Hugging Face's logo

optimum-internal-testing
/
neuron-testing-cache

Model card Files
xet
Community
optimum-internal-testing-user commited on
Commit
4d8b43f
·
verified ·
1 Parent(s): fef0c41

Synchronizing local compiler cache.

Browse files
Files changed (42) hide show
  1. .gitattributes +3 -0
  2. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev0/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/d337db1f2945982d0bc9.json +58 -0
  3. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev0/llama/llamafactory/tiny-random-Llama-3/ada00e76710d2d884cc8.json +62 -0
  4. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev0/llama/unsloth/Llama-3.2-1B-Instruct/1b4bdf074a6b57b52e20.json +63 -0
  5. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev0/llama4_text/tiny-random/llama-4/caf8458124c0c1b8b608.json +81 -0
  6. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev0/mixtral/dacorvo/Mixtral-tiny/9c1a264c4d265060e661.json +58 -0
  7. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev0/phi3/yujiepan/phi-4-tiny-random/af5f9033652bae57657c.json +59 -0
  8. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev0/qwen2/yujiepan/qwen2.5-128k-tiny-random/50dd484535c9c8562031.json +64 -0
  9. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev0/qwen3_moe/optimum-internal-testing/tiny-random-qwen3_moe/39b749f56f6251d4327e.json +65 -0
  10. neuronxcc-2.21.18209.0+043b1bf7/MODULE_0e5b82b136a38372d842+6e4949b4/model.hlo_module.pb +1 -1
  11. neuronxcc-2.21.18209.0+043b1bf7/MODULE_0e5b82b136a38372d842+6e4949b4/model.neff +1 -1
  12. neuronxcc-2.21.18209.0+043b1bf7/MODULE_306829ff9e3811933fe7+a02c3a36/model.neff +1 -1
  13. neuronxcc-2.21.18209.0+043b1bf7/MODULE_306829ff9e3811933fe7+a02c3a36/wrapped_neff.hlo +1 -1
  14. neuronxcc-2.21.18209.0+043b1bf7/MODULE_34362baee2ad4784108e+a02c3a36/model.neff +1 -1
  15. neuronxcc-2.21.18209.0+043b1bf7/MODULE_34362baee2ad4784108e+a02c3a36/wrapped_neff.hlo +1 -1
  16. neuronxcc-2.21.18209.0+043b1bf7/MODULE_3e3c9e9db5ebafa6c488+a02c3a36/model.neff +1 -1
  17. neuronxcc-2.21.18209.0+043b1bf7/MODULE_3e3c9e9db5ebafa6c488+a02c3a36/wrapped_neff.hlo +1 -1
  18. neuronxcc-2.21.18209.0+043b1bf7/MODULE_5b1c456704a89716d745+2dde74c7/model.neff +1 -1
  19. neuronxcc-2.21.18209.0+043b1bf7/MODULE_5b1c456704a89716d745+2dde74c7/wrapped_neff.hlo +1 -1
  20. neuronxcc-2.21.18209.0+043b1bf7/MODULE_5f6f4f62aba29e11d946+24129607/model.hlo_module.pb +1 -1
  21. neuronxcc-2.21.18209.0+043b1bf7/MODULE_5f6f4f62aba29e11d946+24129607/model.neff +1 -1
  22. neuronxcc-2.21.18209.0+043b1bf7/MODULE_9e174a0ad6384712cfa5+24129607/model.hlo_module.pb +1 -1
  23. neuronxcc-2.21.18209.0+043b1bf7/MODULE_9e174a0ad6384712cfa5+24129607/model.neff +1 -1
  24. neuronxcc-2.21.18209.0+043b1bf7/MODULE_b559f9b8396895d936b0+80d05c3f/model.neff +1 -1
  25. neuronxcc-2.21.18209.0+043b1bf7/MODULE_b559f9b8396895d936b0+80d05c3f/wrapped_neff.hlo +1 -1
  26. neuronxcc-2.21.18209.0+043b1bf7/MODULE_c58e0a47eb485c62b0cd+283df001/model.hlo_module.pb +1 -1
  27. neuronxcc-2.21.18209.0+043b1bf7/MODULE_c58e0a47eb485c62b0cd+283df001/model.neff +1 -1
  28. neuronxcc-2.21.18209.0+043b1bf7/MODULE_cedace9d3d693aab8d35+a02c3a36/model.neff +1 -1
  29. neuronxcc-2.21.18209.0+043b1bf7/MODULE_cedace9d3d693aab8d35+a02c3a36/wrapped_neff.hlo +1 -1
  30. neuronxcc-2.21.18209.0+043b1bf7/MODULE_d8cfc8fee2dcbc7833f8+24129607/model.hlo_module.pb +1 -1
  31. neuronxcc-2.21.18209.0+043b1bf7/MODULE_d8cfc8fee2dcbc7833f8+24129607/model.neff +1 -1
  32. neuronxcc-2.21.18209.0+043b1bf7/MODULE_e12a32198da9ddff7d98+24129607/model.hlo_module.pb +1 -1
  33. neuronxcc-2.21.18209.0+043b1bf7/MODULE_e12a32198da9ddff7d98+24129607/model.neff +1 -1
