Hugging Face's logo

optimum-internal-testing
/
neuron-testing-cache

Model card Files
xet
Community
optimum-internal-testing-user commited on
Commit
61c8225
·
verified ·
1 Parent(s): 5bd5bfd

Synchronizing local compiler cache.

Browse files
Files changed (36) hide show
  1. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.3.1.dev7/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/4af54b51abf746d11b6a.json +59 -0
  2. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.3.1.dev7/llama/llamafactory/tiny-random-Llama-3/c153af987ca58d7017b3.json +63 -0
  3. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.3.1.dev7/llama/unsloth/Llama-3.2-1B-Instruct/c6178a14860600d39f62.json +64 -0
  4. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.3.1.dev7/llama4_text/tiny-random/llama-4/97410fe9a0bdba745731.json +82 -0
  5. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.3.1.dev7/mixtral/dacorvo/Mixtral-tiny/6af236681ddeb0e2efe7.json +59 -0
  6. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.3.1.dev7/phi3/yujiepan/phi-4-tiny-random/655c0f6904c3de86d32b.json +60 -0
  7. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.3.1.dev7/qwen2/yujiepan/qwen2.5-128k-tiny-random/17d5f52aa64c5cdcfb59.json +65 -0
  8. neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.3.1.dev7/qwen3_moe/optimum-internal-testing/tiny-random-qwen3_moe/183c9d154bd4a6271376.json +66 -0
  9. neuronxcc-2.21.18209.0+043b1bf7/MODULE_1bcd8486a5476c06ef6c+ed72d204/model.hlo_module.pb +1 -1
  10. neuronxcc-2.21.18209.0+043b1bf7/MODULE_1bcd8486a5476c06ef6c+ed72d204/model.neff +1 -1
  11. neuronxcc-2.21.18209.0+043b1bf7/MODULE_1f955163eec0c0a80988+a9d440f5/model.neff +1 -1
  12. neuronxcc-2.21.18209.0+043b1bf7/MODULE_1f955163eec0c0a80988+a9d440f5/wrapped_neff.hlo +1 -1
  13. neuronxcc-2.21.18209.0+043b1bf7/MODULE_2e7323d127244060130b+a9d440f5/model.neff +1 -1
  14. neuronxcc-2.21.18209.0+043b1bf7/MODULE_2e7323d127244060130b+a9d440f5/wrapped_neff.hlo +1 -1
  15. neuronxcc-2.21.18209.0+043b1bf7/MODULE_61dbdf738ca076ef369a+a9d440f5/model.neff +1 -1
  16. neuronxcc-2.21.18209.0+043b1bf7/MODULE_61dbdf738ca076ef369a+a9d440f5/wrapped_neff.hlo +1 -1
  17. neuronxcc-2.21.18209.0+043b1bf7/MODULE_83cf5c47093549d97b56+ed72d204/model.hlo_module.pb +1 -1
  18. neuronxcc-2.21.18209.0+043b1bf7/MODULE_83cf5c47093549d97b56+ed72d204/model.neff +1 -1
  19. neuronxcc-2.21.18209.0+043b1bf7/MODULE_9413970e765acfb7b93b+ca355898/model.neff +1 -1
  20. neuronxcc-2.21.18209.0+043b1bf7/MODULE_9413970e765acfb7b93b+ca355898/wrapped_neff.hlo +1 -1
  21. neuronxcc-2.21.18209.0+043b1bf7/MODULE_a2ecbfaef5906dad2124+a9d440f5/model.neff +1 -1
  22. neuronxcc-2.21.18209.0+043b1bf7/MODULE_a2ecbfaef5906dad2124+a9d440f5/wrapped_neff.hlo +1 -1
  23. neuronxcc-2.21.18209.0+043b1bf7/MODULE_b37fdf56be00130a14e9+ed72d204/model.hlo_module.pb +1 -1
  24. neuronxcc-2.21.18209.0+043b1bf7/MODULE_b37fdf56be00130a14e9+ed72d204/model.neff +1 -1
  25. neuronxcc-2.21.18209.0+043b1bf7/MODULE_cc93bedf1ffea21b58ba+cd3419b6/model.neff +1 -1
  26. neuronxcc-2.21.18209.0+043b1bf7/MODULE_cc93bedf1ffea21b58ba+cd3419b6/wrapped_neff.hlo +1 -1
  27. neuronxcc-2.21.18209.0+043b1bf7/MODULE_d55f7459a88a4b840dd4+253d6470/model.hlo_module.pb +1 -1
  28. neuronxcc-2.21.18209.0+043b1bf7/MODULE_d55f7459a88a4b840dd4+253d6470/model.neff +1 -1
  29. neuronxcc-2.21.18209.0+043b1bf7/MODULE_d9bfb8a9c556b013ab59+a9d440f5/model.neff +1 -1
