Hugging Face's logo

optimum-internal-testing
/
neuron-testing-cache

Model card Files Community
optimum-internal-testing-user commited on
Commit
cee945d
·
verified ·
1 Parent(s): 386daaf

Synchronizing local compiler cache.

Browse files
Files changed (6) hide show
  1. neuronxcc-2.19.8089.0+8ab9f450/0_REGISTRY/0.3.0.dev3/phi3/microsoft/Phi-3-mini-4k-instruct/1e2d4e1d3b95dfb315e1.json +77 -0
  2. neuronxcc-2.19.8089.0+8ab9f450/MODULE_061f22ddc7af0a5d29f9+84f3e719/model.hlo_module.pb +1 -1
  3. neuronxcc-2.19.8089.0+8ab9f450/MODULE_061f22ddc7af0a5d29f9+84f3e719/model.neff +1 -1
  4. neuronxcc-2.19.8089.0+8ab9f450/MODULE_6ec54d1cb60ff01fc2a1+5be477de/model.neff +1 -1
  5. neuronxcc-2.19.8089.0+8ab9f450/MODULE_6ec54d1cb60ff01fc2a1+5be477de/wrapped_neff.hlo +1 -1
  6. neuronxcc-2.19.8089.0+8ab9f450/MODULE_ef8771b8447d48519ae4+431f5505/model.neff +1 -1
neuronxcc-2.19.8089.0+8ab9f450/0_REGISTRY/0.3.0.dev3/phi3/microsoft/Phi-3-mini-4k-instruct/1e2d4e1d3b95dfb315e1.json ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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": "NxDNeuronConfig",
23
+ "async_mode": false,
24
+ "attn_kernel_enabled": false,
25
+ "batch_size": 4,
26
+ "capacity_factor": null,
27
+ "cc_pipeline_tiling_factor": 2,
28
+ "checkpoint_id": "microsoft/Phi-3-mini-4k-instruct",
29
+ "checkpoint_revision": "0a67737cc96d2554230f90338b163bc6380a2a85",
30
+ "continuous_batching": true,
31
+ "enable_bucketing": false,
32
+ "ep_degree": 1,
33
+ "flash_decoding_enabled": false,
34
+ "fused_qkv": true,
35
+ "glu_mlp": true,
36
+ "is_chunked_prefill": false,
37
+ "local_ranks_size": 2,
38
+ "logical_nc_config": 1,
39
+ "max_batch_size": 4,
40
+ "max_context_length": 4096,
41
+ "max_topk": 256,
42
+ "mlp_kernel_enabled": false,
43
+ "mlp_kernel_fuse_residual_add": false,
44
+ "n_active_tokens": 4096,
45
+ "neuronxcc_version": "2.19.8089.0+8ab9f450",
46
+ "num_cores_per_group": 1,
47
+ "on_device_sampling": false,
48
+ "optimum_neuron_version": "0.3.0.dev3",
49
+ "output_logits": false,
50
+ "padding_side": "right",
51
+ "pp_degree": 1,
52
+ "qk_layernorm": false,
53
+ "qkv_kernel_enabled": false,
54
+ "rpl_reduce_dtype": "bfloat16",
55
+ "sequence_length": 4096,
56
+ "sequence_parallel_enabled": false,
57
+ "speculation_length": 0,
58
+ "start_rank_id": 0,
59
+ "target": null,
60
+ "torch_dtype": "bfloat16",
61
+ "tp_degree": 2,
62
+ "vocab_parallel": false
63
+ },
64
+ "num_attention_heads": 32,
65
+ "num_hidden_layers": 32,
66
+ "num_key_value_heads": 32,
67
+ "original_max_position_embeddings": 4096,
68
+ "partial_rotary_factor": 1.0,
69
+ "resid_pdrop": 0.0,
70
+ "rms_norm_eps": 1e-05,
71
+ "rope_scaling": null,
72
+ "rope_theta": 10000.0,
73
+ "sliding_window": 2047,
74
+ "tie_word_embeddings": false,
75
+ "use_cache": true,
76
+ "vocab_size": 32064
77
+ }
neuronxcc-2.19.8089.0+8ab9f450/MODULE_061f22ddc7af0a5d29f9+84f3e719/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:0677a23025fbd7931d9195daf02d5312e2da1158a7927ddd283ee33bec101024
3
  size 858046
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e5457f5da88c937a6ad1f77bbf6b3a3f33a0cce63ae5b553e6c7d410ad2006a6
3
  size 858046
neuronxcc-2.19.8089.0+8ab9f450/MODULE_061f22ddc7af0a5d29f9+84f3e719/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:e22c01e02854f21979862b09571130be0a7968089cd9166a4240ddcb7d25e926
3
  size 5786624
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f9be2cb3cf6c22ff8a952cfcad2b04253d7e3d2cea415a7f0d932a08392a5034
3
  size 5786624
neuronxcc-2.19.8089.0+8ab9f450/MODULE_6ec54d1cb60ff01fc2a1+5be477de/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:8fc0b3513bc9d5aac316a65397b2ca7cf9324435e864fc965bd48d5b8659044f
3
  size 8725504
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:71bf62faa6b98566edae15f3e2cdebd8e1542b894439de79a47ab3e2648ac6c3
3
  size 8725504
neuronxcc-2.19.8089.0+8ab9f450/MODULE_6ec54d1cb60ff01fc2a1+5be477de/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:97b0bb5c0170043e3f760308c54d4c96cff4c66b0fdc52d19d064e8bbd6ba438
3
  size 8862755
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3e3c343b71c6411b7dd3a7f9cff35f05320c08a368820d18ba06f6a4c818a2da
3
  size 8862755
neuronxcc-2.19.8089.0+8ab9f450/MODULE_ef8771b8447d48519ae4+431f5505/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:2897ded479757c09958822fc780197d52175739e8c36edd6403aa2df5bf5a1e2
3
  size 3974144
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6c6a5423f7358a9e4afb65cff5863bb5fa8ef4452e8112548b7bb3db91cf48e4
3
  size 3974144