diff --git a/.gitattributes b/.gitattributes index 8b650b78dbf58aba35d7f6310829bbacf9e58f07..89de7bfac64c5f889301d40ca372a3740cd33439 100644 --- a/.gitattributes +++ b/.gitattributes @@ -14415,3 +14415,15 @@ neuronxcc-2.21.33363.0+82129205/MODULE_d5ada2fae774253484ca+24129607/model.neff neuronxcc-2.21.33363.0+82129205/MODULE_e973fc6fe6bd9c3d7c03+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text neuronxcc-2.21.33363.0+82129205/MODULE_e973fc6fe6bd9c3d7c03+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text neuronxcc-2.21.33363.0+82129205/MODULE_f81f36ad8e744c8da4f2+24129607/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_3b4b3de0fa211055bd74+47866c66/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_3b4b3de0fa211055bd74+47866c66/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_542c0bd110414da7bca1+f931dc9f/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_542c0bd110414da7bca1+f931dc9f/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_b2bf569c84c54e75eabd+ee54247e/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_b2bf569c84c54e75eabd+ee54247e/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_c6050ed2a2db99f0f447+45a194b0/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_c6050ed2a2db99f0f447+45a194b0/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_df7cef7d783cd9aa888e+2528ca2d/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_df7cef7d783cd9aa888e+2528ca2d/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_fb29e0ab34c0bf999b52+1745662a/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_fb29e0ab34c0bf999b52+1745662a/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/7d021e56337edf8333cc.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/7d021e56337edf8333cc.json new file mode 100644 index 0000000000000000000000000000000000000000..bb53d54a1acdddd873451a2f57fbe10045ca34d5 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/7d021e56337edf8333cc.json @@ -0,0 +1,58 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "hf-internal-testing/tiny-random-GraniteForCausalLM", + "_task": "text-generation", + "architectures": [ + "GraniteForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "attention_multiplier": 1.0, + "dtype": "float32", + "embedding_multiplier": 1.0, + "hidden_act": "silu", + "hidden_size": 32, + "initializer_range": 0.02, + "intermediate_size": 64, + "logits_scaling": 1.0, + "max_position_embeddings": 2048, + "mlp_bias": false, + "model_type": "granite", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 1, + "capacity_factor": null, + "checkpoint_id": "hf-internal-testing/tiny-random-GraniteForCausalLM", + "checkpoint_revision": "c3074ebc0ac2fe545305f5e5f6cce2cc9b2aa0c5", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 1, + "max_context_length": 1024, + "max_topk": 256, + "n_active_tokens": 1024, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.4.dev0", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 1024, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "float32", + "tp_degree": 2 + }, + "num_attention_heads": 4, + "num_hidden_layers": 2, + "num_key_value_heads": 4, + "residual_multiplier": 1.0, + "rms_norm_eps": 1e-06, + "rope_scaling": null, + "rope_theta": 10000.0, + "tie_word_embeddings": false, + "use_cache": true, + "vocab_size": 49152 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/granite/ibm-granite/granite-3.1-2b-instruct/7600b12913a972b64e27.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/granite/ibm-granite/granite-3.1-2b-instruct/7600b12913a972b64e27.json new file mode 100644 index 0000000000000000000000000000000000000000..14c77353adfcec10f7b4e79c4b2a6e95fb5f0d69 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/granite/ibm-granite/granite-3.1-2b-instruct/7600b12913a972b64e27.json @@ -0,0 +1,58 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "ibm-granite/granite-3.1-2b-instruct", + "_task": "text-generation", + "architectures": [ + "GraniteForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.1, + "attention_multiplier": 0.015625, + "dtype": "bfloat16", + "embedding_multiplier": 12.0, + "hidden_act": "silu", + "hidden_size": 2048, + "initializer_range": 0.02, + "intermediate_size": 8192, + "logits_scaling": 8.0, + "max_position_embeddings": 131072, + "mlp_bias": false, + "model_type": "granite", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 4, + "capacity_factor": null, + "checkpoint_id": "ibm-granite/granite-3.1-2b-instruct", + "checkpoint_revision": "bbc2aed595bd38bd770263dc3ab831db9794441d", + "continuous_batching": true, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 4, + "max_context_length": 4096, + "max_topk": 256, + "n_active_tokens": 4096, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.4.dev0", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 4096, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 32, + "num_hidden_layers": 40, + "num_key_value_heads": 8, + "residual_multiplier": 0.22, + "rms_norm_eps": 1e-05, + "rope_scaling": null, + "rope_theta": 5000000.0, + "tie_word_embeddings": true, + "use_cache": true, + "vocab_size": 49155 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/llama/llamafactory/tiny-random-Llama-3/4cf9991cd1aa6c40fc85.