  30. neuronxcc-2.21.18209.0+043b1bf7/MODULE_d9bfb8a9c556b013ab59+a9d440f5/wrapped_neff.hlo +1 -1
  31. neuronxcc-2.21.18209.0+043b1bf7/MODULE_da2f9ef9477a4d66d1a7+ed72d204/model.hlo_module.pb +1 -1
  32. neuronxcc-2.21.18209.0+043b1bf7/MODULE_da2f9ef9477a4d66d1a7+ed72d204/model.neff +1 -1
  33. neuronxcc-2.21.18209.0+043b1bf7/MODULE_e70394de1d791e12c520+ed72d204/model.hlo_module.pb +1 -1
  34. neuronxcc-2.21.18209.0+043b1bf7/MODULE_e70394de1d791e12c520+ed72d204/model.neff +1 -1
  35. neuronxcc-2.21.18209.0+043b1bf7/MODULE_f359722ac2fcd615a216+c2248236/model.hlo_module.pb +1 -1
  36. neuronxcc-2.21.18209.0+043b1bf7/MODULE_f359722ac2fcd615a216+c2248236/model.neff +1 -1
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.3.1.dev7/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/4af54b51abf746d11b6a.json ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "embedding_multiplier": 1.0,
12
+ "hidden_act": "silu",
13
+ "hidden_size": 32,
14
+ "initializer_range": 0.02,
15
+ "intermediate_size": 64,
16
+ "logits_scaling": 1.0,
17
+ "max_position_embeddings": 2048,
18
+ "mlp_bias": false,
19
+ "model_type": "granite",
20
+ "neuron": {
21
+ "_serialized_key": "NxDNeuronConfig",
22
+ "batch_size": 1,
23
+ "capacity_factor": null,
24
+ "checkpoint_id": "hf-internal-testing/tiny-random-GraniteForCausalLM",
25
+ "checkpoint_revision": "c3074ebc0ac2fe545305f5e5f6cce2cc9b2aa0c5",
26
+ "continuous_batching": false,
27
+ "enable_bucketing": false,
28
+ "ep_degree": 1,
29
+ "fused_qkv": true,
30
+ "glu_mlp": true,
31
+ "local_ranks_size": 2,
32
+ "logical_nc_config": 1,
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": "0.3.1.dev7",
40
+ "output_logits": false,
41
+ "pp_degree": 1,
42
+ "sequence_length": 1024,
43
+ "speculation_length": 0,
44
+ "start_rank_id": 0,
45
+ "target": null,
46
+ "torch_dtype": "bfloat16",
47
+ "tp_degree": 2
48
+ },
49
+ "num_attention_heads": 4,
50
+ "num_hidden_layers": 2,
51
+ "num_key_value_heads": 4,
52
+ "residual_multiplier": 1.0,
53
+ "rms_norm_eps": 1e-06,
54
+ "rope_scaling": null,
55
+ "rope_theta": 10000.0,
56
+ "tie_word_embeddings": false,
57
+ "use_cache": true,
58
+ "vocab_size": 49152
59
+ }
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.3.1.dev7/llama/llamafactory/tiny-random-Llama-3/c153af987ca58d7017b3.json ADDED
@@ -0,0 +1,63 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "head_dim": 4,
11
+ "hidden_act": "silu",
12
+ "hidden_size": 16,
13
+ "initializer_range": 0.02,
14
+ "intermediate_size": 64,
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": "llamafactory/tiny-random-Llama-3",
23
+ "checkpoint_revision": "bf2a2e3bf199ad2ee96f02a3c00246c608db22a8",
24
+ "continuous_batching": false,
25
+ "enable_bucketing": false,
26
+ "ep_degree": 1,
27
+ "fused_qkv": true,
28
+ "glu_mlp": true,
29
+ "local_ranks_size": 2,
30
+ "logical_nc_config": 1,
31
+ "max_batch_size": 1,
32
+ "max_context_length": 1024,
33
+ "max_topk": 256,
34
+ "n_active_tokens": 1024,
35
+ "neuronxcc_version": "2.21.18209.0+043b1bf7",
36
+ "on_device_sampling": true,
37
+ "optimum_neuron_version": "0.3.1.dev7",
38
+ "output_logits": false,
39
+ "pp_degree": 1,
40
+ "sequence_length": 1024,
41
+ "speculation_length": 0,
42
+ "start_rank_id": 0,
43
+ "target": null,
44
+ "torch_dtype": "bfloat16",
45
+ "tp_degree": 2
46
+ },
47
+ "num_attention_heads": 4,
48
+ "num_hidden_layers": 2,
