s3recap-p25-b0.1: tau 3.8413 @ block 5, 28.38 tok/s, 1.97x over base AR
Browse files- BEST_SETUP.md +101 -0
- README.md +67 -0
- config.json +170 -0
- eval.log +1101 -0
- head_card.json +123 -0
- mtp_trained.safetensors +3 -0
- train_metrics.json +0 -0
BEST_SETUP.md
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# BEST_SETUP.md — the winning configuration, stated so it can be rebuilt
|
| 2 |
+
|
| 3 |
+
What is served today, why, and the exact commands that reproduce it. Written 2026-08-23, after the
|
| 4 |
+
block-5 draft-head programme closed P2.5, P2.6 and 2.3's training half.
|
| 5 |
+
|
| 6 |
+
The narrative and the arms that lost are in [`wiki/draft-head-finetuning.md`](wiki/draft-head-finetuning.md)
|
| 7 |
+
§9–§10. This file is the operational one: **what to run to get this result back.**
|
| 8 |
+
|
| 9 |
+
## 1. What is served
|
| 10 |
+
|
| 11 |
+
| | |
|
| 12 |
+
|---|---|
|
| 13 |
+
| head | **`s3recap-p25-b0.1`** |
|
| 14 |
+
| suite mean `tau` | **3.8413 / 5** (77 % of the width ceiling) |
|
| 15 |
+
| suite mean tok/s | **28.3825** |
|
| 16 |
+
| base AR | 14.61 tok/s — speculation is **1.97×** |
|
| 17 |
+
| draft block width | **5** (`config.json`'s own `dspark_block_size`) |
|
| 18 |
+
| `config/live_ckpt` | `/home/patrickd/models/ckpt-head-s3recap-p25-b0.1` |
|
| 19 |
+
| engine rev at promotion | `a608d52d76f86d6325b472e34696b15d5b5fb6da` |
|
| 20 |
+
| source sha256 | recorded in `head_card.json`; `tools/verify_head_archive.py --full` checks it |
|
| 21 |
+
|
| 22 |
+
Against the stock head this project started from: **22.66 → 28.38 tok/s, +25.3 %**, every emitted
|
| 23 |
+
token still bit-identical to base AR.
|
| 24 |
+
|
| 25 |
+
## 2. The recipe, exactly
|
| 26 |
+
|
| 27 |
+
One epoch over 1,472 sequences (1,536 generated, 64 reserved as hold-out), three chunks of 491.
|
| 28 |
+
|
| 29 |
+
| knob | value | why this and not another |
|
| 30 |
+
|---|---|---|
|
| 31 |
+
| `--block` | **5** | resolved from `config.json`'s `dspark_block_size`. NOT a tuned value — ladder 2.1 found the served width had drifted to 6 and moving it back was worth +3.91 % at equal `tau`. |
|
| 32 |
+
| `--a-ce` / `--a-tv` | **0.1 / 0.9** | the control's weights. The ce/tv sweep (1.0/0.0 and 0.5/0.5) both lost; F119 had already falsified the CE/TV explanation of F117. |
|
| 33 |
+
| `--deficit` | **on** | weights each block's loss by `min((1-r)/mean(1-r), 3.0)`, `r` = that block's accepted fraction. Spends gradient on the head's worst positions, which is where a draft head's value lives (§8.3). |
|
| 34 |
+
| `--deficit-prior` | 0.4106 | the running mean's prior for the first 32 steps. |
|
| 35 |
+
| `--deficit-clamp` | 3.0 | |
|
| 36 |
+
| `--beta` | **0.1** | KL pull toward the head as it was before the first optimizer step, scaled by `r`. **Bracketed on both sides**: β=0 is short at +2.9 %, β=0.5 lands *below* the incumbent. |
|
| 37 |
+
| `--a-conf` | **0.0** | measured, not defaulted. See §4. |
|
| 38 |
+
| `--hass-from` | **0 (off)** | HASS is retired — `wiki/negative-results.md` §4l. |
|
| 39 |
+
| `--lr` | 5e-5 peak, 73 warmup steps | |
|
| 40 |
+
| `--pos-per-seq` | 16 | |
|
| 41 |
+
| trainable | 517,318,215 params across 72 tensors | |
|
| 42 |
+
|
| 43 |
+
Final chunk's metrics: loss 1.627 → 0.037, tail mean 0.739; ce 0.097, tv 0.029.
|
| 44 |
+
|
| 45 |
+
## 3. Reproducing it
|
| 46 |
+
|
| 47 |
+
```bash
|
| 48 |
+
# 1. the corpus already exists and is retained -- pass 1 (generation) is the expensive stage and
|
| 49 |
+
# does NOT need to be re-run. /home/patrickd/s5-capture/s3/gen.txt, 1536 sequences.
|
| 50 |
+
# 2. run the arm. S5_* map onto the flags in the table above.
|
| 51 |
+
S5_GEN=/home/patrickd/s5-capture/s3/gen.txt S5_HOLDOUT=64 S5_CHUNK=491 S5_KEEP_CAP=1 \
|
| 52 |
+
S5_BLOCK=5 S5_ACE=0.1 S5_ATV=0.9 S5_DEFICIT=1 S5_BETA=0.1 \
|
| 53 |
+
bash scripts/s5_session_p25.sh myhead /home/patrickd/s5-capture/mixed_prompts_s3.txt 1536 512
|
| 54 |
+
```
|
| 55 |
+
|
| 56 |
+
Re-capturing that corpus costs pass 2 + train only, ~0.4 h per 500 sequences. **Generation is what
|
| 57 |
+
dominates a session and it is already done.**
|
| 58 |
+
|
| 59 |
+
To deploy a head once it promotes:
|
| 60 |
+
|
| 61 |
+
```bash
|
| 62 |
+
bash scripts/stage_head.sh <name> --activate # verifies the farm, then repoints config/live_ckpt
|
| 63 |
+
bash scripts/run_server.sh # picks it up on the next start
|
| 64 |
+
```
|
| 65 |
+
|
| 66 |
+
## 4. Two knobs that look like oversights and are measurements
|
| 67 |
+
|
| 68 |
+
**`--a-conf 0`.** F100 measured the confidence term at `conf` ≈ 10034 against `ce` 10.43 under
|
| 69 |
+
teacher forcing and pinned it at 0. Ladder 2.3 then measured the free-running magnitude at **O(1)**
|
| 70 |
+
and trained two arms at `a_conf` 1.0 and 0.1. Both are **worse** than this recipe, because getting
|
| 71 |
+
free-running labels requires HASS and HASS costs −0.046 `tau`. The confidence head gains **+0.018
|
| 72 |
+
AUC** from our fine-tuning, which does not pay for that. Keep it at 0.
|
| 73 |
+
|
| 74 |
+
**The confidence head is still good — it is just not ours.**
|
| 75 |
+
`mtp.2.confidence_head.proj.weight` ships in the base checkpoint, is byte-identical in every
|
| 76 |
+
`a_conf = 0` head including this one, and scores **AUC 0.88** predicting acceptance. That is what
|
| 77 |
+
justifies 2.3's serving-side work, and it needs no training arm at all.
|
| 78 |
+
|
| 79 |
+
**`--hass-from 0`.** Not untried: measured twice, monotone in the wrong direction. See §4l.
|
| 80 |
+
|
| 81 |
+
## 5. Durability — where this lives and how it is checked
|
| 82 |
+
|
| 83 |
+
| artifact | where | recoverable if the disk dies? |
|
| 84 |
+
|---|---|---|
|
| 85 |
+
| engine + every document | git, pushed to `origin/main` | **yes** |
|
| 86 |
+
| head weights (source) | `~/model-backups/heads/<name>/mtp_trained.safetensors`, 1 GB each | **only if uploaded** |
|
| 87 |
+
| head weights (loadable shards) | same directory, ~7 GB each | regenerate from the source |
|
| 88 |
+
| training corpus | `/home/patrickd/s5-capture/s3/gen.txt` | regenerable, but the generate pass is the expensive one |
|
| 89 |
+
| the recipe | **this file**, in git | yes |
|
| 90 |
+
|
| 91 |
+
python3 tools/verify_head_archive.py # presence + size + completeness, seconds
|
| 92 |
+
python3 tools/verify_head_archive.py --full # + sha256 of every file, needs an idle box
|
| 93 |
+
|
| 94 |
+
Every head's `mtp_trained.safetensors` now carries a recorded sha256 in its `head_card.json` (added
|
| 95 |
+
2026-08-23). Before that the archive could not detect corruption of the one file everything else
|
| 96 |
+
regenerates from, which made "rebuildable" an unfalsifiable claim.
|
| 97 |
+
|
| 98 |
+
**Nothing is ever deleted from `~/model-backups/heads/`.** Intermediate per-chunk resume state in
|
| 99 |
+
`s5-capture` (`c0`/`c1` weights and AdamW moments) IS deletable once the session is complete and
|
| 100 |
+
its final `c2` state is archived and hash-verified — that is how 63 GiB was reclaimed on
|
| 101 |
+
2026-08-23 — but the archive itself is append-only.
|
README.md
ADDED
|
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
base_model: 0xSero/DeepSeek-V4-Flash-0731-REAP
|
| 4 |
+
tags:
|
| 5 |
+
- speculative-decoding
|
| 6 |
+
- draft-head
|
| 7 |
+
- dspark
|
| 8 |
+
- jetson
|
| 9 |
+
library_name: safetensors
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# DSpark MTP draft head — `s3recap-p25-b0.1`
|
| 13 |
+
|
| 14 |
+
**Weights and provenance, published as-measured.** This head is a working result from an ongoing
|
| 15 |
+
optimisation programme, not a polished release: it is the current best of eleven measured
|
| 16 |
+
candidates, and it is uploaded so the weights survive the single machine they were produced on.
|
| 17 |
+
Every number below is a measurement from the log included in this repo, not an estimate.
|
| 18 |
+
|
| 19 |
+
Licence follows the base checkpoint,
|
| 20 |
+
[`0xSero/DeepSeek-V4-Flash-0731-REAP`](https://huggingface.co/0xSero/DeepSeek-V4-Flash-0731-REAP);
|
| 21 |
+
this repository adds only the fine-tuned draft-head tensors.
|
| 22 |
+
|
| 23 |
+
## What this is
|
| 24 |
+
|
| 25 |
+
A fine-tuned DSpark MTP draft head for
|
| 26 |
+
[`0xSero/DeepSeek-V4-Flash-0731-REAP`](https://huggingface.co/0xSero/DeepSeek-V4-Flash-0731-REAP)
|
| 27 |
+
(K160, native MXFP4), trained to raise speculative-decode acceptance on a Jetson AGX Thor
|
| 28 |
+
(`sm_110a`) running a from-scratch pure-CUDA inference server.
|
| 29 |
+
|
| 30 |
+
| | measured |
|
| 31 |
+
|---|---|
|
| 32 |
+
| suite mean `tau` | **3.8413 / 5** (77 % of the draft-width ceiling) |
|
| 33 |
+
| suite mean tok/s | **28.3825** |
|
| 34 |
+
| base AR decode | 14.61 tok/s — speculation is **1.97×** |
|
| 35 |
+
| vs the stock shipped head | **22.66 → 28.38 tok/s, +25.3 %** |
|
| 36 |
+
| draft block width | **5** (the checkpoint's own `dspark_block_size`) |
|
| 37 |
+
|
| 38 |
+
Every emitted token is bit-identical to base autoregressive decode, checked on every run. The
|
| 39 |
+
speed-up is lossless by construction: speculation changes how tokens are *produced*, never which
|
| 40 |
+
tokens they are.
|
| 41 |
+
|
| 42 |
+
## Files
|
| 43 |
+
|
| 44 |
+
| file | what it is |
|
| 45 |
+
|---|---|
|
| 46 |
+
| `mtp_trained.safetensors` | **the source of truth.** 72 `mtp.*` tensors. The loadable checkpoint shards regenerate from this deterministically. |
|
| 47 |
+
| `head_card.json` | provenance: base model and revision, engine git rev, per-prompt measurements, recorded sha256 |
|
| 48 |
+
| `train_metrics.json` | the full training history, all three chunks |
|
| 49 |
+
| `eval.log` | the promotion measurement, unedited |
|
| 50 |
+
| `config.json` | the base checkpoint's config, for the architectural constants |
|
| 51 |
+
| `BEST_SETUP.md` | **the recipe** — every hyperparameter with the reason it has that value |
|
| 52 |
+
|
| 53 |
+
## `tau` is not comparable across block widths
|
| 54 |
+
|
| 55 |
+
`tau` counts tokens committed per target forward, and its **ceiling is the draft width**. A head
|
| 56 |
+
scoring 3.84/6 and one scoring 3.84/5 are not the same head measured twice. The same weights that
|
| 57 |
+
read 3.8438 at width 6 read **3.6888** at width 5. Any comparison must fix the width first.
|
| 58 |
+
|
| 59 |
+
## Two hyperparameters that are measurements, not defaults
|
| 60 |
+
|
| 61 |
+
- **`a_conf = 0`** — the confidence-head loss term is off *deliberately*. Training it requires
|
| 62 |
+
free-running acceptance labels, which requires HASS, which costs −0.046 `tau`; the head gains only
|
| 63 |
+
+0.018 AUC from the fine-tuning. The confidence head that ships in the base checkpoint already
|
| 64 |
+
predicts acceptance at **AUC 0.88** without any of our training.
|
| 65 |
+
- **`hass_from = 0`** — HASS was measured twice and is monotone in the wrong direction. Retired.
|
| 66 |
+
|
| 67 |
+
Both are explained in `BEST_SETUP.md`.
|
config.json
ADDED
|
@@ -0,0 +1,170 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"DeepseekV4ForCausalLM"
|
| 4 |
+
],
|
| 5 |
+
"attention_bias": false,
|
| 6 |
+
"attention_dropout": 0.0,
|
| 7 |
+
"bos_token_id": 0,
|
| 8 |
+
"compress_ratios": [
|
| 9 |
+
0,
|
| 10 |
+
0,
|
| 11 |
+
4,
|
| 12 |
+
128,
|
| 13 |
+
4,
|
| 14 |
+
128,
|
| 15 |
+
4,
|
| 16 |
+
128,
|
| 17 |
+
4,
|
| 18 |
+
128,
|
| 19 |
+
4,
|
| 20 |
+
128,
|
| 21 |
+
4,
|
| 22 |
+
128,
|
| 23 |
+
4,
|
| 24 |
+
128,
|
| 25 |
+
4,
|
| 26 |
+
128,
|
| 27 |
+
4,
|
| 28 |
+
128,
|
| 29 |
+
4,
|
| 30 |
+
128,
|
| 31 |
+
4,
|
| 32 |
+
128,
|
| 33 |
+
4,
|
| 34 |
+
128,
|
| 35 |
+
4,
|
| 36 |
+
128,
|
| 37 |
+
4,
|
| 38 |
+
128,
|
| 39 |
+
4,
|
| 40 |
+
128,
|
| 41 |
+
4,
|
| 42 |
+
128,
|
| 43 |
+
4,
|
| 44 |
+
128,
|
| 45 |
+
4,
|
| 46 |
+
128,
|
| 47 |
+
4,
|
| 48 |
+
128,
|
| 49 |
+
4,
|
| 50 |
+
128,
|
| 51 |
+
4,
|
| 52 |
+
0,
|
| 53 |
+
0,
|
| 54 |
+
0
|
| 55 |
+
],
|
| 56 |
+
"compress_rope_theta": 160000,
|
| 57 |
+
"dspark_block_size": 5,
|
| 58 |
+
"dspark_markov_rank": 256,
|
| 59 |
+
"dspark_noise_token_id": 128799,
|
| 60 |
+
"dspark_target_layer_ids": [
|
| 61 |
+
40,
|
| 62 |
+
41,
|
| 63 |
+
42
|
| 64 |
+
],
|
| 65 |
+
"eos_token_id": 1,
|
| 66 |
+
"expert_dtype": "fp4",
|
| 67 |
+
"hc_eps": 1e-06,
|
| 68 |
+
"hc_mult": 4,
|
| 69 |
+
"hc_sinkhorn_iters": 20,
|
| 70 |
+
"head_dim": 512,
|
| 71 |
+
"hidden_act": "silu",
|
| 72 |
+
"hidden_size": 4096,
|
| 73 |
+
"index_head_dim": 128,
|
| 74 |
+
"index_n_heads": 64,
|
| 75 |
+
"index_topk": 512,
|
| 76 |
+
"initializer_range": 0.02,
|
| 77 |
+
"max_position_embeddings": 1048576,
|
| 78 |
+
"model_type": "deepseek_v4",
|
| 79 |
+
"moe_intermediate_size": 2048,
|
| 80 |
+
"n_routed_experts": 160,
|
| 81 |
+
"n_shared_experts": 1,
|
| 82 |
+
"norm_topk_prob": true,
|
| 83 |
+
"num_attention_heads": 64,
|
| 84 |
+
"num_experts_per_tok": 6,
|
| 85 |
+
"num_hash_layers": 3,
|
| 86 |
+
"num_hidden_layers": 43,
|
| 87 |
+
"num_key_value_heads": 1,
|
| 88 |
+
"num_nextn_predict_layers": 1,
|
| 89 |
+
"o_groups": 8,
|
| 90 |
+
"o_lora_rank": 1024,
|
| 91 |
+
"q_lora_rank": 1024,
|
| 92 |
+
"qk_rope_head_dim": 64,
|
| 93 |
+
"quantization_config": {
|
| 94 |
+
"activation_scheme": "dynamic",
|
| 95 |
+
"fmt": "e4m3",
|
| 96 |
+
"quant_method": "fp8",
|
| 97 |
+
"scale_fmt": "ue8m0",
|
| 98 |
+
"weight_block_size": [
|
| 99 |
+
128,
|
| 100 |
+
128
|
| 101 |
+
]
|
| 102 |
+
},
|
| 103 |
+
"reap_pruning": {
|
| 104 |
+
"backbone_layers": 43,
|
| 105 |
+
"built_at_unix": 1785518256,
|
| 106 |
+
"dspark_mtp_target_layers": [
|
| 107 |
+
40,
|
| 108 |
+
41,
|
| 109 |
+
42
|
| 110 |
+
],
|
| 111 |
+
"hash_routed_layers": [
|
| 112 |
+
0,
|
| 113 |
+
1,
|
| 114 |
+
2
|
| 115 |
+
],
|
| 116 |
+
"kept_experts_per_layer": 160,
|
| 117 |
+
"method": "REAP transferred ranking with router identity alignment",
|
| 118 |
+
"original_experts_per_layer": 256,
|
| 119 |
+
"prior_observation_model": "deepseek-ai/DeepSeek-V4-Flash",
|
| 120 |
+
"prior_observation_revision": "60d8d70770c6776ff598c94bb586a859a38244f1",
|
| 121 |
+
"prior_observation_snapshot": {
|
| 122 |
+
"all_experts_observed": true,
|
| 123 |
+
"available_rows_used": 21289,
|
| 124 |
+
"category_counts": {
|
| 125 |
+
"agentic_tool_trajectory": 2253,
|
| 126 |
+
"code_generation": 4096,
|
| 127 |
+
"reasoning_code": 2676,
|
| 128 |
+
"reasoning_math": 4072,
|
| 129 |
+
"reasoning_science": 4096,
|
| 130 |
+
"tool_calling": 4096
|
| 131 |
+
},
|
| 132 |
+
"readme_advertised_full_rows": 24576,
|
| 133 |
+
"repo_id": "0xSero/deepseek-v4-flash-reap-observations-v2",
|
| 134 |
+
"snapshot_label": "partial-v2-21289"
|
| 135 |
+
},
|
| 136 |
+
"prior_plan_sha256": "2837dc8cd0f7512c111bd3a2c9c41278f715ded5965eb01889652ef8e465e1b0",
|
| 137 |
+
"router_alignment": {
|
| 138 |
+
"hash_tid2eid_equal_fraction": [
|
| 139 |
+
1.0,
|
| 140 |
+
1.0,
|
| 141 |
+
1.0
|
| 142 |
+
],
|
| 143 |
+
"identity_nearest_all": true,
|
| 144 |
+
"layers": 43,
|
| 145 |
+
"same_index_cos_mean_min": 0.9923683404922485,
|
| 146 |
+
"same_index_cos_min_global": 0.9460793733596802
|
| 147 |
+
},
|
| 148 |
+
"source_model": "deepseek-ai/DeepSeek-V4-Flash-0731",
|
| 149 |
+
"source_revision": "9e165c30e2704aec5d9d593cce3eebd58bbef1cb"
|
| 150 |
+
},
|
| 151 |
+
"rms_norm_eps": 1e-06,
|
| 152 |
+
"rope_scaling": {
|
| 153 |
+
"beta_fast": 32,
|
| 154 |
+
"beta_slow": 1,
|
| 155 |
+
"factor": 16,
|
| 156 |
+
"original_max_position_embeddings": 65536,
|
| 157 |
+
"type": "yarn"
|
| 158 |
+
},
|
| 159 |
+
"rope_theta": 10000,
|
| 160 |
+
"routed_scaling_factor": 1.5,
|
| 161 |
+
"scoring_func": "sqrtsoftplus",
|
| 162 |
+
"sliding_window": 128,
|
| 163 |
+
"swiglu_limit": 10.0,
|
| 164 |
+
"tie_word_embeddings": false,
|
| 165 |
+
"topk_method": "noaux_tc",
|
| 166 |
+
"torch_dtype": "bfloat16",
|
| 167 |
+
"transformers_version": "4.57.1",
|
| 168 |
+
"use_cache": true,
|
| 169 |
+
"vocab_size": 129280
|
| 170 |
+
}
|
eval.log
ADDED
|
@@ -0,0 +1,1101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[decode] loading /home/patrickd/models/DeepSeek-V4-Flash-0731-REAP ... s=6 NDEC=8 seqmax=297
|
| 2 |
+
[decode] loaded 100.40 GiB, 45821 tensors (weights in MANAGED device-preferred memory)
|
| 3 |
+
[decode] KV layout: fp32, 2048 B/row, 0.024 GiB for 297 rows x 43 layers
|
| 4 |
+
[decode] building 43 layer structs once (persistent)...
