dacorvo HF Staff commited on
Commit
35043c0
·
verified ·
1 Parent(s): efa3e3b

Synchronizing local compiler cache.

Browse files
.gitattributes CHANGED
@@ -12397,3 +12397,6 @@ neuronxcc-2.21.18209.0+043b1bf7/MODULE_e973fc6fe6bd9c3d7c03+a02c3a36/wrapped_nef
12397
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_20cea8c7298be20e3669+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
12398
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_20cea8c7298be20e3669+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
12399
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_d07ca49984f08737bd41+24129607/model.neff filter=lfs diff=lfs merge=lfs -text
 
 
 
 
12397
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_20cea8c7298be20e3669+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
12398
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_20cea8c7298be20e3669+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
12399
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_d07ca49984f08737bd41+24129607/model.neff filter=lfs diff=lfs merge=lfs -text
12400
+ neuronxcc-2.21.18209.0+043b1bf7/MODULE_325229a5236d71ca81f2+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
12401
+ neuronxcc-2.21.18209.0+043b1bf7/MODULE_325229a5236d71ca81f2+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
12402
+ neuronxcc-2.21.18209.0+043b1bf7/MODULE_648617d2920c9c514e3d+24129607/model.neff filter=lfs diff=lfs merge=lfs -text
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.1.dev0/llama/meta-llama/Llama-3.2-3B/79cbf5ef01ee63e6f3b9.json ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "meta-llama/Llama-3.2-3B",
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": 3072,
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": 1,
21
+ "capacity_factor": null,
22
+ "checkpoint_id": "meta-llama/Llama-3.2-3B",
23
+ "checkpoint_revision": "13afe5124825b4f3751f836b40dafda64c1ed062",
24
+ "continuous_batching": false,
25
+ "ep_degree": 1,
26
+ "fused_qkv": true,
27
+ "glu_mlp": true,
28
+ "local_ranks_size": 2,
29
+ "max_batch_size": 1,
30
+ "max_context_length": 4096,
31
+ "max_topk": 256,
32
+ "n_active_tokens": 4096,
33
+ "neuronxcc_version": "2.21.18209.0+043b1bf7",
34
+ "on_device_sampling": true,
35
+ "optimum_neuron_version": "0.4.1.dev0",
36
+ "output_logits": false,
37
+ "pp_degree": 1,
38
+ "sequence_length": 4096,
39
+ "speculation_length": 0,
40
+ "start_rank_id": 0,
41
+ "target": "trn1",
42
+ "torch_dtype": "bfloat16",
43
+ "tp_degree": 2
44
+ },
45
+ "num_attention_heads": 24,
46
+ "num_hidden_layers": 28,
47
+ "num_key_value_heads": 8,
48
+ "pretraining_tp": 1,
49
+ "rms_norm_eps": 1e-05,
50
+ "rope_scaling": {
51
+ "factor": 32.0,
52
+ "high_freq_factor": 4.0,
53
+ "low_freq_factor": 1.0,
54
+ "original_max_position_embeddings": 8192,
55
+ "rope_type": "llama3"
56
+ },
57
+ "rope_theta": 500000.0,
58
+ "tie_word_embeddings": true,
59
+ "use_cache": true,
60
+ "vocab_size": 128256
61
+ }
neuronxcc-2.21.18209.0+043b1bf7/MODULE_325229a5236d71ca81f2+a02c3a36/compile_flags.json ADDED
@@ -0,0 +1 @@
 
 
1
+ ["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]
neuronxcc-2.21.18209.0+043b1bf7/MODULE_325229a5236d71ca81f2+a02c3a36/model.done ADDED
File without changes
neuronxcc-2.21.18209.0+043b1bf7/MODULE_325229a5236d71ca81f2+a02c3a36/model.hlo_module.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fd8cd50ff0ad328a7a06eb9cc974588d97fd50478b32f0f766420bc6ecf649c6
3
+ size 703851
neuronxcc-2.21.18209.0+043b1bf7/MODULE_325229a5236d71ca81f2+a02c3a36/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:73ea32e77ecedd64776c797a3865e34cfd8c052f32788beecbeae0767c9d75d4
3
+ size 5233664
neuronxcc-2.21.18209.0+043b1bf7/MODULE_325229a5236d71ca81f2+a02c3a36/wrapped_neff.hlo ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f583ff63d72e712b27356a9107bcdcec49103911d00173925396c5dce1e243cf
3
+ size 5363109
neuronxcc-2.21.18209.0+043b1bf7/MODULE_648617d2920c9c514e3d+24129607/compile_flags.json ADDED
@@ -0,0 +1 @@
 
 
1
+ ["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"]
neuronxcc-2.21.18209.0+043b1bf7/MODULE_648617d2920c9c514e3d+24129607/model.done ADDED
File without changes
neuronxcc-2.21.18209.0+043b1bf7/MODULE_648617d2920c9c514e3d+24129607/model.hlo_module.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c93d344771cafa1c25164e890fd9f44b6772f41bd02ebe57109e3ebfe587f6e9
3
+ size 800851
neuronxcc-2.21.18209.0+043b1bf7/MODULE_648617d2920c9c514e3d+24129607/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1cfcaf6f3694055467e26a0a99bf921d66545d77ae6a3d9809011aeaa405edef
3
+ size 4219904