Hugging Face's logo

optimum-internal-testing
/
neuron-testing-cache

Model card Files
xet
Community
optimum-internal-testing-user commited on
Commit
2e1f877
·
verified ·
1 Parent(s): 1cba9a3

Synchronizing local compiler cache.

Browse files
Files changed (29) hide show
  1. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/0b211292a3fba95e7dfc.json +58 -0
  2. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/llama/llamafactory/tiny-random-Llama-3/8540d5c414adbbd8b313.json +62 -0
  3. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/llama/unsloth/Llama-3.2-1B-Instruct/d0d54603ce8895958a93.json +63 -0
  4. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/llama4_text/tiny-random/llama-4/93e972ff37cdb6b9211e.json +81 -0
  5. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/mixtral/dacorvo/Mixtral-tiny/952c02b2d9f0624b4779.json +58 -0
  6. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/phi3/yujiepan/phi-4-tiny-random/8938f51ec9548d2b7f4d.json +59 -0
  7. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/qwen2/yujiepan/qwen2.5-128k-tiny-random/315f5ef75143f023ecbc.json +64 -0
  8. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/qwen3_moe/optimum-internal-testing/tiny-random-qwen3_moe/1edce8712c5413c87813.json +65 -0
  9. neuronxcc-2.21.18209.0+043b1bf7/MODULE_0cdc152aff6a7cbf36f6+a02c3a36/model.neff +1 -1
  10. neuronxcc-2.21.18209.0+043b1bf7/MODULE_0cdc152aff6a7cbf36f6+a02c3a36/wrapped_neff.hlo +1 -1
  11. neuronxcc-2.21.18209.0+043b1bf7/MODULE_3e60ddce17a65ac4ee68+24129607/model.neff +1 -1
  12. neuronxcc-2.21.18209.0+043b1bf7/MODULE_410ef55a3aca5cd80fc8+24129607/model.neff +1 -1
  13. neuronxcc-2.21.18209.0+043b1bf7/MODULE_62abb2454c6da8d57ad5+a02c3a36/model.neff +1 -1
  14. neuronxcc-2.21.18209.0+043b1bf7/MODULE_62abb2454c6da8d57ad5+a02c3a36/wrapped_neff.hlo +1 -1
  15. neuronxcc-2.21.18209.0+043b1bf7/MODULE_73587bc7122ba5a81758+80d05c3f/model.neff +1 -1
  16. neuronxcc-2.21.18209.0+043b1bf7/MODULE_73587bc7122ba5a81758+80d05c3f/wrapped_neff.hlo +1 -1
  17. neuronxcc-2.21.18209.0+043b1bf7/MODULE_75329916c4da1120954c+6e4949b4/model.neff +1 -1
  18. neuronxcc-2.21.18209.0+043b1bf7/MODULE_7c590ab98f45f8a8dd15+283df001/model.neff +1 -1
  19. neuronxcc-2.21.18209.0+043b1bf7/MODULE_944763b36d810f4af582+24129607/model.neff +1 -1
  20. neuronxcc-2.21.18209.0+043b1bf7/MODULE_b23271513f9095bde4db+2dde74c7/model.neff +1 -1
  21. neuronxcc-2.21.18209.0+043b1bf7/MODULE_b23271513f9095bde4db+2dde74c7/wrapped_neff.hlo +1 -1
  22. neuronxcc-2.21.18209.0+043b1bf7/MODULE_bdf25cde1a48d420319c+a02c3a36/model.neff +1 -1
  23. neuronxcc-2.21.18209.0+043b1bf7/MODULE_bdf25cde1a48d420319c+a02c3a36/wrapped_neff.hlo +1 -1
  24. neuronxcc-2.21.18209.0+043b1bf7/MODULE_c4737893663f5ff66b53+24129607/model.neff +1 -1
  25. neuronxcc-2.21.18209.0+043b1bf7/MODULE_cde3417e6918fabff4a7+a02c3a36/model.neff +1 -1
  26. neuronxcc-2.21.18209.0+043b1bf7/MODULE_cde3417e6918fabff4a7+a02c3a36/wrapped_neff.hlo +1 -1
  27. neuronxcc-2.21.18209.0+043b1bf7/MODULE_f52fe8e88daf6dc9619a+24129607/model.neff +1 -1
  28. neuronxcc-2.21.18209.0+043b1bf7/MODULE_fef9240b8fbfaca895ba+a02c3a36/model.neff +1 -1
  29. neuronxcc-2.21.18209.0+043b1bf7/MODULE_fef9240b8fbfaca895ba+a02c3a36/wrapped_neff.hlo +1 -1
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/0b211292a3fba95e7dfc.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.dev2",
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.dev2/llama/llamafactory/tiny-random-Llama-3/8540d5c414adbbd8b313.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.dev2",
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.dev2/llama/unsloth/Llama-3.2-1B-Instruct/d0d54603ce8895958a93.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.dev2",
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.dev2/llama4_text/tiny-random/llama-4/93e972ff37cdb6b9211e.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.dev2",
