co/dalle-mini/dalle-mini)), and text-to-image models are now covered by [diffusers](https://github.com/huggingface/diffusers).", "author": "slavakurilyak", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/10935"}, {"key": "issue#10912", "n": 10912, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Summarization length not controlled by max_length, min_length", "action": "closed", "comment": "As @NielsRogge explained (https://github.com/huggingface/transformers/issues/10912#issuecomment-808021158), `max_length`/`min_length` count tokens, not words, and generation can stop at EOS before `max_length`; on current versions prefer `max_new_tokens`/`min_new_tokens`. Closing as expected behavior.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/10912"}, {"key": "issue#10779", "n": 10779, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "EncoderDecoderModel with different model dimensions", "action": "closed", "comment": "This was fixed in #14378: `EncoderDecoderModel` now adds an `enc_to_dec_proj` linear layer when the encoder and decoder hidden sizes differ (see `src/transformers/models/encoder_decoder/modeling_encoder_decoder.py`), so e.g. a bert-large encoder with a gpt2 decoder works. Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/10779"}, {"key": "issue#10365", "n": 10365, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Knowledge Retrieval missing from BlenderBot Implementation", "action": "\u2014", "comment": "", "author": "", "idle_days": 2064, "models": "", "url": "https://github.com/huggingface/transformers/issues/10365"}, {"key": "issue#10340", "n": 10340, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "tokenizer.Tokenizer compatibility with Inference API or Auto* classes", "action": "closed; comment corrected", "comment": "A `tokenizers.Tokenizer` can be wrapped with `PreTrainedTokenizerFast(tokenizer_object=tok)` (see the `tokenizer_object` argument in `src/transformers/tokenization_utils_tokenizers.py`), then `save_pretrained`/`push_to_hub` makes it loadable with `AutoTokenizer`. Docs: https://huggingface.co/docs/transformers/main/en/fast_tokenizers", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/10340"}, {"key": "issue#10234", "n": 10234, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Request to add Switch Transformer", "action": "commented", "comment": "Good news: this is now supported in transformers as `switch_transformers` (added in #19323).", "author": "coderpotter", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/10234"}, {"key": "issue#10201", "n": 10201, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Better Fine-Tuning by Reducing Representational Collapse", "action": "closed", "comment": "Closing this one: no activity since 2022-06-27, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/10201"}, {"key": "issue#10121", "n": 10121, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Allow `do_lower_case=True` for any tokenizer", "action": "closed", "comment": "Since #40936 every tokenizer is backed by the `tokenizers` library, so lowercasing can be enabled for any tokenizer by adding a `Lowercase()` normalizer, e.g. `tok.backend_tokenizer.normalizer = normalizers.Sequence([normalizers.Lowercase(), ...])`, then `save_pretrained`. Closing as this covers the use case.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/10121"}, {"key": "issue#10105", "n": 10105, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "PruneTrain: Fast Neural Network Training by Dynamic Sparse Model Reconfiguration", "action": "commented", "comment": "No public weights on the Hub as of today: PruneTrain is a training technique, not a pretrained model, so it falls outside what transformers adds.", "author": "gaceladri", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/10105"}, {"key": "issue#9965", "n": 9965, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "[trainer] new in pytorch: `torch.optim._multi_tensor` faster optimizers ", "action": "closed", "comment": "The multi-tensor path landed upstream: `torch.optim` optimizers use the foreach implementation by default, and the Trainer now defaults to `optim=\"adamw_torch_fused\"` on torch>=2.8 (see the `optim` docs in `src/transformers/training_args.py`). Nothing left to do on our side.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/9965"}, {"key": "issue#9931", "n": 9931, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "[2D Parallelism] Tracking feasibility", "action": "closed", "comment": "The 2D/N-D parallelism this tracked is now available on main: native tensor parallelism via `tp_plan` (docs/source/en/perf_infer_gpu_multi.md) and N-D parallelism through accelerate's `ParallelismConfig` (`TrainingArguments.parallelism_config` in src/transformers/training_args.py). Closing, thanks for the notes!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/9931"}, {"key": "issue#9925", "n": 9925, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Implementing ELECTRIC training for ELECTRA", "action": "closed", "comment": "Closing this one: no activity since 2021-02-01, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/9925"}, {"key": "issue#9908", "n": 9908, "kind": "issue", "group": "Closed issues", "category": "fixed", "area": "", "type": "", "title": "[seq2seq] some logging for all processes in distributed mode", "action": "closed", "comment": "Per-process logging in distributed runs is handled by `TrainingArguments(log_level_replica=..., log_on_each_node=...)` (added in #11819 and #12276), so replicas can log at INFO without abusing warnings. The `examples/seq2seq` scripts mentioned here no longer exist.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/9908"}, {"key": "issue#9893", "n": 9893, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "rfc: new benchmark tool", "action": "closed", "comment": "The new benchmarking tool now lives in `benchmark_v2/` on main (framework in #40486, overhaul in #41408): reproducible CLI runs via `benchmark_v2/run_benchmarks.py`, JSON results and hardware metrics collection. Closing this RFC in favor of that; feature requests for it are welcome as new issues.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/9893"}, {"key": "issue#9784", "n": 9784, "kind": "issue", "group": "Closed issues", "category": "removed in v5", "area": "", "type": "", "title": "Translation Model in ONNX: Choosable Output Formats", "action": "closed; comment corrected", "comment": "The `convert_graph_to_onnx` script this is about was removed in v5 (#40760); export now goes through `transformers.exporters` (https://huggingface.co/docs/transformers/main/en/exporters). Closing!