  34. neuronxcc-2.21.18209.0+043b1bf7/MODULE_eaa781288c2e1d7ed527+a02c3a36/compile_flags.json +1 -0
  35. neuronxcc-2.21.18209.0+043b1bf7/MODULE_eaa781288c2e1d7ed527+a02c3a36/model.done +0 -0
  36. neuronxcc-2.21.18209.0+043b1bf7/MODULE_eaa781288c2e1d7ed527+a02c3a36/model.hlo_module.pb +3 -0
  37. neuronxcc-2.21.18209.0+043b1bf7/MODULE_eaa781288c2e1d7ed527+a02c3a36/model.neff +3 -0
  38. neuronxcc-2.21.18209.0+043b1bf7/MODULE_eaa781288c2e1d7ed527+a02c3a36/wrapped_neff.hlo +3 -0
  39. neuronxcc-2.21.18209.0+043b1bf7/MODULE_fd32caf35dc9cdf42fac+24129607/compile_flags.json +1 -0
  40. neuronxcc-2.21.18209.0+043b1bf7/MODULE_fd32caf35dc9cdf42fac+24129607/model.done +0 -0
  41. neuronxcc-2.21.18209.0+043b1bf7/MODULE_fd32caf35dc9cdf42fac+24129607/model.hlo_module.pb +3 -0
  42. neuronxcc-2.21.18209.0+043b1bf7/MODULE_fd32caf35dc9cdf42fac+24129607/model.neff +3 -0
.gitattributes CHANGED
@@ -5428,3 +5428,6 @@ neuronxcc-2.21.18209.0+043b1bf7/MODULE_f52fe8e88daf6dc9619a+24129607/model.neff
5428
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_fef9240b8fbfaca895ba+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
5429
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_fef9240b8fbfaca895ba+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
5430
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_e0ce3525251d35c118ed+24129607/model.neff filter=lfs diff=lfs merge=lfs -text
 
 
 
 
5428
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_fef9240b8fbfaca895ba+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
5429
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_fef9240b8fbfaca895ba+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
5430
  neuronxcc-2.21.18209.0+043b1bf7/MODULE_e0ce3525251d35c118ed+24129607/model.neff filter=lfs diff=lfs merge=lfs -text
5431
+ neuronxcc-2.21.18209.0+043b1bf7/MODULE_eaa781288c2e1d7ed527+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
5432
+ neuronxcc-2.21.18209.0+043b1bf7/MODULE_eaa781288c2e1d7ed527+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
5433
+ neuronxcc-2.21.18209.0+043b1bf7/MODULE_fd32caf35dc9cdf42fac+24129607/model.neff filter=lfs diff=lfs merge=lfs -text
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev0/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/d337db1f2945982d0bc9.json ADDED
@@ -0,0 +1,58 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "hf-internal-testing/tiny-random-GraniteForCausalLM",
4
+ "_task": "text-generation",
5
+ "architectures": [
6
+ "GraniteForCausalLM"
7
+ ],
8
+ "attention_bias": false,
9
+ "attention_dropout": 0.0,
10
+ "attention_multiplier": 1.0,
11
+ "dtype": "float32",
12
+ "embedding_multiplier": 1.0,
13
+ "hidden_act": "silu",
14
+ "hidden_size": 32,
15
+ "initializer_range": 0.02,
16
+ "intermediate_size": 64,
17
+ "logits_scaling": 1.0,
18
+ "max_position_embeddings": 2048,
19
+ "mlp_bias": false,
20
+ "model_type": "granite",
21
+ "neuron": {
22
+ "_serialized_key": "NxDNeuronConfig",
23
+ "batch_size": 1,
24
+ "capacity_factor": null,
25
+ "checkpoint_id": "hf-internal-testing/tiny-random-GraniteForCausalLM",
26
+ "checkpoint_revision": "c3074ebc0ac2fe545305f5e5f6cce2cc9b2aa0c5",
27
+ "continuous_batching": false,
28
+ "ep_degree": 1,
29
+ "fused_qkv": true,
30
+ "glu_mlp": true,
31
+ "local_ranks_size": 2,
32
+ "max_batch_size": 1,
33
+ "max_context_length": 1024,
34
+ "max_topk": 256,
35
+ "n_active_tokens": 1024,
36
+ "neuronxcc_version": "2.21.18209.0+043b1bf7",
37
+ "on_device_sampling": true,
38
+ "optimum_neuron_version": "0.4.2.dev0",
39
+ "output_logits": false,
40
+ "pp_degree": 1,
41
+ "sequence_length": 1024,
42
+ "speculation_length": 0,
43
+ "start_rank_id": 0,
44
+ "target": "trn1",
45
+ "torch_dtype": "float32",
46
+ "tp_degree": 2
47
+ },
48
+ "num_attention_heads": 4,
