Hugging Face's logo

aws-neuron
/
optimum-neuron-cache

Model card Files Community
276
dacorvo HF staff commited on
Commit
4769999
•
1 Parent(s): c303216

Synchronizing local compiler cache.

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +16 -0
  2. neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.19/inference/llama/meta-llama/Llama-2-13b-chat-hf/63eb7e01e61523a14109.json +1 -0
  3. neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.19/inference/llama/meta-llama/Llama-2-13b-chat-hf/8772702e5f5223f7de94.json +1 -0
  4. neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.19/inference/llama/meta-llama/Llama-2-13b-chat-hf/d4dddb49c677c531dcd7.json +1 -0
  5. neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.19/inference/llama/meta-llama/Llama-2-13b-chat-hf/f40b584c97bffabc60c2.json +1 -0
  6. neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.19/inference/llama/meta-llama/Llama-2-7b-chat-hf/93dc703e497680b4df4d.json +1 -0
  7. neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.19/inference/llama/meta-llama/Llama-2-7b-chat-hf/b8a63e77583d9b7004b6.json +1 -0
  8. neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.20.dev0/inference/llama/princeton-nlp/Sheared-LLaMA-1.3B/624ef8314775a5c7b63b.json +1 -0
  9. neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.20.dev0/inference/llama/princeton-nlp/Sheared-LLaMA-1.3B/7700052f85791490f5aa.json +1 -0
  10. neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.20.dev0/inference/llama/princeton-nlp/Sheared-LLaMA-1.3B/a737859203da7a2e914f.json +1 -0
  11. neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.20.dev0/inference/llama/princeton-nlp/Sheared-LLaMA-1.3B/f1c71b95ef4e98e06b6a.json +1 -0
  12. neuronxcc-2.12.68.0+4480452af/MODULE_1154b678e32b38df568d+2c2d707e/compile_flags.txt +1 -0
  13. neuronxcc-2.12.68.0+4480452af/MODULE_1154b678e32b38df568d+2c2d707e/model.hlo.pb +3 -0
  14. neuronxcc-2.12.68.0+4480452af/MODULE_1154b678e32b38df568d+2c2d707e/model.neff +3 -0
  15. neuronxcc-2.12.68.0+4480452af/MODULE_13c86804904d0329bc38+2c2d707e/compile_flags.txt +1 -0
  16. neuronxcc-2.12.68.0+4480452af/MODULE_13c86804904d0329bc38+2c2d707e/model.hlo.pb +3 -0
  17. neuronxcc-2.12.68.0+4480452af/MODULE_13c86804904d0329bc38+2c2d707e/model.neff +3 -0
  18. neuronxcc-2.12.68.0+4480452af/MODULE_14fd59d8fbe618033e23+2c2d707e/compile_flags.txt +1 -0
  19. neuronxcc-2.12.68.0+4480452af/MODULE_14fd59d8fbe618033e23+2c2d707e/model.hlo.pb +3 -0
  20. neuronxcc-2.12.68.0+4480452af/MODULE_14fd59d8fbe618033e23+2c2d707e/model.neff +3 -0
  21. neuronxcc-2.12.68.0+4480452af/MODULE_1f4c8e3481cc0ab2ecfe+2c2d707e/compile_flags.txt +1 -0
  22. neuronxcc-2.12.68.0+4480452af/MODULE_1f4c8e3481cc0ab2ecfe+2c2d707e/model.hlo.pb +3 -0
  23. neuronxcc-2.12.68.0+4480452af/MODULE_1f4c8e3481cc0ab2ecfe+2c2d707e/model.neff +3 -0
  24. neuronxcc-2.12.68.0+4480452af/MODULE_22850a3c370343c4e87c+2c2d707e/compile_flags.txt +1 -0
  25. neuronxcc-2.12.68.0+4480452af/MODULE_22850a3c370343c4e87c+2c2d707e/model.hlo.pb +3 -0
  26. neuronxcc-2.12.68.0+4480452af/MODULE_22850a3c370343c4e87c+2c2d707e/model.neff +3 -0
  27. neuronxcc-2.12.68.0+4480452af/MODULE_23be14b918bd101e0307+2c2d707e/compile_flags.txt +1 -0
  28. neuronxcc-2.12.68.0+4480452af/MODULE_23be14b918bd101e0307+2c2d707e/model.hlo.pb +3 -0
