Synchronizing local compiler cache.
Browse files
neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev4/83b64303132921a55cb0a422f2894a8b4d9c6143ff8379ddb5933ecd8a35fb8d/203a23fdeb890743df7c.json
ADDED
|
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_entry_class": "SingleModelCacheEntry",
|
| 3 |
+
"_model_id": "microsoft/phi-4",
|
| 4 |
+
"_task": "text-generation",
|
| 5 |
+
"architectures": [
|
| 6 |
+
"Phi3ForCausalLM"
|
| 7 |
+
],
|
| 8 |
+
"attention_bias": false,
|
| 9 |
+
"attention_dropout": 0.0,
|
| 10 |
+
"dtype": "bfloat16",
|
| 11 |
+
"embd_pdrop": 0.0,
|
| 12 |
+
"hidden_act": "silu",
|
| 13 |
+
"hidden_size": 5120,
|
| 14 |
+
"initializer_range": 0.02,
|
| 15 |
+
"intermediate_size": 17920,
|
| 16 |
+
"max_position_embeddings": 16384,
|
| 17 |
+
"model_type": "phi3",
|
| 18 |
+
"neuron": {
|
| 19 |
+
"_serialized_key": "NxDNeuronConfig",
|
| 20 |
+
"batch_size": 1,
|
| 21 |
+
"capacity_factor": null,
|
| 22 |
+
"checkpoint_id": "microsoft/phi-4",
|
| 23 |
+
"checkpoint_revision": "2db69c1c3e91a05d2c64a3185acfbaf36f744e25",
|
| 24 |
+
"continuous_batching": false,
|
| 25 |
+
"ep_degree": 1,
|
| 26 |
+
"fused_qkv": true,
|
| 27 |
+
"glu_mlp": true,
|
| 28 |
+
"local_ranks_size": 10,
|
| 29 |
+
"max_batch_size": 1,
|
| 30 |
+
"max_context_length": 4096,
|
| 31 |
+
"max_topk": 256,
|
| 32 |
+
"n_active_tokens": 4096,
|
| 33 |
+
"neuronxcc_version": "2.21.33363.0+82129205",
|
| 34 |
+
"on_device_sampling": true,
|
| 35 |
+
"optimum_neuron_version": "0.4.6.dev4",
|
| 36 |
+
"output_logits": false,
|
| 37 |
+
"pp_degree": 1,
|
| 38 |
+
"prefill_chunk_size": 1024,
|
| 39 |
+
"sequence_length": 4096,
|
| 40 |
+
"speculation_length": 0,
|
| 41 |
+
"start_rank_id": 0,
|
| 42 |
+
"target": "trn1",
|
| 43 |
+
"torch_dtype": "bfloat16",
|
| 44 |
+
"tp_degree": 10
|
| 45 |
+
},
|
| 46 |
+
"num_attention_heads": 40,
|
| 47 |
+
"num_hidden_layers": 40,
|
| 48 |
+
"num_key_value_heads": 10,
|
| 49 |
+
"original_max_position_embeddings": 16384,
|
| 50 |
+
"partial_rotary_factor": 1.0,
|
| 51 |
+
"resid_pdrop": 0.0,
|
| 52 |
+
"rms_norm_eps": 1e-05,
|
| 53 |
+
"rope_scaling": null,
|
| 54 |
+
"rope_theta": 250000,
|
| 55 |
+
"sliding_window": null,
|
| 56 |
+
"tie_word_embeddings": false,
|
| 57 |
+
"use_cache": true,
|
| 58 |
+
"vocab_size": 100352
|
| 59 |
+
}
|
neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev4/phi3/microsoft/phi-4/203a23fdeb890743df7c.json
ADDED
|
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_entry_class": "SingleModelCacheEntry",
|
| 3 |
+
"_model_id": "microsoft/phi-4",
|
| 4 |
+
"_task": "text-generation",
|
| 5 |
+
"architectures": [
|
| 6 |
+
"Phi3ForCausalLM"
|
| 7 |
+
],
|
| 8 |
+
"attention_bias": false,
|
| 9 |
+
"attention_dropout": 0.0,
|
| 10 |
+
"dtype": "bfloat16",
|
| 11 |
+
"embd_pdrop": 0.0,
|
| 12 |
+
"hidden_act": "silu",
|
| 13 |
+
"hidden_size": 5120,
|
| 14 |
+
"initializer_range": 0.02,
|
| 15 |
+
"intermediate_size": 17920,
|
| 16 |
+
"max_position_embeddings": 16384,
|
| 17 |
+
"model_type": "phi3",
|
| 18 |
+
"neuron": {
|
| 19 |
+
"_serialized_key": "NxDNeuronConfig",
|
| 20 |
+
"batch_size": 1,
|
| 21 |
+
"capacity_factor": null,
|
| 22 |
+
"checkpoint_id": "microsoft/phi-4",
|
| 23 |
+
"checkpoint_revision": "2db69c1c3e91a05d2c64a3185acfbaf36f744e25",
|
| 24 |
+
"continuous_batching": false,
|
| 25 |
+
"ep_degree": 1,
|
| 26 |
+
"fused_qkv": true,
|
| 27 |
+
"glu_mlp": true,
|
| 28 |
+
"local_ranks_size": 10,
|
| 29 |
+
"max_batch_size": 1,
|
| 30 |
+
"max_context_length": 4096,
|
| 31 |
+
"max_topk": 256,
|
| 32 |
+
"n_active_tokens": 4096,
|
| 33 |
+
"neuronxcc_version": "2.21.33363.0+82129205",
|
| 34 |
+
"on_device_sampling": true,
|
| 35 |
+
"optimum_neuron_version": "0.4.6.dev4",
|
| 36 |
+
"output_logits": false,
|
| 37 |
+
"pp_degree": 1,
|
| 38 |
+
"prefill_chunk_size": 1024,
|
| 39 |
+
"sequence_length": 4096,
|
| 40 |
+
"speculation_length": 0,
|
| 41 |
+
"start_rank_id": 0,
|
| 42 |
+
"target": "trn1",
|
| 43 |
+
"torch_dtype": "bfloat16",
|
| 44 |
+
"tp_degree": 10
|
| 45 |
+
},
|
| 46 |
+
"num_attention_heads": 40,
|
| 47 |
+
"num_hidden_layers": 40,
|
| 48 |
+
"num_key_value_heads": 10,
|
| 49 |
+
"original_max_position_embeddings": 16384,
|
| 50 |
+
"partial_rotary_factor": 1.0,
|
| 51 |
+
"resid_pdrop": 0.0,
|
| 52 |
+
"rms_norm_eps": 1e-05,
|
| 53 |
+
"rope_scaling": null,
|
| 54 |
+
"rope_theta": 250000,
|
| 55 |
+
"sliding_window": null,
|
| 56 |
+
"tie_word_embeddings": false,
|
| 57 |
+
"use_cache": true,
|
| 58 |
+
"vocab_size": 100352
|
| 59 |
+
}
|