49
+ "num_hidden_layers": 2,
50
+ "num_key_value_heads": 4,
51
+ "residual_multiplier": 1.0,
52
+ "rms_norm_eps": 1e-06,
53
+ "rope_scaling": null,
54
+ "rope_theta": 10000.0,
55
+ "tie_word_embeddings": false,
56
+ "use_cache": true,
57
+ "vocab_size": 49152
58
+ }
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev0/llama/llamafactory/tiny-random-Llama-3/ada00e76710d2d884cc8.json ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "llamafactory/tiny-random-Llama-3",
4
+ "_task": "text-generation",
5
+ "architectures": [
6
+ "LlamaForCausalLM"
7
+ ],
8
+ "attention_bias": false,
9
+ "attention_dropout": 0.0,
10
+ "dtype": "float16",
11
+ "head_dim": 4,
12
+ "hidden_act": "silu",
13
+ "hidden_size": 16,
14
+ "initializer_range": 0.02,
15
+ "intermediate_size": 64,
16
+ "max_position_embeddings": 131072,
17
+ "mlp_bias": false,
18
+ "model_type": "llama",
19
+ "neuron": {
20
+ "_serialized_key": "NxDNeuronConfig",
21
+ "batch_size": 1,
22
+ "capacity_factor": null,
23
+ "checkpoint_id": "llamafactory/tiny-random-Llama-3",
24
+ "checkpoint_revision": "bf2a2e3bf199ad2ee96f02a3c00246c608db22a8",
25
+ "continuous_batching": false,
26
+ "ep_degree": 1,
27
+ "fused_qkv": true,
28
+ "glu_mlp": true,
29
+ "local_ranks_size": 2,
30
+ "max_batch_size": 1,
31
+ "max_context_length": 1024,
32
+ "max_topk": 256,
33
+ "n_active_tokens": 1024,
34
+ "neuronxcc_version": "2.21.18209.0+043b1bf7",
35
+ "on_device_sampling": true,
36
+ "optimum_neuron_version": "0.4.2.dev0",
37
+ "output_logits": false,
38
+ "pp_degree": 1,
39
+ "sequence_length": 1024,
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": 4,
47
+ "num_hidden_layers": 2,
48
+ "num_key_value_heads": 4,
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+043b1bf7/0_REGISTRY/0.4.2.dev0/llama/unsloth/Llama-3.2-1B-Instruct/1b4bdf074a6b57b52e20.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
+ "dtype": "bfloat16",
11
+ "head_dim": 64,
12
+ "hidden_act": "silu",
13
+ "hidden_size": 2048,
14
+ "initializer_range": 0.02,
15
+ "intermediate_size": 8192,
16
+ "max_position_embeddings": 131072,
17
+ "mlp_bias": false,
18
+ "model_type": "llama",
19
+ "neuron": {
20
+ "_serialized_key": "NxDNeuronConfig",
21
+ "batch_size": 1,
22
+ "capacity_factor": null,
23
+ "checkpoint_id": "unsloth/Llama-3.2-1B-Instruct",
24
+ "checkpoint_revision": null,
25
+ "continuous_batching": false,
26
+ "ep_degree": 1,
27
+ "fused_qkv": false,
28
+ "glu_mlp": true,
29
+ "local_ranks_size": 2,
30
+ "max_batch_size": 1,
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": false,
36
+ "optimum_neuron_version": "0.4.2.dev0",
37
+ "output_logits": false,
38
+ "pp_degree": 1,
39
+ "sequence_length": 4096,
40
+ "speculation_length": 5,
41
+ "start_rank_id": 0,
42
+ "target": "trn1",
43
+ "torch_dtype": "bfloat16",
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+043b1bf7/0_REGISTRY/0.4.2.dev0/llama4_text/tiny-random/llama-4/caf8458124c0c1b8b608.json ADDED
@@ -0,0 +1,81 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "tiny-random/llama-4",
4
+ "_task": "text-generation",
5
+ "attention_bias": false,
6
+ "attention_chunk_size": 128,
7
+ "attention_dropout": 0.0,
8
+ "attn_scale": 0.1,
9
+ "attn_temperature_tuning": 4,
10
+ "cache_implementation": "hybrid",
11
+ "dtype": "bfloat16",
12
+ "floor_scale": 8192,
13
+ "for_llm_compressor": false,
14
+ "head_dim": 32,
15
+ "hidden_act": "silu",
16
+ "hidden_size": 32,
17
+ "initializer_range": 0.02,
18
+ "interleave_moe_layer_step": 2,
19
+ "intermediate_size": 64,
20
+ "intermediate_size_mlp": 128,
21
+ "layer_types": [
22
+ "chunked_attention",
23
+ "chunked_attention",
24
+ "chunked_attention",
25
+ "full_attention"
26
+ ],
27
+ "max_position_embeddings": 1048576,
