Hugging Face's logo

optimum-internal-testing
/
neuron-testing-cache

Model card Files Community
dacorvo HF Staff commited on
Commit
cc1a689
·
verified ·
1 Parent(s): 47d982b

Synchronizing local compiler cache.

Browse files
Files changed (18) hide show
  1. neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.2.2/granite/ibm-granite/granite-3.1-2b-instruct/72b9e86c6f89db5c69ed.json +73 -0
  2. neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.2.2/llama/unsloth/Llama-3.2-1B-Instruct/b18c1c30fb44cce4ec73.json +78 -0
  3. neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.2.2/qwen2/Qwen/Qwen2.5-0.5B/754fb9fa6aaf85bb444f.json +71 -0
  4. neuronxcc-2.17.194.0+d312836f/MODULE_57b107bd0499cc4986ac+63ea7cf2/model.hlo_module.pb +1 -1
  5. neuronxcc-2.17.194.0+d312836f/MODULE_57b107bd0499cc4986ac+63ea7cf2/model.neff +1 -1
  6. neuronxcc-2.17.194.0+d312836f/MODULE_57b107bd0499cc4986ac+63ea7cf2/wrapped_neff.hlo +1 -1
  7. neuronxcc-2.17.194.0+d312836f/MODULE_64950c85776a119cdf83+b001bea4/model.hlo_module.pb +1 -1
  8. neuronxcc-2.17.194.0+d312836f/MODULE_64950c85776a119cdf83+b001bea4/model.neff +1 -1
  9. neuronxcc-2.17.194.0+d312836f/MODULE_6bb5680c622866b3b45b+63ea7cf2/model.hlo_module.pb +1 -1
  10. neuronxcc-2.17.194.0+d312836f/MODULE_6bb5680c622866b3b45b+63ea7cf2/model.neff +1 -1
  11. neuronxcc-2.17.194.0+d312836f/MODULE_6bb5680c622866b3b45b+63ea7cf2/wrapped_neff.hlo +1 -1
  12. neuronxcc-2.17.194.0+d312836f/MODULE_92d7a6b8bc621dee02b9+b001bea4/model.hlo_module.pb +1 -1
  13. neuronxcc-2.17.194.0+d312836f/MODULE_92d7a6b8bc621dee02b9+b001bea4/model.neff +1 -1
  14. neuronxcc-2.17.194.0+d312836f/MODULE_9a53db93d18e769ee7ea+b001bea4/model.hlo_module.pb +1 -1
  15. neuronxcc-2.17.194.0+d312836f/MODULE_9a53db93d18e769ee7ea+b001bea4/model.neff +1 -1
  16. neuronxcc-2.17.194.0+d312836f/MODULE_ae5ca4b91afd03b04b25+63ea7cf2/model.hlo_module.pb +1 -1
  17. neuronxcc-2.17.194.0+d312836f/MODULE_ae5ca4b91afd03b04b25+63ea7cf2/model.neff +1 -1
  18. neuronxcc-2.17.194.0+d312836f/MODULE_ae5ca4b91afd03b04b25+63ea7cf2/wrapped_neff.hlo +1 -1
neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.2.2/granite/ibm-granite/granite-3.1-2b-instruct/72b9e86c6f89db5c69ed.json ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "embedding_multiplier": 12.0,
12
+ "hidden_act": "silu",
13
+ "hidden_size": 2048,
14
+ "initializer_range": 0.02,
15
+ "intermediate_size": 8192,
16
+ "logits_scaling": 8.0,
17
+ "max_position_embeddings": 131072,
18
+ "mlp_bias": false,
19
+ "model_type": "granite",
20
+ "neuron": {
21
+ "_serialized_key": "NxDNeuronConfig",
22
+ "async_mode": false,
23
+ "attn_kernel_enabled": false,
24
+ "batch_size": 4,
25
+ "capacity_factor": null,
26
+ "cc_pipeline_tiling_factor": 2,
27
+ "checkpoint_id": "ibm-granite/granite-3.1-2b-instruct",
28
+ "checkpoint_revision": "bbc2aed595bd38bd770263dc3ab831db9794441d",
29
+ "continuous_batching": true,
30
+ "enable_bucketing": false,
31
+ "ep_degree": 1,
32
+ "flash_decoding_enabled": false,
33
+ "fused_qkv": true,
34
+ "glu_mlp": true,
35
+ "is_chunked_prefill": false,
36
+ "local_ranks_size": 2,
37
+ "logical_nc_config": 1,
38
+ "max_batch_size": 4,
39
+ "max_context_length": 4096,
40
+ "max_topk": 256,
41
+ "mlp_kernel_enabled": false,
42
+ "mlp_kernel_fuse_residual_add": false,
43
+ "n_active_tokens": 4096,
44
+ "neuronxcc_version": "2.17.194.0+d312836f",