|
| 5 |
+
[decode] structs built. mem 106.3/122.8 GiB
|
| 6 |
+
[decode] prefill 5 positions...
|
| 7 |
+
[f32mk] tile 4x4 (DSV4_F32MK_TILE=unset)
|
| 8 |
+
[decode] PREFILL: 5 positions in 193.4 ms = 25.9 tok/s (38.682 ms/tok)
|
| 9 |
+
[decode] prefill done. caches populated. starting decode.
|
| 10 |
+
step 0 pos 5 -> token 11111 (67.7 ms warmup)
|
| 11 |
+
step 1 pos 6 -> token 16 (67.2 ms)
|
| 12 |
+
step 2 pos 7 -> token 455 (72.8 ms)
|
| 13 |
+
step 3 pos 8 -> token 6102 (66.9 ms)
|
| 14 |
+
step 4 pos 9 -> token 294 (66.6 ms)
|
| 15 |
+
step 5 pos 10 -> token 16603 (66.4 ms)
|
| 16 |
+
step 6 pos 11 -> token 344 (72.6 ms)
|
| 17 |
+
step 7 pos 12 -> token 29168 (66.7 ms)
|
| 18 |
+
|
| 19 |
+
[decode] first decoded token argmax = 11111 (expect 11111) -> GATE PASS
|
| 20 |
+
[decode] generated: 11111 16 455 6102 294 16603 344 29168
|
| 21 |
+
[decode] WARM decode: 68.4 ms/tok = 14.61 tok/s (M=1 steady state, 7-step avg)
|
| 22 |
+
|
| 23 |
+
[spec-verify] M=5 verify in ONE forward: 115.3 ms (= 23.1 ms/tok if all accepted vs 68.4 M=1 -> 2.97x)
|
| 24 |
+
[spec-verify] verify argmax: 11111 16 455 6102 294
|
| 25 |
+
[spec-verify] decode tokens: 11111 16 455 6102 294
|
| 26 |
+
[spec-verify] MATCH 5/5 -> PASS (M=K verify == K sequential decodes; diffs = MoE-atomic near-ties)
|
| 27 |
+
|
| 28 |
+
[spec] loading EXTERNAL DSpark head /home/patrickd/s5-capture/s3recap-p25-b0.1/head ...
|
| 29 |
+
[spec] head loaded 6.53 GiB, 2977 mtp tensors (SEPARATE store)
|
| 30 |
+
[spec] NSTAGE=3 head-experts=160 BLK sweep: 5 5 5 5 5 5 5 5 5
|
| 31 |
+
[spec] head built. mem 113.5/122.8 GiB
|
| 32 |
+
[decode] PREFILL: 5 positions in 132.1 ms = 37.9 tok/s (26.417 ms/tok)
|
| 33 |
+
[spec] decoding 200 tokens (block=5, draft passes=1, adaptK=1.50, prompt=0 s=6)...
|
| 34 |
+
[arena] first draft: draft footprint 210.72 MB, global hwm 210.72 MB / cap 512.00 MB (41.2%), draft path = ARENA
|
| 35 |
+
[margins] 2.05 1.19 0.61 0.58 0.22 K=2
|
| 36 |
+
verify 1: accepted 1/1 (K=2) + correction -> +2 tokens (97.3 ms) cpos=7
|
| 37 |
+
[margins] 0.60 1.60 0.04 0.44 0.35 K=3
|
| 38 |
+
verify 2: accepted 0/2 (K=3) + correction -> +1 tokens (119.7 ms) cpos=8
|
| 39 |
+
[margins] 1.18 0.81 1.95 0.10 1.99 K=2
|
| 40 |
+
verify 3: accepted 1/1 (K=2) + correction -> +2 tokens (96.2 ms) cpos=10
|
| 41 |
+
[margins] 2.89 0.04 4.08 0.51 1.27 K=2
|
| 42 |
+
verify 4: accepted 0/1 (K=2) + correction -> +1 tokens (98.5 ms) cpos=11
|
| 43 |
+
[margins] 6.80 0.12 0.48 2.09 2.52 K=2
|
| 44 |
+
verify 5: accepted 1/1 (K=2) + correction -> +2 tokens (99.0 ms) cpos=13
|
| 45 |
+
[margins] 4.24 3.20 5.22 2.06 0.47 K=5
|
| 46 |
+
verify 6: accepted 4/4 (K=5) + correction -> +5 tokens (139.9 ms) cpos=18
|
| 47 |
+
[margins] 4.93 2.41 2.85 0.84 0.97 K=4
|
| 48 |
+
verify 7: accepted 1/3 (K=4) + correction -> +2 tokens (125.7 ms) cpos=20
|
| 49 |
+
[margins] 2.54 2.12 5.10 2.04 1.74 K=6
|
| 50 |
+
verify 8: accepted 4/5 (K=6) + correction -> +5 tokens (155.0 ms) cpos=25
|
| 51 |
+
[margins] 7.00 2.19 1.93 3.05 2.44 K=6
|
| 52 |
+
verify 9: accepted 1/5 (K=6) + correction -> +2 tokens (156.9 ms) cpos=27
|
| 53 |
+
[margins] 3.58 3.38 6.04 2.45 0.02 K=5
|
| 54 |
+
verify 10: accepted 4/4 (K=5) + correction -> +5 tokens (144.1 ms) cpos=32
|
| 55 |
+
[margins] 7.04 0.68 3.61 1.58 1.68 K=2
|
| 56 |
+
verify 11: accepted 1/1 (K=2) + correction -> +2 tokens (96.4 ms) cpos=34
|
| 57 |
+
[margins] 2.18 3.88 5.15 1.54 0.74 K=5
|
| 58 |
+
verify 12: accepted 4/4 (K=5) + correction -> +5 tokens (141.4 ms) cpos=39
|
| 59 |
+
[margins] 4.79 0.55 4.59 2.46 3.40 K=2
|
| 60 |
+
verify 13: accepted 1/1 (K=2) + correction -> +2 tokens (100.5 ms) cpos=41
|
| 61 |
+
[margins] 3.53 3.69 6.21 1.58 1.31 K=5
|
| 62 |
+
verify 14: accepted 4/4 (K=5) + correction -> +5 tokens (140.3 ms) cpos=46
|
| 63 |
+
[margins] 1.87 0.92 4.00 0.97 3.56 K=2
|
| 64 |
+
verify 15: accepted 1/1 (K=2) + correction -> +2 tokens (100.8 ms) cpos=48
|
| 65 |
+
[margins] 0.03 0.79 1.18 0.36 3.10 K=2
|
| 66 |
+
verify 16: accepted 1/1 (K=2) + correction -> +2 tokens (97.2 ms) cpos=50
|
| 67 |
+
[margins] 4.48 3.88 4.38 1.14 3.07 K=4
|
| 68 |
+
verify 17: accepted 3/3 (K=4) + correction -> +4 tokens (129.5 ms) cpos=54
|
| 69 |
+
[margins] 2.83 1.75 6.72 0.25 2.60 K=4
|
| 70 |
+
verify 18: accepted 3/3 (K=4) + correction -> +4 tokens (128.4 ms) cpos=58
|
| 71 |
+
[margins] 1.31 5.27 5.19 6.26 1.31 K=5
|
| 72 |
+
verify 19: accepted 0/4 (K=5) + correction -> +1 tokens (141.5 ms) cpos=59
|
| 73 |
+
[margins] 5.62 5.99 6.02 0.03 4.35 K=4
|
| 74 |
+
verify 20: accepted 3/3 (K=4) + correction -> +4 tokens (129.0 ms) cpos=63
|
| 75 |
+
[margins] 0.31 0.72 6.97 0.93 3.59 K=2
|
| 76 |
+
verify 21: accepted 0/1 (K=2) + correction -> +1 tokens (101.7 ms) cpos=64
|
| 77 |
+
[margins] 6.56 0.71 3.24 2.43 3.60 K=2
|
| 78 |
+
verify 22: accepted 1/1 (K=2) + correction -> +2 tokens (98.7 ms) cpos=66
|
| 79 |
+
[margins] 4.43 6.02 5.28 1.97 0.03 K=5
|
| 80 |
+
verify 23: accepted 4/4 (K=5) + correction -> +5 tokens (142.8 ms) cpos=71
|
| 81 |
+
[margins] 6.86 1.93 2.12 1.91 1.62 K=6
|
| 82 |
+
verify 24: accepted 5/5 (K=6) + correction -> +6 tokens (166.5 ms) cpos=77
|
| 83 |
+
[margins] 0.98 0.75 6.52 1.74 5.01 K=2
|
| 84 |
+
verify 25: accepted 0/1 (K=2) + correction -> +1 tokens (99.4 ms) cpos=78
|
| 85 |
+
[margins] 6.19 0.44 1.78 5.11 2.83 K=2
|
| 86 |
+
verify 26: accepted 1/1 (K=2) + correction -> +2 tokens (103.6 ms) cpos=80
|
| 87 |
+
[margins] 5.33 6.38 5.57 3.31 1.18 K=5
|
| 88 |
+
verify 27: accepted 4/4 (K=5) + correction -> +5 tokens (141.8 ms) cpos=85
|
| 89 |
+
[margins] 7.43 0.08 1.65 3.24 4.02 K=2
|
| 90 |
+
verify 28: accepted 1/1 (K=2) + correction -> +2 tokens (99.2 ms) cpos=87
|
| 91 |
+
[margins] 6.29 4.56 7.24 2.42 3.20 K=6
|
| 92 |
+
verify 29: accepted 4/5 (K=6) + correction -> +5 tokens (166.4 ms) cpos=92
|
| 93 |
+
[margins] 8.03 1.15 0.71 3.20 4.77 K=2
|
| 94 |
+
verify 30: accepted 1/1 (K=2) + correction -> +2 tokens (100.3 ms) cpos=94
|
| 95 |
+
[margins] 4.84 4.28 4.83 3.96 2.80 K=6
|
| 96 |
+
verify 31: accepted 4/5 (K=6) + correction -> +5 tokens (168.0 ms) cpos=99
|
| 97 |
+
[margins] 6.29 1.55 0.81 3.62 4.81 K=3
|
| 98 |
+
verify 32: accepted 1/2 (K=3) + correction -> +2 tokens (118.1 ms) cpos=101
|
| 99 |
+
[margins] 7.52 5.65 5.46 1.01 2.38 K=4
|
| 100 |
+
verify 33: accepted 3/3 (K=4) + correction -> +4 tokens (131.2 ms) cpos=105
|
| 101 |
+
[margins] 1.67 7.95 0.81 7.12 4.46 K=3
|
| 102 |
+
verify 34: accepted 2/2 (K=3) + correction -> +3 tokens (117.9 ms) cpos=108
|
| 103 |
+
[margins] 7.21 5.82 6.31 2.90 2.43 K=6
|
| 104 |
+
verify 35: accepted 4/5 (K=6) + correction -> +5 tokens (162.2 ms) cpos=113
|
| 105 |
+
[margins] 7.95 1.59 0.57 3.76 2.83 K=3
|
| 106 |
+
verify 36: accepted 1/2 (K=3) + correction -> +2 tokens (116.9 ms) cpos=115
|
| 107 |
+
[margins] 4.16 2.13 6.86 1.97 0.03 K=5
|
| 108 |
+
verify 37: accepted 4/4 (K=5) + correction -> +5 tokens (149.9 ms) cpos=120
|
| 109 |
+
[margins] 7.72 1.12 0.61 4.91 4.33 K=2
|
| 110 |
+
verify 38: accepted 1/1 (K=2) + correction -> +2 tokens (101.4 ms) cpos=122
|
| 111 |
+
[margins] 6.37 3.29 5.57 3.65 1.86 K=6
|
| 112 |
+
verify 39: accepted 4/5 (K=6) + correction -> +5 tokens (183.9 ms) cpos=127
|
| 113 |
+
[margins] 8.02 0.94 1.17 4.06 5.21 K=2
|
| 114 |
+
verify 40: accepted 1/1 (K=2) + correction -> +2 tokens (119.6 ms) cpos=129
|
| 115 |
+
[margins] 8.24 6.40 8.84 1.73 1.27 K=5
|
| 116 |
+
verify 41: accepted 4/4 (K=5) + correction -> +5 tokens (143.4 ms) cpos=134
|
| 117 |
+
[margins] 7.28 1.58 1.56 2.49 4.36 K=6
|
| 118 |
+
verify 42: accepted 1/5 (K=6) + correction -> +2 tokens (168.1 ms) cpos=136
|
| 119 |
+
[margins] 8.18 5.23 6.00 1.81 1.55 K=6
|
| 120 |
+
verify 43: accepted 5/5 (K=6) + correction -> +6 tokens (161.6 ms) cpos=142
|
| 121 |
+
[margins] 9.36 9.46 2.51 2.12 5.31 K=6
|
| 122 |
+
verify 44: accepted 2/5 (K=6) + correction -> +3 tokens (165.2 ms) cpos=145
|
| 123 |
+
[margins] 5.82 6.15 7.36 1.89 2.11 K=6
|
| 124 |
+
verify 45: accepted 4/5 (K=6) + correction -> +5 tokens (160.4 ms) cpos=150
|
| 125 |
+
[margins] 7.07 0.93 2.14 5.72 5.18 K=2
|
| 126 |
+
verify 46: accepted 1/1 (K=2) + correction -> +2 tokens (105.5 ms) cpos=152
|
| 127 |
+
[margins] 14.45 3.20 7.77 8.29 9.15 K=6
|
| 128 |
+
verify 47: accepted 1/5 (K=6) + correction -> +2 tokens (161.7 ms) cpos=154
|
| 129 |
+
[margins] 4.07 8.00 7.85 4.30 2.91 K=6
|
| 130 |
+
verify 48: accepted 4/5 (K=6) + correction -> +5 tokens (166.2 ms) cpos=159
|
| 131 |
+
[margins] 4.15 0.14 3.68 2.71 6.55 K=2
|
| 132 |
+
verify 49: accepted 1/1 (K=2) + correction -> +2 tokens (105.4 ms) cpos=161
|
| 133 |
+
[margins] 9.72 6.61 7.89 7.57 3.35 K=6
|
| 134 |
+
verify 50: accepted 5/5 (K=6) + correction -> +6 tokens (163.6 ms) cpos=167
|
| 135 |
+
[margins] 6.56 1.08 3.04 4.70 4.82 K=2
|
| 136 |
+
verify 51: accepted 1/1 (K=2) + correction -> +2 tokens (105.4 ms) cpos=169
|
| 137 |
+
[margins] 2.76 4.40 6.02 3.43 2.95 K=6
|
| 138 |
+
verify 52: accepted 4/5 (K=6) + correction -> +5 tokens (160.5 ms) cpos=174
|
| 139 |
+
[margins] 3.91 1.91 1.19 4.55 5.20 K=3
|
| 140 |
+
verify 53: accepted 1/2 (K=3) + correction -> +2 tokens (117.4 ms) cpos=176
|
| 141 |
+
[margins] 4.48 5.63 6.55 2.62 0.58 K=5
|
| 142 |
+
verify 54: accepted 4/4 (K=5) + correction -> +5 tokens (144.4 ms) cpos=181
|
| 143 |
+
[margins] 6.69 1.83 0.30 2.10 5.91 K=3
|
| 144 |
+
verify 55: accepted 1/2 (K=3) + correction -> +2 tokens (118.5 ms) cpos=183
|
| 145 |
+
[margins] 4.00 5.26 8.69 2.26 1.00 K=5
|
| 146 |
+
verify 56: accepted 4/4 (K=5) + correction -> +5 tokens (150.9 ms) cpos=188
|
| 147 |
+
[margins] 0.93 0.81 1.43 0.24 6.11 K=2
|
| 148 |
+
verify 57: accepted 1/1 (K=2) + correction -> +2 tokens (102.9 ms) cpos=190
|
| 149 |
+
[margins] 9.52 1.80 1.13 4.33 2.97 K=3
|
| 150 |
+
verify 58: accepted 1/2 (K=3) + correction -> +2 tokens (118.8 ms) cpos=192
|
| 151 |
+
[margins] 5.40 2.31 6.77 6.09 8.68 K=6
|
| 152 |
+
verify 59: accepted 5/5 (K=6) + correction -> +6 tokens (162.1 ms) cpos=198
|
| 153 |
+
[margins] 1.36 6.22 0.44 1.58 3.62 K=3
|
| 154 |
+
verify 60: accepted 2/2 (K=3) + correction -> +3 tokens (119.6 ms) cpos=201
|
| 155 |
+
[margins] 5.79 7.65 8.94 7.61 1.24 K=5
|
| 156 |
+
verify 61: accepted 4/4 (K=5) + correction -> +5 tokens (144.7 ms) cpos=206
|
| 157 |
+
|
| 158 |
+
[spec] generated 201 tokens over 61 verifies: mean tokens/verify = 3.30 (block=5, max 6)
|
| 159 |
+
[spec] tokens: 11111 16 455 6102 294 16603 344 29168 16 455 6102 294 14251 344 16235 16 455 6102 294 29585 344 76405 16 455 6102 294 25062 344 40977 16 455 6102 294 10322 344 17575 16 455 6102 294
|
| 160 |
+
[spec] LOSSLESS GATE: first 8 tokens match base AR -> PASS
|
| 161 |
+
[spec] SPEC-DECODE: 39.6 ms/tok = 25.27 tok/s (vs base M=1 68.4 ms/tok = 14.61 tok/s -> 1.73x)
|
| 162 |
+
[decode] PREFILL: 34 positions in 599.0 ms = 56.8 tok/s (17.618 ms/tok)
|
| 163 |
+
[spec] decoding 200 tokens (block=5, draft passes=1, adaptK=1.50, prompt=1 s=35)...
