dacorvo HF Staff commited on
Commit
c52a69d
·
verified ·
1 Parent(s): 60fdc28

Synchronizing local compiler cache.

Browse files
.gitattributes CHANGED
@@ -12122,3 +12122,6 @@ neuronxcc-2.21.18209.0+043b1bf7/MODULE_d45f3a491b6306d4905c+747527b0/wrapped_nef
12122
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_b9758a1a688ac8ecfcda+747527b0/model.neff filter=lfs diff=lfs merge=lfs -text
12123
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_b9758a1a688ac8ecfcda+747527b0/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
12124
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_dcef693677f812da71c6+877608f3/model.neff filter=lfs diff=lfs merge=lfs -text
 
 
 
 
12122
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_b9758a1a688ac8ecfcda+747527b0/model.neff filter=lfs diff=lfs merge=lfs -text
12123
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_b9758a1a688ac8ecfcda+747527b0/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
12124
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_dcef693677f812da71c6+877608f3/model.neff filter=lfs diff=lfs merge=lfs -text
12125
+ neuronxcc-2.21.18209.0+043b1bf7/MODULE_c542bb2a0e1a607dbfb0+877608f3/model.neff filter=lfs diff=lfs merge=lfs -text
12126
+ neuronxcc-2.21.18209.0+043b1bf7/MODULE_d4bf1ca5d9b73e713e84+747527b0/model.neff filter=lfs diff=lfs merge=lfs -text
12127
+ neuronxcc-2.21.18209.0+043b1bf7/MODULE_d4bf1ca5d9b73e713e84+747527b0/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.1.dev0/llama/meta-llama/Llama-3.3-70B-Instruct/e334bbb82494820aa47c.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": 32,
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": "0.4.1.dev0",
37
+ "output_logits": false,
38
+ "pp_degree": 1,
39
+ "sequence_length": 4096,
40
+ "speculation_length": 0,
41
+ "start_rank_id": 0,
42
+ "target": "trn2",
43
+ "torch_dtype": "bfloat16",
44
+ "tp_degree": 32
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_c542bb2a0e1a607dbfb0+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_c542bb2a0e1a607dbfb0+877608f3/model.done ADDED
File without changes
neuronxcc-2.21.18209.0+043b1bf7/MODULE_c542bb2a0e1a607dbfb0+877608f3/model.hlo_module.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e5ab03a70c57712d8f6a74e1d4294fe4d08f7d791ae95be65aec41c1241840a4
3
+ size 3668245
neuronxcc-2.21.18209.0+043b1bf7/MODULE_c542bb2a0e1a607dbfb0+877608f3/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1bf41bad320a211ff0d275d2d1c61eb5c2abe5dfbd75e09cc743f7e7cb8a0782
3
+ size 2407424
neuronxcc-2.21.18209.0+043b1bf7/MODULE_d4bf1ca5d9b73e713e84+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_d4bf1ca5d9b73e713e84+747527b0/model.done ADDED
File without changes
neuronxcc-2.21.18209.0+043b1bf7/MODULE_d4bf1ca5d9b73e713e84+747527b0/model.hlo_module.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3f1e72acaed68bd1816b27ae849725da0957913844c8fa9e1ca3a2350d176407
3
+ size 4046099
neuronxcc-2.21.18209.0+043b1bf7/MODULE_d4bf1ca5d9b73e713e84+747527b0/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:73ccfc398ec24e3357b5736daf0a69138495bdfe245b77204788b47103003862
3
+ size 21955584
neuronxcc-2.21.18209.0+043b1bf7/MODULE_d4bf1ca5d9b73e713e84+747527b0/wrapped_neff.hlo ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a3062c37a7f5d22a3efb0f36615deea802755c830b690028e559484a4407a136
3
+ size 22326383