Hugging Face's logo

optimum-internal-testing
/
neuron-testing-cache

Model card Files
xet
Community
dacorvo HF Staff commited on
Commit
cc4cf94
·
verified ·
1 Parent(s): 6fb6d15

Synchronizing local compiler cache.

Browse files
Files changed (43) hide show
  1. .gitattributes +12 -0
  2. neuronxcc-2.19.8089.0+8ab9f450/0_REGISTRY/0.3.1.dev5/qwen2/Qwen/Qwen2.5-0.5B-Instruct/b82f205ce639617b2486.json +82 -0
  3. neuronxcc-2.19.8089.0+8ab9f450/0_REGISTRY/0.3.1.dev5/qwen2/Qwen/Qwen2.5-0.5B/149443e088fd21e13be1.json +83 -0
  4. neuronxcc-2.19.8089.0+8ab9f450/0_REGISTRY/0.3.1.dev5/qwen2/Qwen/Qwen2.5-0.5B/69772d57962d7e14e1de.json +83 -0
  5. neuronxcc-2.19.8089.0+8ab9f450/0_REGISTRY/0.3.1.dev5/qwen2/Qwen/Qwen2.5-0.5B/f672420258867c52f0e9.json +83 -0
  6. neuronxcc-2.19.8089.0+8ab9f450/0_REGISTRY/0.3.1.dev6/qwen2/Qwen/Qwen2.5-0.5B/3e1662044ff5af476eb8.json +83 -0
  7. neuronxcc-2.19.8089.0+8ab9f450/MODULE_172c08cd3ccbf1c120d5+ed72d204/compile_flags.json +1 -0
  8. neuronxcc-2.19.8089.0+8ab9f450/MODULE_172c08cd3ccbf1c120d5+ed72d204/model.done +0 -0
  9. neuronxcc-2.19.8089.0+8ab9f450/MODULE_172c08cd3ccbf1c120d5+ed72d204/model.hlo_module.pb +3 -0
  10. neuronxcc-2.19.8089.0+8ab9f450/MODULE_172c08cd3ccbf1c120d5+ed72d204/model.neff +3 -0
  11. neuronxcc-2.19.8089.0+8ab9f450/MODULE_20b86a711b6ece0c7356+a9d440f5/compile_flags.json +1 -0
  12. neuronxcc-2.19.8089.0+8ab9f450/MODULE_20b86a711b6ece0c7356+a9d440f5/model.done +0 -0
  13. neuronxcc-2.19.8089.0+8ab9f450/MODULE_20b86a711b6ece0c7356+a9d440f5/model.hlo_module.pb +3 -0
  14. neuronxcc-2.19.8089.0+8ab9f450/MODULE_20b86a711b6ece0c7356+a9d440f5/model.neff +3 -0
  15. neuronxcc-2.19.8089.0+8ab9f450/MODULE_20b86a711b6ece0c7356+a9d440f5/wrapped_neff.hlo +3 -0
  16. neuronxcc-2.19.8089.0+8ab9f450/MODULE_304d7ac91c99b6581390+ed72d204/compile_flags.json +1 -0
  17. neuronxcc-2.19.8089.0+8ab9f450/MODULE_304d7ac91c99b6581390+ed72d204/model.done +0 -0
  18. neuronxcc-2.19.8089.0+8ab9f450/MODULE_304d7ac91c99b6581390+ed72d204/model.hlo_module.pb +3 -0
  19. neuronxcc-2.19.8089.0+8ab9f450/MODULE_304d7ac91c99b6581390+ed72d204/model.neff +3 -0
  20. neuronxcc-2.19.8089.0+8ab9f450/MODULE_90181ce17be0f8510277+ed72d204/compile_flags.json +1 -0
  21. neuronxcc-2.19.8089.0+8ab9f450/MODULE_90181ce17be0f8510277+ed72d204/model.done +0 -0
  22. neuronxcc-2.19.8089.0+8ab9f450/MODULE_90181ce17be0f8510277+ed72d204/model.hlo_module.pb +3 -0
  23. neuronxcc-2.19.8089.0+8ab9f450/MODULE_90181ce17be0f8510277+ed72d204/model.neff +3 -0
  24. neuronxcc-2.19.8089.0+8ab9f450/MODULE_974960a065697d53ebbb+ed72d204/compile_flags.json +1 -0
  25. neuronxcc-2.19.8089.0+8ab9f450/MODULE_974960a065697d53ebbb+ed72d204/model.done +0 -0
  26. neuronxcc-2.19.8089.0+8ab9f450/MODULE_974960a065697d53ebbb+ed72d204/model.hlo_module.pb +3 -0
  27. neuronxcc-2.19.8089.0+8ab9f450/MODULE_974960a065697d53ebbb+ed72d204/model.neff +3 -0
  28. neuronxcc-2.19.8089.0+8ab9f450/MODULE_ae633976c414e74f7634+ed72d204/model.hlo_module.pb +1 -1
  29. neuronxcc-2.19.8089.0+8ab9f450/MODULE_bcec2cc4ba44dbe255aa+a9d440f5/compile_flags.json +1 -0
  30. neuronxcc-2.19.8089.0+8ab9f450/MODULE_bcec2cc4ba44dbe255aa+a9d440f5/model.done +0 -0
