optimum-internal-testing-user commited on
Commit
bec72d7
·
verified ·
1 Parent(s): ac447b9

Synchronizing local compiler cache.

Browse files
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev0/phi3/microsoft/Phi-3-mini-4k-instruct/996d84f500e96cfd7606.json ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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": "configuration_phi3.Phi3Config",
12
+ "AutoModelForCausalLM": "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": "NxDNeuronConfig",
23
+ "batch_size": 4,
24
+ "capacity_factor": null,
25
+ "checkpoint_id": "microsoft/Phi-3-mini-4k-instruct",
26
+ "checkpoint_revision": "0a67737cc96d2554230f90338b163bc6380a2a85",
27
+ "continuous_batching": true,
28
+ "dtype": "bfloat16",
29
+ "ep_degree": 1,
30
+ "fused_qkv": true,
31
+ "glu_mlp": true,
32
+ "local_ranks_size": 2,
33
+ "max_batch_size": 4,
34
+ "max_context_length": 4096,
35
+ "max_topk": 256,
36
+ "n_active_tokens": 4096,
37
+ "neuronxcc_version": "2.21.18209.0+043b1bf7",
38
+ "on_device_sampling": true,
39
+ "optimum_neuron_version": "0.4.2.dev0",
40
+ "output_logits": false,
41
+ "pp_degree": 1,
42
+ "sequence_length": 4096,
43
+ "speculation_length": 0,
44
+ "start_rank_id": 0,
45
+ "target": "trn1",
46
+ "tp_degree": 2
47
+ },
48
+ "num_attention_heads": 32,
49
+ "num_hidden_layers": 32,
50
+ "num_key_value_heads": 32,
51
+ "original_max_position_embeddings": 4096,
52
+ "partial_rotary_factor": 1.0,
53
+ "resid_pdrop": 0.0,
54
+ "rms_norm_eps": 1e-05,
55
+ "rope_scaling": null,
56
+ "rope_theta": 10000.0,
57
+ "sliding_window": 2047,
58
+ "tie_word_embeddings": false,
59
+ "use_cache": true,
60
+ "vocab_size": 32064
61
+ }