|
| 164 |
+
[margins] 1.87 2.62 1.63 1.76 1.55 K=6
|
| 165 |
+
verify 1: accepted 0/5 (K=6) + correction -> +1 tokens (156.1 ms) cpos=35
|
| 166 |
+
[margins] 2.73 2.31 1.91 2.31 5.21 K=6
|
| 167 |
+
verify 2: accepted 0/5 (K=6) + correction -> +1 tokens (158.9 ms) cpos=36
|
| 168 |
+
[margins] 1.33 1.64 0.43 3.75 0.82 K=3
|
| 169 |
+
verify 3: accepted 0/2 (K=3) + correction -> +1 tokens (106.6 ms) cpos=37
|
| 170 |
+
[margins] 1.68 1.94 1.88 1.41 0.83 K=4
|
| 171 |
+
verify 4: accepted 2/3 (K=4) + correction -> +3 tokens (122.7 ms) cpos=40
|
| 172 |
+
[margins] 0.74 0.54 0.88 0.15 1.23 K=2
|
| 173 |
+
verify 5: accepted 0/1 (K=2) + correction -> +1 tokens (96.9 ms) cpos=41
|
| 174 |
+
[margins] 1.03 3.73 0.19 1.00 1.07 K=3
|
| 175 |
+
verify 6: accepted 2/2 (K=3) + correction -> +3 tokens (112.1 ms) cpos=44
|
| 176 |
+
[margins] 0.29 6.09 9.75 2.47 0.80 K=5
|
| 177 |
+
verify 7: accepted 4/4 (K=5) + correction -> +5 tokens (137.8 ms) cpos=49
|
| 178 |
+
[margins] 5.70 1.82 0.96 1.49 0.84 K=3
|
| 179 |
+
verify 8: accepted 2/2 (K=3) + correction -> +3 tokens (110.9 ms) cpos=52
|
| 180 |
+
[margins] 4.83 5.61 7.76 2.10 0.53 K=5
|
| 181 |
+
verify 9: accepted 3/4 (K=5) + correction -> +4 tokens (141.0 ms) cpos=56
|
| 182 |
+
[margins] 3.79 1.77 0.73 2.72 2.23 K=3
|
| 183 |
+
verify 10: accepted 2/2 (K=3) + correction -> +3 tokens (108.0 ms) cpos=59
|
| 184 |
+
[margins] 0.59 0.66 3.38 0.43 2.03 K=2
|
| 185 |
+
verify 11: accepted 0/1 (K=2) + correction -> +1 tokens (102.1 ms) cpos=60
|
| 186 |
+
[margins] 0.79 1.64 5.09 0.92 0.75 K=4
|
| 187 |
+
verify 12: accepted 0/3 (K=4) + correction -> +1 tokens (126.8 ms) cpos=61
|
| 188 |
+
[margins] 0.73 3.49 0.57 1.46 0.12 K=3
|
| 189 |
+
verify 13: accepted 1/2 (K=3) + correction -> +2 tokens (114.7 ms) cpos=63
|
| 190 |
+
[margins] 2.13 4.18 0.45 0.13 0.76 K=3
|
| 191 |
+
verify 14: accepted 0/2 (K=3) + correction -> +1 tokens (115.4 ms) cpos=64
|
| 192 |
+
[margins] 4.35 0.17 0.45 6.38 2.66 K=2
|
| 193 |
+
verify 15: accepted 1/1 (K=2) + correction -> +2 tokens (97.6 ms) cpos=66
|
| 194 |
+
[margins] 1.28 1.42 3.12 5.18 0.17 K=2
|
| 195 |
+
verify 16: accepted 1/1 (K=2) + correction -> +2 tokens (104.8 ms) cpos=68
|
| 196 |
+
[margins] 0.28 6.20 0.10 0.38 1.08 K=3
|
| 197 |
+
verify 17: accepted 2/2 (K=3) + correction -> +3 tokens (112.1 ms) cpos=71
|
| 198 |
+
[margins] 2.65 1.84 1.17 5.35 0.13 K=3
|
| 199 |
+
verify 18: accepted 2/2 (K=3) + correction -> +3 tokens (114.6 ms) cpos=74
|
| 200 |
+
[margins] 7.70 1.81 3.22 2.41 3.05 K=6
|
| 201 |
+
verify 19: accepted 5/5 (K=6) + correction -> +6 tokens (160.7 ms) cpos=80
|
| 202 |
+
[margins] 5.53 4.96 4.74 2.16 6.22 K=6
|
| 203 |
+
verify 20: accepted 5/5 (K=6) + correction -> +6 tokens (158.7 ms) cpos=86
|
| 204 |
+
[margins] 8.18 4.23 2.66 7.50 0.23 K=5
|
| 205 |
+
verify 21: accepted 4/4 (K=5) + correction -> +5 tokens (143.5 ms) cpos=91
|
| 206 |
+
[margins] 6.92 2.93 3.23 8.78 1.03 K=5
|
| 207 |
+
verify 22: accepted 4/4 (K=5) + correction -> +5 tokens (147.0 ms) cpos=96
|
| 208 |
+
[margins] 8.82 5.34 3.88 8.44 3.00 K=6
|
| 209 |
+
verify 23: accepted 5/5 (K=6) + correction -> +6 tokens (158.6 ms) cpos=102
|
| 210 |
+
[margins] 10.25 3.15 7.76 7.18 6.10 K=6
|
| 211 |
+
verify 24: accepted 5/5 (K=6) + correction -> +6 tokens (165.8 ms) cpos=108
|
| 212 |
+
[margins] 2.42 1.08 1.03 0.27 4.20 K=2
|
| 213 |
+
verify 25: accepted 0/1 (K=2) + correction -> +1 tokens (102.1 ms) cpos=109
|
| 214 |
+
[margins] 9.15 6.37 5.26 6.03 5.92 K=6
|
| 215 |
+
verify 26: accepted 5/5 (K=6) + correction -> +6 tokens (159.7 ms) cpos=115
|
| 216 |
+
[margins] 9.38 8.70 9.40 3.03 7.29 K=6
|
| 217 |
+
verify 27: accepted 5/5 (K=6) + correction -> +6 tokens (167.3 ms) cpos=121
|
| 218 |
+
[margins] 9.71 5.22 3.71 9.28 0.14 K=5
|
| 219 |
+
verify 28: accepted 4/4 (K=5) + correction -> +5 tokens (147.3 ms) cpos=126
|
| 220 |
+
[margins] 8.92 6.87 7.15 0.45 2.55 K=4
|
| 221 |
+
verify 29: accepted 3/3 (K=4) + correction -> +4 tokens (148.9 ms) cpos=130
|
| 222 |
+
[margins] 8.44 7.58 5.90 2.99 8.70 K=6
|
| 223 |
+
verify 30: accepted 5/5 (K=6) + correction -> +6 tokens (166.7 ms) cpos=136
|
| 224 |
+
[margins] 7.69 5.87 6.46 8.88 2.04 K=6
|
| 225 |
+
verify 31: accepted 5/5 (K=6) + correction -> +6 tokens (162.2 ms) cpos=142
|
| 226 |
+
[margins] 13.34 6.28 10.18 11.72 8.81 K=6
|
| 227 |
+
verify 32: accepted 5/5 (K=6) + correction -> +6 tokens (168.5 ms) cpos=148
|
| 228 |
+
[margins] 3.10 0.60 7.60 5.15 3.57 K=2
|
| 229 |
+
verify 33: accepted 0/1 (K=2) + correction -> +1 tokens (103.5 ms) cpos=149
|
| 230 |
+
[margins] 8.80 10.30 7.82 7.31 4.41 K=6
|
| 231 |
+
verify 34: accepted 5/5 (K=6) + correction -> +6 tokens (160.8 ms) cpos=155
|
| 232 |
+
[margins] 10.07 8.34 8.21 2.85 8.51 K=6
|
| 233 |
+
verify 35: accepted 5/5 (K=6) + correction -> +6 tokens (167.7 ms) cpos=161
|
| 234 |
+
[margins] 10.31 6.76 5.52 9.18 8.00 K=6
|
| 235 |
+
verify 36: accepted 5/5 (K=6) + correction -> +6 tokens (162.4 ms) cpos=167
|
| 236 |
+
[margins] 14.30 5.71 10.55 11.28 10.76 K=6
|
| 237 |
+
verify 37: accepted 5/5 (K=6) + correction -> +6 tokens (168.8 ms) cpos=173
|
| 238 |
+
[margins] 2.89 9.13 9.39 7.16 6.62 K=6
|
| 239 |
+
verify 38: accepted 5/5 (K=6) + correction -> +6 tokens (160.4 ms) cpos=179
|
| 240 |
+
[margins] 9.76 10.16 7.36 8.07 5.32 K=6
|
| 241 |
+
verify 39: accepted 5/5 (K=6) + correction -> +6 tokens (165.7 ms) cpos=185
|
| 242 |
+
[margins] 10.29 10.56 10.13 8.05 10.46 K=6
|
| 243 |
+
verify 40: accepted 5/5 (K=6) + correction -> +6 tokens (161.8 ms) cpos=191
|
| 244 |
+
[margins] 12.38 9.61 5.61 9.00 0.17 K=5
|
| 245 |
+
verify 41: accepted 4/4 (K=5) + correction -> +5 tokens (152.3 ms) cpos=196
|
| 246 |
+
[margins] 12.75 8.45 9.73 9.40 7.51 K=6
|
| 247 |
+
verify 42: accepted 5/5 (K=6) + correction -> +6 tokens (162.9 ms) cpos=202
|
| 248 |
+
[margins] 14.13 6.45 8.08 7.39 6.97 K=6
|
| 249 |
+
verify 43: accepted 5/5 (K=6) + correction -> +6 tokens (169.0 ms) cpos=208
|
| 250 |
+
[margins] 5.07 0.18 0.68 0.30 0.44 K=2
|
| 251 |
+
verify 44: accepted 0/1 (K=2) + correction -> +1 tokens (104.5 ms) cpos=209
|
| 252 |
+
[margins] 9.65 6.74 7.61 7.77 8.57 K=6
|
| 253 |
+
verify 45: accepted 5/5 (K=6) + correction -> +6 tokens (162.2 ms) cpos=215
|
| 254 |
+
[margins] 9.36 11.92 10.06 9.12 8.62 K=6
|
| 255 |
+
verify 46: accepted 5/5 (K=6) + correction -> +6 tokens (169.6 ms) cpos=221
|
| 256 |
+
[margins] 11.22 4.15 9.98 9.97 0.23 K=5
|
| 257 |
+
verify 47: accepted 4/4 (K=5) + correction -> +5 tokens (149.1 ms) cpos=226
|
| 258 |
+
[margins] 13.41 6.37 7.86 7.28 7.47 K=6
|
| 259 |
+
verify 48: accepted 5/5 (K=6) + correction -> +6 tokens (169.4 ms) cpos=232
|
| 260 |
+
[margins] 13.79 6.37 11.11 9.68 9.13 K=6
|
| 261 |
+
verify 49: accepted 5/5 (K=6) + correction -> +6 tokens (165.9 ms) cpos=238
|
| 262 |
+
|
| 263 |
+
[spec] generated 204 tokens over 49 verifies: mean tokens/verify = 4.16 (block=5, max 6)
|
| 264 |
+
[spec] tokens: 4147 554 1926 270 4105 339 43 1114 1926 270 1178 65 50219 4105 418 270 9517 28 4593 1281 106239 3305 1 1 14 223 19 16 23 14 223 20 16 23 14 223 21 16 23 14
|
| 265 |
+
[spec] SPEC-DECODE: 33.5 ms/tok = 29.87 tok/s (vs base M=1 68.4 ms/tok = 14.61 tok/s -> 2.04x)
|
| 266 |
+
[decode] PREFILL: 41 positions in 688.3 ms = 59.6 tok/s (16.787 ms/tok)
|
| 267 |
+
[spec] decoding 200 tokens (block=5, draft passes=1, adaptK=1.50, prompt=2 s=42)...
|
| 268 |
+
[margins] 6.99 0.29 1.49 2.97 0.65 K=2
|
| 269 |
+
verify 1: accepted 1/1 (K=2) + correction -> +2 tokens (100.2 ms) cpos=43
|
| 270 |
+
[margins] 0.40 3.75 1.97 1.82 0.13 K=5
|
| 271 |
+
verify 2: accepted 4/4 (K=5) + correction -> +5 tokens (133.3 ms) cpos=48
|
| 272 |
+
[margins] 3.86 3.70 0.01 1.31 1.54 K=3
|
| 273 |
+
verify 3: accepted 0/2 (K=3) + correction -> +1 tokens (110.1 ms) cpos=49
|
| 274 |
+
[margins] 2.26 7.15 0.51 1.47 3.11 K=3
|
| 275 |
+
verify 4: accepted 0/2 (K=3) + correction -> +1 tokens (114.6 ms) cpos=50
|
| 276 |
+
[margins] 0.95 0.64 0.97 1.57 2.26 K=2
|
| 277 |
+
verify 5: accepted 0/1 (K=2) + correction -> +1 tokens (95.2 ms) cpos=51
|
| 278 |
+
[margins] 6.49 3.76 0.33 0.88 2.06 K=3
|
| 279 |
+
verify 6: accepted 1/2 (K=3) + correction -> +2 tokens (109.6 ms) cpos=53
|
| 280 |
+
[margins] 0.35 4.34 5.45 0.01 1.63 K=4
|
| 281 |
+
verify 7: accepted 0/3 (K=4) + correction -> +1 tokens (127.0 ms) cpos=54
|
| 282 |
+
[margins] 1.20 6.26 3.23 0.61 0.04 K=4
|
| 283 |
+
verify 8: accepted 2/3 (K=4) + correction -> +3 tokens (123.1 ms) cpos=57
|
| 284 |
+
[margins] 0.87 4.69 7.47 1.37 4.19 K=4
|
| 285 |
+
verify 9: accepted 0/3 (K=4) + correction -> +1 tokens (126.9 ms) cpos=58
|
| 286 |
+
[margins] 4.54 3.82 0.43 3.32 0.75 K=3
|
| 287 |
+
verify 10: accepted 2/2 (K=3) + correction -> +3 tokens (113.9 ms) cpos=61
|
| 288 |
+
[margins] 3.34 6.38 8.08 1.97 5.03 K=6
|
| 289 |
+
verify 11: accepted 3/5 (K=6) + correction -> +4 tokens (153.3 ms) cpos=65
|
| 290 |
+
[margins] 1.88 7.60 5.31 0.34 6.16 K=4
|
| 291 |
+
verify 12: accepted 3/3 (K=4) + correction -> +4 tokens (129.1 ms) cpos=69
|
| 292 |
+
[margins] 6.20 7.42 6.27 1.40 2.32 K=4
|
| 293 |
+
verify 13: accepted 2/3 (K=4) + correction -> +3 tokens (129.4 ms) cpos=72
|
| 294 |
+
[margins] 3.59 0.84 1.43 2.95 0.58 K=2
|
| 295 |
+
verify 14: accepted 1/1 (K=2) + correction -> +2 tokens (97.7 ms) cpos=74
|
| 296 |
+
[margins] 3.52 5.70 2.60 0.25 5.15 K=4
|
| 297 |
+
verify 15: accepted 1/3 (K=4) + correction -> +2 tokens (126.2 ms) cpos=76
|
| 298 |
+
[margins] 9.64 4.96 5.76 3.60 1.23 K=5
|
| 299 |
+
verify 16: accepted 2/4 (K=5) + correction -> +3 tokens (140.5 ms) cpos=79
|
| 300 |
+
[margins] 15.81 11.05 6.33 5.32 0.21 K=5
|
| 301 |
+
verify 17: accepted 4/4 (K=5) + correction -> +5 tokens (145.7 ms) cpos=84
|
| 302 |
+
[margins] 19.32 9.15 2.23 0.30 10.07 K=4
|
| 303 |
+
verify 18: accepted 2/3 (K=4) + correction -> +3 tokens (130.1 ms) cpos=87
|
| 304 |
+
[margins] 11.12 16.11 13.05 3.37 11.52 K=6
|
| 305 |
+
verify 19: accepted 5/5 (K=6) + correction -> +6 tokens (161.6 ms) cpos=93
|
| 306 |
+
[margins] 9.36 10.67 12.22 8.95 5.20 K=6
|
| 307 |
+
verify 20: accepted 4/5 (K=6) + correction -> +5 tokens (156.9 ms) cpos=98
|
| 308 |
+
[margins] 9.14 7.97 6.20 3.79 1.33 K=5
|
| 309 |
+
verify 21: accepted 4/4 (K=5) + correction -> +5 tokens (141.1 ms) cpos=103
|
| 310 |
+
[margins] 15.77 15.43 10.25 2.62 1.86 K=6
|
| 311 |
+
verify 22: accepted 5/5 (K=6) + correction -> +6 tokens (164.9 ms) cpos=109
|
| 312 |
+
[margins] 11.68 3.03 3.34 8.70 3.87 K=6
|
| 313 |
+
verify 23: accepted 5/5 (K=6) + correction -> +6 tokens (154.4 ms) cpos=115
|
| 314 |
+
[margins] 10.41 9.14 11.31 0.64 1.40 K=4
|
| 315 |
+
verify 24: accepted 3/3 (K=4) + correction -> +4 tokens (130.9 ms) cpos=119
|
| 316 |
+
[margins] 5.04 8.75 2.47 4.20 0.97 K=5
|
| 317 |
+
verify 25: accepted 4/4 (K=5) + correction -> +5 tokens (148.6 ms) cpos=124
|
| 318 |
+
[margins] 13.10 15.41 10.11 9.14 0.21 K=5
|
| 319 |
+
verify 26: accepted 4/4 (K=5) + correction -> +5 tokens (157.1 ms) cpos=129
|
| 320 |
+
[margins] 8.65 8.81 9.26 8.13 0.16 K=5
|
| 321 |
+
verify 27: accepted 4/4 (K=5) + correction -> +5 tokens (145.4 ms) cpos=134
|
| 322 |
+
[margins] 1.61 0.94 11.13 4.49 7.83 K=2
|
| 323 |
+
verify 28: accepted 0/1 (K=2) + correction -> +1 tokens (108.7 ms) cpos=135
|
| 324 |
+
[margins] 14.85 9.20 1.64 7.69 5.38 K=6
|
| 325 |
+
verify 29: accepted 5/5 (K=6) + correction -> +6 tokens (166.2 ms) cpos=141
|
| 326 |
+
[margins] 14.67 12.63 13.23 13.92 3.10 K=6
|
| 327 |
+
verify 30: accepted 5/5 (K=6) + correction -> +6 tokens (158.8 ms) cpos=147
|
| 328 |
+
[margins] 6.00 15.42 9.30 4.91 2.08 K=6
|
| 329 |
+
verify 31: accepted 5/5 (K=6) + correction -> +6 tokens (166.7 ms) cpos=153
|
| 330 |
+
[margins] 1.86 9.44 12.10 7.51 5.37 K=6
|
| 331 |
+
verify 32: accepted 4/5 (K=6) + correction -> +5 tokens (161.0 ms) cpos=158
|
| 332 |
+
[margins] 12.44 8.12 9.44 0.88 0.00 K=4
|
| 333 |
+
verify 33: accepted 3/3 (K=4) + correction -> +4 tokens (129.3 ms) cpos=162
|
| 334 |
+
[margins] 14.71 1.03 2.82 4.59 2.89 K=2
|
| 335 |
+
verify 34: accepted 1/1 (K=2) + correction -> +2 tokens (104.2 ms) cpos=164
|
| 336 |
+
[margins] 13.82 8.41 2.18 0.29 0.94 K=4
|
| 337 |
+
verify 35: accepted 2/3 (K=4) + correction -> +3 tokens (132.0 ms) cpos=167
|
| 338 |
+
[margins] 9.48 13.05 11.56 8.47 0.51 K=5
|
| 339 |
+
verify 36: accepted 4/4 (K=5) + correction -> +5 tokens (150.0 ms) cpos=172
|
| 340 |
+
[margins] 15.95 8.50 12.14 6.82 0.77 K=5
|
| 341 |
+
verify 37: accepted 4/4 (K=5) + correction -> +5 tokens (145.0 ms) cpos=177
|
| 342 |
+
[margins] 16.12 14.93 14.08 9.59 4.59 K=6
|
| 343 |
+
verify 38: accepted 5/5 (K=6) + correction -> +6 tokens (160.7 ms) cpos=183
|
| 344 |
+
[margins] 14.24 10.81 9.40 2.24 6.53 K=6
|
| 345 |
+
verify 39: accepted 5/5 (K=6) + correction -> +6 tokens (167.4 ms) cpos=189
|
| 346 |
+
[margins] 18.12 10.86 2.50 12.57 5.58 K=6
|
| 347 |
+
verify 40: accepted 5/5 (K=6) + correction -> +6 tokens (159.4 ms) cpos=195
|
| 348 |
+
[margins] 15.22 12.62 10.71 6.06 8.57 K=6
|
| 349 |
+
verify 41: accepted 5/5 (K=6) + correction -> +6 tokens (169.1 ms) cpos=201