28
+ "model_type": "llama4_text",
29
+ "moe_layers": [
30
+ 1,
31
+ 3
32
+ ],
33
+ "neuron": {
34
+ "_serialized_key": "NxDNeuronConfig",
35
+ "batch_size": 1,
36
+ "capacity_factor": null,
37
+ "checkpoint_id": "tiny-random/llama-4",
38
+ "checkpoint_revision": "9e716f5d4d1ffe0a44a15f46f4a12b840439aba4",
39
+ "continuous_batching": false,
40
+ "ep_degree": 1,
41
+ "fused_qkv": false,
42
+ "glu_mlp": true,
43
+ "local_ranks_size": 2,
44
+ "max_batch_size": 1,
45
+ "max_context_length": 1024,
46
+ "max_topk": 256,
47
+ "n_active_tokens": 1024,
48
+ "neuronxcc_version": "2.21.18209.0+043b1bf7",
49
+ "on_device_sampling": true,
50
+ "optimum_neuron_version": "0.4.2.dev0",
51
+ "output_logits": false,
52
+ "pp_degree": 1,
53
+ "sequence_length": 1024,
54
+ "speculation_length": 0,
55
+ "start_rank_id": 0,
56
+ "target": "trn1",
57
+ "torch_dtype": "bfloat16",
58
+ "tp_degree": 2
59
+ },
60
+ "no_rope_layers": [
61
+ 1,
62
+ 1,
63
+ 1,
64
+ 0
65
+ ],
66
+ "num_attention_heads": 1,
67
+ "num_experts_per_tok": 1,
68
+ "num_hidden_layers": 4,
69
+ "num_key_value_heads": 1,
70
+ "num_local_experts": 8,
71
+ "output_router_logits": false,
72
+ "rms_norm_eps": 1e-05,
73
+ "rope_scaling": null,
74
+ "rope_theta": 500000.0,
75
+ "router_aux_loss_coef": 0.001,
76
+ "router_jitter_noise": 0.0,
77
+ "tie_word_embeddings": true,
78
+ "use_cache": true,
79
+ "use_qk_norm": true,
80
+ "vocab_size": 202048
81
+ }
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev0/mixtral/dacorvo/Mixtral-tiny/9c1a264c4d265060e661.json ADDED
@@ -0,0 +1,58 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "dacorvo/Mixtral-tiny",
4
+ "_task": "text-generation",
5
+ "architectures": [
6
+ "MixtralForCausalLM"
7
+ ],
8
+ "attention_dropout": 0.0,
9
+ "dtype": "float16",
10
+ "head_dim": 32,
11
+ "hidden_act": "silu",
12
+ "hidden_size": 1024,
13
+ "initializer_range": 0.02,
14
+ "intermediate_size": 3584,
15
+ "max_position_embeddings": 1024,
16
+ "model_type": "mixtral",
17
+ "neuron": {
18
+ "_serialized_key": "NxDNeuronConfig",
19
+ "batch_size": 1,
20
+ "capacity_factor": null,
21
+ "checkpoint_id": "dacorvo/Mixtral-tiny",
22
+ "checkpoint_revision": "c557ba205ddff6ea911f4719e0d543d6c08356b6",
23
+ "continuous_batching": false,
24
+ "ep_degree": 1,
25
+ "fused_qkv": false,
26
+ "glu_mlp": true,
27
+ "local_ranks_size": 2,
28
+ "max_batch_size": 1,
29
+ "max_context_length": 1024,
30
+ "max_topk": 256,
31
+ "n_active_tokens": 1024,
32
+ "neuronxcc_version": "2.21.18209.0+043b1bf7",
33
+ "on_device_sampling": false,
34
+ "optimum_neuron_version": "0.4.2.dev0",
35
+ "output_logits": false,
36
+ "pp_degree": 1,
37
+ "sequence_length": 1024,
38
+ "speculation_length": 0,
39
+ "start_rank_id": 0,
40
+ "target": "trn1",
41
+ "torch_dtype": "float16",
42
+ "tp_degree": 2
43
+ },
44
+ "num_attention_heads": 32,
45
+ "num_experts_per_tok": 2,
46
+ "num_hidden_layers": 2,
47
+ "num_key_value_heads": 8,
48
+ "num_local_experts": 8,
49
+ "output_router_logits": false,
50
+ "rms_norm_eps": 1e-05,
51
+ "rope_theta": 10000.0,
52
+ "router_aux_loss_coef": 0.001,
53
+ "router_jitter_noise": 0.0,
54
+ "sliding_window": 4096,
55
+ "tie_word_embeddings": false,
56
+ "use_cache": true,
57
+ "vocab_size": 32000
58
+ }
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev0/phi3/yujiepan/phi-4-tiny-random/af5f9033652bae57657c.json ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "yujiepan/phi-4-tiny-random",
4
+ "_task": "text-generation",
5
+ "architectures": [
6
+ "Phi3ForCausalLM"
7
+ ],
8
+ "attention_bias": false,
9
+ "attention_dropout": 0.0,
10
+ "auto_map": {},
11
+ "dtype": "bfloat16",
12
+ "embd_pdrop": 0.0,
13
+ "hidden_act": "silu",
14
+ "hidden_size": 16,
15
+ "initializer_range": 0.02,
