dacorvo HF Staff commited on
Commit
55b928a
·
verified ·
1 Parent(s): 6bd48f6

Synchronizing local compiler cache.

Browse files
neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.2.0.dev2/inference/llama/meta-llama/Meta-Llama-3.1-8B/63ae9940e985694a6de1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "meta-llama/Meta-Llama-3.1-8B",
4
+ "_task": "text-generation",
5
+ "architectures": [
6
+ "LlamaForCausalLM"
7
+ ],
8
+ "attention_bias": false,
9
+ "attention_dropout": 0.0,
10
+ "head_dim": 128,
11
+ "hidden_act": "silu",
12
+ "hidden_size": 4096,
13
+ "initializer_range": 0.02,
14
+ "intermediate_size": 14336,
15
+ "max_position_embeddings": 131072,
16
+ "mlp_bias": false,
17
+ "model_type": "llama",
18
+ "neuron": {
19
+ "auto_cast_type": "bf16",
20
+ "batch_size": 32,
21
+ "checkpoint_id": "meta-llama/Meta-Llama-3.1-8B",
22
+ "checkpoint_revision": "d04e592bb4f6aa9cfee91e2e20afa771667e1d4b",
23
+ "compiler_type": "neuronx-cc",
24
+ "compiler_version": "2.17.194.0+d312836f",
25
+ "num_cores": 8,
26
+ "sequence_length": 4096,
27
+ "task": "text-generation"
28
+ },
29
+ "num_attention_heads": 32,
30
+ "num_hidden_layers": 32,
31
+ "num_key_value_heads": 8,
32
+ "pretraining_tp": 1,
33
+ "rms_norm_eps": 1e-05,
34
+ "rope_scaling": {
35
+ "factor": 8.0,
36
+ "high_freq_factor": 4.0,
37
+ "low_freq_factor": 1.0,
38
+ "original_max_position_embeddings": 8192,
39
+ "rope_type": "llama3"
40
+ },
41
+ "rope_theta": 500000.0,
42
+ "tie_word_embeddings": false,
43
+ "use_cache": true,
44
+ "vocab_size": 128256
45
+ }