|
| 350 |
+
[margins] 8.17 11.50 3.28 8.80 0.53 K=5
|
| 351 |
+
verify 42: accepted 4/4 (K=5) + correction -> +5 tokens (147.5 ms) cpos=206
|
| 352 |
+
[margins] 4.11 14.29 2.45 6.76 0.40 K=5
|
| 353 |
+
verify 43: accepted 4/4 (K=5) + correction -> +5 tokens (140.9 ms) cpos=211
|
| 354 |
+
[margins] 24.84 12.95 9.93 8.33 0.29 K=5
|
| 355 |
+
verify 44: accepted 4/4 (K=5) + correction -> +5 tokens (150.0 ms) cpos=216
|
| 356 |
+
[margins] 27.53 14.72 12.11 6.63 6.23 K=6
|
| 357 |
+
verify 45: accepted 5/5 (K=6) + correction -> +6 tokens (161.5 ms) cpos=222
|
| 358 |
+
[margins] 12.68 17.25 11.22 5.13 9.75 K=6
|
| 359 |
+
verify 46: accepted 5/5 (K=6) + correction -> +6 tokens (167.9 ms) cpos=228
|
| 360 |
+
[margins] 23.77 15.82 4.02 14.35 3.29 K=6
|
| 361 |
+
verify 47: accepted 5/5 (K=6) + correction -> +6 tokens (163.4 ms) cpos=234
|
| 362 |
+
[margins] 20.02 11.07 25.28 8.58 12.51 K=6
|
| 363 |
+
verify 48: accepted 5/5 (K=6) + correction -> +6 tokens (171.2 ms) cpos=240
|
| 364 |
+
[margins] 6.66 22.12 2.37 3.77 3.15 K=6
|
| 365 |
+
verify 49: accepted 5/5 (K=6) + correction -> +6 tokens (164.4 ms) cpos=246
|
| 366 |
+
|
| 367 |
+
[spec] generated 205 tokens over 49 verifies: mean tokens/verify = 4.18 (block=5, max 6)
|
| 368 |
+
[spec] tokens: 1 1 1 1 1 1 50 1697 1 16 1 50 1697 16 1 50 1697 16 1 50 1697 16 1 50 1697 16 1 50 1697 16 20692 50 1697 16 791 50 1697 4394 1015 50
|
| 369 |
+
[spec] SPEC-DECODE: 33.6 ms/tok = 29.80 tok/s (vs base M=1 68.4 ms/tok = 14.61 tok/s -> 2.04x)
|
| 370 |
+
[decode] PREFILL: 26 positions in 455.9 ms = 57.0 tok/s (17.536 ms/tok)
|
| 371 |
+
[spec] decoding 200 tokens (block=5, draft passes=1, adaptK=1.50, prompt=3 s=27)...
|
| 372 |
+
[margins] 1.84 1.32 1.83 0.59 0.95 K=2
|
| 373 |
+
verify 1: accepted 1/1 (K=2) + correction -> +2 tokens (102.6 ms) cpos=28
|
| 374 |
+
[margins] 4.68 0.68 1.05 0.97 0.14 K=2
|
| 375 |
+
verify 2: accepted 1/1 (K=2) + correction -> +2 tokens (99.1 ms) cpos=30
|
| 376 |
+
[margins] 3.33 0.65 0.10 0.00 0.03 K=2
|
| 377 |
+
verify 3: accepted 1/1 (K=2) + correction -> +2 tokens (103.2 ms) cpos=32
|
| 378 |
+
[margins] 2.45 1.70 1.84 0.62 1.31 K=4
|
| 379 |
+
verify 4: accepted 3/3 (K=4) + correction -> +4 tokens (118.3 ms) cpos=36
|
| 380 |
+
[margins] 2.42 5.03 1.31 0.87 1.23 K=3
|
| 381 |
+
verify 5: accepted 2/2 (K=3) + correction -> +3 tokens (105.7 ms) cpos=39
|
| 382 |
+
[margins] 5.28 2.98 4.94 1.94 0.37 K=5
|
| 383 |
+
verify 6: accepted 4/4 (K=5) + correction -> +5 tokens (135.4 ms) cpos=44
|
| 384 |
+
[margins] 0.07 12.34 2.06 4.97 2.32 K=6
|
| 385 |
+
verify 7: accepted 0/5 (K=6) + correction -> +1 tokens (148.6 ms) cpos=45
|
| 386 |
+
[margins] 6.92 2.61 7.02 1.66 4.38 K=6
|
| 387 |
+
verify 8: accepted 5/5 (K=6) + correction -> +6 tokens (144.2 ms) cpos=51
|
| 388 |
+
[margins] 6.55 2.28 6.99 1.83 4.00 K=6
|
| 389 |
+
verify 9: accepted 5/5 (K=6) + correction -> +6 tokens (150.4 ms) cpos=57
|
| 390 |
+
[margins] 7.11 1.97 8.02 2.18 3.55 K=6
|
| 391 |
+
verify 10: accepted 5/5 (K=6) + correction -> +6 tokens (145.3 ms) cpos=63
|
| 392 |
+
[margins] 5.67 2.47 6.72 3.00 3.84 K=6
|
| 393 |
+
verify 11: accepted 5/5 (K=6) + correction -> +6 tokens (149.9 ms) cpos=69
|
| 394 |
+
[margins] 7.19 2.49 6.70 1.91 3.66 K=6
|
| 395 |
+
verify 12: accepted 5/5 (K=6) + correction -> +6 tokens (146.0 ms) cpos=75
|
| 396 |
+
[margins] 8.87 1.73 4.90 1.69 2.72 K=6
|
| 397 |
+
verify 13: accepted 0/5 (K=6) + correction -> +1 tokens (155.7 ms) cpos=76
|
| 398 |
+
[margins] 1.95 5.01 0.87 1.45 0.16 K=3
|
| 399 |
+
verify 14: accepted 0/2 (K=3) + correction -> +1 tokens (110.4 ms) cpos=77
|
| 400 |
+
[margins] 1.06 2.86 2.70 1.37 1.71 K=4
|
| 401 |
+
verify 15: accepted 0/3 (K=4) + correction -> +1 tokens (125.9 ms) cpos=78
|
| 402 |
+
[margins] 2.07 2.52 2.00 3.07 2.03 K=6
|
| 403 |
+
verify 16: accepted 2/5 (K=6) + correction -> +3 tokens (159.8 ms) cpos=81
|
| 404 |
+
[margins] 2.05 0.03 0.64 0.61 1.08 K=2
|
| 405 |
+
verify 17: accepted 1/1 (K=2) + correction -> +2 tokens (97.1 ms) cpos=83
|
| 406 |
+
[margins] 0.65 0.54 0.78 0.48 0.14 K=2
|
| 407 |
+
verify 18: accepted 1/1 (K=2) + correction -> +2 tokens (101.6 ms) cpos=85
|
| 408 |
+
[margins] 4.80 1.80 0.45 0.81 1.49 K=3
|
| 409 |
+
verify 19: accepted 1/2 (K=3) + correction -> +2 tokens (112.2 ms) cpos=87
|
| 410 |
+
[margins] 4.06 3.25 1.36 2.58 0.91 K=3
|
| 411 |
+
verify 20: accepted 1/2 (K=3) + correction -> +2 tokens (115.5 ms) cpos=89
|
| 412 |
+
[margins] 1.68 3.68 0.53 1.37 0.46 K=3
|
| 413 |
+
verify 21: accepted 2/2 (K=3) + correction -> +3 tokens (114.5 ms) cpos=92
|
| 414 |
+
[margins] 2.78 0.87 0.84 2.74 3.37 K=2
|
| 415 |
+
verify 22: accepted 1/1 (K=2) + correction -> +2 tokens (99.1 ms) cpos=94
|
| 416 |
+
[margins] 0.09 8.19 0.69 1.53 2.98 K=3
|
| 417 |
+
verify 23: accepted 2/2 (K=3) + correction -> +3 tokens (113.2 ms) cpos=97
|
| 418 |
+
[margins] 1.96 2.39 0.12 3.89 0.25 K=3
|
| 419 |
+
verify 24: accepted 2/2 (K=3) + correction -> +3 tokens (117.4 ms) cpos=100
|
| 420 |
+
[margins] 0.05 0.36 2.87 0.49 0.56 K=2
|
| 421 |
+
verify 25: accepted 1/1 (K=2) + correction -> +2 tokens (97.4 ms) cpos=102
|
| 422 |
+
[margins] 0.45 0.10 1.09 3.26 8.66 K=2
|
| 423 |
+
verify 26: accepted 0/1 (K=2) + correction -> +1 tokens (101.7 ms) cpos=103
|
| 424 |
+
[margins] 6.22 7.23 1.38 4.23 7.05 K=3
|
| 425 |
+
verify 27: accepted 2/2 (K=3) + correction -> +3 tokens (112.4 ms) cpos=106
|
| 426 |
+
[margins] 7.38 5.75 0.44 5.45 2.23 K=3
|
| 427 |
+
verify 28: accepted 2/2 (K=3) + correction -> +3 tokens (113.1 ms) cpos=109
|
| 428 |
+
[margins] 9.62 1.91 1.73 1.46 1.79 K=4
|
| 429 |
+
verify 29: accepted 3/3 (K=4) + correction -> +4 tokens (127.8 ms) cpos=113
|
| 430 |
+
[margins] 0.80 3.80 2.23 1.56 2.72 K=6
|
| 431 |
+
verify 30: accepted 0/5 (K=6) + correction -> +1 tokens (155.8 ms) cpos=114
|
| 432 |
+
[margins] 0.86 10.01 2.28 2.04 2.54 K=6
|
| 433 |
+
verify 31: accepted 2/5 (K=6) + correction -> +3 tokens (161.1 ms) cpos=117
|
| 434 |
+
[margins] 0.97 5.42 4.29 2.79 9.29 K=6
|
| 435 |
+
verify 32: accepted 5/5 (K=6) + correction -> +6 tokens (161.4 ms) cpos=123
|
| 436 |
+
[margins] 10.15 10.95 5.23 12.02 4.51 K=6
|
| 437 |
+
verify 33: accepted 5/5 (K=6) + correction -> +6 tokens (177.2 ms) cpos=129
|
| 438 |
+
[margins] 10.92 12.66 12.00 9.61 3.53 K=6
|
| 439 |
+
verify 34: accepted 5/5 (K=6) + correction -> +6 tokens (160.0 ms) cpos=135
|
| 440 |
+
[margins] 3.24 1.08 0.51 0.73 3.56 K=2
|
| 441 |
+
verify 35: accepted 1/1 (K=2) + correction -> +2 tokens (107.6 ms) cpos=137
|
| 442 |
+
[margins] 1.27 1.95 1.46 0.63 3.37 K=3
|
| 443 |
+
verify 36: accepted 0/2 (K=3) + correction -> +1 tokens (115.7 ms) cpos=138
|
| 444 |
+
[margins] 0.62 1.50 0.38 2.67 3.92 K=3
|
| 445 |
+
verify 37: accepted 1/2 (K=3) + correction -> +2 tokens (115.1 ms) cpos=140
|
| 446 |
+
[margins] 3.63 1.63 0.30 0.31 0.53 K=3
|
| 447 |
+
verify 38: accepted 0/2 (K=3) + correction -> +1 tokens (112.4 ms) cpos=141
|
| 448 |
+
[margins] 1.36 0.21 0.93 0.02 0.90 K=2
|
| 449 |
+
verify 39: accepted 1/1 (K=2) + correction -> +2 tokens (99.0 ms) cpos=143
|
| 450 |
+
[margins] 0.59 0.77 0.09 0.26 0.70 K=2
|
| 451 |
+
verify 40: accepted 0/1 (K=2) + correction -> +1 tokens (104.8 ms) cpos=144
|
| 452 |
+
[margins] 1.59 3.16 1.43 2.97 0.01 K=3
|
| 453 |
+
verify 41: accepted 1/2 (K=3) + correction -> +2 tokens (116.0 ms) cpos=146
|
| 454 |
+
[margins] 5.34 0.73 1.33 1.94 0.19 K=2
|
| 455 |
+
verify 42: accepted 1/1 (K=2) + correction -> +2 tokens (104.1 ms) cpos=148
|
| 456 |
+
[margins] 0.09 1.38 2.17 0.02 0.62 K=2
|
| 457 |
+
verify 43: accepted 0/1 (K=2) + correction -> +1 tokens (100.4 ms) cpos=149
|
| 458 |
+
[margins] 0.77 2.74 2.83 0.93 1.31 K=4
|
| 459 |
+
verify 44: accepted 0/3 (K=4) + correction -> +1 tokens (128.7 ms) cpos=150
|
| 460 |
+
[margins] 1.77 0.66 0.41 2.16 0.03 K=2
|
| 461 |
+
verify 45: accepted 1/1 (K=2) + correction -> +2 tokens (102.2 ms) cpos=152
|
| 462 |
+
[margins] 4.15 0.57 0.39 1.68 1.24 K=2
|
| 463 |
+
verify 46: accepted 1/1 (K=2) + correction -> +2 tokens (99.1 ms) cpos=154
|
| 464 |
+
[margins] 0.83 0.63 3.02 0.17 0.58 K=2
|
| 465 |
+
verify 47: accepted 1/1 (K=2) + correction -> +2 tokens (104.9 ms) cpos=156
|
| 466 |
+
[margins] 4.59 1.77 1.01 1.07 0.08 K=3
|
| 467 |
+
verify 48: accepted 1/2 (K=3) + correction -> +2 tokens (110.2 ms) cpos=158
|
| 468 |
+
[margins] 3.51 1.79 1.82 0.65 0.87 K=4
|
| 469 |
+
verify 49: accepted 3/3 (K=4) + correction -> +4 tokens (131.2 ms) cpos=162
|
| 470 |
+
[margins] 5.80 2.99 0.94 0.29 0.81 K=3
|
| 471 |
+
verify 50: accepted 2/2 (K=3) + correction -> +3 tokens (114.6 ms) cpos=165
|
| 472 |
+
[margins] 3.65 1.88 2.35 0.71 2.28 K=4
|
| 473 |
+
verify 51: accepted 2/3 (K=4) + correction -> +3 tokens (129.0 ms) cpos=168
|
| 474 |
+
[margins] 4.78 2.00 7.66 2.62 0.50 K=5
|
| 475 |
+
verify 52: accepted 4/4 (K=5) + correction -> +5 tokens (144.0 ms) cpos=173
|
| 476 |
+
[margins] 0.31 0.45 3.16 0.57 0.07 K=2
|
| 477 |
+
verify 53: accepted 0/1 (K=2) + correction -> +1 tokens (101.3 ms) cpos=174
|
| 478 |
+
[margins] 0.21 2.33 0.27 4.13 3.47 K=3
|
| 479 |
+
verify 54: accepted 0/2 (K=3) + correction -> +1 tokens (118.5 ms) cpos=175
|
| 480 |
+
[margins] 0.25 1.81 1.56 0.06 0.32 K=4
|
| 481 |
+
verify 55: accepted 1/3 (K=4) + correction -> +2 tokens (129.3 ms) cpos=177
|
| 482 |
+
[margins] 4.65 1.15 2.82 2.02 3.33 K=2
|
| 483 |
+
verify 56: accepted 1/1 (K=2) + correction -> +2 tokens (101.2 ms) cpos=179
|
| 484 |
+
[margins] 3.64 2.98 2.40 1.98 0.85 K=5
|
| 485 |
+
verify 57: accepted 1/4 (K=5) + correction -> +2 tokens (148.9 ms) cpos=181
|
| 486 |
+
[margins] 4.91 0.33 0.21 0.40 1.31 K=2
|
| 487 |
+
verify 58: accepted 1/1 (K=2) + correction -> +2 tokens (99.5 ms) cpos=183
|
| 488 |
+
[margins] 2.91 4.45 5.28 3.03 3.57 K=6
|
| 489 |
+
verify 59: accepted 2/5 (K=6) + correction -> +3 tokens (160.3 ms) cpos=186
|
| 490 |
+
[margins] 2.42 3.35 0.77 0.06 1.24 K=3
|
| 491 |
+
verify 60: accepted 2/2 (K=3) + correction -> +3 tokens (117.7 ms) cpos=189
|
| 492 |
+
[margins] 3.32 1.23 0.27 1.84 0.24 K=2
|
| 493 |
+
verify 61: accepted 1/1 (K=2) + correction -> +2 tokens (103.6 ms) cpos=191
|
| 494 |
+
[margins] 8.24 0.76 1.48 1.50 2.27 K=2
|
| 495 |
+
verify 62: accepted 1/1 (K=2) + correction -> +2 tokens (106.4 ms) cpos=193
|
| 496 |
+
[margins] 0.58 1.27 1.05 1.84 1.66 K=2
|
| 497 |
+
verify 63: accepted 1/1 (K=2) + correction -> +2 tokens (100.3 ms) cpos=195
|
| 498 |
+
[margins] 2.48 1.09 2.33 4.29 0.60 K=2
|
| 499 |
+
verify 64: accepted 1/1 (K=2) + correction -> +2 tokens (104.0 ms) cpos=197
|
| 500 |
+
[margins] 0.97 0.58 6.75 0.96 3.33 K=2
|
| 501 |
+
verify 65: accepted 0/1 (K=2) + correction -> +1 tokens (99.5 ms) cpos=198
|
| 502 |
+
[margins] 1.36 0.79 0.41 4.39 1.36 K=2
|
| 503 |
+
verify 66: accepted 0/1 (K=2) + correction -> +1 tokens (105.7 ms) cpos=199
|
| 504 |
+
[margins] 3.82 6.69 7.98 2.25 5.82 K=6
|
| 505 |
+
verify 67: accepted 0/5 (K=6) + correction -> +1 tokens (165.8 ms) cpos=200
|
| 506 |
+
[margins] 7.29 2.77 0.29 2.77 3.45 K=3
|
| 507 |
+
verify 68: accepted 2/2 (K=3) + correction -> +3 tokens (112.8 ms) cpos=203
|
| 508 |
+
[margins] 4.37 3.27 3.24 2.47 1.33 K=5
|
| 509 |
+
verify 69: accepted 1/4 (K=5) + correction -> +2 tokens (143.5 ms) cpos=205
|
| 510 |
+
[margins] 2.44 1.10 0.01 1.41 1.43 K=2
|
| 511 |
+
verify 70: accepted 1/1 (K=2) + correction -> +2 tokens (99.6 ms) cpos=207
|
| 512 |
+
[margins] 0.11 1.82 1.60 3.92 4.61 K=6
|
| 513 |
+
verify 71: accepted 0/5 (K=6) + correction -> +1 tokens (165.5 ms) cpos=208
|
| 514 |
+
[margins] 11.36 0.64 1.96 4.54 0.31 K=2
|
| 515 |
+
verify 72: accepted 1/1 (K=2) + correction -> +2 tokens (101.1 ms) cpos=210
|
| 516 |
+
[margins] 2.98 5.85 2.25 1.83 5.40 K=6
|
| 517 |
+
verify 73: accepted 2/5 (K=6) + correction -> +3 tokens (159.9 ms) cpos=213
|
| 518 |
+
[margins] 0.78 0.08 0.16 1.09 0.40 K=2
|
| 519 |
+
verify 74: accepted 1/1 (K=2) + correction -> +2 tokens (101.0 ms) cpos=215
|
| 520 |
+
[margins] 3.72 2.06 5.09 4.52 2.61 K=6
|
| 521 |
+
verify 75: accepted 4/5 (K=6) + correction -> +5 tokens (171.6 ms) cpos=220
|
| 522 |
+
[margins] 4.17 1.96 2.68 0.63 1.84 K=4
|
| 523 |
+
verify 76: accepted 0/3 (K=4) + correction -> +1 tokens (137.5 ms) cpos=221
|
| 524 |
+
[margins] 11.70 2.40 2.92 0.75 5.34 K=4
|
| 525 |
+
verify 77: accepted 3/3 (K=4) + correction -> +4 tokens (134.9 ms) cpos=225
|
| 526 |
+
[margins] 11.74 6.58 1.21 1.96 0.07 K=3
|
| 527 |
+
verify 78: accepted 2/2 (K=3) + correction -> +3 tokens (122.3 ms) cpos=228
|
| 528 |
+
|
| 529 |
+
[spec] generated 202 tokens over 78 verifies: mean tokens/verify = 2.59 (block=5, max 6)
|
| 530 |
+
[spec] tokens: 1 16 1 16 1 16 1 16 1 16 1 16 1 16 1 16 1 16 1 16 1 16 1 16 1 16 1 16 1 16 1 16 1 16 1 16 1 16 1 16
|
| 531 |
+
[spec] SPEC-DECODE: 47.4 ms/tok = 21.08 tok/s (vs base M=1 68.4 ms/tok = 14.61 tok/s -> 1.44x)
|
| 532 |
+
[decode] PREFILL: 14 positions in 281.1 ms = 49.8 tok/s (20.080 ms/tok)
|
| 533 |
+
[spec] decoding 200 tokens (block=5, draft passes=1, adaptK=1.50, prompt=4 s=15)...