  31. neuronxcc-2.19.8089.0+8ab9f450/MODULE_bcec2cc4ba44dbe255aa+a9d440f5/model.hlo_module.pb +3 -0
  32. neuronxcc-2.19.8089.0+8ab9f450/MODULE_bcec2cc4ba44dbe255aa+a9d440f5/model.neff +3 -0
  33. neuronxcc-2.19.8089.0+8ab9f450/MODULE_bcec2cc4ba44dbe255aa+a9d440f5/wrapped_neff.hlo +3 -0
  34. neuronxcc-2.19.8089.0+8ab9f450/MODULE_d379451a3eeeb9a21ff6+a9d440f5/compile_flags.json +1 -0
  35. neuronxcc-2.19.8089.0+8ab9f450/MODULE_d379451a3eeeb9a21ff6+a9d440f5/model.done +0 -0
  36. neuronxcc-2.19.8089.0+8ab9f450/MODULE_d379451a3eeeb9a21ff6+a9d440f5/model.hlo_module.pb +3 -0
  37. neuronxcc-2.19.8089.0+8ab9f450/MODULE_d379451a3eeeb9a21ff6+a9d440f5/model.neff +3 -0
  38. neuronxcc-2.19.8089.0+8ab9f450/MODULE_d379451a3eeeb9a21ff6+a9d440f5/wrapped_neff.hlo +3 -0
  39. neuronxcc-2.19.8089.0+8ab9f450/MODULE_d7030396f798ba07e1a9+a9d440f5/compile_flags.json +1 -0
  40. neuronxcc-2.19.8089.0+8ab9f450/MODULE_d7030396f798ba07e1a9+a9d440f5/model.done +0 -0
  41. neuronxcc-2.19.8089.0+8ab9f450/MODULE_d7030396f798ba07e1a9+a9d440f5/model.hlo_module.pb +3 -0
  42. neuronxcc-2.19.8089.0+8ab9f450/MODULE_d7030396f798ba07e1a9+a9d440f5/model.neff +3 -0
  43. neuronxcc-2.19.8089.0+8ab9f450/MODULE_d7030396f798ba07e1a9+a9d440f5/wrapped_neff.hlo +3 -0
.gitattributes CHANGED
@@ -4558,3 +4558,15 @@ neuronxcc-2.19.8089.0+8ab9f450/MODULE_efeb7f7f6d73497d3fc5+a9d440f5/model.neff f
4558
  neuronxcc-2.19.8089.0+8ab9f450/MODULE_efeb7f7f6d73497d3fc5+a9d440f5/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
4559
  neuronxcc-2.19.8089.0+8ab9f450/MODULE_f1ce9fc7b3c25b7b2459+ed72d204/model.neff filter=lfs diff=lfs merge=lfs -text
4560
  neuronxcc-2.19.8089.0+8ab9f450/MODULE_f25ef7f12c02216593d0+ed72d204/model.neff filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
 
 
 
 
4558
  neuronxcc-2.19.8089.0+8ab9f450/MODULE_efeb7f7f6d73497d3fc5+a9d440f5/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
4559
  neuronxcc-2.19.8089.0+8ab9f450/MODULE_f1ce9fc7b3c25b7b2459+ed72d204/model.neff filter=lfs diff=lfs merge=lfs -text
4560
  neuronxcc-2.19.8089.0+8ab9f450/MODULE_f25ef7f12c02216593d0+ed72d204/model.neff filter=lfs diff=lfs merge=lfs -text
4561
+ neuronxcc-2.19.8089.0+8ab9f450/MODULE_172c08cd3ccbf1c120d5+ed72d204/model.neff filter=lfs diff=lfs merge=lfs -text
4562
+ neuronxcc-2.19.8089.0+8ab9f450/MODULE_20b86a711b6ece0c7356+a9d440f5/model.neff filter=lfs diff=lfs merge=lfs -text
4563
+ neuronxcc-2.19.8089.0+8ab9f450/MODULE_20b86a711b6ece0c7356+a9d440f5/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
4564
+ neuronxcc-2.19.8089.0+8ab9f450/MODULE_304d7ac91c99b6581390+ed72d204/model.neff filter=lfs diff=lfs merge=lfs -text
4565
+ neuronxcc-2.19.8089.0+8ab9f450/MODULE_90181ce17be0f8510277+ed72d204/model.neff filter=lfs diff=lfs merge=lfs -text
4566
+ neuronxcc-2.19.8089.0+8ab9f450/MODULE_974960a065697d53ebbb+ed72d204/model.neff filter=lfs diff=lfs merge=lfs -text
4567
+ neuronxcc-2.19.8089.0+8ab9f450/MODULE_bcec2cc4ba44dbe255aa+a9d440f5/model.neff filter=lfs diff=lfs merge=lfs -text
4568
+ neuronxcc-2.19.8089.0+8ab9f450/MODULE_bcec2cc4ba44dbe255aa+a9d440f5/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
4569
