dacorvo HF Staff commited on
Commit
705780b
·
verified ·
1 Parent(s): 32ad413

Synchronizing local compiler cache.

Browse files
.gitattributes CHANGED
@@ -11752,3 +11752,6 @@ neuronxcc-2.21.18209.0+043b1bf7/MODULE_e289b7d80056996226e7+d0d57c8a/wrapped_nef
11752
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_455832c0f106feb96c4d+564b7b5b/model.neff filter=lfs diff=lfs merge=lfs -text
11753
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_68378177b17fd9218535+d0d57c8a/model.neff filter=lfs diff=lfs merge=lfs -text
11754
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_68378177b17fd9218535+d0d57c8a/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
 
 
 
 
11752
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_455832c0f106feb96c4d+564b7b5b/model.neff filter=lfs diff=lfs merge=lfs -text
11753
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_68378177b17fd9218535+d0d57c8a/model.neff filter=lfs diff=lfs merge=lfs -text
11754
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_68378177b17fd9218535+d0d57c8a/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
11755
+ neuronxcc-2.21.18209.0+043b1bf7/MODULE_41ffa4569452a6aba7d2+877608f3/model.neff filter=lfs diff=lfs merge=lfs -text
11756
+ neuronxcc-2.21.18209.0+043b1bf7/MODULE_c3015b480968e039faa6+747527b0/model.neff filter=lfs diff=lfs merge=lfs -text
11757
+ neuronxcc-2.21.18209.0+043b1bf7/MODULE_c3015b480968e039faa6+747527b0/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.0.dev0/llama/meta-llama/Llama-3.3-70B-Instruct/cc08a9b6ed1f8cca22f1.json ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "meta-llama/Llama-3.3-70B-Instruct",
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": 8192,
13
+ "initializer_range": 0.02,
14
+ "intermediate_size": 28672,
15
+ "max_position_embeddings": 131072,
16
+ "mlp_bias": false,
17
+ "model_type": "llama",
18
+ "neuron": {
19
+ "_serialized_key": "NxDNeuronConfig",
20
+ "batch_size": 32,
21
+ "capacity_factor": null,
22
+ "checkpoint_id": "meta-llama/Llama-3.3-70B-Instruct",
23
+ "checkpoint_revision": "6f6073b423013f6a7d4d9f39144961bfbfbc386b",
24
+ "continuous_batching": true,
25
+ "enable_bucketing": false,
26
+ "ep_degree": 1,
27
+ "fused_qkv": true,
28
+ "glu_mlp": true,
29
+ "local_ranks_size": 64,
30
+ "max_batch_size": 32,
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": "0.4.0.dev0",
37
+ "output_logits": false,
38
+ "pp_degree": 1,
39
+ "sequence_length": 4096,
40
+ "speculation_length": 0,
41
+ "start_rank_id": 0,
42
+ "target": "trn2",
43
+ "torch_dtype": "bfloat16",
44
+ "tp_degree": 64
45
+ },
46
+ "num_attention_heads": 64,
47
+ "num_hidden_layers": 80,
48
+ "num_key_value_heads": 8,
49
+ "pretraining_tp": 1,
50
+ "rms_norm_eps": 1e-05,
51
+ "rope_scaling": {
52
+ "factor": 8.0,
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": false,
60
+ "use_cache": true,
61
+ "vocab_size": 128256
62
+ }
neuronxcc-2.21.18209.0+043b1bf7/MODULE_41ffa4569452a6aba7d2+877608f3/compile_flags.json ADDED
@@ -0,0 +1 @@
 
 
1
+ ["--target=trn2", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=2", "--logfile=/tmp/nxd_model/context_encoding_model/_tp0_bk0/log-neuron-cc.txt"]
neuronxcc-2.21.18209.0+043b1bf7/MODULE_41ffa4569452a6aba7d2+877608f3/model.done ADDED
File without changes
neuronxcc-2.21.18209.0+043b1bf7/MODULE_41ffa4569452a6aba7d2+877608f3/model.hlo_module.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:365936646dcebc05a9a6355f2a5ab790edaa89e60b1b84888a546fe95cd7f05a
3
+ size 3584622
neuronxcc-2.21.18209.0+043b1bf7/MODULE_41ffa4569452a6aba7d2+877608f3/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a4c5529e1a24440786790d44d0216f3a2b6b6e451c85a003235d8ba925161823
3
+ size 1956864
neuronxcc-2.21.18209.0+043b1bf7/MODULE_c3015b480968e039faa6+747527b0/compile_flags.json ADDED
@@ -0,0 +1 @@
 
 
1
+ ["--target=trn2", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=2", "--logfile=/tmp/nxd_model/token_generation_model/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]
neuronxcc-2.21.18209.0+043b1bf7/MODULE_c3015b480968e039faa6+747527b0/model.done ADDED
File without changes
neuronxcc-2.21.18209.0+043b1bf7/MODULE_c3015b480968e039faa6+747527b0/model.hlo_module.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6782db84ec9645e702039f7ab9e3efbf5b5bef8c78957a7eb4c1738f02eea599
3
+ size 4561977
neuronxcc-2.21.18209.0+043b1bf7/MODULE_c3015b480968e039faa6+747527b0/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2e77858b8d606fdb02f32560d94318a6829a0d713962f31e6126add34811ccc6
3
+ size 18965504
neuronxcc-2.21.18209.0+043b1bf7/MODULE_c3015b480968e039faa6+747527b0/wrapped_neff.hlo ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ca469455d7752594a3bf2040d5407a48ddab47356f1e083b52453f32103c0b62
3
+ size 19334623