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/llama/llamafactory/tiny-random-Llama-3/4cf9991cd1aa6c40fc85.json new file mode 100644 index 0000000000000000000000000000000000000000..58f3a38b488a50af3511365b81f0f6a7bc54accc --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/llama/llamafactory/tiny-random-Llama-3/4cf9991cd1aa6c40fc85.json @@ -0,0 +1,62 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "llamafactory/tiny-random-Llama-3", + "_task": "text-generation", + "architectures": [ + "LlamaForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "dtype": "float16", + "head_dim": 4, + "hidden_act": "silu", + "hidden_size": 16, + "initializer_range": 0.02, + "intermediate_size": 64, + "max_position_embeddings": 131072, + "mlp_bias": false, + "model_type": "llama", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 2, + "capacity_factor": null, + "checkpoint_id": "llamafactory/tiny-random-Llama-3", + "checkpoint_revision": "bf2a2e3bf199ad2ee96f02a3c00246c608db22a8", + "continuous_batching": true, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 2, + "max_context_length": 128, + "max_topk": 256, + "n_active_tokens": 128, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.4.dev0", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 128, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn2", + "torch_dtype": "float16", + "tp_degree": 2 + }, + "num_attention_heads": 4, + "num_hidden_layers": 2, + "num_key_value_heads": 4, + "pretraining_tp": 1, + "rms_norm_eps": 1e-05, + "rope_scaling": { + "factor": 8.0, + "high_freq_factor": 4.0, + "low_freq_factor": 1.0, + "original_max_position_embeddings": 8192, + "rope_type": "llama3" + }, + "rope_theta": 500000.0, + "tie_word_embeddings": false, + "use_cache": true, + "vocab_size": 128256 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/llama/llamafactory/tiny-random-Llama-3/d191e10af9fe9bcbb39c.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/llama/llamafactory/tiny-random-Llama-3/d191e10af9fe9bcbb39c.json new file mode 100644 index 0000000000000000000000000000000000000000..c53753b4579d8ab372476b98109411d1e9f12063 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/llama/llamafactory/tiny-random-Llama-3/d191e10af9fe9bcbb39c.json @@ -0,0 +1,62 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "llamafactory/tiny-random-Llama-3", + "_task": "text-generation", + "architectures": [ + "LlamaForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "dtype": "float16", + "head_dim": 4, + "hidden_act": "silu", + "hidden_size": 16, + "initializer_range": 0.02, + "intermediate_size": 64, + "max_position_embeddings": 131072, + "mlp_bias": false, + "model_type": "llama", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 1, + "capacity_factor": null, + "checkpoint_id": "llamafactory/tiny-random-Llama-3", + "checkpoint_revision": "bf2a2e3bf199ad2ee96f02a3c00246c608db22a8", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 1, + "max_context_length": 512, + "max_topk": 256, + "n_active_tokens": 512, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.4.dev0", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 512, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "float16", + "tp_degree": 2 + }, + "num_attention_heads": 4, + "num_hidden_layers": 2, + "num_key_value_heads": 4, + "pretraining_tp": 1, + "rms_norm_eps": 1e-05, + "rope_scaling": { + "factor": 8.0, + "high_freq_factor": 4.0, + "low_freq_factor": 1.0, + "original_max_position_embeddings": 8192, + "rope_type": "llama3" + }, + "rope_theta": 500000.0, + "tie_word_embeddings": false, + "use_cache": true, + "vocab_size": 128256 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/llama/llamafactory/tiny-random-Llama-3/dab3aaee34c5e26fea75.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/llama/llamafactory/tiny-random-Llama-3/dab3aaee34c5e26fea75.json new file mode 100644 index 0000000000000000000000000000000000000000..cfcaaf3c30a7a68da4d1483b2134a6664890a8dd --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/llama/llamafactory/tiny-random-Llama-3/dab3aaee34c5e26fea75.json @@ -0,0 +1,62 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "llamafactory/tiny-random-Llama-3", + "_task": "text-generation", + "architectures": [ + "LlamaForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "dtype": "float16", + "head_dim": 4, + "hidden_act": "silu", + "hidden_size": 16, + "initializer_range": 0.02, + "intermediate_size": 64, + "max_position_embeddings": 131072, + "mlp_bias": false, + "model_type": "llama", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 1, + "capacity_factor": null, + "checkpoint_id": "llamafactory/tiny-random-Llama-3", + "checkpoint_revision": "bf2a2e3bf199ad2ee96f02a3c00246c608db22a8", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 1, + "max_context_length": 1024, + "max_topk": 256, + "n_active_tokens": 1024, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.4.dev0", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 1024, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "float16", + "tp_degree": 2 + }, + "num_attention_heads": 4, + "num_hidden_layers": 2, + "num_key_value_heads": 4, + "pretraining_tp": 1, + "rms_norm_eps": 1e-05, + "rope_scaling": { + "factor": 8.0, + "high_freq_factor": 4.0, + "low_freq_factor": 1.0, + "original_max_position_embeddings": 8192, + "rope_type": "llama3" + }, + "rope_theta": 500000.0, + "tie_word_embeddings": false, + "use_cache": true, + "vocab_size": 128256 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/llama/unsloth/Llama-3.2-1B-Instruct/002a772f46d3cc90fc0f.