+ neuronxcc-2.19.8089.0+8ab9f450/MODULE_d379451a3eeeb9a21ff6+a9d440f5/model.neff filter=lfs diff=lfs merge=lfs -text
4570
+ neuronxcc-2.19.8089.0+8ab9f450/MODULE_d379451a3eeeb9a21ff6+a9d440f5/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
4571
+ neuronxcc-2.19.8089.0+8ab9f450/MODULE_d7030396f798ba07e1a9+a9d440f5/model.neff filter=lfs diff=lfs merge=lfs -text
4572
+ neuronxcc-2.19.8089.0+8ab9f450/MODULE_d7030396f798ba07e1a9+a9d440f5/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
neuronxcc-2.19.8089.0+8ab9f450/0_REGISTRY/0.3.1.dev5/qwen2/Qwen/Qwen2.5-0.5B-Instruct/b82f205ce639617b2486.json ADDED
@@ -0,0 +1,82 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "Qwen/Qwen2.5-0.5B-Instruct",
4
+ "_task": "text-generation",
5
+ "architectures": [
6
+ "Qwen2ForCausalLM"
7
+ ],
8
+ "attention_dropout": 0.0,
9
+ "hidden_act": "silu",
10
+ "hidden_size": 896,
11
+ "initializer_range": 0.02,
12
+ "intermediate_size": 4864,
13
+ "layer_types": [
14
+ "full_attention",
15
+ "full_attention",
16
+ "full_attention",
17
+ "full_attention",
18
+ "full_attention",
19
+ "full_attention",
20
+ "full_attention",
21
+ "full_attention",
22
+ "full_attention",
23
+ "full_attention",
24
+ "full_attention",
25
+ "full_attention",
26
+ "full_attention",
27
+ "full_attention",
28
+ "full_attention",
29
+ "full_attention",
30
+ "full_attention",
31
+ "full_attention",
32
+ "full_attention",
33
+ "full_attention",
34
+ "full_attention",
35
+ "full_attention",
36
+ "full_attention",
37
+ "full_attention"
38
+ ],
39
+ "max_position_embeddings": 32768,
40
+ "max_window_layers": 21,
41
+ "model_type": "qwen2",
42
+ "neuron": {
43
+ "_serialized_key": "NxDNeuronConfig",
44
+ "batch_size": 1,
45
+ "capacity_factor": null,
46
+ "checkpoint_id": "Qwen/Qwen2.5-0.5B-Instruct",
47
+ "checkpoint_revision": "7ae557604adf67be50417f59c2c2f167def9a775",
48
+ "continuous_batching": false,
49
+ "enable_bucketing": false,
50
+ "ep_degree": 1,
51
+ "fused_qkv": false,
52
+ "glu_mlp": true,
53
+ "local_ranks_size": 24,
54
+ "logical_nc_config": 1,
55
+ "max_batch_size": 1,
56
+ "max_context_length": 4096,
57
+ "max_topk": 256,
58
+ "n_active_tokens": 4096,
59
+ "neuronxcc_version": "2.19.8089.0+8ab9f450",
60
+ "on_device_sampling": true,
61
+ "optimum_neuron_version": "0.3.1.dev5",
62
+ "output_logits": false,
63
+ "pp_degree": 1,
64
+ "sequence_length": 4096,
65
+ "speculation_length": 0,
66
+ "start_rank_id": 0,
67
+ "target": null,
68
+ "torch_dtype": "float32",
69
+ "tp_degree": 24
70
+ },
71
+ "num_attention_heads": 14,
72
+ "num_hidden_layers": 24,
73
+ "num_key_value_heads": 2,
74
+ "rms_norm_eps": 1e-06,
75
+ "rope_scaling": null,
76
+ "rope_theta": 1000000.0,
77
+ "sliding_window": null,
78
+ "tie_word_embeddings": true,
79
+ "use_cache": true,
80
+ "use_sliding_window": false,
81
+ "vocab_size": 151936
82
+ }
neuronxcc-2.19.8089.0+8ab9f450/0_REGISTRY/0.3.1.dev5/qwen2/Qwen/Qwen2.5-0.5B/149443e088fd21e13be1.json ADDED
@@ -0,0 +1,83 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "Qwen/Qwen2.5-0.5B",
4
+ "_task": "text-generation",
5
+ "architectures": [
6
+ "Qwen2ForCausalLM"
7
+ ],
8
+ "attention_dropout": 0.0,
9
+ "hidden_act": "silu",
10