49
+ "num_key_value_heads": 4,
50
+ "pretraining_tp": 1,
51
+ "rms_norm_eps": 1e-05,
52
+ "rope_scaling": {
53
+ "factor": 8.0,
54
+ "high_freq_factor": 4.0,
55
+ "low_freq_factor": 1.0,
56
+ "original_max_position_embeddings": 8192,
57
+ "rope_type": "llama3"
58
+ },
59
+ "rope_theta": 500000.0,
60
+ "tie_word_embeddings": false,
61
+ "use_cache": true,
62
+ "vocab_size": 128256
63
+ }
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.3.1.dev7/llama/unsloth/Llama-3.2-1B-Instruct/c6178a14860600d39f62.json ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "batch_size": 1,
21
+ "capacity_factor": null,
22
+ "checkpoint_id": "unsloth/Llama-3.2-1B-Instruct",
23
+ "checkpoint_revision": null,
24
+ "continuous_batching": false,
25
+ "enable_bucketing": false,
26
+ "ep_degree": 1,
27
+ "fused_qkv": false,
28
+ "glu_mlp": true,
29
+ "local_ranks_size": 2,
30
+ "logical_nc_config": 1,
31
+ "max_batch_size": 1,
32
+ "max_context_length": 4096,
33
+ "max_topk": 256,
34
+ "n_active_tokens": 4096,
35
+ "neuronxcc_version": "2.21.18209.0+043b1bf7",
36
+ "on_device_sampling": false,
37
+ "optimum_neuron_version": "0.3.1.dev7",
38
+ "output_logits": false,
39
+ "pp_degree": 1,
40
+ "sequence_length": 4096,
41
+ "speculation_length": 5,
42
+ "start_rank_id": 0,
43
+ "target": null,
44
+ "torch_dtype": "bfloat16",
45
+ "tp_degree": 2
46
+ },
47
+ "num_attention_heads": 32,
48
+ "num_hidden_layers": 16,
49
+ "num_key_value_heads": 8,
50
+ "pretraining_tp": 1,
51
+ "rms_norm_eps": 1e-05,
52
+ "rope_scaling": {
53
+ "factor": 32.0,
54
+ "high_freq_factor": 4.0,
55
+ "low_freq_factor": 1.0,
56
+ "original_max_position_embeddings": 8192,
57
+ "rope_type": "llama3"
58
+ },
59
+ "rope_theta": 500000.0,
60
+ "tie_word_embeddings": true,
61
+ "unsloth_fixed": true,
62
+ "use_cache": true,
63
+ "vocab_size": 128256
64
+ }
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.3.1.dev7/llama4_text/tiny-random/llama-4/97410fe9a0bdba745731.json ADDED
@@ -0,0 +1,82 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "floor_scale": 8192,
12
+ "for_llm_compressor": false,
13
+ "head_dim": 32,
14
+ "hidden_act": "silu",
15
+ "hidden_size": 32,
16
+ "initializer_range": 0.02,
17
+ "interleave_moe_layer_step": 2,
18
+ "intermediate_size": 64,
19
+ "intermediate_size_mlp": 128,
20
+ "layer_types": [
21
+ "chunked_attention",
22
+ "chunked_attention",
23
+ "chunked_attention",
24
+ "full_attention"
25
+ ],
26
+ "max_position_embeddings": 1048576,
27
+ "model_type": "llama4_text",
28
+ "moe_layers": [
29
+ 1,
30
+ 3
31
+ ],
32
+ "neuron": {
33
+ "_serialized_key": "NxDNeuronConfig",
34
+ "batch_size": 1,
35
+ "capacity_factor": null,
36
+ "checkpoint_id": "tiny-random/llama-4",
37
+ "checkpoint_revision": "9e716f5d4d1ffe0a44a15f46f4a12b840439aba4",
38
+ "continuous_batching": false,
39
+ "enable_bucketing": false,
40
+ "ep_degree": 1,
41
+ "fused_qkv": false,
42
+ "glu_mlp": true,
43
+ "local_ranks_size": 2,
44
+ "logical_nc_config": 1,
45
+ "max_batch_size": 1,
46
+ "max_context_length": 1024,
47
+ "max_topk": 256,
48
+ "n_active_tokens": 1024,
49
+ "neuronxcc_version": "2.21.18209.0+043b1bf7",
50
+ "on_device_sampling": true,
51
+ "optimum_neuron_version": "0.3.1.dev7",
52
+ "output_logits": false,
53
+ "pp_degree": 1,
54
+ "sequence_length": 1024,
55
+ "speculation_length": 0,
56
+ "start_rank_id": 0,
57
+ "target": null,
58
