Synchronizing local compiler cache.
Browse files- neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.2.0.dev3/inference/phi3/microsoft/Phi-3-mini-4k-instruct/c5f0ac0391211fa5aba4.json +55 -0
- neuronxcc-2.17.194.0+d312836f/MODULE_2db40b5b23a523e16b87+613edded/model.neff +1 -1
- neuronxcc-2.17.194.0+d312836f/MODULE_967d83f73b8fe30e4d14+613edded/model.neff +1 -1
neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.2.0.dev3/inference/phi3/microsoft/Phi-3-mini-4k-instruct/c5f0ac0391211fa5aba4.json
ADDED
|
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_entry_class": "SingleModelCacheEntry",
|
| 3 |
+
"_model_id": "microsoft/Phi-3-mini-4k-instruct",
|
| 4 |
+
"_task": "text-generation",
|
| 5 |
+
"architectures": [
|
| 6 |
+
"Phi3ForCausalLM"
|
| 7 |
+
],
|
| 8 |
+
"attention_bias": false,
|
| 9 |
+
"attention_dropout": 0.0,
|
| 10 |
+
"auto_map": {
|
| 11 |
+
"AutoConfig": "microsoft/Phi-3-mini-4k-instruct--configuration_phi3.Phi3Config",
|
| 12 |
+
"AutoModelForCausalLM": "microsoft/Phi-3-mini-4k-instruct--modeling_phi3.Phi3ForCausalLM"
|
| 13 |
+
},
|
| 14 |
+
"embd_pdrop": 0.0,
|
| 15 |
+
"hidden_act": "silu",
|
| 16 |
+
"hidden_size": 3072,
|
| 17 |
+
"initializer_range": 0.02,
|
| 18 |
+
"intermediate_size": 8192,
|
| 19 |
+
"max_position_embeddings": 4096,
|
| 20 |
+
"model_type": "phi3",
|
| 21 |
+
"neuron": {
|
| 22 |
+
"_serialized_key": "HloNeuronConfig",
|
| 23 |
+
"all_reduce_dtype": null,
|
| 24 |
+
"allow_flash_attention": false,
|
| 25 |
+
"attention_layout": "HSB",
|
| 26 |
+
"attn_output_transposed": false,
|
| 27 |
+
"auto_cast_type": "bf16",
|
| 28 |
+
"batch_size": 4,
|
| 29 |
+
"checkpoint_id": "microsoft/Phi-3-mini-4k-instruct",
|
| 30 |
+
"checkpoint_revision": "0a67737cc96d2554230f90338b163bc6380a2a85",
|
| 31 |
+
"collectives_layout": "HSB",
|
| 32 |
+
"continuous_batching": true,
|
| 33 |
+
"fuse_qkv": true,
|
| 34 |
+
"group_query_attention": null,
|
| 35 |
+
"log_softmax_scores": false,
|
| 36 |
+
"neuronxcc_version": "2.17.194.0+d312836f",
|
| 37 |
+
"optimum_neuron_version": "0.2.0.dev3",
|
| 38 |
+
"output_all_logits": false,
|
| 39 |
+
"sequence_length": 4096,
|
| 40 |
+
"tp_degree": 2
|
| 41 |
+
},
|
| 42 |
+
"num_attention_heads": 32,
|
| 43 |
+
"num_hidden_layers": 32,
|
| 44 |
+
"num_key_value_heads": 32,
|
| 45 |
+
"original_max_position_embeddings": 4096,
|
| 46 |
+
"partial_rotary_factor": 1.0,
|
| 47 |
+
"resid_pdrop": 0.0,
|
| 48 |
+
"rms_norm_eps": 1e-05,
|
| 49 |
+
"rope_scaling": null,
|
| 50 |
+
"rope_theta": 10000.0,
|
| 51 |
+
"sliding_window": 2047,
|
| 52 |
+
"tie_word_embeddings": false,
|
| 53 |
+
"use_cache": true,
|
| 54 |
+
"vocab_size": 32064
|
| 55 |
+
}
|
neuronxcc-2.17.194.0+d312836f/MODULE_2db40b5b23a523e16b87+613edded/model.neff
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 6677504
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:536abc8484f3de7b1579992b2d54d3dddd8ae66a4445772fb21a0c919af6d674
|
| 3 |
size 6677504
|
neuronxcc-2.17.194.0+d312836f/MODULE_967d83f73b8fe30e4d14+613edded/model.neff
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 14900224
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:375c4dc2f07e67d19e754400c381a8b5d67fb29f3cc941eff3527a8e9f5c89cc
|
| 3 |
size 14900224
|