|
| 534 |
+
[margins] 0.29 0.41 1.67 1.63 0.25 K=2
|
| 535 |
+
verify 1: accepted 0/1 (K=2) + correction -> +1 tokens (102.2 ms) cpos=15
|
| 536 |
+
[margins] 0.09 0.07 2.91 2.91 0.81 K=2
|
| 537 |
+
verify 2: accepted 0/1 (K=2) + correction -> +1 tokens (100.2 ms) cpos=16
|
| 538 |
+
[margins] 0.85 0.90 0.63 1.42 0.43 K=2
|
| 539 |
+
verify 3: accepted 0/1 (K=2) + correction -> +1 tokens (97.3 ms) cpos=17
|
| 540 |
+
[margins] 0.32 0.50 0.15 1.86 1.44 K=2
|
| 541 |
+
verify 4: accepted 0/1 (K=2) + correction -> +1 tokens (95.3 ms) cpos=18
|
| 542 |
+
[margins] 0.84 0.49 0.16 0.19 1.50 K=2
|
| 543 |
+
verify 5: accepted 1/1 (K=2) + correction -> +2 tokens (99.7 ms) cpos=20
|
| 544 |
+
[margins] 0.76 1.06 0.58 0.59 1.44 K=2
|
| 545 |
+
verify 6: accepted 1/1 (K=2) + correction -> +2 tokens (94.1 ms) cpos=22
|
| 546 |
+
[margins] 1.84 0.96 0.71 0.37 0.35 K=2
|
| 547 |
+
verify 7: accepted 1/1 (K=2) + correction -> +2 tokens (100.9 ms) cpos=24
|
| 548 |
+
[margins] 0.28 0.35 0.18 0.36 3.59 K=2
|
| 549 |
+
verify 8: accepted 1/1 (K=2) + correction -> +2 tokens (96.6 ms) cpos=26
|
| 550 |
+
[margins] 0.27 0.04 0.32 0.21 0.70 K=2
|
| 551 |
+
verify 9: accepted 0/1 (K=2) + correction -> +1 tokens (100.8 ms) cpos=27
|
| 552 |
+
[margins] 0.24 0.04 0.32 0.86 0.15 K=2
|
| 553 |
+
verify 10: accepted 1/1 (K=2) + correction -> +2 tokens (103.7 ms) cpos=29
|
| 554 |
+
[margins] 3.10 2.37 0.56 3.24 0.89 K=3
|
| 555 |
+
verify 11: accepted 2/2 (K=3) + correction -> +3 tokens (112.8 ms) cpos=32
|
| 556 |
+
[margins] 1.10 1.31 0.21 0.04 1.30 K=2
|
| 557 |
+
verify 12: accepted 1/1 (K=2) + correction -> +2 tokens (94.6 ms) cpos=34
|
| 558 |
+
[margins] 1.42 0.91 0.29 1.34 0.10 K=2
|
| 559 |
+
verify 13: accepted 1/1 (K=2) + correction -> +2 tokens (99.7 ms) cpos=36
|
| 560 |
+
[margins] 0.46 0.04 3.31 2.09 0.11 K=2
|
| 561 |
+
verify 14: accepted 1/1 (K=2) + correction -> +2 tokens (96.2 ms) cpos=38
|
| 562 |
+
[margins] 0.25 2.22 1.24 1.49 1.36 K=3
|
| 563 |
+
verify 15: accepted 1/2 (K=3) + correction -> +2 tokens (111.0 ms) cpos=40
|
| 564 |
+
[margins] 3.35 3.29 2.89 0.90 0.85 K=4
|
| 565 |
+
verify 16: accepted 1/3 (K=4) + correction -> +2 tokens (127.0 ms) cpos=42
|
| 566 |
+
[margins] 0.61 0.04 0.69 0.25 1.32 K=2
|
| 567 |
+
verify 17: accepted 1/1 (K=2) + correction -> +2 tokens (101.2 ms) cpos=44
|
| 568 |
+
[margins] 1.75 8.28 0.77 1.66 1.84 K=3
|
| 569 |
+
verify 18: accepted 2/2 (K=3) + correction -> +3 tokens (109.9 ms) cpos=47
|
| 570 |
+
[margins] 3.67 7.43 2.84 1.51 2.83 K=6
|
| 571 |
+
verify 19: accepted 3/5 (K=6) + correction -> +4 tokens (160.8 ms) cpos=51
|
| 572 |
+
[margins] 2.57 0.32 0.49 2.08 0.13 K=2
|
| 573 |
+
verify 20: accepted 0/1 (K=2) + correction -> +1 tokens (101.4 ms) cpos=52
|
| 574 |
+
[margins] 1.68 1.98 1.30 1.17 3.86 K=3
|
| 575 |
+
verify 21: accepted 0/2 (K=3) + correction -> +1 tokens (108.4 ms) cpos=53
|
| 576 |
+
[margins] 7.66 3.20 2.17 0.89 2.44 K=4
|
| 577 |
+
verify 22: accepted 2/3 (K=4) + correction -> +3 tokens (129.4 ms) cpos=56
|
| 578 |
+
[margins] 4.57 1.62 4.31 3.17 1.27 K=5
|
| 579 |
+
verify 23: accepted 4/4 (K=5) + correction -> +5 tokens (136.4 ms) cpos=61
|
| 580 |
+
[margins] 1.33 1.58 0.21 1.13 0.04 K=3
|
| 581 |
+
verify 24: accepted 0/2 (K=3) + correction -> +1 tokens (112.8 ms) cpos=62
|
| 582 |
+
[margins] 1.18 0.61 1.90 1.57 2.89 K=2
|
| 583 |
+
verify 25: accepted 0/1 (K=2) + correction -> +1 tokens (100.2 ms) cpos=63
|
| 584 |
+
[margins] 1.26 1.57 1.06 1.37 0.26 K=3
|
| 585 |
+
verify 26: accepted 2/2 (K=3) + correction -> +3 tokens (114.8 ms) cpos=66
|
| 586 |
+
[margins] 0.90 3.00 0.57 0.06 0.01 K=3
|
| 587 |
+
verify 27: accepted 0/2 (K=3) + correction -> +1 tokens (113.0 ms) cpos=67
|
| 588 |
+
[margins] 2.15 0.72 0.77 2.15 1.43 K=2
|
| 589 |
+
verify 28: accepted 1/1 (K=2) + correction -> +2 tokens (103.1 ms) cpos=69
|
| 590 |
+
[margins] 0.13 0.42 0.62 0.05 0.94 K=2
|
| 591 |
+
verify 29: accepted 0/1 (K=2) + correction -> +1 tokens (99.0 ms) cpos=70
|
| 592 |
+
[margins] 1.86 0.99 0.79 0.56 0.43 K=2
|
| 593 |
+
verify 30: accepted 0/1 (K=2) + correction -> +1 tokens (101.6 ms) cpos=71
|
| 594 |
+
[margins] 0.26 0.72 1.96 0.20 0.00 K=2
|
| 595 |
+
verify 31: accepted 0/1 (K=2) + correction -> +1 tokens (103.2 ms) cpos=72
|
| 596 |
+
[margins] 0.34 0.12 2.20 0.60 0.04 K=2
|
| 597 |
+
verify 32: accepted 0/1 (K=2) + correction -> +1 tokens (97.9 ms) cpos=73
|
| 598 |
+
[margins] 5.45 2.08 1.05 0.43 1.28 K=3
|
| 599 |
+
verify 33: accepted 2/2 (K=3) + correction -> +3 tokens (113.1 ms) cpos=76
|
| 600 |
+
[margins] 2.30 3.95 1.90 1.08 1.43 K=4
|
| 601 |
+
verify 34: accepted 0/3 (K=4) + correction -> +1 tokens (127.7 ms) cpos=77
|
| 602 |
+
[margins] 3.77 1.77 0.11 2.50 0.75 K=3
|
| 603 |
+
verify 35: accepted 2/2 (K=3) + correction -> +3 tokens (115.7 ms) cpos=80
|
| 604 |
+
[margins] 1.70 5.19 1.25 2.27 8.57 K=3
|
| 605 |
+
verify 36: accepted 2/2 (K=3) + correction -> +3 tokens (110.4 ms) cpos=83
|
| 606 |
+
[margins] 1.92 11.06 2.73 2.86 5.42 K=6
|
| 607 |
+
verify 37: accepted 2/5 (K=6) + correction -> +3 tokens (162.5 ms) cpos=86
|
| 608 |
+
[margins] 4.76 5.08 3.62 2.21 0.39 K=5
|
| 609 |
+
verify 38: accepted 2/4 (K=5) + correction -> +3 tokens (129.4 ms) cpos=89
|
| 610 |
+
[margins] 7.00 4.01 0.96 0.81 4.75 K=3
|
| 611 |
+
verify 39: accepted 1/2 (K=3) + correction -> +2 tokens (112.5 ms) cpos=91
|
| 612 |
+
[margins] 4.60 3.17 1.28 1.48 3.69 K=3
|
| 613 |
+
verify 40: accepted 1/2 (K=3) + correction -> +2 tokens (112.0 ms) cpos=93
|
| 614 |
+
[margins] 3.88 2.19 5.17 3.24 2.46 K=6
|
| 615 |
+
verify 41: accepted 1/5 (K=6) + correction -> +2 tokens (150.4 ms) cpos=95
|
| 616 |
+
[margins] 3.93 1.64 6.55 2.88 3.76 K=6
|
| 617 |
+
verify 42: accepted 1/5 (K=6) + correction -> +2 tokens (156.4 ms) cpos=97
|
| 618 |
+
[margins] 2.71 2.44 2.26 1.59 6.31 K=6
|
| 619 |
+
verify 43: accepted 3/5 (K=6) + correction -> +4 tokens (151.4 ms) cpos=101
|
| 620 |
+
[margins] 2.38 4.93 2.49 1.58 3.37 K=6
|
| 621 |
+
verify 44: accepted 5/5 (K=6) + correction -> +6 tokens (147.5 ms) cpos=107
|
| 622 |
+
[margins] 2.12 5.64 2.45 2.72 3.05 K=6
|
| 623 |
+
verify 45: accepted 5/5 (K=6) + correction -> +6 tokens (153.8 ms) cpos=113
|
| 624 |
+
[margins] 2.68 4.65 2.43 1.74 2.67 K=6
|
| 625 |
+
verify 46: accepted 5/5 (K=6) + correction -> +6 tokens (148.2 ms) cpos=119
|
| 626 |
+
[margins] 3.29 5.01 2.90 1.93 3.11 K=6
|
| 627 |
+
verify 47: accepted 5/5 (K=6) + correction -> +6 tokens (156.4 ms) cpos=125
|
| 628 |
+
[margins] 3.24 4.75 3.58 0.79 3.10 K=4
|
| 629 |
+
verify 48: accepted 3/3 (K=4) + correction -> +4 tokens (137.8 ms) cpos=129
|
| 630 |
+
[margins] 3.02 4.76 3.99 1.66 3.97 K=6
|
| 631 |
+
verify 49: accepted 5/5 (K=6) + correction -> +6 tokens (150.8 ms) cpos=135
|
| 632 |
+
[margins] 3.34 6.17 2.59 2.87 3.69 K=6
|
| 633 |
+
verify 50: accepted 5/5 (K=6) + correction -> +6 tokens (154.8 ms) cpos=141
|
| 634 |
+
[margins] 3.22 6.03 3.09 2.03 2.79 K=6
|
| 635 |
+
verify 51: accepted 5/5 (K=6) + correction -> +6 tokens (150.7 ms) cpos=147
|
| 636 |
+
[margins] 4.42 5.60 2.38 0.70 2.62 K=4
|
| 637 |
+
verify 52: accepted 3/3 (K=4) + correction -> +4 tokens (126.6 ms) cpos=151
|
| 638 |
+
[margins] 3.33 6.65 2.86 2.75 3.48 K=6
|
| 639 |
+
verify 53: accepted 4/5 (K=6) + correction -> +5 tokens (156.4 ms) cpos=156
|
| 640 |
+
[margins] 5.36 2.38 4.10 1.63 1.26 K=5
|
| 641 |
+
verify 54: accepted 1/4 (K=5) + correction -> +2 tokens (138.8 ms) cpos=158
|
| 642 |
+
[margins] 4.50 1.43 1.05 0.20 7.23 K=2
|
| 643 |
+
verify 55: accepted 0/1 (K=2) + correction -> +1 tokens (104.3 ms) cpos=159
|
| 644 |
+
[margins] 7.98 0.21 0.14 1.75 0.16 K=2
|
| 645 |
+
verify 56: accepted 1/1 (K=2) + correction -> +2 tokens (104.8 ms) cpos=161
|
| 646 |
+
[margins] 7.96 5.46 1.12 4.48 1.86 K=3
|
| 647 |
+
verify 57: accepted 2/2 (K=3) + correction -> +3 tokens (115.9 ms) cpos=164
|
| 648 |
+
[margins] 7.77 5.28 3.26 3.80 2.22 K=6
|
| 649 |
+
verify 58: accepted 2/5 (K=6) + correction -> +3 tokens (156.5 ms) cpos=167
|
| 650 |
+
[margins] 10.58 3.92 8.42 8.55 4.19 K=6
|
| 651 |
+
verify 59: accepted 5/5 (K=6) + correction -> +6 tokens (166.3 ms) cpos=173
|
| 652 |
+
[margins] 5.62 0.59 0.90 0.92 0.87 K=2
|
| 653 |
+
verify 60: accepted 1/1 (K=2) + correction -> +2 tokens (99.4 ms) cpos=175
|
| 654 |
+
[margins] 7.42 8.20 0.27 1.52 4.52 K=3
|
| 655 |
+
verify 61: accepted 2/2 (K=3) + correction -> +3 tokens (118.2 ms) cpos=178
|
| 656 |
+
[margins] 16.47 13.13 10.55 10.02 9.22 K=6
|
| 657 |
+
verify 62: accepted 5/5 (K=6) + correction -> +6 tokens (167.3 ms) cpos=184
|
| 658 |
+
[margins] 6.22 7.81 10.06 12.91 5.08 K=6
|
| 659 |
+
verify 63: accepted 5/5 (K=6) + correction -> +6 tokens (162.3 ms) cpos=190
|
| 660 |
+
[margins] 10.95 5.20 9.73 7.75 12.23 K=6
|
| 661 |
+
verify 64: accepted 5/5 (K=6) + correction -> +6 tokens (169.7 ms) cpos=196
|
| 662 |
+
[margins] 13.24 9.02 7.71 8.96 3.00 K=6
|
| 663 |
+
verify 65: accepted 4/5 (K=6) + correction -> +5 tokens (165.2 ms) cpos=201
|
| 664 |
+
[margins] 5.83 10.10 10.99 15.46 7.52 K=6
|
| 665 |
+
verify 66: accepted 5/5 (K=6) + correction -> +6 tokens (163.6 ms) cpos=207
|
| 666 |
+
[margins] 10.50 5.98 9.04 11.20 12.32 K=6
|
| 667 |
+
verify 67: accepted 5/5 (K=6) + correction -> +6 tokens (167.4 ms) cpos=213
|
| 668 |
+
[margins] 12.32 15.19 12.04 9.92 3.43 K=6
|
| 669 |
+
verify 68: accepted 4/5 (K=6) + correction -> +5 tokens (163.6 ms) cpos=218
|
| 670 |
+
|
| 671 |
+
[spec] generated 204 tokens over 68 verifies: mean tokens/verify = 3.00 (block=5, max 6)
|
| 672 |
+
[spec] tokens: 5715 14 295 834 10175 14 8470 3939 436 344 2239 362 4197 339 35 126664 23809 16765 270 3077 305 1990 23714 538 9479 21121 35977 2184 47270 34788 14 832 851 588 366 99782 1082 20351 14319 35977
|
| 673 |
+
[spec] SPEC-DECODE: 41.4 ms/tok = 24.14 tok/s (vs base M=1 68.4 ms/tok = 14.61 tok/s -> 1.65x)
|
| 674 |
+
[decode] PREFILL: 82 positions in 1289.2 ms = 63.6 tok/s (15.722 ms/tok)
|
| 675 |
+
[spec] decoding 200 tokens (block=5, draft passes=1, adaptK=1.50, prompt=5 s=83)...