+ "torch_dtype": "bfloat16",
59
+ "tp_degree": 2
60
+ },
61
+ "no_rope_layers": [
62
+ 1,
63
+ 1,
64
+ 1,
65
+ 0
66
+ ],
67
+ "num_attention_heads": 1,
68
+ "num_experts_per_tok": 1,
69
+ "num_hidden_layers": 4,
70
+ "num_key_value_heads": 1,
71
+ "num_local_experts": 8,
72
+ "output_router_logits": false,
73
+ "rms_norm_eps": 1e-05,
74
+ "rope_scaling": null,
75
+ "rope_theta": 500000.0,
76
+ "router_aux_loss_coef": 0.001,
77
+ "router_jitter_noise": 0.0,
78
+ "tie_word_embeddings": true,
79
+ "use_cache": true,
80
+ "use_qk_norm": true,
81
+ "vocab_size": 202048
82
+ }
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.3.1.dev7/mixtral/dacorvo/Mixtral-tiny/6af236681ddeb0e2efe7.json ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "head_dim": 32,
10
+ "hidden_act": "silu",
11
+ "hidden_size": 1024,
12
+ "initializer_range": 0.02,
13
+ "intermediate_size": 3584,
14
+ "max_position_embeddings": 1024,
15
+ "model_type": "mixtral",
16
+ "neuron": {
17
+ "_serialized_key": "NxDNeuronConfig",
18
+ "batch_size": 1,
19
+ "capacity_factor": null,
20
+ "checkpoint_id": "dacorvo/Mixtral-tiny",
21
+ "checkpoint_revision": "c557ba205ddff6ea911f4719e0d543d6c08356b6",
22
+ "continuous_batching": false,
23
+ "enable_bucketing": false,
24
+ "ep_degree": 1,
25
+ "fused_qkv": false,
26
+ "glu_mlp": true,
27
+ "local_ranks_size": 2,
28
+ "logical_nc_config": 1,
29
+ "max_batch_size": 1,
30
+ "max_context_length": 1024,
31
+ "max_topk": 256,
32
+ "n_active_tokens": 1024,
33
+ "neuronxcc_version": "2.21.18209.0+043b1bf7",
34
+ "on_device_sampling": false,
35
+ "optimum_neuron_version": "0.3.1.dev7",
36
+ "output_logits": false,
37
+ "pp_degree": 1,
38
+ "sequence_length": 1024,
39
+ "speculation_length": 0,
40
+ "start_rank_id": 0,
41
+ "target": null,
42
+ "torch_dtype": "bfloat16",
43
+ "tp_degree": 2
44
+ },
45
+ "num_attention_heads": 32,
46
+ "num_experts_per_tok": 2,
47
+ "num_hidden_layers": 2,
48
+ "num_key_value_heads": 8,
49
+ "num_local_experts": 8,
50
+ "output_router_logits": false,
51
+ "rms_norm_eps": 1e-05,
52
+ "rope_theta": 10000.0,
53
+ "router_aux_loss_coef": 0.001,
54
+ "router_jitter_noise": 0.0,
55
+ "sliding_window": 4096,
56
+ "tie_word_embeddings": false,
57
+ "use_cache": true,
58
+ "vocab_size": 32000
59
+ }
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.3.1.dev7/phi3/yujiepan/phi-4-tiny-random/655c0f6904c3de86d32b.json ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "embd_pdrop": 0.0,
12
+ "hidden_act": "silu",
13
+ "hidden_size": 16,
14
+ "initializer_range": 0.02,
15
+ "intermediate_size": 32,
16
+ "max_position_embeddings": 16384,
17
+ "model_type": "phi3",
18
+ "neuron": {
19
+ "_serialized_key": "NxDNeuronConfig",
20
+ "batch_size": 1,
21
+ "capacity_factor": null,
22
+ "checkpoint_id": "yujiepan/phi-4-tiny-random",
23
+ "checkpoint_revision": "18a9a1168dc97ac6d128f811925670c275610f5a",
24
+ "continuous_batching": false,
25
+ "enable_bucketing": false,
26
+ "ep_degree": 1,
27
+ "fused_qkv": true,
28
+ "glu_mlp": true,
29
+ "local_ranks_size": 2,
30
+ "logical_nc_config": 1,
31
+ "max_batch_size": 1,
32
+ "max_context_length": 1024,
33
+ "max_topk": 256,
34
+ "n_active_tokens": 1024,
35
+ "neuronxcc_version": "2.21.18209.0+043b1bf7",
36
+ "on_device_sampling": true,
37
+ "optimum_neuron_version": "0.3.1.dev7",
38
+ "output_logits": false,
39
+ "pp_degree": 1,
40
+ "sequence_length": 1024,
41