+ "hidden_size": 896,
11
+ "initializer_range": 0.02,
12
+ "intermediate_size": 4864,
13
+ "layer_types": [
14
+ "full_attention",
15
+ "full_attention",
16
+ "full_attention",
17
+ "full_attention",
18
+ "full_attention",
19
+ "full_attention",
20
+ "full_attention",
21
+ "full_attention",
22
+ "full_attention",
23
+ "full_attention",
24
+ "full_attention",
25
+ "full_attention",
26
+ "full_attention",
27
+ "full_attention",
28
+ "full_attention",
29
+ "full_attention",
30
+ "full_attention",
31
+ "full_attention",
32
+ "full_attention",
33
+ "full_attention",
34
+ "full_attention",
35
+ "full_attention",
36
+ "full_attention",
37
+ "full_attention"
38
+ ],
39
+ "max_position_embeddings": 32768,
40
+ "max_window_layers": 24,
41
+ "model_type": "qwen2",
42
+ "neuron": {
43
+ "_serialized_key": "NxDNeuronConfig",
44
+ "batch_size": 2,
45
+ "capacity_factor": null,
46
+ "checkpoint_id": "Qwen/Qwen2.5-0.5B",
47
+ "checkpoint_revision": "060db6499f32faf8b98477b0a26969ef7d8b9987",
48
+ "continuous_batching": true,
49
+ "enable_bucketing": false,
50
+ "ep_degree": 1,
51
+ "fused_qkv": false,
52
+ "glu_mlp": true,
53
+ "local_ranks_size": 2,
54
+ "logical_nc_config": 1,
55
+ "max_batch_size": 2,
56
+ "max_context_length": 1024,
57
+ "max_topk": 256,
58
+ "n_active_tokens": 1024,
59
+ "neuronxcc_version": "2.19.8089.0+8ab9f450",
60
+ "on_device_sampling": false,
61
+ "optimum_neuron_version": "0.3.1.dev5",
62
+ "output_logits": false,
63
+ "pp_degree": 1,
64
+ "sequence_length": 1024,
65
+ "speculation_length": 0,
66
+ "start_rank_id": 0,
67
+ "target": null,
68
+ "torch_dtype": "float16",
69
+ "tp_degree": 2
70
+ },
71
+ "num_attention_heads": 14,
72
+ "num_hidden_layers": 24,
73
+ "num_key_value_heads": 2,
74
+ "rms_norm_eps": 1e-06,
75
+ "rope_scaling": null,
76
+ "rope_theta": 1000000.0,
77
+ "sliding_window": null,
78
+ "tie_word_embeddings": true,
79
+ "use_cache": true,
80
+ "use_mrope": false,
81
+ "use_sliding_window": false,
82
+ "vocab_size": 151936
83
+ }
neuronxcc-2.19.8089.0+8ab9f450/0_REGISTRY/0.3.1.dev5/qwen2/Qwen/Qwen2.5-0.5B/69772d57962d7e14e1de.json ADDED
@@ -0,0 +1,83 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "Qwen/Qwen2.5-0.5B",
4
+ "_task": "text-generation",
5
+ "architectures": [
6
+ "Qwen2ForCausalLM"
7
+ ],
8
+ "attention_dropout": 0.0,
9
+ "hidden_act": "silu",
10
+ "hidden_size": 896,
11
+ "initializer_range": 0.02,
12
+ "intermediate_size": 4864,
13
+ "layer_types": [
14
+ "full_attention",
15
+ "full_attention",
16
+ "full_attention",
17
+ "full_attention",
18
+ "full_attention",
19
+ "full_attention",
20
+ "full_attention",
21
+ "full_attention",
22
+ "full_attention",
23
+ "full_attention",
24
+ "full_attention",
25
+ "full_attention",
26
+ "full_attention",
27
+ "full_attention",
28
+ "full_attention",
29
+ "full_attention",
30
+ "full_attention",
31
+ "full_attention",
32
+ "full_attention",
33
+ "full_attention",
34
+ "full_attention",
35
+ "full_attention",
36
+ "full_attention",
37
+ "full_attention"
38
+ ],
39
+ "max_position_embeddings": 32768,
40
+ "max_window_layers": 24,
41
+ "model_type": "qwen2",
42
+ "neuron": {
43
+ "_serialized_key": "NxDNeuronConfig",
44