45
+ "num_cores_per_group": 1,
46
+ "on_device_sampling": false,
47
+ "optimum_neuron_version": "0.2.2",
48
+ "output_logits": false,
49
+ "padding_side": "right",
50
+ "pp_degree": 1,
51
+ "qk_layernorm": false,
52
+ "qkv_kernel_enabled": false,
53
+ "rpl_reduce_dtype": "bfloat16",
54
+ "sequence_length": 4096,
55
+ "sequence_parallel_enabled": false,
56
+ "speculation_length": 0,
57
+ "start_rank_id": 0,
58
+ "target": null,
59
+ "torch_dtype": "bfloat16",
60
+ "tp_degree": 2,
61
+ "vocab_parallel": false
62
+ },
63
+ "num_attention_heads": 32,
64
+ "num_hidden_layers": 40,
65
+ "num_key_value_heads": 8,
66
+ "residual_multiplier": 0.22,
67
+ "rms_norm_eps": 1e-05,
68
+ "rope_scaling": null,
69
+ "rope_theta": 5000000.0,
70
+ "tie_word_embeddings": true,
71
+ "use_cache": true,
72
+ "vocab_size": 49155
73
+ }
neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.2.2/llama/unsloth/Llama-3.2-1B-Instruct/b18c1c30fb44cce4ec73.json ADDED
@@ -0,0 +1,78 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "async_mode": false,
21
+ "attn_kernel_enabled": false,
22
+ "batch_size": 4,
23
+ "capacity_factor": null,
24
+ "cc_pipeline_tiling_factor": 2,
25
+ "checkpoint_id": "unsloth/Llama-3.2-1B-Instruct",
26
+ "checkpoint_revision": "5a8abab4a5d6f164389b1079fb721cfab8d7126c",
27
+ "continuous_batching": true,
28
+ "enable_bucketing": false,
29
+ "ep_degree": 1,
30
+ "flash_decoding_enabled": false,
31
+ "fused_qkv": true,
32
+ "glu_mlp": true,
33
+ "is_chunked_prefill": false,
34
+ "local_ranks_size": 2,
35
+ "logical_nc_config": 1,
36
+ "max_batch_size": 4,
37
+ "max_context_length": 4096,
38
+ "max_topk": 256,
39
+ "mlp_kernel_enabled": false,
40
+ "mlp_kernel_fuse_residual_add": false,
41
+ "n_active_tokens": 4096,
42
+ "neuronxcc_version": "2.17.194.0+d312836f",
43
+ "num_cores_per_group": 1,
44
+ "on_device_sampling": false,
45
+ "optimum_neuron_version": "0.2.2",
46
+ "output_logits": false,
47
+ "padding_side": "right",
48
+ "pp_degree": 1,
49
+ "qk_layernorm": false,
50
+ "qkv_kernel_enabled": false,
51
+ "rpl_reduce_dtype": "float16",
52
+ "sequence_length": 4096,
53
+ "sequence_parallel_enabled": false,
54
+ "speculation_length": 0,
55
+ "start_rank_id": 0,
56
+ "target": null,
57
+ "torch_dtype": "float16",
58
+ "tp_degree": 2,
59
+ "vocab_parallel": false
60
+ },
61
+ "num_attention_heads": 32,
62
+ "num_hidden_layers": 16,
63
+ "num_key_value_heads": 8,
64
+ "pretraining_tp": 1,
65
+ "rms_norm_eps": 1e-05,
66
+ "rope_scaling": {
67
+ "factor": 32.0,
68
+ "high_freq_factor": 4.0,
69
+ "low_freq_factor": 1.0,
70
+ "original_max_position_embeddings": 8192,
71
+ "rope_type": "llama3"
72
+ },
73
+ "rope_theta": 500000.0,
74
+ "tie_word_embeddings": true,
75
+ "unsloth_fixed": true,
76
+ "use_cache": true,
77
+ "vocab_size": 128256
78
+ }
neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.2.2/qwen2/Qwen/Qwen2.5-0.5B/754fb9fa6aaf85bb444f.json ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "Qwen/Qwen2.5-0.5B",
4
+ "_task": "text-generation",
5
+ "architectures": [
6
+ "Qwen2ForCausalLM"
7
+ ],
8
+ "attention_dropout": 0.0,
9
+ "hidden_act": "silu",
10
+ "hidden_size": 896,
11
+ "initializer_range": 0.02,
12
+ "intermediate_size": 4864,
13
+ "max_position_embeddings": 32768,
14
+ "max_window_layers": 24,
15
+ "model_type": "qwen2",
16
+ "neuron": {
17