  29. neuronxcc-2.12.68.0+4480452af/MODULE_23be14b918bd101e0307+2c2d707e/model.neff +3 -0
  30. neuronxcc-2.12.68.0+4480452af/MODULE_439cb9059d1a4720adfb+2c2d707e/compile_flags.txt +1 -0
  31. neuronxcc-2.12.68.0+4480452af/MODULE_439cb9059d1a4720adfb+2c2d707e/model.hlo.pb +3 -0
  32. neuronxcc-2.12.68.0+4480452af/MODULE_439cb9059d1a4720adfb+2c2d707e/model.neff +3 -0
  33. neuronxcc-2.12.68.0+4480452af/MODULE_441b50063b845da3c878+2c2d707e/compile_flags.txt +1 -0
  34. neuronxcc-2.12.68.0+4480452af/MODULE_441b50063b845da3c878+2c2d707e/model.hlo.pb +3 -0
  35. neuronxcc-2.12.68.0+4480452af/MODULE_441b50063b845da3c878+2c2d707e/model.neff +3 -0
  36. neuronxcc-2.12.68.0+4480452af/MODULE_62ebf9bdd149ff8c97d9+2c2d707e/compile_flags.txt +1 -0
  37. neuronxcc-2.12.68.0+4480452af/MODULE_62ebf9bdd149ff8c97d9+2c2d707e/model.hlo.pb +3 -0
  38. neuronxcc-2.12.68.0+4480452af/MODULE_62ebf9bdd149ff8c97d9+2c2d707e/model.neff +3 -0
  39. neuronxcc-2.12.68.0+4480452af/MODULE_9278928fd84a21a91a59+2c2d707e/compile_flags.txt +1 -0
  40. neuronxcc-2.12.68.0+4480452af/MODULE_9278928fd84a21a91a59+2c2d707e/model.hlo.pb +3 -0
  41. neuronxcc-2.12.68.0+4480452af/MODULE_9b8a4b44cab7fd6a36b6+2c2d707e/compile_flags.txt +1 -0
  42. neuronxcc-2.12.68.0+4480452af/MODULE_9b8a4b44cab7fd6a36b6+2c2d707e/model.hlo.pb +3 -0
  43. neuronxcc-2.12.68.0+4480452af/MODULE_9b8a4b44cab7fd6a36b6+2c2d707e/model.neff +3 -0
  44. neuronxcc-2.12.68.0+4480452af/MODULE_bd284532da36dbde1d9d+2c2d707e/compile_flags.txt +1 -0
  45. neuronxcc-2.12.68.0+4480452af/MODULE_bd284532da36dbde1d9d+2c2d707e/model.hlo.pb +3 -0
  46. neuronxcc-2.12.68.0+4480452af/MODULE_bd284532da36dbde1d9d+2c2d707e/model.neff +3 -0
  47. neuronxcc-2.12.68.0+4480452af/MODULE_dd3fcd08d9d80a1eff7d+2c2d707e/compile_flags.txt +1 -0
  48. neuronxcc-2.12.68.0+4480452af/MODULE_dd3fcd08d9d80a1eff7d+2c2d707e/model.hlo.pb +3 -0
  49. neuronxcc-2.12.68.0+4480452af/MODULE_dd3fcd08d9d80a1eff7d+2c2d707e/model.neff +3 -0
  50. neuronxcc-2.12.68.0+4480452af/MODULE_e342da015657e3ad5e30+2c2d707e/compile_flags.txt +1 -0
.gitattributes CHANGED
@@ -1119,3 +1119,19 @@ neuronxcc-2.12.54.0+f631c2365/MODULE_96c74602c6595c924750+2c2d707e/model.neff fi
1119
  neuronxcc-2.12.54.0+f631c2365/MODULE_b731fc0c29ae932dea47+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1120
  neuronxcc-2.12.54.0+f631c2365/MODULE_9c7a263166e0f6950fe0+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1121
  neuronxcc-2.12.54.0+f631c2365/MODULE_b0880fb3960d07f95631+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1119
  neuronxcc-2.12.54.0+f631c2365/MODULE_b731fc0c29ae932dea47+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1120
  neuronxcc-2.12.54.0+f631c2365/MODULE_9c7a263166e0f6950fe0+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1121
  neuronxcc-2.12.54.0+f631c2365/MODULE_b0880fb3960d07f95631+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1122
+ neuronxcc-2.12.68.0+4480452af/MODULE_1154b678e32b38df568d+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1123
+ neuronxcc-2.12.68.0+4480452af/MODULE_13c86804904d0329bc38+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1124
+ neuronxcc-2.12.68.0+4480452af/MODULE_14fd59d8fbe618033e23+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1125