+ "batch_size": 2,
45
+ "capacity_factor": null,
46
+ "checkpoint_id": "Qwen/Qwen2.5-0.5B",
47
+ "checkpoint_revision": "060db6499f32faf8b98477b0a26969ef7d8b9987",
48
+ "continuous_batching": true,
49
+ "enable_bucketing": false,
50
+ "ep_degree": 1,
51
+ "fused_qkv": false,
52
+ "glu_mlp": true,
53
+ "local_ranks_size": 2,
54
+ "logical_nc_config": 1,
55
+ "max_batch_size": 2,
56
+ "max_context_length": 128,
57
+ "max_topk": 256,
58
+ "n_active_tokens": 128,
59
+ "neuronxcc_version": "2.19.8089.0+8ab9f450",
60
+ "on_device_sampling": false,
61
+ "optimum_neuron_version": "0.3.1.dev5",
62
+ "output_logits": false,
63
+ "pp_degree": 1,
64
+ "sequence_length": 128,
65
+ "speculation_length": 0,
66
+ "start_rank_id": 0,
67
+ "target": null,
68
+ "torch_dtype": "float16",
69
+ "tp_degree": 2
70
+ },
71
+ "num_attention_heads": 14,
72
+ "num_hidden_layers": 24,
73
+ "num_key_value_heads": 2,
74
+ "rms_norm_eps": 1e-06,
75
+ "rope_scaling": null,
76
+ "rope_theta": 1000000.0,
77
+ "sliding_window": null,
78
+ "tie_word_embeddings": true,
79
+ "use_cache": true,
80
+ "use_mrope": false,
81
+ "use_sliding_window": false,
82
+ "vocab_size": 151936
83
+ }
neuronxcc-2.19.8089.0+8ab9f450/0_REGISTRY/0.3.1.dev5/qwen2/Qwen/Qwen2.5-0.5B/f672420258867c52f0e9.json ADDED
@@ -0,0 +1,83 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "Qwen/Qwen2.5-0.5B",
4
+ "_task": "text-generation",
5
+ "architectures": [
6
+ "Qwen2ForCausalLM"
7
+ ],
8
+ "attention_dropout": 0.0,
9
+ "hidden_act": "silu",
10
+ "hidden_size": 896,
11
+ "initializer_range": 0.02,
12
+ "intermediate_size": 4864,
13
+ "layer_types": [
14
+ "full_attention",
15
+ "full_attention",
16
+ "full_attention",
17
+ "full_attention",
18
+ "full_attention",
19
+ "full_attention",
20
+ "full_attention",
21
+ "full_attention",
22
+ "full_attention",
23
+ "full_attention",
24
+ "full_attention",
25
+ "full_attention",
26
+ "full_attention",
27
+ "full_attention",
28
+ "full_attention",
29
+ "full_attention",
30
+ "full_attention",
31
+ "full_attention",
32
+ "full_attention",
33
+ "full_attention",
34
+ "full_attention",
35
+ "full_attention",
36
+ "full_attention",
37
+ "full_attention"
38
+ ],
39
+ "max_position_embeddings": 32768,
40
+ "max_window_layers": 24,
41
+ "model_type": "qwen2",
42
+ "neuron": {
43
+ "_serialized_key": "NxDNeuronConfig",
44
+ "batch_size": 1,
45
+ "capacity_factor": null,
46
+ "checkpoint_id": "Qwen/Qwen2.5-0.5B",
47
+ "checkpoint_revision": "060db6499f32faf8b98477b0a26969ef7d8b9987",
48
+ "continuous_batching": false,
49
+ "enable_bucketing": false,
50
+ "ep_degree": 1,
51
+ "fused_qkv": false,
52
+ "glu_mlp": true,
53
+ "local_ranks_size": 2,
54
+ "logical_nc_config": 1,
55
+ "max_batch_size": 1,
56
+ "max_context_length": 4096,
57
+ "max_topk": 256,
58
+ "n_active_tokens": 4096,
59
+ "neuronxcc_version": "2.19.8089.0+8ab9f450",
60
+ "on_device_sampling": true,
61
+ "optimum_neuron_version": "0.3.1.dev5",
62
+ "output_logits": false,
63
+ "pp_degree": 1,
64
+ "sequence_length": 4096,
65
+ "speculation_length": 0,
66
+ "start_rank_id": 0,
67
+ "target": null,
68
+ "torch_dtype": "float32",
69
+ "tp_degree": 2
70
+ },
71
+ "num_attention_heads": 14,