|
| 676 |
+
[margins] 7.79 4.69 1.65 0.12 3.01 K=4
|
| 677 |
+
verify 1: accepted 1/3 (K=4) + correction -> +2 tokens (130.2 ms) cpos=84
|
| 678 |
+
[margins] 0.02 0.35 0.33 1.11 0.90 K=2
|
| 679 |
+
verify 2: accepted 0/1 (K=2) + correction -> +1 tokens (99.2 ms) cpos=85
|
| 680 |
+
[margins] 0.28 0.01 0.19 0.58 1.06 K=2
|
| 681 |
+
verify 3: accepted 1/1 (K=2) + correction -> +2 tokens (99.0 ms) cpos=87
|
| 682 |
+
[margins] 6.16 2.19 4.10 1.30 0.10 K=4
|
| 683 |
+
verify 4: accepted 3/3 (K=4) + correction -> +4 tokens (128.6 ms) cpos=91
|
| 684 |
+
[margins] 11.91 7.31 6.91 3.86 3.69 K=6
|
| 685 |
+
verify 5: accepted 5/5 (K=6) + correction -> +6 tokens (163.0 ms) cpos=97
|
| 686 |
+
[margins] 8.85 5.59 3.54 0.57 2.74 K=4
|
| 687 |
+
verify 6: accepted 3/3 (K=4) + correction -> +4 tokens (128.4 ms) cpos=101
|
| 688 |
+
[margins] 4.90 0.75 2.43 0.34 0.06 K=2
|
| 689 |
+
verify 7: accepted 1/1 (K=2) + correction -> +2 tokens (102.6 ms) cpos=103
|
| 690 |
+
[margins] 2.70 1.98 2.44 0.48 1.46 K=4
|
| 691 |
+
verify 8: accepted 3/3 (K=4) + correction -> +4 tokens (129.3 ms) cpos=107
|
| 692 |
+
[margins] 7.35 8.07 9.98 7.46 3.32 K=6
|
| 693 |
+
verify 9: accepted 5/5 (K=6) + correction -> +6 tokens (165.6 ms) cpos=113
|
| 694 |
+
[margins] 8.11 11.81 8.12 9.51 14.11 K=6
|
| 695 |
+
verify 10: accepted 5/5 (K=6) + correction -> +6 tokens (159.0 ms) cpos=119
|
| 696 |
+
[margins] 4.58 2.22 0.04 1.40 5.07 K=3
|
| 697 |
+
verify 11: accepted 2/2 (K=3) + correction -> +3 tokens (117.8 ms) cpos=122
|
| 698 |
+
[margins] 8.36 3.92 8.33 2.79 2.22 K=6
|
| 699 |
+
verify 12: accepted 3/5 (K=6) + correction -> +4 tokens (181.1 ms) cpos=126
|
| 700 |
+
[margins] 13.05 5.46 4.64 9.82 7.69 K=6
|
| 701 |
+
verify 13: accepted 5/5 (K=6) + correction -> +6 tokens (180.9 ms) cpos=132
|
| 702 |
+
[margins] 12.36 14.29 0.31 9.08 10.48 K=3
|
| 703 |
+
verify 14: accepted 2/2 (K=3) + correction -> +3 tokens (114.8 ms) cpos=135
|
| 704 |
+
[margins] 13.21 15.45 10.09 7.63 0.68 K=5
|
| 705 |
+
verify 15: accepted 4/4 (K=5) + correction -> +5 tokens (148.5 ms) cpos=140
|
| 706 |
+
[margins] 7.49 2.35 5.92 1.60 4.92 K=6
|
| 707 |
+
verify 16: accepted 0/5 (K=6) + correction -> +1 tokens (155.7 ms) cpos=141
|
| 708 |
+
[margins] 7.00 5.91 1.69 0.60 1.10 K=4
|
| 709 |
+
verify 17: accepted 2/3 (K=4) + correction -> +3 tokens (128.6 ms) cpos=144
|
| 710 |
+
[margins] 10.09 5.58 4.79 6.12 0.76 K=5
|
| 711 |
+
verify 18: accepted 4/4 (K=5) + correction -> +5 tokens (144.7 ms) cpos=149
|
| 712 |
+
[margins] 13.47 11.96 5.75 5.81 11.44 K=6
|
| 713 |
+
verify 19: accepted 5/5 (K=6) + correction -> +6 tokens (160.7 ms) cpos=155
|
| 714 |
+
[margins] 15.73 11.82 11.61 8.26 0.59 K=5
|
| 715 |
+
verify 20: accepted 4/4 (K=5) + correction -> +5 tokens (148.4 ms) cpos=160
|
| 716 |
+
[margins] 8.56 2.17 10.45 6.32 4.86 K=6
|
| 717 |
+
verify 21: accepted 0/5 (K=6) + correction -> +1 tokens (156.3 ms) cpos=161
|
| 718 |
+
[margins] 10.05 6.32 0.26 6.77 0.09 K=3
|
| 719 |
+
verify 22: accepted 2/2 (K=3) + correction -> +3 tokens (116.4 ms) cpos=164
|
| 720 |
+
[margins] 10.36 7.85 4.09 8.54 7.71 K=6
|
| 721 |
+
verify 23: accepted 5/5 (K=6) + correction -> +6 tokens (161.2 ms) cpos=170
|
| 722 |
+
[margins] 1.03 7.49 11.44 9.91 0.05 K=5
|
| 723 |
+
verify 24: accepted 4/4 (K=5) + correction -> +5 tokens (145.3 ms) cpos=175
|
| 724 |
+
[margins] 15.57 13.29 11.19 7.43 1.69 K=6
|
| 725 |
+
verify 25: accepted 4/5 (K=6) + correction -> +5 tokens (167.6 ms) cpos=180
|
| 726 |
+
[margins] 16.39 16.18 6.71 4.13 9.00 K=6
|
| 727 |
+
verify 26: accepted 5/5 (K=6) + correction -> +6 tokens (160.5 ms) cpos=186
|
| 728 |
+
[margins] 5.71 9.35 3.61 6.79 2.12 K=6
|
| 729 |
+
verify 27: accepted 5/5 (K=6) + correction -> +6 tokens (166.8 ms) cpos=192
|
| 730 |
+
[margins] 13.02 16.02 8.84 10.13 11.85 K=6
|
| 731 |
+
verify 28: accepted 5/5 (K=6) + correction -> +6 tokens (161.2 ms) cpos=198
|
| 732 |
+
[margins] 8.75 10.62 12.35 13.49 3.14 K=6
|
| 733 |
+
verify 29: accepted 5/5 (K=6) + correction -> +6 tokens (163.5 ms) cpos=204
|
| 734 |
+
[margins] 9.36 4.03 3.58 6.70 4.04 K=6
|
| 735 |
+
verify 30: accepted 5/5 (K=6) + correction -> +6 tokens (160.8 ms) cpos=210
|
| 736 |
+
[margins] 9.00 8.90 8.63 9.22 0.44 K=5
|
| 737 |
+
verify 31: accepted 4/4 (K=5) + correction -> +5 tokens (144.5 ms) cpos=215
|
| 738 |
+
[margins] 16.68 14.18 9.93 7.88 0.75 K=5
|
| 739 |
+
verify 32: accepted 4/4 (K=5) + correction -> +5 tokens (150.9 ms) cpos=220
|
| 740 |
+
[margins] 17.34 13.13 3.34 1.87 6.64 K=6
|
| 741 |
+
verify 33: accepted 5/5 (K=6) + correction -> +6 tokens (159.8 ms) cpos=226
|
| 742 |
+
[margins] 9.85 4.39 1.41 10.63 1.18 K=3
|
| 743 |
+
verify 34: accepted 2/2 (K=3) + correction -> +3 tokens (120.3 ms) cpos=229
|
| 744 |
+
[margins] 18.61 10.57 8.22 10.92 10.89 K=6
|
| 745 |
+
verify 35: accepted 5/5 (K=6) + correction -> +6 tokens (165.1 ms) cpos=235
|
| 746 |
+
[margins] 8.86 11.06 7.87 5.90 0.37 K=5
|
| 747 |
+
verify 36: accepted 4/4 (K=5) + correction -> +5 tokens (151.8 ms) cpos=240
|
| 748 |
+
[margins] 17.85 14.64 5.60 3.33 7.27 K=6
|
| 749 |
+
verify 37: accepted 5/5 (K=6) + correction -> +6 tokens (162.4 ms) cpos=246
|
| 750 |
+
[margins] 11.20 4.54 0.04 11.09 0.52 K=3
|
| 751 |
+
verify 38: accepted 2/2 (K=3) + correction -> +3 tokens (121.7 ms) cpos=249
|
| 752 |
+
[margins] 14.78 12.09 7.36 12.65 11.04 K=6
|
| 753 |
+
verify 39: accepted 5/5 (K=6) + correction -> +6 tokens (161.0 ms) cpos=255
|
| 754 |
+
[margins] 15.39 12.54 11.17 6.72 6.81 K=6
|
| 755 |
+
verify 40: accepted 5/5 (K=6) + correction -> +6 tokens (183.3 ms) cpos=261
|
| 756 |
+
[margins] 15.10 9.57 6.60 8.14 0.20 K=5
|
| 757 |
+
verify 41: accepted 4/4 (K=5) + correction -> +5 tokens (145.9 ms) cpos=266
|
| 758 |
+
[margins] 12.79 10.48 1.20 11.18 4.69 K=3
|
| 759 |
+
verify 42: accepted 2/2 (K=3) + correction -> +3 tokens (120.3 ms) cpos=269
|
| 760 |
+
[margins] 14.59 13.24 7.37 12.78 9.13 K=6
|
| 761 |
+
verify 43: accepted 5/5 (K=6) + correction -> +6 tokens (165.2 ms) cpos=275
|
| 762 |
+
[margins] 13.77 15.80 10.92 8.70 8.60 K=6
|
| 763 |
+
verify 44: accepted 5/5 (K=6) + correction -> +6 tokens (168.5 ms) cpos=281
|
| 764 |
+
[margins] 14.68 6.75 6.35 8.40 0.40 K=5
|
| 765 |
+
verify 45: accepted 4/4 (K=5) + correction -> +5 tokens (145.3 ms) cpos=286
|
| 766 |
+
|
| 767 |
+
[spec] generated 204 tokens over 45 verifies: mean tokens/verify = 4.53 (block=5, max 6)
|
| 768 |
+
[spec] tokens: 1 50 201 671 15020 4181 362 2775 14817 14922 344 223 25 22765 7975 55 15 3180 16 1 50 201 671 15020 4181 362 2775 14817 14922 344 223 25 22765 7975 55 15 3180 16 1 50
|
| 769 |
+
[spec] SPEC-DECODE: 32.1 ms/tok = 31.17 tok/s (vs base M=1 68.4 ms/tok = 14.61 tok/s -> 2.13x)
|
| 770 |
+
[decode] PREFILL: 14 positions in 282.7 ms = 49.5 tok/s (20.194 ms/tok)
|
| 771 |
+
[spec] decoding 200 tokens (block=5, draft passes=1, adaptK=1.50, prompt=6 s=15)...
|
| 772 |
+
[margins] 0.44 5.98 0.81 0.96 0.33 K=3
|
| 773 |
+
verify 1: accepted 1/2 (K=3) + correction -> +2 tokens (115.3 ms) cpos=16
|
| 774 |
+
[margins] 3.37 0.07 0.85 1.35 4.17 K=2
|
| 775 |
+
verify 2: accepted 1/1 (K=2) + correction -> +2 tokens (98.5 ms) cpos=18
|
| 776 |
+
[margins] 1.75 0.62 1.76 2.15 3.91 K=2
|
| 777 |
+
verify 3: accepted 1/1 (K=2) + correction -> +2 tokens (103.0 ms) cpos=20
|
| 778 |
+
[margins] 1.68 1.26 3.74 2.98 4.03 K=2
|
| 779 |
+
verify 4: accepted 1/1 (K=2) + correction -> +2 tokens (98.1 ms) cpos=22
|
| 780 |
+
[margins] 3.00 0.68 4.07 3.16 3.92 K=2
|
| 781 |
+
verify 5: accepted 1/1 (K=2) + correction -> +2 tokens (101.6 ms) cpos=24
|
| 782 |
+
[margins] 1.77 1.17 1.47 1.66 4.31 K=2
|
| 783 |
+
verify 6: accepted 1/1 (K=2) + correction -> +2 tokens (98.6 ms) cpos=26
|
| 784 |
+
[margins] 1.94 1.14 2.31 1.17 5.37 K=2
|
| 785 |
+
verify 7: accepted 1/1 (K=2) + correction -> +2 tokens (102.5 ms) cpos=28
|
| 786 |
+
[margins] 1.79 1.25 2.52 0.86 6.53 K=2
|
| 787 |
+
verify 8: accepted 1/1 (K=2) + correction -> +2 tokens (97.8 ms) cpos=30
|
| 788 |
+
[margins] 1.51 2.18 2.69 0.16 7.43 K=4
|
| 789 |
+
verify 9: accepted 0/3 (K=4) + correction -> +1 tokens (118.4 ms) cpos=31
|
| 790 |
+
[margins] 0.36 8.55 2.26 0.64 7.31 K=4
|
| 791 |
+
verify 10: accepted 0/3 (K=4) + correction -> +1 tokens (125.2 ms) cpos=32
|
| 792 |
+
[margins] 7.07 0.17 1.61 7.74 2.07 K=2
|
| 793 |
+
verify 11: accepted 1/1 (K=2) + correction -> +2 tokens (96.7 ms) cpos=34
|
| 794 |
+
[margins] 1.84 6.07 0.21 1.03 0.64 K=3
|
| 795 |
+
verify 12: accepted 2/2 (K=3) + correction -> +3 tokens (110.6 ms) cpos=37
|
| 796 |
+
[margins] 12.33 7.09 4.03 3.92 6.71 K=6
|
| 797 |
+
verify 13: accepted 5/5 (K=6) + correction -> +6 tokens (148.5 ms) cpos=43
|
| 798 |
+
[margins] 14.84 8.16 8.92 10.39 7.32 K=6
|
| 799 |
+
verify 14: accepted 5/5 (K=6) + correction -> +6 tokens (155.8 ms) cpos=49
|
| 800 |
+
[margins] 16.51 10.08 10.84 10.01 10.41 K=6
|
| 801 |
+
verify 15: accepted 5/5 (K=6) + correction -> +6 tokens (151.8 ms) cpos=55
|
| 802 |
+
[margins] 14.95 10.17 10.91 7.50 9.57 K=6
|
| 803 |
+
verify 16: accepted 5/5 (K=6) + correction -> +6 tokens (154.8 ms) cpos=61
|
| 804 |
+
[margins] 18.71 10.35 12.55 9.78 10.69 K=6
|
| 805 |
+
verify 17: accepted 5/5 (K=6) + correction -> +6 tokens (152.0 ms) cpos=67
|
| 806 |
+
[margins] 17.24 10.49 9.29 9.66 9.53 K=6
|
| 807 |
+
verify 18: accepted 5/5 (K=6) + correction -> +6 tokens (159.2 ms) cpos=73
|
| 808 |
+
[margins] 16.06 10.60 11.92 9.02 10.50 K=6
|
| 809 |
+
verify 19: accepted 5/5 (K=6) + correction -> +6 tokens (152.9 ms) cpos=79
|
| 810 |
+
[margins] 16.47 10.91 13.05 11.30 12.43 K=6
|
| 811 |
+
verify 20: accepted 5/5 (K=6) + correction -> +6 tokens (157.7 ms) cpos=85
|
| 812 |
+
[margins] 16.59 10.89 10.83 11.70 9.96 K=6
|
| 813 |
+
verify 21: accepted 5/5 (K=6) + correction -> +6 tokens (155.3 ms) cpos=91
|
| 814 |
+
[margins] 17.15 10.88 12.35 8.96 11.04 K=6
|
| 815 |
+
verify 22: accepted 5/5 (K=6) + correction -> +6 tokens (161.9 ms) cpos=97
|
| 816 |
+
[margins] 18.14 11.52 14.10 10.76 12.14 K=6
|
| 817 |
+
verify 23: accepted 5/5 (K=6) + correction -> +6 tokens (153.1 ms) cpos=103
|
| 818 |
+
[margins] 14.97 10.97 13.31 11.55 10.32 K=6
|
| 819 |
+
verify 24: accepted 5/5 (K=6) + correction -> +6 tokens (161.8 ms) cpos=109
|
| 820 |
+
[margins] 17.27 10.76 13.91 10.87 10.76 K=6
|
| 821 |
+
verify 25: accepted 5/5 (K=6) + correction -> +6 tokens (156.2 ms) cpos=115
|
| 822 |
+
[margins] 14.70 10.42 13.75 11.38 10.40 K=6
|
| 823 |
+
verify 26: accepted 5/5 (K=6) + correction -> +6 tokens (160.5 ms) cpos=121
|
| 824 |
+
[margins] 16.55 10.81 13.15 8.58 11.41 K=6
|
| 825 |
+
verify 27: accepted 5/5 (K=6) + correction -> +6 tokens (157.4 ms) cpos=127
|
| 826 |
+
[margins] 17.05 10.93 13.96 10.78 10.77 K=6
|
| 827 |
+
verify 28: accepted 5/5 (K=6) + correction -> +6 tokens (175.9 ms) cpos=133
|
| 828 |
+
[margins] 16.57 9.20 12.73 7.65 9.87 K=6
|
| 829 |
+
verify 29: accepted 5/5 (K=6) + correction -> +6 tokens (154.0 ms) cpos=139
|
| 830 |
+
[margins] 15.10 9.66 13.36 9.46 10.78 K=6
|
| 831 |
+
verify 30: accepted 5/5 (K=6) + correction -> +6 tokens (160.8 ms) cpos=145
|
| 832 |
+
[margins] 16.90 10.03 12.96 10.83 9.70 K=6
|
| 833 |
+
verify 31: accepted 5/5 (K=6) + correction -> +6 tokens (155.8 ms) cpos=151
|
| 834 |
+
[margins] 14.83 10.28 14.34 10.70 10.47 K=6
|
| 835 |
+
verify 32: accepted 5/5 (K=6) + correction -> +6 tokens (160.1 ms) cpos=157
|
| 836 |
+
[margins] 16.21 9.60 12.10 12.39 10.58 K=6
|
| 837 |
+
verify 33: accepted 5/5 (K=6) + correction -> +6 tokens (159.5 ms) cpos=163
|
| 838 |
+
[margins] 18.54 9.38 13.03 10.65 9.80 K=6
|
| 839 |
+
verify 34: accepted 5/5 (K=6) + correction -> +6 tokens (160.9 ms) cpos=169
|
| 840 |
+
[margins] 15.89 10.59 13.07 7.85 12.17 K=6
|
| 841 |
+
verify 35: accepted 5/5 (K=6) + correction -> +6 tokens (158.6 ms) cpos=175
|
| 842 |
+
[margins] 17.30 9.59 12.80 8.52 10.71 K=6
|
| 843 |
+
verify 36: accepted 5/5 (K=6) + correction -> +6 tokens (162.2 ms) cpos=181
|
| 844 |
+
[margins] 18.13 10.37 12.62 8.69 10.32 K=6
|
| 845 |
+
verify 37: accepted 5/5 (K=6) + correction -> +6 tokens (159.0 ms) cpos=187
|
| 846 |
+
[margins] 18.11 10.80 14.03 12.35 9.50 K=6
|
| 847 |
+
verify 38: accepted 5/5 (K=6) + correction -> +6 tokens (163.9 ms) cpos=193
|
| 848 |
+
[margins] 17.53 9.79 13.56 11.00 8.84 K=6
|
| 849 |
+
verify 39: accepted 5/5 (K=6) + correction -> +6 tokens (158.0 ms) cpos=199
|
| 850 |
+
[margins] 19.79 9.86 13.45 11.04 9.32 K=6
|
| 851 |
+
verify 40: accepted 5/5 (K=6) + correction -> +6 tokens (163.1 ms) cpos=205
|
| 852 |
+
[margins] 17.05 10.57 13.28 11.31 10.40 K=6
|
| 853 |
+
verify 41: accepted 5/5 (K=6) + correction -> +6 tokens (158.2 ms) cpos=211
|
| 854 |
+
[margins] 19.39 9.73 12.74 7.52 9.74 K=6
|
| 855 |
+
verify 42: accepted 5/5 (K=6) + correction -> +6 tokens (163.4 ms) cpos=217
|
| 856 |
+
|
| 857 |
+
[spec] generated 203 tokens over 42 verifies: mean tokens/verify = 4.83 (block=5, max 6)
|
| 858 |
+
[spec] tokens: 1 16 1 16 1 16 1 16 1 16 1 16 1 16 1 16 403 31310 16 403 31310 16 403 31310 16 403 31310 16 403 31310 16 403 31310 16 403 31310 16 403 31310 16
|
| 859 |
+
[spec] SPEC-DECODE: 29.4 ms/tok = 34.05 tok/s (vs base M=1 68.4 ms/tok = 14.61 tok/s -> 2.33x)
|
| 860 |
+
[decode] PREFILL: 37 positions in 640.7 ms = 57.7 tok/s (17.317 ms/tok)
|
| 861 |
+
[spec] decoding 200 tokens (block=5, draft passes=1, adaptK=1.50, prompt=7 s=38)...