+ neuronxcc-2.12.68.0+4480452af/MODULE_1f4c8e3481cc0ab2ecfe+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1126
+ neuronxcc-2.12.68.0+4480452af/MODULE_22850a3c370343c4e87c+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1127
+ neuronxcc-2.12.68.0+4480452af/MODULE_23be14b918bd101e0307+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1128
+ neuronxcc-2.12.68.0+4480452af/MODULE_439cb9059d1a4720adfb+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1129
+ neuronxcc-2.12.68.0+4480452af/MODULE_441b50063b845da3c878+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1130
+ neuronxcc-2.12.68.0+4480452af/MODULE_62ebf9bdd149ff8c97d9+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1131
+ neuronxcc-2.12.68.0+4480452af/MODULE_9b8a4b44cab7fd6a36b6+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1132
+ neuronxcc-2.12.68.0+4480452af/MODULE_bd284532da36dbde1d9d+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1133
+ neuronxcc-2.12.68.0+4480452af/MODULE_dd3fcd08d9d80a1eff7d+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1134
+ neuronxcc-2.12.68.0+4480452af/MODULE_e342da015657e3ad5e30+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1135
+ neuronxcc-2.12.68.0+4480452af/MODULE_e6d595b44e212a97a91e+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1136
+ neuronxcc-2.12.68.0+4480452af/MODULE_eb5999be2069b95cacd6+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
1137
+ neuronxcc-2.12.68.0+4480452af/MODULE_f0ced8f5974d435da438+2c2d707e/model.neff filter=lfs diff=lfs merge=lfs -text
neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.19/inference/llama/meta-llama/Llama-2-13b-chat-hf/63eb7e01e61523a14109.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"vocab_size": 32000, "max_position_embeddings": 4096, "hidden_size": 5120, "intermediate_size": 13824, "num_hidden_layers": 40, "num_attention_heads": 40, "num_key_value_heads": 40, "hidden_act": "silu", "initializer_range": 0.02, "rms_norm_eps": 1e-05, "pretraining_tp": 1, "use_cache": true, "rope_theta": 10000.0, "rope_scaling": null, "attention_bias": false, "attention_dropout": 0.0, "torch_dtype": "float16", "tie_word_embeddings": false, "architectures": ["LlamaForCausalLM"], "bos_token_id": 1, "eos_token_id": 2, "model_type": "llama", "neuron": {"task": "text-generation", "batch_size": 16, "num_cores": 24, "auto_cast_type": "fp16", "sequence_length": 4096, "compiler_type": "neuronx-cc", "compiler_version": "2.12.68.0+4480452af", "checkpoint_id": "meta-llama/Llama-2-13b-chat-hf", "checkpoint_revision": "c2f3ec81aac798ae26dcc57799a994dfbf521496"}}
neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.19/inference/llama/meta-llama/Llama-2-13b-chat-hf/8772702e5f5223f7de94.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"vocab_size": 32000, "max_position_embeddings": 4096, "hidden_size": 5120, "intermediate_size": 13824, "num_hidden_layers": 40, "num_attention_heads": 40, "num_key_value_heads": 40, "hidden_act": "silu", "initializer_range": 0.02, "rms_norm_eps": 1e-05, "pretraining_tp": 1, "use_cache": true, "rope_theta": 10000.0, "rope_scaling": null, "attention_bias": false, "attention_dropout": 0.0, "torch_dtype": "float16", "tie_word_embeddings": false, "architectures": ["LlamaForCausalLM"], "bos_token_id": 1, "eos_token_id": 2, "model_type": "llama", "neuron": {"auto_cast_type": "fp16", "batch_size": 4, "checkpoint_id": "meta-llama/Llama-2-13b-chat-hf", "checkpoint_revision": "c2f3ec81aac798ae26dcc57799a994dfbf521496", "compiler_type": "neuronx-cc", "compiler_version": "2.12.68.0+4480452af", "num_cores": 24, "sequence_length": 4096, "task": "text-generation"}}
neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.19/inference/llama/meta-llama/Llama-2-13b-chat-hf/d4dddb49c677c531dcd7.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"vocab_size": 32000, "max_position_embeddings": 4096, "hidden_size": 5120, "intermediate_size": 13824, "num_hidden_layers": 40, "num_attention_heads": 40, "num_key_value_heads": 40, "hidden_act": "silu", "initializer_range": 0.02, "rms_norm_eps": 1e-05, "pretraining_tp": 1, "use_cache": true, "rope_theta": 10000.0, "rope_scaling": null, "attention_bias": false, "attention_dropout": 0.0, "torch_dtype": "float16", "tie_word_embeddings": false, "architectures": ["LlamaForCausalLM"], "bos_token_id": 1, "eos_token_id": 2, "model_type": "llama", "neuron": {"auto_cast_type": "fp16", "batch_size": 1, "checkpoint_id": "meta-llama/Llama-2-13b-chat-hf", "checkpoint_revision": "c2f3ec81aac798ae26dcc57799a994dfbf521496", "compiler_type": "neuronx-cc", "compiler_version": "2.12.68.0+4480452af", "num_cores": 24, "sequence_length": 4096, "task": "text-generation"}}
neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.19/inference/llama/meta-llama/Llama-2-13b-chat-hf/f40b584c97bffabc60c2.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"vocab_size": 32000, "max_position_embeddings": 4096, "hidden_size": 5120, "intermediate_size": 13824, "num_hidden_layers": 40, "num_attention_heads": 40, "num_key_value_heads": 40, "hidden_act": "silu", "initializer_range": 0.02, "rms_norm_eps": 1e-05, "pretraining_tp": 1, "use_cache": true, "rope_theta": 10000.0, "rope_scaling": null, "attention_bias": false, "attention_dropout": 0.0, "torch_dtype": "float16", "tie_word_embeddings": false, "architectures": ["LlamaForCausalLM"], "bos_token_id": 1, "eos_token_id": 2, "model_type": "llama", "neuron": {"auto_cast_type": "fp16", "batch_size": 16, "checkpoint_id": "meta-llama/Llama-2-13b-chat-hf", "checkpoint_revision": "c2f3ec81aac798ae26dcc57799a994dfbf521496", "compiler_type": "neuronx-cc", "compiler_version": "2.12.68.0+4480452af", "num_cores": 24, "sequence_length": 4096, "task": "text-generation"}}
neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.19/inference/llama/meta-llama/Llama-2-7b-chat-hf/93dc703e497680b4df4d.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"vocab_size": 32000, "max_position_embeddings": 4096, "hidden_size": 4096, "intermediate_size": 11008, "num_hidden_layers": 32, "num_attention_heads": 32, "num_key_value_heads": 32, "hidden_act": "silu", "initializer_range": 0.02, "rms_norm_eps": 1e-05, "pretraining_tp": 1, "use_cache": true, "rope_theta": 10000.0, "rope_scaling": null, "attention_bias": false, "attention_dropout": 0.0, "torch_dtype": "float16", "tie_word_embeddings": false, "architectures": ["LlamaForCausalLM"], "bos_token_id": 1, "eos_token_id": 2, "model_type": "llama", "neuron": {"auto_cast_type": "fp16", "batch_size": 4, "checkpoint_id": "meta-llama/Llama-2-7b-chat-hf", "checkpoint_revision": "c1b0db933684edbfe29a06fa47eb19cc48025e93", "compiler_type": "neuronx-cc", "compiler_version": "2.12.68.0+4480452af", "num_cores": 24, "sequence_length": 4096, "task": "text-generation"}}
neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.19/inference/llama/meta-llama/Llama-2-7b-chat-hf/b8a63e77583d9b7004b6.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"vocab_size": 32000, "max_position_embeddings": 4096, "hidden_size": 4096, "intermediate_size": 11008, "num_hidden_layers": 32, "num_attention_heads": 32, "num_key_value_heads": 32, "hidden_act": "silu", "initializer_range": 0.02, "rms_norm_eps": 1e-05, "pretraining_tp": 1, "use_cache": true, "rope_theta": 10000.0, "rope_scaling": null, "attention_bias": false, "attention_dropout": 0.0, "torch_dtype": "float16", "tie_word_embeddings": false, "architectures": ["LlamaForCausalLM"], "bos_token_id": 1, "eos_token_id": 2, "model_type": "llama", "neuron": {"auto_cast_type": "fp16", "batch_size": 1, "checkpoint_id": "meta-llama/Llama-2-7b-chat-hf", "checkpoint_revision": "c1b0db933684edbfe29a06fa47eb19cc48025e93", "compiler_type": "neuronx-cc", "compiler_version": "2.12.68.0+4480452af", "num_cores": 24, "sequence_length": 4096, "task": "text-generation"}}
neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.20.dev0/inference/llama/princeton-nlp/Sheared-LLaMA-1.3B/624ef8314775a5c7b63b.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"vocab_size": 32000, "max_position_embeddings": 4096, "hidden_size": 2048, "intermediate_size": 5504, "num_hidden_layers": 24, "num_attention_heads": 16, "num_key_value_heads": 16, "hidden_act": "silu", "initializer_range": 0.02, "rms_norm_eps": 1e-05, "pretraining_tp": 1, "use_cache": true, "rope_theta": 10000.0, "rope_scaling": null, "attention_bias": false, "attention_dropout": 0.0, "torch_dtype": "float32", "tie_word_embeddings": false, "architectures": ["LlamaForCausalLM"], "bos_token_id": 1, "pad_token_id": 0, "eos_token_id": 2, "model_type": "llama", "neuron": {"auto_cast_type": "fp32", "batch_size": 4, "checkpoint_id": "princeton-nlp/Sheared-LLaMA-1.3B", "checkpoint_revision": "a4b76938edbf571ea7d7d9904861cbdca08809b4", "compiler_type": "neuronx-cc", "compiler_version": "2.12.68.0+4480452af", "num_cores": 2, "sequence_length": 1024, "task": "text-generation"}}
neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.20.dev0/inference/llama/princeton-nlp/Sheared-LLaMA-1.3B/7700052f85791490f5aa.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"vocab_size": 32000, "max_position_embeddings": 4096, "hidden_size": 2048, "intermediate_size": 5504, "num_hidden_layers": 24, "num_attention_heads": 16, "num_key_value_heads": 16, "hidden_act": "silu", "initializer_range": 0.02, "rms_norm_eps": 1e-05, "pretraining_tp": 1, "use_cache": true, "rope_theta": 10000.0, "rope_scaling": null, "attention_bias": false, "attention_dropout": 0.0, "torch_dtype": "float32", "tie_word_embeddings": false, "architectures": ["LlamaForCausalLM"], "bos_token_id": 1, "pad_token_id": 0, "eos_token_id": 2, "model_type": "llama", "neuron": {"task": "text-generation", "batch_size": 1, "num_cores": 2, "auto_cast_type": "bf16", "sequence_length": 1024, "compiler_type": "neuronx-cc", "compiler_version": "2.12.68.0+4480452af", "checkpoint_id": "princeton-nlp/Sheared-LLaMA-1.3B", "checkpoint_revision": "a4b76938edbf571ea7d7d9904861cbdca08809b4"}}
neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.20.dev0/inference/llama/princeton-nlp/Sheared-LLaMA-1.3B/a737859203da7a2e914f.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"vocab_size": 32000, "max_position_embeddings": 4096, "hidden_size": 2048, "intermediate_size": 5504, "num_hidden_layers": 24, "num_attention_heads": 16, "num_key_value_heads": 16, "hidden_act": "silu", "initializer_range": 0.02, "rms_norm_eps": 1e-05, "pretraining_tp": 1, "use_cache": true, "rope_theta": 10000.0, "rope_scaling": null, "attention_bias": false, "attention_dropout": 0.0, "torch_dtype": "float32", "tie_word_embeddings": false, "architectures": ["LlamaForCausalLM"], "bos_token_id": 1, "pad_token_id": 0, "eos_token_id": 2, "model_type": "llama", "neuron": {"task": "text-generation", "batch_size": 4, "num_cores": 2, "auto_cast_type": "fp16", "sequence_length": 1024, "compiler_type": "neuronx-cc", "compiler_version": "2.12.68.0+4480452af", "checkpoint_id": "princeton-nlp/Sheared-LLaMA-1.3B", "checkpoint_revision": "a4b76938edbf571ea7d7d9904861cbdca08809b4"}}
neuronxcc-2.12.68.0+4480452af/0_REGISTRY/0.0.20.dev0/inference/llama/princeton-nlp/Sheared-LLaMA-1.3B/f1c71b95ef4e98e06b6a.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"vocab_size": 32000, "max_position_embeddings": 4096, "hidden_size": 2048, "intermediate_size": 5504, "num_hidden_layers": 24, "num_attention_heads": 16, "num_key_value_heads": 16, "hidden_act": "silu", "initializer_range": 0.02, "rms_norm_eps": 1e-05, "pretraining_tp": 1, "use_cache": true, "rope_theta": 10000.0, "rope_scaling": null, "attention_bias": false, "attention_dropout": 0.0, "torch_dtype": "float32", "tie_word_embeddings": false, "architectures": ["LlamaForCausalLM"], "bos_token_id": 1, "pad_token_id": 0, "eos_token_id": 2, "model_type": "llama", "neuron": {"task": "text-generation", "batch_size": 4, "num_cores": 2, "auto_cast_type": "fp32", "sequence_length": 1024, "compiler_type": "neuronx-cc", "compiler_version": "2.12.68.0+4480452af", "checkpoint_id": "princeton-nlp/Sheared-LLaMA-1.3B", "checkpoint_revision": "a4b76938edbf571ea7d7d9904861cbdca08809b4"}}
neuronxcc-2.12.68.0+4480452af/MODULE_1154b678e32b38df568d+2c2d707e/compile_flags.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ --model-type=transformer --auto-cast=none
neuronxcc-2.12.68.0+4480452af/MODULE_1154b678e32b38df568d+2c2d707e/model.hlo.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0a060fccd716fe8f5683c4f86250fbeb8b28943ac48b2e578534dcbf655785f2
3
+ size 437888
neuronxcc-2.12.68.0+4480452af/MODULE_1154b678e32b38df568d+2c2d707e/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ef145ab8e588dddd7332b13131940c9eb328571e94d023b69de380392c236a9f
3
+ size 28232704
neuronxcc-2.12.68.0+4480452af/MODULE_13c86804904d0329bc38+2c2d707e/compile_flags.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ --model-type=transformer --auto-cast=none
neuronxcc-2.12.68.0+4480452af/MODULE_13c86804904d0329bc38+2c2d707e/model.hlo.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:70e150a71819706f9ed33715d6826037f41a6721e23464e8d39484d0d3f89114
3
+ size 261709
neuronxcc-2.12.68.0+4480452af/MODULE_13c86804904d0329bc38+2c2d707e/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b7385093a9a2686451785e49e7fa14bf437f3fc81ed4173dc652e41d9a4e7b97