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/9784"}, {"key": "issue#9648", "n": 9648, "kind": "issue", "group": "Closed issues", "category": "superseded", "area": "", "type": "", "title": "Easier perplexity computation", "action": "closed", "comment": "Perplexity computation is documented in `docs/source/en/perplexity.md` (https://huggingface.co/docs/transformers/main/en/perplexity). As @LysandreJik suggested (https://github.com/huggingface/transformers/issues/9648#issuecomment-762197639), a docs PR adding a batched variant to that guide is welcome. Closing.", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/9648"}, {"key": "issue#9606", "n": 9606, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "[DeepSpeed] Features to integrate / Optimizations to add / Experiments to do", "action": "closed", "comment": "Closing this one: no activity since 2023-06-15, it was reported against v4 (v5 is very different). If it still applies on the latest version, please open a new issue with a reproducer!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/9606"}, {"key": "issue#9526", "n": 9526, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Siamese Multi-depth Transformer-based Hierarchical Encoder", "action": "commented", "comment": "Some details for anyone who'd like to pick this up: no official SMITH checkpoints are on the Hugging Face Hub as of today, so adding it would first need converted weights.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "lalitpagaria", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/9526"}, {"key": "issue#9061", "n": 9061, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "CharacterBERT", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: [helboukkouri/character-bert](https://huggingface.co/helboukkouri/character-bert), [helboukkouri/character-bert-medical](https://huggingface.co/helboukkouri/character-bert-medical)\n- Downloads (last 30d): 274 \u00b7 Likes: 2 \u00b7 Params: ~110M (BERT-base size) \u00b7 License: not specified\n- Released: 2022-03 \u00b7 Remote code on the Hub: yes\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "helboukkouri", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/9061"}, {"key": "issue#8454", "n": 8454, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add POINTER model", "action": "commented", "comment": "Some details for anyone who'd like to pick this up:\n- Hub: no POINTER checkpoint on the Hub; weights are only linked from the [original GitHub repo](https://github.com/dreasysnail/POINTER)\n- Released: 2020-11 (paper) \u00b7 Remote code on the Hub: no\n- Note: it's BERT-large with one extra token, so the weights should load into `BertForMaskedLM`; only the iterative insertion decoding is missing.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "dreasysnail", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/8454"}, {"key": "issue#7951", "n": 7951, "kind": "issue", "group": "Open issues", "category": "bug", "area": "tokenization", "type": "bug", "title": "Unexpected/wrong handling of added special tokens in special_tokens_mask (GPT1, BERT, possibly others)", "action": "commented; labelled: bug", "comment": "This still reproduces on main: with `openai-community/openai-gpt` and added `<bos>/<sep>/<eos>`, `return_special_tokens_mask=True` returns all zeros. The mask for non-formatted inputs comes from each tokenizer's `get_special_tokens_mask` (only template tokens like CLS/SEP), while `PreTrainedTokenizerBase.get_special_tokens_mask` in `src/transformers/tokenization_utils_base.py` already has an `all_special_ids` membership path for `already_has_special_tokens=True`. A fix would OR that membership check into the mask built in `tokenization_python.py` / `tokenization_utils_tokenizers.py`; a PR with a test in `tests/test_tokenization_common.py` would be welcome.", "author": "matejklemen", "idle_days": -1, "models": "bert", "url": "https://github.com/huggingface/transformers/issues/7951"}, {"key": "issue#7675", "n": 7675, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add FAVOR+ / Performer attention", "action": "commented", "comment": "Some details for anyone who'd like to pick this up: no official Performer checkpoints are on the Hugging Face Hub as of today, so adding it would first need converted weights.\n\nTo add it, follow the [modular guide](https://huggingface.co/docs/transformers/main/en/modular_transformers) and the [Transformers tenets](https://huggingface.co/spaces/transformers-community/Transformers-tenets).", "author": "marrrcin", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/7675"}, {"key": "issue#7397", "n": 7397, "kind": "issue", "group": "Closed issues", "category": "stale", "area": "", "type": "", "title": "Add DistilBERTGeneration comparable to BertGeneration", "action": "closed", "comment": "We won't add a causal-LM/decoder head to DistilBERT: it has no cross-attention, so using it as an `EncoderDecoderModel` decoder would need architecture changes on an older model. Small modern decoder checkpoints or `BertGenerationDecoder` with a distilled BERT checkpoint are better options today. Closing as not planned + way too old!", "author": "", "idle_days": null, "models": "", "url": "https://github.com/huggingface/transformers/issues/7397"}, {"key": "issue#7370", "n": 7370, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "Add new PET Model", "action": "commented", "comment": "No public weights on the Hub as of today. PET is a training method on top of existing MLMs (e.g. RoBERTa/ALBERT), not a new architecture.", "author": "sagarreddypatil", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/7370"}, {"key": "issue#6497", "n": 6497, "kind": "issue", "group": "Open issues", "category": "new model", "area": "models", "type": "new model", "title": "BERT and SpanBERT for Coreference Resolution", "action": "commented", "comment": "SpanBERT already loads as `bert` in transformers:\n\n- Hub: [SpanBERT/spanbert-large-cased](https://huggingface.co/SpanBERT/spanbert-large-cased)\n- Downloads (last 30d): 94,759 \u00b7 Likes: 14\n\nThe coreference head from the paper isn't on the Hub. For coreference today, see e.g. [biu-nlp/f-coref](https://huggingface.co/biu-nlp/f-coref) (fastcoref).", "author": "techscribr", "idle_days": 0, "models": "", "url": "https://github.com/huggingface/transformers/issues/6497"}]