72
+ "num_hidden_layers": 24,
73
+ "num_key_value_heads": 2,
74
+ "rms_norm_eps": 1e-06,
75
+ "rope_scaling": null,
76
+ "rope_theta": 1000000.0,
77
+ "sliding_window": null,
78
+ "tie_word_embeddings": true,
79
+ "use_cache": true,
80
+ "use_mrope": false,
81
+ "use_sliding_window": false,
82
+ "vocab_size": 151936
83
+ }
neuronxcc-2.19.8089.0+8ab9f450/0_REGISTRY/0.3.1.dev6/qwen2/Qwen/Qwen2.5-0.5B/3e1662044ff5af476eb8.json ADDED
@@ -0,0 +1,83 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "Qwen/Qwen2.5-0.5B",
4
+ "_task": "text-generation",
5
+ "architectures": [
6
+ "Qwen2ForCausalLM"
7
+ ],
8
+ "attention_dropout": 0.0,
9
+ "hidden_act": "silu",
10
+ "hidden_size": 896,
11
+ "initializer_range": 0.02,
12
+ "intermediate_size": 4864,
13
+ "layer_types": [
14
+ "full_attention",
15
+ "full_attention",
16
+ "full_attention",
17
+ "full_attention",
18
+ "full_attention",
19
+ "full_attention",
20
+ "full_attention",
21
+ "full_attention",
22
+ "full_attention",
23
+ "full_attention",
24
+ "full_attention",
25
+ "full_attention",
26
+ "full_attention",
27
+ "full_attention",
28
+ "full_attention",
29
+ "full_attention",
30
+ "full_attention",
31
+ "full_attention",
32
+ "full_attention",
33
+ "full_attention",
34
+ "full_attention",
35
+ "full_attention",
36
+ "full_attention",
37
+ "full_attention"
38
+ ],
39
+ "max_position_embeddings": 32768,
40
+ "max_window_layers": 24,
41
+ "model_type": "qwen2",
42
+ "neuron": {
43
+ "_serialized_key": "NxDNeuronConfig",
44
+ "batch_size": 1,
45
+ "capacity_factor": null,
46
+ "checkpoint_id": "Qwen/Qwen2.5-0.5B",
47
+ "checkpoint_revision": "060db6499f32faf8b98477b0a26969ef7d8b9987",
48
+ "continuous_batching": false,
49
+ "enable_bucketing": false,
50
+ "ep_degree": 1,
51
+ "fused_qkv": false,
52
+ "glu_mlp": true,
53
+ "local_ranks_size": 2,
54
+ "logical_nc_config": 1,
55
+ "max_batch_size": 1,
56
+ "max_context_length": 4096,
57
+ "max_topk": 256,
58
+ "n_active_tokens": 4096,
59
+ "neuronxcc_version": "2.19.8089.0+8ab9f450",
60
+ "on_device_sampling": true,
61
+ "optimum_neuron_version": "0.3.1.dev6",
62
+ "output_logits": false,
63
+ "pp_degree": 1,
64
+ "sequence_length": 4096,
65
+ "speculation_length": 0,
66
+ "start_rank_id": 0,
67
+ "target": null,
68
+ "torch_dtype": "bfloat16",
69
+ "tp_degree": 2
70
+ },
71
+ "num_attention_heads": 14,
72
+ "num_hidden_layers": 24,
73
+ "num_key_value_heads": 2,
74
+ "rms_norm_eps": 1e-06,
75
+ "rope_scaling": null,
76
+ "rope_theta": 1000000.0,
77
+ "sliding_window": null,
78
+ "tie_word_embeddings": true,
79
+ "use_cache": true,
80
+ "use_mrope": false,
81
+ "use_sliding_window": false,
82
+ "vocab_size": 151936
83
+ }
neuronxcc-2.19.8089.0+8ab9f450/MODULE_172c08cd3ccbf1c120d5+ed72d204/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_model/_tp0_bk0/log-neuron-cc.txt"]
neuronxcc-2.19.8089.0+8ab9f450/MODULE_172c08cd3ccbf1c120d5+ed72d204/model.done ADDED
File without changes
neuronxcc-2.19.8089.0+8ab9f450/MODULE_172c08cd3ccbf1c120d5+ed72d204/model.hlo_module.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:99a6eb22f0b96f5eed0d5990710ddf13d86bc77a63c7b69828d7056ee66fb0cb
3
+ size 498526