+ "_serialized_key": "NxDNeuronConfig",
18
+ "async_mode": false,
19
+ "attn_kernel_enabled": false,
20
+ "batch_size": 4,
21
+ "capacity_factor": null,
22
+ "cc_pipeline_tiling_factor": 2,
23
+ "checkpoint_id": "Qwen/Qwen2.5-0.5B",
24
+ "checkpoint_revision": "060db6499f32faf8b98477b0a26969ef7d8b9987",
25
+ "continuous_batching": true,
26
+ "enable_bucketing": false,
27
+ "ep_degree": 1,
28
+ "flash_decoding_enabled": false,
29
+ "fused_qkv": false,
30
+ "glu_mlp": true,
31
+ "is_chunked_prefill": false,
32
+ "local_ranks_size": 2,
33
+ "logical_nc_config": 1,
34
+ "max_batch_size": 4,
35
+ "max_context_length": 4096,
36
+ "max_topk": 256,
37
+ "mlp_kernel_enabled": false,
38
+ "mlp_kernel_fuse_residual_add": false,
39
+ "n_active_tokens": 4096,
40
+ "neuronxcc_version": "2.17.194.0+d312836f",
41
+ "num_cores_per_group": 1,
42
+ "on_device_sampling": false,
43
+ "optimum_neuron_version": "0.2.2",
44
+ "output_logits": false,
45
+ "padding_side": "right",
46
+ "pp_degree": 1,
47
+ "qk_layernorm": false,
48
+ "qkv_kernel_enabled": false,
49
+ "rpl_reduce_dtype": "float16",
50
+ "sequence_length": 4096,
51
+ "sequence_parallel_enabled": false,
52
+ "speculation_length": 0,
53
+ "start_rank_id": 0,
54
+ "target": null,
55
+ "torch_dtype": "float16",
56
+ "tp_degree": 2,
57
+ "vocab_parallel": false
58
+ },
59
+ "num_attention_heads": 14,
60
+ "num_hidden_layers": 24,
61
+ "num_key_value_heads": 2,
62
+ "rms_norm_eps": 1e-06,
63
+ "rope_scaling": null,
64
+ "rope_theta": 1000000.0,
65
+ "sliding_window": 32768,
66
+ "tie_word_embeddings": true,
67
+ "use_cache": true,
68
+ "use_mrope": false,
69
+ "use_sliding_window": false,
70
+ "vocab_size": 151936
71
+ }
neuronxcc-2.17.194.0+d312836f/MODULE_57b107bd0499cc4986ac+63ea7cf2/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:c80a392f2b88bd697653c255c2844cbd5429b70c71397c2996ec1d4fa671b508
3
  size 68277
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a914031f26bb159d35140c96a1bdb74370696834acf754811e40b142622a02a4
3
  size 68277
neuronxcc-2.17.194.0+d312836f/MODULE_57b107bd0499cc4986ac+63ea7cf2/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:59822b9ac234fe96b1378ef14006c9a57bc68b0f64b85c13cf716263221b30b3
3
  size 257024
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c868ec53737adc2fe50d3708f1a3cedfc2e769e9e5d707555376ee2a9a153028
3
  size 257024
neuronxcc-2.17.194.0+d312836f/MODULE_57b107bd0499cc4986ac+63ea7cf2/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:3f23d4992dbbde57ebc1da486fa7c73ce0c7850327b93063de3f53e19e100364
3
  size 268322
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9eee4b03603e48bfdd262029efc30ecd40c1dbfcb10b94f627cf4e6696aa82f6
3
  size 268322
neuronxcc-2.17.194.0+d312836f/MODULE_64950c85776a119cdf83+b001bea4/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:dcee907a68aababb046a67c95c8902d62468d2ab04a7992937e104f5796ce005
3
  size 82456
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d161deceb48f8f73a0aa6227a14f23c73b193b8c2f54ec986673930820610f16
3
  size 82456
neuronxcc-2.17.194.0+d312836f/MODULE_64950c85776a119cdf83+b001bea4/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:92948386ead7279817b9212fe4a41783244e24813282c3e7ebc4627de442a6a5
3
  size 420864
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:06018e131b181177574a276f6294a56b20726b844088fb5bc1b9799f8bc4eef5
3
  size 420864