+ "speculation_length": 0,
42
+ "start_rank_id": 0,
43
+ "target": null,
44
+ "torch_dtype": "bfloat16",
45
+ "tp_degree": 2
46
+ },
47
+ "num_attention_heads": 2,
48
+ "num_hidden_layers": 2,
49
+ "num_key_value_heads": 1,
50
+ "original_max_position_embeddings": 16384,
51
+ "partial_rotary_factor": 1.0,
52
+ "resid_pdrop": 0.0,
53
+ "rms_norm_eps": 1e-05,
54
+ "rope_scaling": null,
55
+ "rope_theta": 250000,
56
+ "sliding_window": null,
57
+ "tie_word_embeddings": false,
58
+ "use_cache": true,
59
+ "vocab_size": 100352
60
+ }
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.3.1.dev7/qwen2/yujiepan/qwen2.5-128k-tiny-random/17d5f52aa64c5cdcfb59.json ADDED
@@ -0,0 +1,65 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "hidden_act": "silu",
10
+ "hidden_size": 8,
11
+ "initializer_range": 0.02,
12
+ "intermediate_size": 16,
13
+ "layer_types": [
14
+ "full_attention",
15
+ "full_attention"
16
+ ],
17
+ "max_position_embeddings": 32768,
18
+ "max_window_layers": 1,
19
+ "model_type": "qwen2",
20
+ "neuron": {
21
+ "_serialized_key": "NxDNeuronConfig",
22
+ "batch_size": 1,
23
+ "capacity_factor": null,
24
+ "checkpoint_id": "yujiepan/qwen2.5-128k-tiny-random",
25
+ "checkpoint_revision": "c8296d4ca3f87782876d2382fbb6481d1beb8ef0",
26
+ "continuous_batching": false,
27
+ "enable_bucketing": false,
28
+ "ep_degree": 1,
29
+ "fused_qkv": false,
30
+ "glu_mlp": true,
31
+ "local_ranks_size": 2,
32
+ "logical_nc_config": 1,
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": "0.3.1.dev7",
40
+ "output_logits": false,
41
+ "pp_degree": 1,
42
+ "sequence_length": 1024,
43
+ "speculation_length": 0,
44
+ "start_rank_id": 0,
45
+ "target": null,
46
+ "torch_dtype": "bfloat16",
47
+ "tp_degree": 2
48
+ },
49
+ "num_attention_heads": 4,
50
+ "num_hidden_layers": 2,
51
+ "num_key_value_heads": 2,
52
+ "rms_norm_eps": 1e-06,
53
+ "rope_scaling": {
54
+ "factor": 4.0,
55
+ "original_max_position_embeddings": 32768,
56
+ "rope_type": "yarn",
57
+ "type": "yarn"
58
+ },
59
+ "rope_theta": 1000000.0,
60
+ "sliding_window": null,
61
+ "tie_word_embeddings": false,
62
+ "use_cache": true,
63
+ "use_sliding_window": false,
64
+ "vocab_size": 152064
65
+ }
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.3.1.dev7/qwen3_moe/optimum-internal-testing/tiny-random-qwen3_moe/183c9d154bd4a6271376.json ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ "head_dim": 32,
12
+ "hidden_act": "silu",
13
+ "hidden_size": 64,
14
+ "initializer_range": 0.02,
15
+ "intermediate_size": 128,
16
+ "max_position_embeddings": 40960,
17
+ "max_window_layers": 1,
18
+ "mlp_only_layers": [],
19
+ "model_type": "qwen3_moe",
20
+ "moe_intermediate_size": 128,
21
+ "neuron": {
22
+ "_serialized_key": "NxDNeuronConfig",
23
+ "batch_size": 1,
24
+ "capacity_factor": null,
25
+ "checkpoint_id": "optimum-internal-testing/tiny-random-qwen3_moe",
26
+ "checkpoint_revision": "e0230be2839556b44b7400a233c73c74b4abb7af",
27
+ "continuous_batching": false,
28
+ "enable_bucketing": false,
29
+ "ep_degree": 1,
30
+ "fused_qkv": false,
31
+ "glu_mlp": true,
32
+ "local_ranks_size": 2,
33
+ "logical_nc_config": 1,
34
+ "max_batch_size": 1,
35
+ "max_context_length": 1024,
36
+ "max_topk": 256,
37
+ "n_active_tokens": 1024,
38
+ "neuronxcc_version": "2.21.18209.0+043b1bf7",
39
+ "on_device_sampling": true,
40
+ "optimum_neuron_version": "0.3.1.dev7",
41
+ "output_logits": false,