neuronxcc-2.19.8089.0+8ab9f450/MODULE_172c08cd3ccbf1c120d5+ed72d204/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f992ce1ac61902555c28228adf7e398266d423bb9a7bf80e89e5529ed4b9bb7b
3
+ size 4250624
neuronxcc-2.19.8089.0+8ab9f450/MODULE_20b86a711b6ece0c7356+a9d440f5/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_model/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]
neuronxcc-2.19.8089.0+8ab9f450/MODULE_20b86a711b6ece0c7356+a9d440f5/model.done ADDED
File without changes
neuronxcc-2.19.8089.0+8ab9f450/MODULE_20b86a711b6ece0c7356+a9d440f5/model.hlo_module.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fe33969857c35b10d949a92acee4f59fd0a6b47fea748c3c54e7eeb8a3953d6e
3
+ size 573043
neuronxcc-2.19.8089.0+8ab9f450/MODULE_20b86a711b6ece0c7356+a9d440f5/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:81a9c2de0feb308d68e3f6223c99901a93c1c18b245887597848193718bb0ee8
3
+ size 1444864
neuronxcc-2.19.8089.0+8ab9f450/MODULE_20b86a711b6ece0c7356+a9d440f5/wrapped_neff.hlo ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8f21f00ed2b5e9d2659f12c7ae9501d7306ae67f3ae20e37c1149ca452b179da
3
+ size 1589890
neuronxcc-2.19.8089.0+8ab9f450/MODULE_304d7ac91c99b6581390+ed72d204/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_model/_tp0_bk0/log-neuron-cc.txt"]
neuronxcc-2.19.8089.0+8ab9f450/MODULE_304d7ac91c99b6581390+ed72d204/model.done ADDED
File without changes
neuronxcc-2.19.8089.0+8ab9f450/MODULE_304d7ac91c99b6581390+ed72d204/model.hlo_module.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c8b48addb1bc3d932e8c31773aad6ad55403341cbb47e87889c773528ac3a0c7
3
+ size 581174
neuronxcc-2.19.8089.0+8ab9f450/MODULE_304d7ac91c99b6581390+ed72d204/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1709a073dd962ffd51f1deb4073bf181d321d8f33a8527cac577e57dbce4031e
3
+ size 2724864
neuronxcc-2.19.8089.0+8ab9f450/MODULE_90181ce17be0f8510277+ed72d204/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_model/_tp0_bk0/log-neuron-cc.txt"]
neuronxcc-2.19.8089.0+8ab9f450/MODULE_90181ce17be0f8510277+ed72d204/model.done ADDED
File without changes
neuronxcc-2.19.8089.0+8ab9f450/MODULE_90181ce17be0f8510277+ed72d204/model.hlo_module.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:99a16435815f55018f92dc9158264d64baf957ff4ccf27672582a0afbda8eddb
3
+ size 581174
neuronxcc-2.19.8089.0+8ab9f450/MODULE_90181ce17be0f8510277+ed72d204/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f7ae6d3db9f395d9e5cd74a6fd5a96bc23c20049e829399934444e7d447a3d2a
3
+ size 1117184
neuronxcc-2.19.8089.0+8ab9f450/MODULE_974960a065697d53ebbb+ed72d204/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_model/_tp0_bk0/log-neuron-cc.txt"]
neuronxcc-2.19.8089.0+8ab9f450/MODULE_974960a065697d53ebbb+ed72d204/model.done ADDED
File without changes
neuronxcc-2.19.8089.0+8ab9f450/MODULE_974960a065697d53ebbb+ed72d204/model.hlo_module.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:80f6fbf505ffbb78c6468a746a6a6e231ec29ab87746a41d3578dabf2c74d649
3
+ size 523884
neuronxcc-2.19.8089.0+8ab9f450/MODULE_974960a065697d53ebbb+ed72d204/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:12c12be1cc0eac518f0ba9ce2f3fa89541527a5c43800b051bde5e7da0c2c7a3