|
| 862 |
+
[margins] 0.98 3.15 0.17 0.18 0.29 K=3
|
| 863 |
+
verify 1: accepted 0/2 (K=3) + correction -> +1 tokens (116.5 ms) cpos=38
|
| 864 |
+
[margins] 0.04 6.06 0.38 1.71 0.08 K=3
|
| 865 |
+
verify 2: accepted 0/2 (K=3) + correction -> +1 tokens (111.5 ms) cpos=39
|
| 866 |
+
[margins] 1.75 0.47 5.09 0.60 0.70 K=2
|
| 867 |
+
verify 3: accepted 1/1 (K=2) + correction -> +2 tokens (99.8 ms) cpos=41
|
| 868 |
+
[margins] 3.00 0.30 1.07 6.03 0.53 K=2
|
| 869 |
+
verify 4: accepted 1/1 (K=2) + correction -> +2 tokens (94.6 ms) cpos=43
|
| 870 |
+
[margins] 0.99 1.90 0.23 0.25 0.25 K=3
|
| 871 |
+
verify 5: accepted 0/2 (K=3) + correction -> +1 tokens (111.5 ms) cpos=44
|
| 872 |
+
[margins] 0.62 0.13 0.94 0.37 0.49 K=2
|
| 873 |
+
verify 6: accepted 1/1 (K=2) + correction -> +2 tokens (95.3 ms) cpos=46
|
| 874 |
+
[margins] 1.13 1.12 1.91 0.20 1.86 K=2
|
| 875 |
+
verify 7: accepted 0/1 (K=2) + correction -> +1 tokens (101.1 ms) cpos=47
|
| 876 |
+
[margins] 2.20 2.68 0.15 2.05 1.09 K=3
|
| 877 |
+
verify 8: accepted 1/2 (K=3) + correction -> +2 tokens (113.7 ms) cpos=49
|
| 878 |
+
[margins] 8.38 6.72 5.84 2.82 0.90 K=5
|
| 879 |
+
verify 9: accepted 4/4 (K=5) + correction -> +5 tokens (137.9 ms) cpos=54
|
| 880 |
+
[margins] 0.41 0.33 1.23 0.98 3.08 K=2
|
| 881 |
+
verify 10: accepted 0/1 (K=2) + correction -> +1 tokens (102.6 ms) cpos=55
|
| 882 |
+
[margins] 1.18 0.68 1.05 3.04 3.35 K=2
|
| 883 |
+
verify 11: accepted 1/1 (K=2) + correction -> +2 tokens (102.2 ms) cpos=57
|
| 884 |
+
[margins] 2.96 0.56 0.60 1.65 0.16 K=2
|
| 885 |
+
verify 12: accepted 1/1 (K=2) + correction -> +2 tokens (96.9 ms) cpos=59
|
| 886 |
+
[margins] 4.36 2.50 0.67 0.18 0.89 K=3
|
| 887 |
+
verify 13: accepted 2/2 (K=3) + correction -> +3 tokens (113.5 ms) cpos=62
|
| 888 |
+
[margins] 1.49 0.06 2.81 1.53 3.46 K=2
|
| 889 |
+
verify 14: accepted 1/1 (K=2) + correction -> +2 tokens (102.3 ms) cpos=64
|
| 890 |
+
[margins] 6.41 8.88 4.29 3.34 4.84 K=6
|
| 891 |
+
verify 15: accepted 3/5 (K=6) + correction -> +4 tokens (155.6 ms) cpos=68
|
| 892 |
+
[margins] 5.74 2.84 0.62 0.36 0.77 K=3
|
| 893 |
+
verify 16: accepted 2/2 (K=3) + correction -> +3 tokens (108.6 ms) cpos=71
|
| 894 |
+
[margins] 3.41 2.09 3.91 0.45 0.04 K=4
|
| 895 |
+
verify 17: accepted 3/3 (K=4) + correction -> +4 tokens (125.2 ms) cpos=75
|
| 896 |
+
[margins] 2.86 8.92 0.37 3.52 3.38 K=3
|
| 897 |
+
verify 18: accepted 2/2 (K=3) + correction -> +3 tokens (111.8 ms) cpos=78
|
| 898 |
+
[margins] 8.21 7.80 5.97 7.72 9.54 K=6
|
| 899 |
+
verify 19: accepted 5/5 (K=6) + correction -> +6 tokens (160.4 ms) cpos=84
|
| 900 |
+
[margins] 7.61 8.53 7.24 3.85 1.11 K=5
|
| 901 |
+
verify 20: accepted 4/4 (K=5) + correction -> +5 tokens (141.2 ms) cpos=89
|
| 902 |
+
[margins] 8.57 6.63 4.54 0.84 0.51 K=4
|
| 903 |
+
verify 21: accepted 3/3 (K=4) + correction -> +4 tokens (128.3 ms) cpos=93
|
| 904 |
+
[margins] 4.97 3.27 2.03 1.77 0.14 K=5
|
| 905 |
+
verify 22: accepted 3/4 (K=5) + correction -> +4 tokens (138.0 ms) cpos=97
|
| 906 |
+
[margins] 10.11 13.21 11.04 9.22 4.44 K=6
|
| 907 |
+
verify 23: accepted 5/5 (K=6) + correction -> +6 tokens (159.8 ms) cpos=103
|
| 908 |
+
[margins] 9.91 4.52 3.55 1.38 9.39 K=4
|
| 909 |
+
verify 24: accepted 3/3 (K=4) + correction -> +4 tokens (130.7 ms) cpos=107
|
| 910 |
+
[margins] 4.59 3.50 10.03 0.41 3.62 K=4
|
| 911 |
+
verify 25: accepted 3/3 (K=4) + correction -> +4 tokens (130.7 ms) cpos=111
|
| 912 |
+
[margins] 10.64 6.39 8.54 4.05 7.51 K=6
|
| 913 |
+
verify 26: accepted 5/5 (K=6) + correction -> +6 tokens (167.0 ms) cpos=117
|
| 914 |
+
[margins] 10.84 8.75 3.83 6.28 5.89 K=6
|
| 915 |
+
verify 27: accepted 5/5 (K=6) + correction -> +6 tokens (161.6 ms) cpos=123
|
| 916 |
+
[margins] 14.08 3.75 8.48 8.62 0.33 K=5
|
| 917 |
+
verify 28: accepted 4/4 (K=5) + correction -> +5 tokens (162.0 ms) cpos=128
|
| 918 |
+
[margins] 5.39 5.29 3.61 3.03 2.75 K=6
|
| 919 |
+
verify 29: accepted 3/5 (K=6) + correction -> +4 tokens (160.5 ms) cpos=132
|
| 920 |
+
[margins] 5.59 2.27 4.07 1.79 0.89 K=5
|
| 921 |
+
verify 30: accepted 4/4 (K=5) + correction -> +5 tokens (141.7 ms) cpos=137
|
| 922 |
+
[margins] 2.81 0.02 4.72 2.99 3.71 K=2
|
| 923 |
+
verify 31: accepted 1/1 (K=2) + correction -> +2 tokens (101.7 ms) cpos=139
|
| 924 |
+
[margins] 5.89 0.87 6.36 7.94 5.74 K=2
|
| 925 |
+
verify 32: accepted 1/1 (K=2) + correction -> +2 tokens (104.9 ms) cpos=141
|
| 926 |
+
[margins] 9.70 8.20 5.71 1.66 0.84 K=5
|
| 927 |
+
verify 33: accepted 4/4 (K=5) + correction -> +5 tokens (145.4 ms) cpos=146
|
| 928 |
+
[margins] 7.92 10.33 11.05 11.82 1.44 K=5
|
| 929 |
+
verify 34: accepted 4/4 (K=5) + correction -> +5 tokens (146.6 ms) cpos=151
|
| 930 |
+
[margins] 2.78 0.61 0.52 7.90 0.49 K=2
|
| 931 |
+
verify 35: accepted 1/1 (K=2) + correction -> +2 tokens (109.6 ms) cpos=153
|
| 932 |
+
[margins] 3.04 3.05 1.01 1.23 4.13 K=3
|
| 933 |
+
verify 36: accepted 0/2 (K=3) + correction -> +1 tokens (109.1 ms) cpos=154
|
| 934 |
+
[margins] 6.18 2.77 0.01 2.97 1.83 K=3
|
| 935 |
+
verify 37: accepted 1/2 (K=3) + correction -> +2 tokens (115.2 ms) cpos=156
|
| 936 |
+
[margins] 6.23 2.96 0.79 2.18 1.64 K=3
|
| 937 |
+
verify 38: accepted 1/2 (K=3) + correction -> +2 tokens (110.0 ms) cpos=158
|
| 938 |
+
[margins] 4.63 0.67 3.90 2.17 0.87 K=2
|
| 939 |
+
verify 39: accepted 1/1 (K=2) + correction -> +2 tokens (107.5 ms) cpos=160
|
| 940 |
+
[margins] 4.03 1.64 3.13 1.23 5.07 K=4
|
| 941 |
+
verify 40: accepted 3/3 (K=4) + correction -> +4 tokens (124.0 ms) cpos=164
|
| 942 |
+
[margins] 4.01 4.98 3.85 1.57 3.63 K=6
|
| 943 |
+
verify 41: accepted 5/5 (K=6) + correction -> +6 tokens (149.1 ms) cpos=170
|
| 944 |
+
[margins] 3.12 6.34 3.15 3.00 3.52 K=6
|
| 945 |
+
verify 42: accepted 5/5 (K=6) + correction -> +6 tokens (154.6 ms) cpos=176
|
| 946 |
+
[margins] 2.78 7.00 2.67 4.39 4.24 K=6
|
| 947 |
+
verify 43: accepted 5/5 (K=6) + correction -> +6 tokens (150.2 ms) cpos=182
|
| 948 |
+
[margins] 1.79 7.10 3.55 4.95 4.65 K=6
|
| 949 |
+
verify 44: accepted 5/5 (K=6) + correction -> +6 tokens (155.3 ms) cpos=188
|
| 950 |
+
[margins] 1.04 6.72 3.47 4.56 4.81 K=6
|
| 951 |
+
verify 45: accepted 5/5 (K=6) + correction -> +6 tokens (150.2 ms) cpos=194
|
| 952 |
+
[margins] 0.56 6.97 3.20 4.39 4.46 K=6
|
| 953 |
+
verify 46: accepted 5/5 (K=6) + correction -> +6 tokens (155.5 ms) cpos=200
|
| 954 |
+
[margins] 0.61 6.10 3.30 3.80 4.05 K=6
|
| 955 |
+
verify 47: accepted 5/5 (K=6) + correction -> +6 tokens (150.2 ms) cpos=206
|
| 956 |
+
[margins] 1.74 6.41 3.04 3.37 4.34 K=6
|
| 957 |
+
verify 48: accepted 5/5 (K=6) + correction -> +6 tokens (156.5 ms) cpos=212
|
| 958 |
+
[margins] 1.86 5.92 2.73 3.44 4.26 K=6
|
| 959 |
+
verify 49: accepted 5/5 (K=6) + correction -> +6 tokens (148.8 ms) cpos=218
|
| 960 |
+
[margins] 1.87 6.02 3.82 2.22 4.47 K=6
|
| 961 |
+
verify 50: accepted 5/5 (K=6) + correction -> +6 tokens (156.2 ms) cpos=224
|
| 962 |
+
[margins] 2.17 6.53 3.51 2.29 4.51 K=6
|
| 963 |
+
verify 51: accepted 5/5 (K=6) + correction -> +6 tokens (149.9 ms) cpos=230
|
| 964 |
+
[margins] 1.96 6.28 4.35 1.47 4.68 K=4
|
| 965 |
+
verify 52: accepted 0/3 (K=4) + correction -> +1 tokens (124.3 ms) cpos=231
|
| 966 |
+
[margins] 8.23 1.25 3.15 4.04 0.00 K=2
|
| 967 |
+
verify 53: accepted 0/1 (K=2) + correction -> +1 tokens (104.9 ms) cpos=232
|
| 968 |
+
[margins] 5.68 3.22 7.48 1.16 1.78 K=4
|
| 969 |
+
verify 54: accepted 1/3 (K=4) + correction -> +2 tokens (126.9 ms) cpos=234
|
| 970 |
+
[margins] 3.70 1.35 2.88 0.65 0.50 K=2
|
| 971 |
+
verify 55: accepted 1/1 (K=2) + correction -> +2 tokens (104.2 ms) cpos=236
|
| 972 |
+
[margins] 3.42 0.74 6.13 0.48 0.85 K=2
|
| 973 |
+
verify 56: accepted 0/1 (K=2) + correction -> +1 tokens (101.0 ms) cpos=237
|
| 974 |
+
|
| 975 |
+
[spec] generated 200 tokens over 56 verifies: mean tokens/verify = 3.57 (block=5, max 6)
|
| 976 |
+
[spec] tokens: 343 12888 28 566 344 260 19075 3417 9508 5718 270 2753 294 270 6616 366 1512 1527 27819 13459 16 5715 270 12919 6329 1512 1527 940 223 19 27819 13459 16 455 3452 2753 344 979 7490 201
|
| 977 |
+
[spec] SPEC-DECODE: 35.6 ms/tok = 28.12 tok/s (vs base M=1 68.4 ms/tok = 14.61 tok/s -> 1.92x)
|
| 978 |
+
[decode] PREFILL: 12 positions in 250.6 ms = 47.9 tok/s (20.881 ms/tok)
|
| 979 |
+
[spec] decoding 200 tokens (block=5, draft passes=1, adaptK=1.50, prompt=8 s=13)...