42
+ "pp_degree": 1,
43
+ "sequence_length": 1024,
44
+ "speculation_length": 0,
45
+ "start_rank_id": 0,
46
+ "target": null,
47
+ "torch_dtype": "bfloat16",
48
+ "tp_degree": 2
49
+ },
50
+ "norm_topk_prob": true,
51
+ "num_attention_heads": 2,
52
+ "num_experts": 8,
53
+ "num_experts_per_tok": 2,
54
+ "num_hidden_layers": 2,
55
+ "num_key_value_heads": 1,
56
+ "output_router_logits": false,
57
+ "rms_norm_eps": 1e-06,
58
+ "rope_scaling": null,
59
+ "rope_theta": 1000000.0,
60
+ "router_aux_loss_coef": 0.001,
61
+ "sliding_window": null,
62
+ "tie_word_embeddings": true,
63
+ "use_cache": true,
64
+ "use_sliding_window": false,
65
+ "vocab_size": 151936
66
+ }
neuronxcc-2.21.18209.0+043b1bf7/MODULE_1bcd8486a5476c06ef6c+ed72d204/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:c4dc03e16062b1663ff5165bc8a89267d6fd336b6c7e071e7407b74fb98c700c
3
  size 86123
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a6ef80ef3cb60d75adb0ca4623dbc94521657e3d2c87edcc3dafa307e2018e44
3
  size 86123
neuronxcc-2.21.18209.0+043b1bf7/MODULE_1bcd8486a5476c06ef6c+ed72d204/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:5924e3b77731a29d1fcfd621b509f821bee84c4da050992d6c8ae0f2fe62a96d
3
  size 257024
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bf5e594f2fc800091ee01be5fdd2d0520d10556608360285e64f3559c71827f2
3
  size 257024
neuronxcc-2.21.18209.0+043b1bf7/MODULE_1f955163eec0c0a80988+a9d440f5/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:e070ed42641ef2f47dbe21548d4d5eb645012bde0712d5ca5857e88eb8355a8b
3
  size 277504
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:33ebd3c8b6fe8e42e02557d6bd4f0e42f6e9495a880d9f0df063be8b480d8fb7
3
  size 277504
neuronxcc-2.21.18209.0+043b1bf7/MODULE_1f955163eec0c0a80988+a9d440f5/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:afda52dbba53bfca29b2dad8dc3c2d460c63ce8afc651b95c4eda744b31762c3
3
  size 289031
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:388f72f4c1762ffad0b22d77aede0960740405ee7e30e3bcc5f02525fe1245c3
3
  size 289031
neuronxcc-2.21.18209.0+043b1bf7/MODULE_2e7323d127244060130b+a9d440f5/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:2a578825ef8646927536b47cdb3fa276198a366020af795418bedda1f4afcc8a
3
  size 543744
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3d9ae26135029118e6981ba882bbad31be8e04ccb3d72b1a23f9c3b05623efa4
3
  size 543744
neuronxcc-2.21.18209.0+043b1bf7/MODULE_2e7323d127244060130b+a9d440f5/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:70c652f8b81492de9a2083a79be8f432f65d183bc5b87b8b37778641cda530b2
3
  size 563380
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:96553b9d89b36bec9f2f8b66c463a4ef3bf236b352362a6de720451fce35c262
3
  size 563380
neuronxcc-2.21.18209.0+043b1bf7/MODULE_61dbdf738ca076ef369a+a9d440f5/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:770fd1206a5629984460e7429939d19ad42d3ae6bf81b225a80642effe8b18d3
3
  size 216064
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:993c0762582a5a9d8b30f793072bfb135a59c5398c311b1d2bfbeab365d4470a
3
  size 216064
neuronxcc-2.21.18209.0+043b1bf7/MODULE_61dbdf738ca076ef369a+a9d440f5/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:0c0052f1d39a3951738cecaafefde0bb52a3f92ea636adc35066d09160a8feea
3
  size 224412
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b8f56d2fabb9364bee5ba0490059eda72b9e4daf841f4a8b673704f76156bf0e
3
  size 224412