16
+ "intermediate_size": 32,
17
+ "max_position_embeddings": 16384,
18
+ "model_type": "phi3",
19
+ "neuron": {
20
+ "_serialized_key": "NxDNeuronConfig",
21
+ "batch_size": 1,
22
+ "capacity_factor": null,
23
+ "checkpoint_id": "yujiepan/phi-4-tiny-random",
24
+ "checkpoint_revision": "18a9a1168dc97ac6d128f811925670c275610f5a",
25
+ "continuous_batching": false,
26
+ "ep_degree": 1,
27
+ "fused_qkv": true,
28
+ "glu_mlp": true,
29
+ "local_ranks_size": 2,
30
+ "max_batch_size": 1,
31
+ "max_context_length": 1024,
32
+ "max_topk": 256,
33
+ "n_active_tokens": 1024,
34
+ "neuronxcc_version": "2.21.18209.0+043b1bf7",
35
+ "on_device_sampling": true,
36
+ "optimum_neuron_version": "0.4.2.dev0",
37
+ "output_logits": false,
38
+ "pp_degree": 1,
39
+ "sequence_length": 1024,
40
+ "speculation_length": 0,
41
+ "start_rank_id": 0,
42
+ "target": "trn1",
43
+ "torch_dtype": "bfloat16",
44
+ "tp_degree": 2
45
+ },
46
+ "num_attention_heads": 2,
47
+ "num_hidden_layers": 2,
48
+ "num_key_value_heads": 1,
49
+ "original_max_position_embeddings": 16384,
50
+ "partial_rotary_factor": 1.0,
51
+ "resid_pdrop": 0.0,
52
+ "rms_norm_eps": 1e-05,
53
+ "rope_scaling": null,
54
+ "rope_theta": 250000,
55
+ "sliding_window": null,
56
+ "tie_word_embeddings": false,
57
+ "use_cache": true,
58
+ "vocab_size": 100352
59
+ }
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev0/qwen2/yujiepan/qwen2.5-128k-tiny-random/50dd484535c9c8562031.json ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "yujiepan/qwen2.5-128k-tiny-random",
4
+ "_task": "text-generation",
5
+ "architectures": [
6
+ "Qwen2ForCausalLM"
7
+ ],
8
+ "attention_dropout": 0.0,
9
+ "dtype": "bfloat16",
10
+ "hidden_act": "silu",
11
+ "hidden_size": 8,
12
+ "initializer_range": 0.02,
13
+ "intermediate_size": 16,
14
+ "layer_types": [
15
+ "full_attention",
16
+ "full_attention"
17
+ ],
18
+ "max_position_embeddings": 32768,
19
+ "max_window_layers": 1,
20
+ "model_type": "qwen2",
21
+ "neuron": {
22
+ "_serialized_key": "NxDNeuronConfig",
23
+ "batch_size": 1,
24
+ "capacity_factor": null,
25
+ "checkpoint_id": "yujiepan/qwen2.5-128k-tiny-random",
26
+ "checkpoint_revision": "c8296d4ca3f87782876d2382fbb6481d1beb8ef0",
27
+ "continuous_batching": false,
28
+ "ep_degree": 1,
29
+ "fused_qkv": false,
30
+ "glu_mlp": true,
31
+ "local_ranks_size": 2,
32
+ "max_batch_size": 1,
33
+ "max_context_length": 1024,
34
+ "max_topk": 256,
35
+ "n_active_tokens": 1024,
36
+ "neuronxcc_version": "2.21.18209.0+043b1bf7",
37
+ "on_device_sampling": true,
38
+ "optimum_neuron_version": "0.4.2.dev0",
39
+ "output_logits": false,
40
+ "pp_degree": 1,
41
+ "sequence_length": 1024,
42
+ "speculation_length": 0,
43
+ "start_rank_id": 0,
44
+ "target": "trn1",
45
+ "torch_dtype": "bfloat16",
46
+ "tp_degree": 2
47
+ },
48
+ "num_attention_heads": 4,
49
+ "num_hidden_layers": 2,
50
+ "num_key_value_heads": 2,
51
+ "rms_norm_eps": 1e-06,
52
+ "rope_scaling": {
53
+ "factor": 4.0,
54
+ "original_max_position_embeddings": 32768,
55
+ "rope_type": "yarn",
56
+ "type": "yarn"
57
+ },
58
+ "rope_theta": 1000000.0,
59
+ "sliding_window": null,
60
+ "tie_word_embeddings": false,
61
+ "use_cache": true,
62
+ "use_sliding_window": false,
63
+ "vocab_size": 152064
64
+ }
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev0/qwen3_moe/optimum-internal-testing/tiny-random-qwen3_moe/39b749f56f6251d4327e.json ADDED
@@ -0,0 +1,65 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "optimum-internal-testing/tiny-random-qwen3_moe",
4
+ "_task": "text-generation",
5
+ "architectures": [
6
+ "Qwen3MoeForCausalLM"
7
+ ],
8
+ "attention_bias": false,
9
+ "attention_dropout": 0.0,
10
+ "decoder_sparse_step": 2,
11
+ "dtype": "float32",
12