3
+ size 17798144
neuronxcc-2.19.8089.0+8ab9f450/MODULE_ae633976c414e74f7634+ed72d204/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:58b78643a150fb95e6783be77452d355a39021ef66872e666b849586b20f155d
3
  size 529164
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:32b41832549a6bc6b6ac34d56c008f42bee1769ebb42280684d6a18d4b14ab0d
3
  size 529164
neuronxcc-2.19.8089.0+8ab9f450/MODULE_bcec2cc4ba44dbe255aa+a9d440f5/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_model/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]
neuronxcc-2.19.8089.0+8ab9f450/MODULE_bcec2cc4ba44dbe255aa+a9d440f5/model.done ADDED
File without changes
neuronxcc-2.19.8089.0+8ab9f450/MODULE_bcec2cc4ba44dbe255aa+a9d440f5/model.hlo_module.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0340ea9c9f3757f6701e2467620e00001e3bac281a5067da7491f5debf45c9a2
3
+ size 514401
neuronxcc-2.19.8089.0+8ab9f450/MODULE_bcec2cc4ba44dbe255aa+a9d440f5/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d3dc15944d3e940ac9ba6a8536562a42eb50b3233c85ebd9bc81764817c53c2e
3
+ size 646144
neuronxcc-2.19.8089.0+8ab9f450/MODULE_bcec2cc4ba44dbe255aa+a9d440f5/wrapped_neff.hlo ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:83b38f7cb10576814869756d20f8e829cd01d2410c19c9926989559a9ba13f15
3
+ size 780434
neuronxcc-2.19.8089.0+8ab9f450/MODULE_d379451a3eeeb9a21ff6+a9d440f5/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_model/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]
neuronxcc-2.19.8089.0+8ab9f450/MODULE_d379451a3eeeb9a21ff6+a9d440f5/model.done ADDED
File without changes
neuronxcc-2.19.8089.0+8ab9f450/MODULE_d379451a3eeeb9a21ff6+a9d440f5/model.hlo_module.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0ca069c4b82b5ec5bcda4470d4248f42760f3265cc8a8cf28d9e750e4bef5bc6
3
+ size 582901
neuronxcc-2.19.8089.0+8ab9f450/MODULE_d379451a3eeeb9a21ff6+a9d440f5/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:56f42d47d44d96481748c731c4e3f74b21cc4c775d32b0cfcfe9e805134f247b
3
+ size 1147904
neuronxcc-2.19.8089.0+8ab9f450/MODULE_d379451a3eeeb9a21ff6+a9d440f5/wrapped_neff.hlo ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8f61e4840df55936d5ba716f891116f24852d16aa6e85bd285e7b4b99a9a9037
3
+ size 1292815
neuronxcc-2.19.8089.0+8ab9f450/MODULE_d7030396f798ba07e1a9+a9d440f5/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_model/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]
neuronxcc-2.19.8089.0+8ab9f450/MODULE_d7030396f798ba07e1a9+a9d440f5/model.done ADDED
File without changes
neuronxcc-2.19.8089.0+8ab9f450/MODULE_d7030396f798ba07e1a9+a9d440f5/model.hlo_module.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e40dee86f8546d2afe31775a13de5f73b8c014b99a04ee6708f21e668715008b
3
+ size 582901
neuronxcc-2.19.8089.0+8ab9f450/MODULE_d7030396f798ba07e1a9+a9d440f5/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bed06cae0ac580b3e5cf99225769ab97a2b47611d3b1eca03dde45d3ff1f89f7
3
+ size 1209344
neuronxcc-2.19.8089.0+8ab9f450/MODULE_d7030396f798ba07e1a9+a9d440f5/wrapped_neff.hlo ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c63adabb8457539ec6b44286e7dbddc2c339f70531eec4653ed09612ed3e55ad
3
+ size 1354255