neuronxcc-2.21.18209.0+043b1bf7/MODULE_83cf5c47093549d97b56+ed72d204/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:151194d0fb53ed69ccb9ed2e6e778ca3c4e378f7dc502bd714607def47154e99
3
  size 694128
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c2b33c4c0522fe9947a0c68ef64814fde72dde4f6ed9d34d5d5619cf11339c8d
3
  size 694128
neuronxcc-2.21.18209.0+043b1bf7/MODULE_83cf5c47093549d97b56+ed72d204/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:10ea3d05a936f21d4c9a03a7754b654b6cf95be775bf68ce1d227f694fcd9ee8
3
  size 625664
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6db6c89372f665a38fde37407d067c84f3617ec0cccb884737bb0db99382b8ae
3
  size 625664
neuronxcc-2.21.18209.0+043b1bf7/MODULE_9413970e765acfb7b93b+ca355898/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b100a7ac40df374b83f0a6cb2473ee1ded71cecea8eb3a7a2538e2d47eb1e1d8
3
  size 277504
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6b63d5da99352233ddc5e5579b600ac439c3433f01e6273ff1d0ae1efac0affe
3
  size 277504
neuronxcc-2.21.18209.0+043b1bf7/MODULE_9413970e765acfb7b93b+ca355898/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:877022416bcdaab2e25ce13b0e9f9274914d12ecd836d4524cc6520426bb8c1e
3
  size 289571
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9a0dbab9e8c7009010fb3250478bcbbd9e0b411792d8ff0bbec2b3d7c5ee750f
3
  size 289571
neuronxcc-2.21.18209.0+043b1bf7/MODULE_a2ecbfaef5906dad2124+a9d440f5/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:030072481fd6c7d80b617b0b351e1e6957b78b96a3e29d0d230bd28fcedc30e7
3
  size 277504
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:de15f33c1abb9a2ee624aff6daf275532359f2da5d71b960f24b41b5c2b0d029
3
  size 277504
neuronxcc-2.21.18209.0+043b1bf7/MODULE_a2ecbfaef5906dad2124+a9d440f5/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:517a87a2493dfb014bd5f68fcab45118b1f28c460345b025e4b99df8c4007ed4
3
  size 285854
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c3afd9e68296adbc0ca6fc6da5368d6a91faab57a468f319d37411a9bb4236aa
3
  size 285854
neuronxcc-2.21.18209.0+043b1bf7/MODULE_b37fdf56be00130a14e9+ed72d204/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d065f8d4d6f1bf1cea666bfce7464a2aca12bf44873244b9157c1caffd720f0d
3
  size 82772
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:06fd8e452936ab55254a1235f09adc57bc5a456f714ad3f56c18ad9d1a7a796c
3
  size 82772
neuronxcc-2.21.18209.0+043b1bf7/MODULE_b37fdf56be00130a14e9+ed72d204/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:44d9e8112524094d41e817259f3f0fccc05a86667c23170541243a02720108a5
3
  size 267264
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:89aee54905f5adeeca417d37958c06599a2b31312652db385873b1805f71b68d
3
  size 267264
neuronxcc-2.21.18209.0+043b1bf7/MODULE_cc93bedf1ffea21b58ba+cd3419b6/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:65f3473fa2b14f940a23bff32205b58f2aa9173a612162f12dbc3a761eaa0bcb
3
  size 328704
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9d6681ecb6abfa562adfb83ccf2ec0e646996f810414c2a6c35ca1a0365cd559
3
  size 328704
neuronxcc-2.21.18209.0+043b1bf7/MODULE_cc93bedf1ffea21b58ba+cd3419b6/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ccc6e31c8d6dee70e7daa0174d82deedb51d0ae9c68e73bb2796b3b817c10351
3
  size 338402
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:36fbb5ce321757902d5ad391c66411caf80e4d12da10df009f9b3df255615b2d
3
  size 338402