+ "head_dim": 32,
13
+ "hidden_act": "silu",
14
+ "hidden_size": 64,
15
+ "initializer_range": 0.02,
16
+ "intermediate_size": 128,
17
+ "max_position_embeddings": 40960,
18
+ "max_window_layers": 1,
19
+ "mlp_only_layers": [],
20
+ "model_type": "qwen3_moe",
21
+ "moe_intermediate_size": 128,
22
+ "neuron": {
23
+ "_serialized_key": "NxDNeuronConfig",
24
+ "batch_size": 1,
25
+ "capacity_factor": null,
26
+ "checkpoint_id": "optimum-internal-testing/tiny-random-qwen3_moe",
27
+ "checkpoint_revision": "e0230be2839556b44b7400a233c73c74b4abb7af",
28
+ "continuous_batching": false,
29
+ "ep_degree": 1,
30
+ "fused_qkv": false,
31
+ "glu_mlp": true,
32
+ "local_ranks_size": 2,
33
+ "max_batch_size": 1,
34
+ "max_context_length": 1024,
35
+ "max_topk": 256,
36
+ "n_active_tokens": 1024,
37
+ "neuronxcc_version": "2.21.18209.0+043b1bf7",
38
+ "on_device_sampling": true,
39
+ "optimum_neuron_version": "0.4.2.dev0",
40
+ "output_logits": false,
41
+ "pp_degree": 1,
42
+ "sequence_length": 1024,
43
+ "speculation_length": 0,
44
+ "start_rank_id": 0,
45
+ "target": "trn1",
46
+ "torch_dtype": "float32",
47
+ "tp_degree": 2
48
+ },
49
+ "norm_topk_prob": true,
50
+ "num_attention_heads": 2,
51
+ "num_experts": 8,
52
+ "num_experts_per_tok": 2,
53
+ "num_hidden_layers": 2,
54
+ "num_key_value_heads": 1,
55
+ "output_router_logits": false,
56
+ "rms_norm_eps": 1e-06,
57
+ "rope_scaling": null,
58
+ "rope_theta": 1000000.0,
59
+ "router_aux_loss_coef": 0.001,
60
+ "sliding_window": null,
61
+ "tie_word_embeddings": true,
62
+ "use_cache": true,
63
+ "use_sliding_window": false,
64
+ "vocab_size": 151936
65
+ }
neuronxcc-2.21.18209.0+043b1bf7/MODULE_0e5b82b136a38372d842+6e4949b4/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:a10e0e7fab67d52e1d86800cd31328b9c683741b76775e5f053ae7439ad948e2
3
  size 84807
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a4ca5f9d821d1c473132183e17a3329f8a5e917eba12398fae359f735a592fcc
3
  size 84807
neuronxcc-2.21.18209.0+043b1bf7/MODULE_0e5b82b136a38372d842+6e4949b4/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:acdb2b92a248d6f414d51aed82b439e4ea6824126680668738cd99e83490ef47
3
  size 646144
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3795ce2d676301e229b44c2f40e341b53aa3820c669b2286ecf05a2d7749deb7
3
  size 646144
neuronxcc-2.21.18209.0+043b1bf7/MODULE_306829ff9e3811933fe7+a02c3a36/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:cd05fe6bf94ef2f336d720f7c7e1340990d09ac509619d39951378cca0455955
3
  size 277504
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b7c605196e0d8b015ab990546dae23b633f7e81d08ceddcd19366df190f1589a
3
  size 277504
neuronxcc-2.21.18209.0+043b1bf7/MODULE_306829ff9e3811933fe7+a02c3a36/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ca11c405716ec625d51c9411aff6ab27669f655504e692e6f1b8bfa39b360dcc
3
  size 285854
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ad8364c6b31b739b8254554fe4215378d7120fa340f06039910f44bb10d61c2c
3
  size 285854
neuronxcc-2.21.18209.0+043b1bf7/MODULE_34362baee2ad4784108e+a02c3a36/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:7deb3eddbbb70e1a96bda4ddbb30dd168f54c901f9253a0b5d7c2ba5243775f4
3
  size 246784
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b056f907d1197885453eb7ff1a5718d8e2a96acf470d6ada139786c8bcb8cc83
3
  size 246784
neuronxcc-2.21.18209.0+043b1bf7/MODULE_34362baee2ad4784108e+a02c3a36/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b9d79579f809a3b6c9da5f3353e3470729c5431d3969ed20fd3d2786df7c8cf8
3
  size 255104
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:38327f2b91e2112f1a9cb6b2c8b15cbd110ceefa42d6b36fca864ee9d5b0d27d
3
  size 255104