3
+ size 4629504
neuronxcc-2.12.68.0+4480452af/MODULE_14fd59d8fbe618033e23+2c2d707e/compile_flags.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ --model-type=transformer --auto-cast=none
neuronxcc-2.12.68.0+4480452af/MODULE_14fd59d8fbe618033e23+2c2d707e/model.hlo.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ab4f1189b9b16998156cc3358fb2cd5c3b562e75d0e957a923b819b7f063943d
3
+ size 225998
neuronxcc-2.12.68.0+4480452af/MODULE_14fd59d8fbe618033e23+2c2d707e/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8ac14cdf2344ed1500dd809ebba3a0a63c1d521ae596bbef8f57210013e555c8
3
+ size 1823744
neuronxcc-2.12.68.0+4480452af/MODULE_1f4c8e3481cc0ab2ecfe+2c2d707e/compile_flags.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ --model-type=transformer --auto-cast=none
neuronxcc-2.12.68.0+4480452af/MODULE_1f4c8e3481cc0ab2ecfe+2c2d707e/model.hlo.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1d9545060a46d7deeda64f76eec5868a6e800eb05f08299260c1dfb8eb328f24
3
+ size 256261
neuronxcc-2.12.68.0+4480452af/MODULE_1f4c8e3481cc0ab2ecfe+2c2d707e/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0f7bfa2cef67985f8e93c0549ba0a45dc9da2d628cc0311d75374f3c6cd5e169
3
+ size 7681024
neuronxcc-2.12.68.0+4480452af/MODULE_22850a3c370343c4e87c+2c2d707e/compile_flags.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ --model-type=transformer --auto-cast=none
neuronxcc-2.12.68.0+4480452af/MODULE_22850a3c370343c4e87c+2c2d707e/model.hlo.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5e6326e41992e691fa8133fa973fc23ff5a6a1d0ccdc7fcca4b53ff80986da5b
3
+ size 223497
neuronxcc-2.12.68.0+4480452af/MODULE_22850a3c370343c4e87c+2c2d707e/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4123837c77ddb4716ad5674d772cf2f8887dd2bee22f4a3bb3b7caf4cdfe719d
3
+ size 2571264
neuronxcc-2.12.68.0+4480452af/MODULE_23be14b918bd101e0307+2c2d707e/compile_flags.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ --model-type=transformer --auto-cast=none
neuronxcc-2.12.68.0+4480452af/MODULE_23be14b918bd101e0307+2c2d707e/model.hlo.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5298d063e0b0217d531e500f2a003e597b37881244b029ca276dc30c632ab0ac
3
+ size 256261
neuronxcc-2.12.68.0+4480452af/MODULE_23be14b918bd101e0307+2c2d707e/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a92fdc60f880041fb004c6337777bd150002245d1c7f7a49d50101c86731bba5
3
+ size 7681024
neuronxcc-2.12.68.0+4480452af/MODULE_439cb9059d1a4720adfb+2c2d707e/compile_flags.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ --model-type=transformer --auto-cast=none
neuronxcc-2.12.68.0+4480452af/MODULE_439cb9059d1a4720adfb+2c2d707e/model.hlo.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ffd2dc6439c7e93d0de24d09aeede49422a2702868b539af8aab3e719e9dc150
3
+ size 295537
neuronxcc-2.12.68.0+4480452af/MODULE_439cb9059d1a4720adfb+2c2d707e/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:592ecc0f2e26c2b716fa32a3cef40b6ce7ce4e9211cea329d01e2b05ebbda324
3
+ size 2622464
neuronxcc-2.12.68.0+4480452af/MODULE_441b50063b845da3c878+2c2d707e/compile_flags.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ --model-type=transformer --auto-cast=none