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.dev2/mixtral/dacorvo/Mixtral-tiny/952c02b2d9f0624b4779.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.dev2",
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.dev2/phi3/yujiepan/phi-4-tiny-random/8938f51ec9548d2b7f4d.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.dev2",
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.dev2/qwen2/yujiepan/qwen2.5-128k-tiny-random/315f5ef75143f023ecbc.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.dev2",
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.dev2/qwen3_moe/optimum-internal-testing/tiny-random-qwen3_moe/1edce8712c5413c87813.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.dev2",
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_0cdc152aff6a7cbf36f6+a02c3a36/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d739b84d4fa0670e473bf91678f85de5c84fb7c8be40e9a74971aae19ec6675d
3
  size 277504
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:67e09c44a15069dd37c7d65a7e542fb2bcc2572c7fb74a99849cd2366e654696
3
  size 277504
neuronxcc-2.21.18209.0+043b1bf7/MODULE_0cdc152aff6a7cbf36f6+a02c3a36/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:207ade7804ce4d1d0f2deeee2e754a62391bfe290a49d17d55e1e4fc41250519
3
  size 289031
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e7369f5cff77f0c13ba9ddde8986adbbb27b73e5708b095274f89e31f35a8fd3
3
  size 289031
neuronxcc-2.21.18209.0+043b1bf7/MODULE_3e60ddce17a65ac4ee68+24129607/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:437d7b44097169b961fbc8717ef10d3f95ea638a0f6aa1438c7424a87719d133
3
  size 267264
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bbec9f813756324b2eba269b2d116489343495cc9ba37a0781848bcd1fd9e79a
3
  size 267264
neuronxcc-2.21.18209.0+043b1bf7/MODULE_410ef55a3aca5cd80fc8+24129607/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:aa1bdeb5c8a482c88f0fc4864061a0205a01bfd033bf83dec7b3699a8b3cfb76
3
  size 267264
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3cf2be9da8594eaa5a105457213149fa45406fd21f0520fd7a4c85e02a49a94d
3
  size 267264
neuronxcc-2.21.18209.0+043b1bf7/MODULE_62abb2454c6da8d57ad5+a02c3a36/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:9be266b8f1fde950f5494a506f680db11039381a0394b3f7feff596049128590
3
  size 543744
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:717b6d19c8732bde44db15f49ef215eba0ace5ef06c847ce2f5d2be87241a2d6
3
  size 543744
neuronxcc-2.21.18209.0+043b1bf7/MODULE_62abb2454c6da8d57ad5+a02c3a36/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:57b15690253dc88f3fa4dd35d1249a712e46d773ffb81541b675504d98e07a57
3
  size 563380
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7810674e248cb7be00a5c030a2618a679130073031477fff3fa6cc718dbdf575
3
  size 563380
neuronxcc-2.21.18209.0+043b1bf7/MODULE_73587bc7122ba5a81758+80d05c3f/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:fd65cfc28b5490f918f6b2aa4e0eeb688b82cdc49e4dec6872923521d6fef18d
3
  size 277504
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:20277f8fa6c9d9b73b350f8bd2185dcacfb6a14e40d29b979aa4df87370f5a23
3
  size 277504
neuronxcc-2.21.18209.0+043b1bf7/MODULE_73587bc7122ba5a81758+80d05c3f/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:e5430cc5b5cab84a5965fa959cd8a8eae6acfd6edeb45d544cea462af8a43baf
3
  size 289571
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0e20ae142ae0fa5cb43ce1efd93f34384ab7f74d6841845e0d0b258a5a88a1b6
3
  size 289571
neuronxcc-2.21.18209.0+043b1bf7/MODULE_75329916c4da1120954c+6e4949b4/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:1e09232be137868b2d8cfd22f270e71d10612acbe309447a40c430b392330504
3
  size 646144
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:87453a63a0e61f0ee4aa33fac4a5226ae229be32a8dc97197a3d74c7ca6181ac
3
  size 646144
neuronxcc-2.21.18209.0+043b1bf7/MODULE_7c590ab98f45f8a8dd15+283df001/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:54d6c8a2fb33e791d6f85e9706e36e9c541940ff0756b07f528aaed5f6bd9a3a
