Synchronizing local compiler cache.
Browse files
neuronxcc-2.18.121.0+9e31e41a/0_REGISTRY/0.3.0.dev1/qwen2/Qwen/Qwen2.5-0.5B/bc6f5fb886c4795b2c9f.json
ADDED
|
@@ -0,0 +1,49 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_entry_class": "SingleModelCacheEntry",
|
| 3 |
+
"_model_id": "Qwen/Qwen2.5-0.5B",
|
| 4 |
+
"_task": "text-generation",
|
| 5 |
+
"architectures": [
|
| 6 |
+
"Qwen2ForCausalLM"
|
| 7 |
+
],
|
| 8 |
+
"attention_dropout": 0.0,
|
| 9 |
+
"hidden_act": "silu",
|
| 10 |
+
"hidden_size": 896,
|
| 11 |
+
"initializer_range": 0.02,
|
| 12 |
+
"intermediate_size": 4864,
|
| 13 |
+
"max_position_embeddings": 32768,
|
| 14 |
+
"max_window_layers": 24,
|
| 15 |
+
"model_type": "qwen2",
|
| 16 |
+
"neuron": {
|
| 17 |
+
"_serialized_key": "HloNeuronConfig",
|
| 18 |
+
"all_reduce_dtype": null,
|
| 19 |
+
"allow_flash_attention": true,
|
| 20 |
+
"attention_layout": "HSB",
|
| 21 |
+
"attn_output_transposed": false,
|
| 22 |
+
"auto_cast_type": "fp16",
|
| 23 |
+
"batch_size": 4,
|
| 24 |
+
"checkpoint_id": "Qwen/Qwen2.5-0.5B",
|
| 25 |
+
"checkpoint_revision": "060db6499f32faf8b98477b0a26969ef7d8b9987",
|
| 26 |
+
"collectives_layout": "HSB",
|
| 27 |
+
"continuous_batching": true,
|
| 28 |
+
"fuse_qkv": false,
|
| 29 |
+
"group_query_attention": "shard-over-heads",
|
| 30 |
+
"log_softmax_scores": false,
|
| 31 |
+
"neuronxcc_version": "2.18.121.0+9e31e41a",
|
| 32 |
+
"optimum_neuron_version": "0.3.0.dev1",
|
| 33 |
+
"output_all_logits": false,
|
| 34 |
+
"sequence_length": 4096,
|
| 35 |
+
"tp_degree": 2
|
| 36 |
+
},
|
| 37 |
+
"num_attention_heads": 14,
|
| 38 |
+
"num_hidden_layers": 24,
|
| 39 |
+
"num_key_value_heads": 2,
|
| 40 |
+
"rms_norm_eps": 1e-06,
|
| 41 |
+
"rope_scaling": null,
|
| 42 |
+
"rope_theta": 1000000.0,
|
| 43 |
+
"sliding_window": 32768,
|
| 44 |
+
"tie_word_embeddings": true,
|
| 45 |
+
"use_cache": true,
|
| 46 |
+
"use_mrope": false,
|
| 47 |
+
"use_sliding_window": false,
|
| 48 |
+
"vocab_size": 151936
|
| 49 |
+
}
|
neuronxcc-2.18.121.0+9e31e41a/MODULE_2e229618015e416964c5+613edded/model.neff
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 1414144
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7833961dbd69e13a714a7091714095cb43e5af77e8500d657f648e2c95f9f2fb
|
| 3 |
size 1414144
|
neuronxcc-2.18.121.0+9e31e41a/MODULE_4b31f40da9553c6566ed+613edded/model.neff
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 18494464
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:cd6a04a4b8450f24da1a32b614a4d76cbbdae27c943d5c5b5552ca459ebab863
|
| 3 |
size 18494464
|