|
| 980 |
+
[margins] 2.44 1.52 0.04 0.77 1.28 K=3
|
| 981 |
+
verify 1: accepted 1/2 (K=3) + correction -> +2 tokens (111.1 ms) cpos=14
|
| 982 |
+
[margins] 0.29 4.20 3.29 0.70 5.14 K=4
|
| 983 |
+
verify 2: accepted 0/3 (K=4) + correction -> +1 tokens (127.4 ms) cpos=15
|
| 984 |
+
[margins] 0.25 1.22 2.10 0.12 2.03 K=2
|
| 985 |
+
verify 3: accepted 1/1 (K=2) + correction -> +2 tokens (101.4 ms) cpos=17
|
| 986 |
+
[margins] 0.67 0.11 0.19 2.77 0.22 K=2
|
| 987 |
+
verify 4: accepted 1/1 (K=2) + correction -> +2 tokens (97.8 ms) cpos=19
|
| 988 |
+
[margins] 2.75 0.04 1.11 4.17 0.44 K=2
|
| 989 |
+
verify 5: accepted 1/1 (K=2) + correction -> +2 tokens (101.8 ms) cpos=21
|
| 990 |
+
[margins] 2.65 0.01 0.40 4.64 0.22 K=2
|
| 991 |
+
verify 6: accepted 1/1 (K=2) + correction -> +2 tokens (98.5 ms) cpos=23
|
| 992 |
+
[margins] 3.23 0.47 0.26 0.09 4.19 K=2
|
| 993 |
+
verify 7: accepted 1/1 (K=2) + correction -> +2 tokens (102.1 ms) cpos=25
|
| 994 |
+
[margins] 3.85 0.34 1.23 0.08 0.34 K=2
|
| 995 |
+
verify 8: accepted 1/1 (K=2) + correction -> +2 tokens (98.2 ms) cpos=27
|
| 996 |
+
[margins] 4.80 0.16 1.80 0.31 5.11 K=2
|
| 997 |
+
verify 9: accepted 1/1 (K=2) + correction -> +2 tokens (102.1 ms) cpos=29
|
| 998 |
+
[margins] 5.91 0.54 2.23 0.39 4.64 K=2
|
| 999 |
+
verify 10: accepted 1/1 (K=2) + correction -> +2 tokens (98.6 ms) cpos=31
|
| 1000 |
+
[margins] 5.88 0.02 0.23 4.73 0.47 K=2
|
| 1001 |
+
verify 11: accepted 1/1 (K=2) + correction -> +2 tokens (105.9 ms) cpos=33
|
| 1002 |
+
[margins] 5.77 0.26 4.23 0.50 4.10 K=2
|
| 1003 |
+
verify 12: accepted 1/1 (K=2) + correction -> +2 tokens (99.8 ms) cpos=35
|
| 1004 |
+
[margins] 6.27 0.16 6.26 0.64 5.05 K=2
|
| 1005 |
+
verify 13: accepted 1/1 (K=2) + correction -> +2 tokens (102.6 ms) cpos=37
|
| 1006 |
+
[margins] 5.87 0.20 5.98 0.48 4.20 K=2
|
| 1007 |
+
verify 14: accepted 1/1 (K=2) + correction -> +2 tokens (98.9 ms) cpos=39
|
| 1008 |
+
[margins] 7.18 0.36 6.91 0.44 5.32 K=2
|
| 1009 |
+
verify 15: accepted 1/1 (K=2) + correction -> +2 tokens (102.5 ms) cpos=41
|
| 1010 |
+
[margins] 6.17 0.25 5.47 0.70 4.26 K=2
|
| 1011 |
+
verify 16: accepted 1/1 (K=2) + correction -> +2 tokens (99.3 ms) cpos=43
|
| 1012 |
+
[margins] 6.99 0.16 6.11 0.91 4.58 K=2
|
| 1013 |
+
verify 17: accepted 1/1 (K=2) + correction -> +2 tokens (103.4 ms) cpos=45
|
| 1014 |
+
[margins] 6.22 1.07 4.81 0.94 3.66 K=2
|
| 1015 |
+
verify 18: accepted 0/1 (K=2) + correction -> +1 tokens (96.6 ms) cpos=46
|
| 1016 |
+
[margins] 5.46 3.83 6.46 0.26 0.44 K=4
|
| 1017 |
+
verify 19: accepted 3/3 (K=4) + correction -> +4 tokens (126.2 ms) cpos=50
|
| 1018 |
+
[margins] 3.10 2.18 1.22 1.44 3.51 K=3
|
| 1019 |
+
verify 20: accepted 2/2 (K=3) + correction -> +3 tokens (113.7 ms) cpos=53
|
| 1020 |
+
[margins] 8.07 8.31 0.08 1.85 1.79 K=3
|
| 1021 |
+
verify 21: accepted 2/2 (K=3) + correction -> +3 tokens (112.5 ms) cpos=56
|
| 1022 |
+
[margins] 2.53 0.68 3.34 1.65 1.10 K=2
|
| 1023 |
+
verify 22: accepted 1/1 (K=2) + correction -> +2 tokens (100.0 ms) cpos=58
|
| 1024 |
+
[margins] 0.39 6.92 1.00 3.40 2.10 K=3
|
| 1025 |
+
verify 23: accepted 0/2 (K=3) + correction -> +1 tokens (110.9 ms) cpos=59
|
| 1026 |
+
[margins] 2.48 2.61 3.86 5.46 0.06 K=5
|
| 1027 |
+
verify 24: accepted 0/4 (K=5) + correction -> +1 tokens (139.9 ms) cpos=60
|
| 1028 |
+
[margins] 5.96 3.96 6.63 7.37 3.06 K=6
|
| 1029 |
+
verify 25: accepted 5/5 (K=6) + correction -> +6 tokens (154.5 ms) cpos=66
|
| 1030 |
+
[margins] 6.95 11.49 7.99 1.82 1.64 K=6
|
| 1031 |
+
verify 26: accepted 3/5 (K=6) + correction -> +4 tokens (163.2 ms) cpos=70
|
| 1032 |
+
[margins] 2.32 3.24 0.18 5.08 3.38 K=3
|
| 1033 |
+
verify 27: accepted 0/2 (K=3) + correction -> +1 tokens (115.4 ms) cpos=71
|
| 1034 |
+
[margins] 0.74 3.67 2.71 2.81 1.92 K=6
|
| 1035 |
+
verify 28: accepted 0/5 (K=6) + correction -> +1 tokens (159.3 ms) cpos=72
|
| 1036 |
+
[margins] 2.68 2.66 4.14 7.25 2.44 K=6
|
| 1037 |
+
verify 29: accepted 5/5 (K=6) + correction -> +6 tokens (155.5 ms) cpos=78
|
| 1038 |
+
[margins] 6.06 7.28 4.55 0.38 2.30 K=4
|
| 1039 |
+
verify 30: accepted 3/3 (K=4) + correction -> +4 tokens (128.3 ms) cpos=82
|
| 1040 |
+
[margins] 3.51 5.61 3.94 4.67 1.81 K=6
|
| 1041 |
+
verify 31: accepted 5/5 (K=6) + correction -> +6 tokens (162.0 ms) cpos=88
|
| 1042 |
+
[margins] 4.23 5.80 2.42 6.25 3.07 K=6
|
| 1043 |
+
verify 32: accepted 5/5 (K=6) + correction -> +6 tokens (157.3 ms) cpos=94
|
| 1044 |
+
[margins] 4.43 8.42 8.01 6.00 3.67 K=6
|
| 1045 |
+
verify 33: accepted 5/5 (K=6) + correction -> +6 tokens (165.0 ms) cpos=100
|
| 1046 |
+
[margins] 4.95 7.89 5.02 7.22 3.50 K=6
|
| 1047 |
+
verify 34: accepted 5/5 (K=6) + correction -> +6 tokens (159.4 ms) cpos=106
|
| 1048 |
+
[margins] 6.15 9.13 9.10 6.58 4.66 K=6
|
| 1049 |
+
verify 35: accepted 5/5 (K=6) + correction -> +6 tokens (165.6 ms) cpos=112
|
| 1050 |
+
[margins] 4.21 6.38 0.19 8.24 3.77 K=3
|
| 1051 |
+
verify 36: accepted 2/2 (K=3) + correction -> +3 tokens (113.4 ms) cpos=115
|
| 1052 |
+
[margins] 4.44 14.08 2.18 5.66 8.55 K=6
|
| 1053 |
+
verify 37: accepted 5/5 (K=6) + correction -> +6 tokens (165.4 ms) cpos=121
|
| 1054 |
+
[margins] 7.43 9.86 6.73 5.51 8.31 K=6
|
| 1055 |
+
verify 38: accepted 5/5 (K=6) + correction -> +6 tokens (158.6 ms) cpos=127
|
| 1056 |
+
[margins] 4.89 14.41 2.53 5.46 8.41 K=6
|
| 1057 |
+
verify 39: accepted 5/5 (K=6) + correction -> +6 tokens (182.2 ms) cpos=133
|
| 1058 |
+
[margins] 8.73 12.42 8.24 4.11 6.61 K=6
|
| 1059 |
+
verify 40: accepted 5/5 (K=6) + correction -> +6 tokens (160.9 ms) cpos=139
|
| 1060 |
+
[margins] 4.31 10.66 2.83 5.70 6.43 K=6
|
| 1061 |
+
verify 41: accepted 5/5 (K=6) + correction -> +6 tokens (166.7 ms) cpos=145
|
| 1062 |
+
[margins] 10.95 11.72 8.29 1.74 7.14 K=6
|
| 1063 |
+
verify 42: accepted 5/5 (K=6) + correction -> +6 tokens (160.5 ms) cpos=151
|
| 1064 |
+
[margins] 5.12 13.21 2.30 7.44 7.80 K=6
|
| 1065 |
+
verify 43: accepted 5/5 (K=6) + correction -> +6 tokens (166.2 ms) cpos=157
|
| 1066 |
+
[margins] 12.23 11.57 7.75 4.38 7.57 K=6
|
| 1067 |
+
verify 44: accepted 5/5 (K=6) + correction -> +6 tokens (162.1 ms) cpos=163
|
| 1068 |
+
[margins] 6.11 13.46 2.69 7.09 6.96 K=6
|
| 1069 |
+
verify 45: accepted 5/5 (K=6) + correction -> +6 tokens (166.8 ms) cpos=169
|
| 1070 |
+
[margins] 11.20 7.96 8.31 5.28 2.96 K=6
|
| 1071 |
+
verify 46: accepted 5/5 (K=6) + correction -> +6 tokens (160.5 ms) cpos=175
|
| 1072 |
+
[margins] 5.26 12.48 2.78 7.88 6.14 K=6
|
| 1073 |
+
verify 47: accepted 5/5 (K=6) + correction -> +6 tokens (165.5 ms) cpos=181
|
| 1074 |
+
[margins] 12.07 13.01 7.48 4.19 11.50 K=6
|
| 1075 |
+
verify 48: accepted 5/5 (K=6) + correction -> +6 tokens (161.7 ms) cpos=187
|
| 1076 |
+
[margins] 5.17 12.36 3.18 6.97 7.40 K=6
|
| 1077 |
+
verify 49: accepted 5/5 (K=6) + correction -> +6 tokens (165.6 ms) cpos=193
|
| 1078 |
+
[margins] 10.63 11.96 8.13 3.63 8.90 K=6
|
| 1079 |
+
verify 50: accepted 5/5 (K=6) + correction -> +6 tokens (162.8 ms) cpos=199
|
| 1080 |
+
[margins] 5.70 12.84 3.31 7.10 8.07 K=6
|
| 1081 |
+
verify 51: accepted 5/5 (K=6) + correction -> +6 tokens (165.8 ms) cpos=205
|
| 1082 |
+
[margins] 10.57 12.38 8.12 5.53 9.04 K=6
|
| 1083 |
+
verify 52: accepted 5/5 (K=6) + correction -> +6 tokens (162.9 ms) cpos=211
|
| 1084 |
+
[margins] 7.19 13.13 3.43 7.41 9.02 K=6
|
| 1085 |
+
verify 53: accepted 5/5 (K=6) + correction -> +6 tokens (167.4 ms) cpos=217
|
| 1086 |
+
|
| 1087 |
+
[spec] generated 205 tokens over 53 verifies: mean tokens/verify = 3.87 (block=5, max 6)
|
| 1088 |
+
[spec] tokens: 1 201 1 201 1 201 1 201 1 201 1 201 1 201 1 201 1 201 1 201 1 201 1 201 1 201 1 201 1 201 1 201 3085 344 270 6102 294 8760 33 9361
|
| 1089 |
+
[spec] SPEC-DECODE: 34.7 ms/tok = 28.83 tok/s (vs base M=1 68.4 ms/tok = 14.61 tok/s -> 1.97x)
|
| 1090 |
+
|
| 1091 |
+
[blksweep] BLK passes adaptK->eff prompt | mean tok/verify | ms/tok | tok/s | vs base 14.61 tok/s
|
| 1092 |
+
[blksweep] 5 1 1.50 1.50 0 | 3.30 | 39.6 | 25.27 | 1.73x
|
| 1093 |
+
[blksweep] 5 1 1.50 1.50 1 | 4.16 | 33.5 | 29.87 | 2.04x
|
| 1094 |
+
[blksweep] 5 1 1.50 1.50 2 | 4.18 | 33.6 | 29.80 | 2.04x
|
| 1095 |
+
[blksweep] 5 1 1.50 1.50 3 | 2.59 | 47.4 | 21.08 | 1.44x
|
| 1096 |
+
[blksweep] 5 1 1.50 1.50 4 | 3.00 | 41.4 | 24.14 | 1.65x
|
| 1097 |
+
[blksweep] 5 1 1.50 1.50 5 | 4.53 | 32.1 | 31.17 | 2.13x
|
| 1098 |
+
[blksweep] 5 1 1.50 1.50 6 | 4.83 | 29.4 | 34.05 | 2.33x
|
| 1099 |
+
[blksweep] 5 1 1.50 1.50 7 | 3.57 | 35.6 | 28.12 | 1.92x
|
| 1100 |
+
[blksweep] 5 1 1.50 1.50 8 | 3.87 | 34.7 | 28.83 | 1.97x
|
| 1101 |
+
[decode] mem 106.5/122.8 GiB
|
head_card.json
ADDED
|
@@ -0,0 +1,123 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"name": "s3recap-p25-b0.1",
|
| 3 |
+
"promoted_utc": "2026-08-22T12:19:06Z",
|
| 4 |
+
"engine_git_rev": "a608d52d76f86d6325b472e34696b15d5b5fb6da",
|
| 5 |
+
"engine_tree_dirty": true,
|
| 6 |
+
"base_model": "0xSero/DeepSeek-V4-Flash-0731-REAP",
|
| 7 |
+
"base_source_model": "deepseek-ai/DeepSeek-V4-Flash-0731",
|
| 8 |
+
"base_source_revision": "9e165c30e2704aec5d9d593cce3eebd58bbef1cb",
|
| 9 |
+
"reap": {
|
| 10 |
+
"original_experts_per_layer": 256,
|
| 11 |
+
"kept_experts_per_layer": 160
|
| 12 |
+
},
|
| 13 |
+
"measurement": {
|
| 14 |
+
"protocol": "8-prompt suite (scripts/prompt_suite.json, one per category), NGEN0>=200 so tau is past the drafter's 128-token sliding window (F92); block 6, adaptK 1.50; clean run; canonical 6-token prompt is a CONTROL, excluded from the mean (F96).",
|
| 15 |
+
"suite_mean_tau": 3.8413,
|
| 16 |
+
"suite_mean_tok_s": 28.3825,
|
| 17 |
+
"base_ar_tok_s": 14.61,
|
| 18 |
+
"n_suite_prompts": 8,
|
| 19 |
+
"per_prompt": [
|
| 20 |
+
{
|
| 21 |
+
"block": 5,
|
| 22 |
+
"prompt": 0,
|
| 23 |
+
"tau": 3.3,
|
| 24 |
+
"ms_tok": 39.6,
|
| 25 |
+
"tok_s": 25.27
|
| 26 |
+
},
|
| 27 |
+
{
|
| 28 |
+
"block": 5,
|
| 29 |
+
"prompt": 1,
|
| 30 |
+
"tau": 4.16,
|
| 31 |
+
"ms_tok": 33.5,
|
| 32 |
+
"tok_s": 29.87
|
| 33 |
+
},
|
| 34 |
+
{
|
| 35 |
+
"block": 5,
|
| 36 |
+
"prompt": 2,
|
| 37 |
+
"tau": 4.18,
|
| 38 |
+
"ms_tok": 33.6,
|
| 39 |
+
"tok_s": 29.8
|
| 40 |
+
},
|
| 41 |
+
{
|
| 42 |
+
"block": 5,
|
| 43 |
+
"prompt": 3,
|
| 44 |
+
"tau": 2.59,
|
| 45 |
+
"ms_tok": 47.4,
|
| 46 |
+
"tok_s": 21.08
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"block": 5,
|
| 50 |
+
"prompt": 4,
|
| 51 |
+
"tau": 3.0,
|
| 52 |
+
"ms_tok": 41.4,
|
| 53 |
+
"tok_s": 24.14
|
| 54 |
+
},
|
| 55 |
+
{
|
| 56 |
+
"block": 5,
|
| 57 |
+
"prompt": 5,
|
| 58 |
+
"tau": 4.53,
|
| 59 |
+
"ms_tok": 32.1,
|
| 60 |
+
"tok_s": 31.17
|
| 61 |
+
},
|
| 62 |
+
{
|
| 63 |
+
"block": 5,
|
| 64 |
+
"prompt": 6,
|
| 65 |
+
"tau": 4.83,
|
| 66 |
+
"ms_tok": 29.4,
|
| 67 |
+
"tok_s": 34.05
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"block": 5,
|
| 71 |
+
"prompt": 7,
|
| 72 |
+
"tau": 3.57,
|
| 73 |
+
"ms_tok": 35.6,
|
| 74 |
+
"tok_s": 28.12
|
| 75 |
+
},
|
| 76 |
+
{
|
| 77 |
+
"block": 5,
|
| 78 |
+
"prompt": 8,
|
| 79 |
+
"tau": 3.87,
|
| 80 |
+
"ms_tok": 34.7,
|
| 81 |
+
"tok_s": 28.83
|
| 82 |
+
}
|
| 83 |
+
],
|
| 84 |
+
"lossless_gate": "PASS"
|
| 85 |
+
},
|
| 86 |
+
"incumbent_at_promotion": null,
|
| 87 |
+
"notes": "",
|
| 88 |
+
"files": [
|
| 89 |
+
{
|
| 90 |
+
"file": "config.json",
|
| 91 |
+
"bytes": 3602,
|
| 92 |
+
"sha256": "de9ea59b8bba4cc49f67b4d6a30cebb148ab465aebc57225b355158477a56c8e"
|
| 93 |
+
},
|
| 94 |
+
{
|
| 95 |
+
"file": "model-00046-of-00048.safetensors",
|
| 96 |
+
"bytes": 2326148920,
|
| 97 |
+
"sha256": "790139019c605eb055ec01f2c5d4b7e7fc0143cf4f6ac9b64d0ed6001a2248ae"
|
| 98 |
+
},
|
| 99 |
+
{
|
| 100 |
+
"file": "model-00047-of-00048.safetensors",
|
| 101 |
+
"bytes": 2275805696,
|
| 102 |
+
"sha256": "335d72f9624fe2dcc1aac25c63b6ade4c5797ab36a78908c673713c78f789cdc"
|
| 103 |
+
},
|
| 104 |
+
{
|
| 105 |
+
"file": "model-00048-of-00048.safetensors",
|
| 106 |
+
"bytes": 2408468996,
|
| 107 |
+
"sha256": "80f11aa77095970528df0eab5db2ac0bd540e99b5e1c429131468229968505e1"
|
| 108 |
+
},
|
| 109 |
+
{
|
| 110 |
+
"file": "model.safetensors.index.json",
|
| 111 |
+
"bytes": 213205,
|
| 112 |
+
"sha256": "3c6ca0116218bd358ad3b8ca8ab067194585edb187d594aef9f42986ec5fe50f"
|
| 113 |
+
},
|
| 114 |
+
{
|
| 115 |
+
"file": "mtp_trained.safetensors",
|
| 116 |
+
"bytes": 1034643326,
|
| 117 |
+
"sha256": "42552974a69adc8a212b6198eacb5de0acf48c1f9327758922bf38bcc9a280ec"
|
| 118 |
+
}
|
| 119 |
+
],
|
| 120 |
+
"archive_notes": [
|
| 121 |
+
"mtp_trained.safetensors sha256 recorded 2026-08-23. It is the SOURCE the loadable shards regenerate from; without a recorded hash the archive could not detect its corruption, which made the 'rebuildable' claim unfalsifiable."
|
| 122 |
+
]
|
| 123 |
+
}
|
mtp_trained.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:42552974a69adc8a212b6198eacb5de0acf48c1f9327758922bf38bcc9a280ec
|
| 3 |
+
size 1034643326
|
train_metrics.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|