3
  size 410624
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5d1bba671be448d435c50e4127ec5fba1ba16046c08f8989bbcf523be7c008f0
3
  size 410624
neuronxcc-2.21.18209.0+043b1bf7/MODULE_944763b36d810f4af582+24129607/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:92dd3b28236dcc312133af918524457b58d3edd938a7215e9182986096f412cf
3
  size 328704
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ddd8d3551d47f323a91bf25ad839a3dca973388be862d715c63cdb354af0205b
3
  size 328704
neuronxcc-2.21.18209.0+043b1bf7/MODULE_b23271513f9095bde4db+2dde74c7/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:292a69a9148f9c05f59edcf90bedc36df7f6b4ad4c9809803417a2722d2fcbe4
3
  size 369664
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ceeef9279bb8e51e7a0d62471f12c7f6c3b1251ab67c4c1ddf4176ce0792ccc2
3
  size 369664
neuronxcc-2.21.18209.0+043b1bf7/MODULE_b23271513f9095bde4db+2dde74c7/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:fa540a0feade5d8e05fe273feb36a195264ef8bc3894cee3435a1855f5dd1a86
3
  size 379362
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6261673e4357b7cc8be86c50d68b2dec0391fdcf82f47ddbd7f87957d8a80ab6
3
  size 379362
neuronxcc-2.21.18209.0+043b1bf7/MODULE_bdf25cde1a48d420319c+a02c3a36/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:3a4c1cad6cc2f4091267dbddfa3483e0246772dfcaf064f47f6badd4fa8a7007
3
  size 246784
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:064181ba521191c47841fa65dfaa6d03a0888e4a5ee8dc02c69c752f260e87d4
3
  size 246784
neuronxcc-2.21.18209.0+043b1bf7/MODULE_bdf25cde1a48d420319c+a02c3a36/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:00bbb2afc046d4c3cbbbfe3850c81e7e3f937aeb30a2d4a3d11a1183307d5df8
3
  size 255104
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:159ae54b26f84dc04db91de2e57f8129840bd7d4d9b5eacc2be5d8861d68ba66
3
  size 255104
neuronxcc-2.21.18209.0+043b1bf7/MODULE_c4737893663f5ff66b53+24129607/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:1af38a8a6754c625afb316538337a813719c32cb5ad266a92973a197f4e5b7e5
3
  size 359424
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6f2fc165c86186746a174e43e90d8175dcfc08a352edbd260d94323c49ff822d
3
  size 359424
neuronxcc-2.21.18209.0+043b1bf7/MODULE_cde3417e6918fabff4a7+a02c3a36/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:64919d62dcbd34034657e41c929a84e1c0f96c218f37a4016a38c99d5485a545
3
  size 277504
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:05348be5d75c411834a603897c254cec02819e459d1f28ec3220ee12f482e066
3
  size 277504
neuronxcc-2.21.18209.0+043b1bf7/MODULE_cde3417e6918fabff4a7+a02c3a36/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:9cd83cdc12477b3a9cdba56d570ba892e03757d68ce4bff8fd3b6b76bf942c12
3
  size 285854
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:193be0e696b96269db8603af550ccfd1f46ff3e540ed5a5b4df6a7754864490f
3
  size 285854
neuronxcc-2.21.18209.0+043b1bf7/MODULE_f52fe8e88daf6dc9619a+24129607/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:e4d8f8136bbefdd7a815b2493c64d37cc3a33c9bdc4b0cae8d8b979c26547eb7
3
  size 625664
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1b4b3605c28a6dc8aebf8fce2eab1c88030ee1772ae2ee91211e1bc57a033923
3
  size 625664
neuronxcc-2.21.18209.0+043b1bf7/MODULE_fef9240b8fbfaca895ba+a02c3a36/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ca053ed57e065c4e562e8d3d57c87291ae798920c6dd5071fb790b206161b694
3
  size 216064
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dae5ae0f9d16e9871be2fcef3cd5d4444e47b277935ec78f2a7a4e60d1cc40c9
3
  size 216064
neuronxcc-2.21.18209.0+043b1bf7/MODULE_fef9240b8fbfaca895ba+a02c3a36/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:9a913279d78182f991b552ba0bffdedaacf106eefd8272400efea91e13e93987
3
  size 224412
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:979ade80e54f22a5d9eff2409e8f9534de75ddec1a951c87a3ee2f9af1d32d65
3
  size 224412