optimum-internal-testing-user commited on
Commit
682bfb9
·
verified ·
1 Parent(s): 025c4f6

Synchronizing local compiler cache.

Browse files
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.1.dev0/llama/unsloth/Llama-3.2-1B-Instruct/eeb8b7f2495863c099b1.json ADDED
@@ -0,0 +1,63 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "unsloth/Llama-3.2-1B-Instruct",
4
+ "_task": "text-generation",
5
+ "architectures": [
6
+ "LlamaForCausalLM"
7
+ ],
8
+ "attention_bias": false,
9
+ "attention_dropout": 0.0,
10
+ "head_dim": 64,
11
+ "hidden_act": "silu",
12
+ "hidden_size": 2048,
13
+ "initializer_range": 0.02,
14
+ "intermediate_size": 8192,
15
+ "max_position_embeddings": 131072,
16
+ "mlp_bias": false,
17
+ "model_type": "llama",
18
+ "neuron": {
19
+ "_serialized_key": "NxDNeuronConfig",
20
+ "batch_size": 4,
21
+ "capacity_factor": null,
22
+ "checkpoint_id": "unsloth/Llama-3.2-1B-Instruct",
23
+ "checkpoint_revision": "5a8abab4a5d6f164389b1079fb721cfab8d7126c",
24
+ "continuous_batching": true,
25
+ "enable_bucketing": false,
26
+ "ep_degree": 1,
27
+ "fused_qkv": true,
28
+ "glu_mlp": true,
29
+ "local_ranks_size": 2,
30
+ "max_batch_size": 4,
31
+ "max_context_length": 4096,
32
+ "max_topk": 256,
33
+ "n_active_tokens": 4096,
34
+ "neuronxcc_version": "2.21.18209.0+043b1bf7",
35
+ "on_device_sampling": true,
36
+ "optimum_neuron_version": "0.4.1.dev0",
37
+ "output_logits": false,
38
+ "pp_degree": 1,
39
+ "sequence_length": 4096,
40
+ "speculation_length": 0,
41
+ "start_rank_id": 0,
42
+ "target": "trn1",
43
+ "torch_dtype": "float16",
44
+ "tp_degree": 2
45
+ },
46
+ "num_attention_heads": 32,
47
+ "num_hidden_layers": 16,
48
+ "num_key_value_heads": 8,
49
+ "pretraining_tp": 1,
50
+ "rms_norm_eps": 1e-05,
51
+ "rope_scaling": {
52
+ "factor": 32.0,
53
+ "high_freq_factor": 4.0,
54
+ "low_freq_factor": 1.0,
55
+ "original_max_position_embeddings": 8192,
56
+ "rope_type": "llama3"
57
+ },
58
+ "rope_theta": 500000.0,
59
+ "tie_word_embeddings": true,
60
+ "unsloth_fixed": true,
61
+ "use_cache": true,
62
+ "vocab_size": 128256
63
+ }
neuronxcc-2.21.18209.0+043b1bf7/MODULE_4be33b117e68e2de3326+a9d440f5/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:85d1017d9072684d2d1a12173c4d18b8708d8ba23c73d51518e3800474ee4648
3
  size 3093504
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d70d650de28d7b13defbb5f9d78b06f99ec63ba26cda803845b298cbca3d0581
3
  size 3093504
neuronxcc-2.21.18209.0+043b1bf7/MODULE_4be33b117e68e2de3326+a9d440f5/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:0002e7f94df7d6ce285a7661a9c52a16f7ba6401ed6e2acd13b0771f6c96eaed
3
  size 3167748
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8d0d1b1e40a092a3eec560354aeb1191196e8b39c2312249a4d6f358c7ef2a51
3
  size 3167748
neuronxcc-2.21.18209.0+043b1bf7/MODULE_73b50e8e61409e1f8fe1+ed72d204/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:7ba540c2173774df5d8327bd084a3643c64d105ab3c83eceb0643ed402cf6282
3
  size 509252
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:441fac3e96e30c894231d7f6dd20d4e89e6b4cf1295ed681a96fb69af4ea7582
3
  size 509252
neuronxcc-2.21.18209.0+043b1bf7/MODULE_73b50e8e61409e1f8fe1+ed72d204/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:9e0decb7641b1d291b671a33734d2176cd8df441e3eaa6d5ae0d411f30b20f7d
3
  size 41585664
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ccfc71e4eaf13a28e93ac88d5f63f2d222b919a9db2e9f4a966422a8f787f591
3
  size 41585664