neuronxcc-2.21.18209.0+043b1bf7/MODULE_3e3c9e9db5ebafa6c488+a02c3a36/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:404de5af59e6f94a61f81cde8b8cda46c83233cc4057beff6f8c3acd63366dab
3
  size 216064
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a747955754cb2f158d067db4d137c06c2651614cf130a168162ec7f666684246
3
  size 216064
neuronxcc-2.21.18209.0+043b1bf7/MODULE_3e3c9e9db5ebafa6c488+a02c3a36/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ebea0a97eea7252733aad3fdad6698aa9bd37a943854e12c490192c68349b867
3
  size 224412
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:805754439bde8a6aae29418d51cce0b1e6180fe6b73ebbe87570a812ab52c316
3
  size 224412
neuronxcc-2.21.18209.0+043b1bf7/MODULE_5b1c456704a89716d745+2dde74c7/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ddd02000f2d84a371d61c097d65091631c1fa8146c813cdc745073fd6412d332
3
  size 369664
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:16ea009145d02ce85defd8134c2e85fc366d1adad322510b8afb4bca265a21a5
3
  size 369664
neuronxcc-2.21.18209.0+043b1bf7/MODULE_5b1c456704a89716d745+2dde74c7/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:631a3150ffdb269763aa833c922a697f1d28ee8ddcd1f13cc7e2821c4ccafa7f
3
  size 379362
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:861e848e500f5e1153ddf1d65d1730f065aea4fb1eeef54433d5b454dc34c746
3
  size 379362
neuronxcc-2.21.18209.0+043b1bf7/MODULE_5f6f4f62aba29e11d946+24129607/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:0c682387ff5e856c8633edf6652ad9964a9d39ca3f28626e81265e2317477b76
3
  size 694128
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9fbbf3f40bbbab6effc92f66a43674d834f17fe76fe444b0bf1a7cc8ab75d6e3
3
  size 694128
neuronxcc-2.21.18209.0+043b1bf7/MODULE_5f6f4f62aba29e11d946+24129607/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ef74edece0e3f2a9983a5704ed0ec9e1b393a35e78cb4341bab95e292e4f8102
3
  size 625664
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8a72b9302ceec8914862eb67a16fb28397b675f44e24c4b5420bdde75064c4d2
3
  size 625664
neuronxcc-2.21.18209.0+043b1bf7/MODULE_9e174a0ad6384712cfa5+24129607/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:a372687cccc6c3b2dfaebf019aecb7448ab1010f4e6a4e27d9bf9113796003d9
3
  size 83504
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2da8e96ddb7d196e62e21c9dbd1c83c51f547f7062c7383d64c73e294b0e2b77
3
  size 83504
neuronxcc-2.21.18209.0+043b1bf7/MODULE_9e174a0ad6384712cfa5+24129607/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:277597d33e2e8fa5c13c602ba19c9080ceccf75c029ff9a4b04762378e9f7f00
3
  size 328704
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f974a5484276100ae1453119a4c0156862cb54fa72a21a09b206bcf49eee8f1c
3
  size 328704
neuronxcc-2.21.18209.0+043b1bf7/MODULE_b559f9b8396895d936b0+80d05c3f/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:462ad0770217814c8f9c67bba35f1a690256836beb3fcd75a2b3bd9a2632acc8
3
  size 277504
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ff996d67415aeef6fad3bf8539db44dbe050bbd08aa9355b0c7d0327b59397cf
3
  size 277504
neuronxcc-2.21.18209.0+043b1bf7/MODULE_b559f9b8396895d936b0+80d05c3f/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:0acdc44fea39bd93e9535966f66d6af4e1e10e1b74d94160cdf5db24df175969
3
  size 289571
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7a7e3924eccb81d810e49bab9904fed3163cd2feff36be2b32b94a57f5cc457c
3
  size 289571
neuronxcc-2.21.18209.0+043b1bf7/MODULE_c58e0a47eb485c62b0cd+283df001/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:350bb53b1fa1ad9a168087af81f22d3e2294d8975672ab464e2a17aa9ad76955
3
  size 97794
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:280af5ff0d37eea7029b83af8ee546d88a000d8ad0b29e2ce631fac588ef41a3
3
  size 97794
