optimum-internal-testing-user commited on
Commit
1e89e67
·
verified ·
1 Parent(s): a6ef304

Synchronizing local compiler cache.

Browse files
neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev3/granite/ibm-granite/granite-3.1-2b-instruct/31629530ce956e3eb55e.json ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "ibm-granite/granite-3.1-2b-instruct",
4
+ "_task": "text-generation",
5
+ "architectures": [
6
+ "GraniteForCausalLM"
7
+ ],
8
+ "attention_bias": false,
9
+ "attention_dropout": 0.1,
10
+ "attention_multiplier": 0.015625,
11
+ "dtype": "bfloat16",
12
+ "embedding_multiplier": 12.0,
13
+ "hidden_act": "silu",
14
+ "hidden_size": 2048,
15
+ "initializer_range": 0.02,
16
+ "intermediate_size": 8192,
17
+ "logits_scaling": 8.0,
18
+ "max_position_embeddings": 131072,
19
+ "mlp_bias": false,
20
+ "model_type": "granite",
21
+ "neuron": {
22
+ "_serialized_key": "NxDNeuronConfig",
23
+ "batch_size": 1,
24
+ "capacity_factor": null,
25
+ "checkpoint_id": "ibm-granite/granite-3.1-2b-instruct",
26
+ "checkpoint_revision": "bbc2aed595bd38bd770263dc3ab831db9794441d",
27
+ "continuous_batching": false,
28
+ "ep_degree": 1,
29
+ "fused_qkv": true,
30
+ "glu_mlp": true,
31
+ "local_ranks_size": 2,
32
+ "max_batch_size": 1,
33
+ "max_context_length": 8192,
34
+ "max_topk": 256,
35
+ "n_active_tokens": 8192,
36
+ "neuronxcc_version": "2.21.33363.0+82129205",
37
+ "on_device_sampling": true,
38
+ "optimum_neuron_version": "0.4.6.dev3",
39
+ "output_logits": false,
40
+ "pp_degree": 1,
41
+ "prefill_chunk_size": 1024,
42
+ "sequence_length": 8192,
43
+ "speculation_length": 0,
44
+ "start_rank_id": 0,
45
+ "target": "trn1",
46
+ "torch_dtype": "bfloat16",
47
+ "tp_degree": 2
48
+ },
49
+ "num_attention_heads": 32,
50
+ "num_hidden_layers": 40,
51
+ "num_key_value_heads": 8,
52
+ "residual_multiplier": 0.22,
53
+ "rms_norm_eps": 1e-05,
54
+ "rope_scaling": null,
55
+ "rope_theta": 5000000.0,
56
+ "tie_word_embeddings": true,
57
+ "use_cache": true,
58
+ "vocab_size": 49155
59
+ }
neuronxcc-2.21.33363.0+82129205/MODULE_da7543b73fbd9c77b695+6170d8e1/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:6d205571fe7e17384578352c7c49dc0c2d095ae2bdb745edb1db783c94df73fc
3
  size 1538064
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8039f37cf8f74df32eb6821df951e715cb9be5a7b6c9374dafd414d5ed8545bd
3
  size 1538064
neuronxcc-2.21.33363.0+82129205/MODULE_da7543b73fbd9c77b695+6170d8e1/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:98d1b4384f2a614ccd9b9a1bd6dfc1ee06f0277e6dd173cf6795f23e60fc3fb3
3
  size 111862784
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d1d31257822d52e37e1fc1b5d7630208f8c8b75bbff6ad6594c32a1335526791
3
  size 111862784