neuronxcc-2.12.68.0+4480452af/MODULE_441b50063b845da3c878+2c2d707e/model.hlo.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8043d11cce1bc62c812cb9df49922f8d59fc1b247b0e3b7dd86262388faf81c6
3
+ size 220730
neuronxcc-2.12.68.0+4480452af/MODULE_441b50063b845da3c878+2c2d707e/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d86b555fa1eb898a9ec0d4244a4ddea536bbbaade14712509d56fc3705a736f1
3
+ size 1793024
neuronxcc-2.12.68.0+4480452af/MODULE_62ebf9bdd149ff8c97d9+2c2d707e/compile_flags.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ --model-type=transformer --auto-cast=none
neuronxcc-2.12.68.0+4480452af/MODULE_62ebf9bdd149ff8c97d9+2c2d707e/model.hlo.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:39bf29b9c97232e2e1bbdeb1bd25b9011cd0aee2a5fed300fea50a39c2228c54
3
+ size 287274
neuronxcc-2.12.68.0+4480452af/MODULE_62ebf9bdd149ff8c97d9+2c2d707e/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ab8b786fdb966254b9d9ecdfac8b4dab2d98eb89d56d1bbac419af1e4e307d5d
3
+ size 3103744
neuronxcc-2.12.68.0+4480452af/MODULE_9278928fd84a21a91a59+2c2d707e/compile_flags.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ --model-type=transformer --auto-cast=none
neuronxcc-2.12.68.0+4480452af/MODULE_9278928fd84a21a91a59+2c2d707e/model.hlo.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:07dd146c5284a257d6b40fdaf686ea6111b4f605a9984a5b61e0d65094e77127
3
+ size 295537
neuronxcc-2.12.68.0+4480452af/MODULE_9b8a4b44cab7fd6a36b6+2c2d707e/compile_flags.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ --model-type=transformer --auto-cast=none
neuronxcc-2.12.68.0+4480452af/MODULE_9b8a4b44cab7fd6a36b6+2c2d707e/model.hlo.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:523773fbfb5fcb65958793eef10cede18b4f29eb0000dd8b5eadfa28d08a9c0c
3
+ size 225998
neuronxcc-2.12.68.0+4480452af/MODULE_9b8a4b44cab7fd6a36b6+2c2d707e/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:70590c26bb648b4fcbaa924459c803db721af42cbd8d63ed8ee6badd36ec0100
3
+ size 1731584
neuronxcc-2.12.68.0+4480452af/MODULE_bd284532da36dbde1d9d+2c2d707e/compile_flags.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ --model-type=transformer --auto-cast=none
neuronxcc-2.12.68.0+4480452af/MODULE_bd284532da36dbde1d9d+2c2d707e/model.hlo.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d02643af119934dfbf6104863ca64d217c8374d48d944802f5a74714b8566011
3
+ size 225998
neuronxcc-2.12.68.0+4480452af/MODULE_bd284532da36dbde1d9d+2c2d707e/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a786ade92464a476f97e1b945c349ac991ab5bf4591753e34878b472b18eb31f
3
+ size 1680384
neuronxcc-2.12.68.0+4480452af/MODULE_dd3fcd08d9d80a1eff7d+2c2d707e/compile_flags.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ --model-type=transformer --auto-cast=none
neuronxcc-2.12.68.0+4480452af/MODULE_dd3fcd08d9d80a1eff7d+2c2d707e/model.hlo.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8eee6818db4e307c3ed592714a5c1a21a4d696ab4d3f9650bcc3160b93bef0a9
3
+ size 493747
neuronxcc-2.12.68.0+4480452af/MODULE_dd3fcd08d9d80a1eff7d+2c2d707e/model.neff ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d45ec0ae33707e270c0fd8b1ac7484acb9028a204eba040441c761ea8e32b3fc
3
+ size 7465984
neuronxcc-2.12.68.0+4480452af/MODULE_e342da015657e3ad5e30+2c2d707e/compile_flags.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ --model-type=transformer --auto-cast=none