neuronxcc-2.21.18209.0+043b1bf7/MODULE_c58e0a47eb485c62b0cd+283df001/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:a8d9dc7e3ca715c4f55c2a75db9efab18bdaf41c94e99c38f7535ba42e264401
3
  size 410624
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e388a63d97aebb5b1d1f4709e583b69231c4e4ba30f7fe82a52f52bb36da44d1
3
  size 410624
neuronxcc-2.21.18209.0+043b1bf7/MODULE_cedace9d3d693aab8d35+a02c3a36/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:8155a25abdce92432d6801929169663a9e6a232dac8d6c74132694d5121bd5e9
3
  size 543744
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6c60be0a0e4105c57c1ec7cb3290ce15d2eb241660272a2f9930a202c2984e24
3
  size 543744
neuronxcc-2.21.18209.0+043b1bf7/MODULE_cedace9d3d693aab8d35+a02c3a36/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d5c6226722db571990de1c61a426bf004d0ffddb6dff79e081f44c35b1b6287d
3
  size 563380
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:07fa3ac04729061af42c6b0885b7dfb9ce2f32095660307f3cea0ababe53d8ee
3
  size 563380
neuronxcc-2.21.18209.0+043b1bf7/MODULE_d8cfc8fee2dcbc7833f8+24129607/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:cae3e272e287603df0c86eb4ddc0a7c585caa5fbe0ed740e6de302228334505c
3
  size 82772
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bd6b7dce1fa2edf9077c5b0efedf2d380f4bbe5ef4940ab0b1c97cc711f93302
3
  size 82772
neuronxcc-2.21.18209.0+043b1bf7/MODULE_d8cfc8fee2dcbc7833f8+24129607/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:1c5e898d34facaa2d5e67d092b5c278d4006e2087f8872b5afeff96d2b12d1ff
3
  size 267264
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c589aa85c9efee52de2359e9634b23f64f66a338d2c0d10332f2e4d57be35af0
3
  size 267264
neuronxcc-2.21.18209.0+043b1bf7/MODULE_e12a32198da9ddff7d98+24129607/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:bd0e3dc962ab353e73f16bb06291bedc7a2915ee5f9c906e61d51ba150cecff6
3
  size 81516
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:427ee18b73570263ea4cf4d3b074e045016a97c69e93d1d5005be46a919c1061
3
  size 81516
neuronxcc-2.21.18209.0+043b1bf7/MODULE_e12a32198da9ddff7d98+24129607/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ccb67735de77fbbb5f3904ec5dbd6ad863c5b3929e342f426d56ae36685cf01d
3
  size 267264
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:18ad6fb1ac301cc5e03ccb854a092556d4cbcac30c2cf5b775044bb636eea700
3
  size 267264
neuronxcc-2.21.18209.0+043b1bf7/MODULE_eaa781288c2e1d7ed527+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_eaa781288c2e1d7ed527+a02c3a36/model.done ADDED
File without changes
neuronxcc-2.21.18209.0+043b1bf7/MODULE_eaa781288c2e1d7ed527+a02c3a36/model.hlo_module.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f0c908a5ddbf03970315b73168cfd6d3b14f39328f2b700c5d1e9c4144323034
3
+ size 91147
neuronxcc-2.21.18209.0+043b1bf7/MODULE_eaa781288c2e1d7ed527+a02c3a36/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:11a62551dbb738d0fc74120a4343226488bc719d5adf6e1a6e79877a68832a25
3
+ size 277504
neuronxcc-2.21.18209.0+043b1bf7/MODULE_eaa781288c2e1d7ed527+a02c3a36/wrapped_neff.hlo ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e6b92d0da3e57f6299c6a620caa146be7066219fd2f0e8cdfea3cd0de73b41bc
3
+ size 289031
neuronxcc-2.21.18209.0+043b1bf7/MODULE_fd32caf35dc9cdf42fac+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_fd32caf35dc9cdf42fac+24129607/model.done ADDED
File without changes
neuronxcc-2.21.18209.0+043b1bf7/MODULE_fd32caf35dc9cdf42fac+24129607/model.hlo_module.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4a9a1cd5febd5ac3bf2b41d9cbde6ccbae36214d9913a3da808c7e4ae0c199c5
3
+ size 90382
neuronxcc-2.21.18209.0+043b1bf7/MODULE_fd32caf35dc9cdf42fac+24129607/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:124ad6c4c3bb8fdad29022fb046c403aa68804f155dad560c58e6a651f6d4f6c
3
+ size 359424