neuronxcc-2.21.18209.0+043b1bf7/MODULE_d55f7459a88a4b840dd4+253d6470/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:35b5fc2c042152ca44f6d9d602f0e045c6817a0584e944e7e73ed828d193880b
3
  size 104026
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1879a7a85dc136634486007b5c6bb4edba879cb0315f90f6bed1b69b015f0582
3
  size 104026
neuronxcc-2.21.18209.0+043b1bf7/MODULE_d55f7459a88a4b840dd4+253d6470/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:a589d111f6c2cfbea8866546021d7a1aa9b826b573d0d63e21b241076e20f504
3
  size 369664
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:da8a935db2944e68bc2469f0fa9d69805a342c5bec1466ac387e4abc7b10176e
3
  size 369664
neuronxcc-2.21.18209.0+043b1bf7/MODULE_d9bfb8a9c556b013ab59+a9d440f5/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:4066dd46e1d54dabdf88f312013ad294c7fc88c77a7fbb231922407f052488aa
3
  size 246784
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3f651aeeb5c33f515c93ea1dbcc94cfe02117a442ac1809bb67ffa8a5b51913a
3
  size 246784
neuronxcc-2.21.18209.0+043b1bf7/MODULE_d9bfb8a9c556b013ab59+a9d440f5/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:37ced1bab1fd78821049fa662e57d160be710f7662fcc47b5fdf639b3ea68506
3
  size 255104
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2f19e222ca12644edfa90789c4d91dc1a2ef8a09b320611b9edc2f836577ed36
3
  size 255104
neuronxcc-2.21.18209.0+043b1bf7/MODULE_da2f9ef9477a4d66d1a7+ed72d204/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:afa0ba4e2d28a0f71e13ca970f64df79176ee70fd0b970cf30ed05823b1d8a87
3
  size 83504
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5f564b21f936be2f7c37157d6dd54ad7e1ea7bb1ef4a98065c35b9069bb10c70
3
  size 83504
neuronxcc-2.21.18209.0+043b1bf7/MODULE_da2f9ef9477a4d66d1a7+ed72d204/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ce5ebc88d61c444b20f154f4a45dcf148a9d6a8e782d55f93957aa93706df119
3
  size 328704
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dfe96800a0f8eac88577ab03ecd27703d5555f255a79aa235006f1d09d9d56a9
3
  size 328704
neuronxcc-2.21.18209.0+043b1bf7/MODULE_e70394de1d791e12c520+ed72d204/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:7e323d3dbdc3856b221fc463f0cea182e2c70423c291e9beb5c2a93e844508c6
3
  size 90382
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c13ef0e8f1b71f0013908b62623b6e90fb093b3a681f9582a6b058faa94ed0f0
3
  size 90382
neuronxcc-2.21.18209.0+043b1bf7/MODULE_e70394de1d791e12c520+ed72d204/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:dca15053b670a78cda44e3c2bea4d3c7903ed75006f6e5b9f2e3c4e7b0f2e50b
3
  size 359424
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6bec00e84950bd9045bd7bb9f181a9e33b16bb74fbfcf2110d06261a149b1fa1
3
  size 359424
neuronxcc-2.21.18209.0+043b1bf7/MODULE_f359722ac2fcd615a216+c2248236/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:c9328e5ec6c47a08605ba682901125ec76d723364f8fca10f81cb2ea8269cbdc
3
  size 84809
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f7fc4ab93d81a4a7b6feb4355dccc795e4cbb3883b168355dcc2949065b171be
3
  size 84809
neuronxcc-2.21.18209.0+043b1bf7/MODULE_f359722ac2fcd615a216+c2248236/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:7f6e231017b67f7a56c58f65de6b3cf098859bb41fd2dfbda4905f86ae3ab033
3
  size 656384
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c55dfa098e09e651f3227db4fd8285b14df9f436170232cb1cc332026762d919
3
  size 656384