neuronxcc-2.17.194.0+d312836f/MODULE_6bb5680c622866b3b45b+63ea7cf2/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:3919b87cd300d92efea625d9191475bf962567fa7cc582dc03fbaa7123aaef00
3
  size 79431
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:12aceb27cdf723d91c999dfa8673d18f8cdabe145b276c11445d19e86a5f5145
3
  size 79431
neuronxcc-2.17.194.0+d312836f/MODULE_6bb5680c622866b3b45b+63ea7cf2/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:53662264c93fa3d6ec8da672571667788df0b89378229fea74b79ab52e7e5f98
3
  size 318464
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6cc031ed02ba0c6ba5aa974a455517d4c142fb3079c57115af3c0a4117b4e9f4
3
  size 318464
neuronxcc-2.17.194.0+d312836f/MODULE_6bb5680c622866b3b45b+63ea7cf2/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:80b4fbbd8686981be9c999a65558ea26f26602f5f85881e75da131f45d626abb
3
  size 329762
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:85e3ceea8c3aeb3881b637508b4f0bd79debc800f3a70ebf651b7c68f3ee8665
3
  size 329762
neuronxcc-2.17.194.0+d312836f/MODULE_92d7a6b8bc621dee02b9+b001bea4/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:cc0680fc90d24969529eec0e979e03103686155af2f67c5ed6429f88b001437f
3
  size 81609
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6d96dab97289c7e8f0f86606e3e28fe6ba499bac93101c635261281f994837bf
3
  size 81609
neuronxcc-2.17.194.0+d312836f/MODULE_92d7a6b8bc621dee02b9+b001bea4/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:a10fa64dc00260eedd31ea9a0cda060e7602e936bcfbf564fcce9689adb86170
3
  size 359424
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6283c2b8221af2ea89f2bedbae48a1b2f8d7fb42b4e630fc3ad3e16519d01b14
3
  size 359424
neuronxcc-2.17.194.0+d312836f/MODULE_9a53db93d18e769ee7ea+b001bea4/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d6163c5d320ba6bd9f87fca643b0afc17b7e71fe5ed4c5adcd75ef4be45e5d94
3
  size 81607
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:316aa6d670ec51cbf810083a84be88f91b1bb9cbc07460058f55b9f20cc9dfc2
3
  size 81607
neuronxcc-2.17.194.0+d312836f/MODULE_9a53db93d18e769ee7ea+b001bea4/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:fbe8af19d362170ce971c490fa35cee4329dc0f0bf81170b494cac33b3e14dd6
3
  size 359424
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6613c5481ae06b85bb8abb953bc30cc50c4addff26b3478d9e71c68116482f9d
3
  size 359424
neuronxcc-2.17.194.0+d312836f/MODULE_ae5ca4b91afd03b04b25+63ea7cf2/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:9d389505d85aee196b80a9662c4ea75f8f2f86b10ce7e2897d4bc3996728988c
3
  size 68279
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7bf1de332dd0f1615aa9bb39e8b50034c3b02b64240d386640900156271b83e8
3
  size 68279
neuronxcc-2.17.194.0+d312836f/MODULE_ae5ca4b91afd03b04b25+63ea7cf2/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:a628bd4de2e0ab1cf7d7212d721b66d45f44c06b8ea21d6e6727fc48d2cde98d
3
  size 257024
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c809a8eebc804bd1bd9bddf1d8d93146a057e5c1d57af08032b900a9ab6d0830
3
  size 257024
neuronxcc-2.17.194.0+d312836f/MODULE_ae5ca4b91afd03b04b25+63ea7cf2/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:0470d5d9e68d41b0269bdb1d15f7bf8c19f7583400fddde0ddfa8fb323d14a97
3
  size 268322
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e3a45648500fccd8d6bada22f3ccf68bafd048f7a678b92125321c7667da7b02
3
  size 268322