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/llama/unsloth/Llama-3.2-1B-Instruct/002a772f46d3cc90fc0f.json new file mode 100644 index 0000000000000000000000000000000000000000..4a4d48b7a6e1c44008f71321044696114573a201 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/llama/unsloth/Llama-3.2-1B-Instruct/002a772f46d3cc90fc0f.json @@ -0,0 +1,63 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "unsloth/Llama-3.2-1B-Instruct", + "_task": "text-generation", + "architectures": [ + "LlamaForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "dtype": "bfloat16", + "head_dim": 64, + "hidden_act": "silu", + "hidden_size": 2048, + "initializer_range": 0.02, + "intermediate_size": 8192, + "max_position_embeddings": 131072, + "mlp_bias": false, + "model_type": "llama", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 1, + "capacity_factor": null, + "checkpoint_id": "unsloth/Llama-3.2-1B-Instruct", + "checkpoint_revision": null, + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": false, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 1, + "max_context_length": 4096, + "max_topk": 256, + "n_active_tokens": 4096, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": false, + "optimum_neuron_version": "0.4.4.dev0", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 4096, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 32, + "num_hidden_layers": 16, + "num_key_value_heads": 8, + "pretraining_tp": 1, + "rms_norm_eps": 1e-05, + "rope_scaling": { + "factor": 32.0, + "high_freq_factor": 4.0, + "low_freq_factor": 1.0, + "original_max_position_embeddings": 8192, + "rope_type": "llama3" + }, + "rope_theta": 500000.0, + "tie_word_embeddings": true, + "unsloth_fixed": true, + "use_cache": true, + "vocab_size": 128256 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/llama/unsloth/Llama-3.2-1B-Instruct/71268e0716d858cef753.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/llama/unsloth/Llama-3.2-1B-Instruct/71268e0716d858cef753.json new file mode 100644 index 0000000000000000000000000000000000000000..bc286598c9e1c67c0661acd2d0bdda06cb2dc7d5 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/llama/unsloth/Llama-3.2-1B-Instruct/71268e0716d858cef753.json @@ -0,0 +1,63 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "unsloth/Llama-3.2-1B-Instruct", + "_task": "text-generation", + "architectures": [ + "LlamaForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "dtype": "bfloat16", + "head_dim": 64, + "hidden_act": "silu", + "hidden_size": 2048, + "initializer_range": 0.02, + "intermediate_size": 8192, + "max_position_embeddings": 131072, + "mlp_bias": false, + "model_type": "llama", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 1, + "capacity_factor": null, + "checkpoint_id": "unsloth/Llama-3.2-1B-Instruct", + "checkpoint_revision": null, + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": false, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 1, + "max_context_length": 4096, + "max_topk": 256, + "n_active_tokens": 4096, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": false, + "optimum_neuron_version": "0.4.4.dev0", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 4096, + "speculation_length": 5, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 32, + "num_hidden_layers": 16, + "num_key_value_heads": 8, + "pretraining_tp": 1, + "rms_norm_eps": 1e-05, + "rope_scaling": { + "factor": 32.0, + "high_freq_factor": 4.0, + "low_freq_factor": 1.0, + "original_max_position_embeddings": 8192, + "rope_type": "llama3" + }, + "rope_theta": 500000.0, + "tie_word_embeddings": true, + "unsloth_fixed": true, + "use_cache": true, + "vocab_size": 128256 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/llama4_text/tiny-random/llama-4/993bc3e2a3a56af4961a.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/llama4_text/tiny-random/llama-4/993bc3e2a3a56af4961a.json new file mode 100644 index 0000000000000000000000000000000000000000..7fa510c93dee075513b5282eb8ef2365b8371a7c --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/llama4_text/tiny-random/llama-4/993bc3e2a3a56af4961a.json @@ -0,0 +1,81 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "tiny-random/llama-4", + "_task": "text-generation", + "attention_bias": false, + "attention_chunk_size": 128, + "attention_dropout": 0.0, + "attn_scale": 0.1, + "attn_temperature_tuning": 4, + "cache_implementation": "hybrid", + "dtype": "bfloat16", + "floor_scale": 8192, + "for_llm_compressor": false, + "head_dim": 32, + "hidden_act": "silu", + "hidden_size": 32, + "initializer_range": 0.02, + "interleave_moe_layer_step": 2, + "intermediate_size": 64, + "intermediate_size_mlp": 128, + "layer_types": [ + "chunked_attention", + "chunked_attention", + "chunked_attention", + "full_attention" + ], + "max_position_embeddings": 1048576, + "model_type": "llama4_text", + "moe_layers": [ + 1, + 3 + ], + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 1, + "capacity_factor": null, + "checkpoint_id": "tiny-random/llama-4", + "checkpoint_revision": "9e716f5d4d1ffe0a44a15f46f4a12b840439aba4", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": false, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 1, + "max_context_length": 1024, + "max_topk": 256, + "n_active_tokens": 1024, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.4.dev0", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 1024, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "no_rope_layers": [ + 1, + 1, + 1, + 0 + ], + "num_attention_heads": 1, + "num_experts_per_tok": 1, + "num_hidden_layers": 4, + "num_key_value_heads": 1, + "num_local_experts": 8, + "output_router_logits": false, + "rms_norm_eps": 1e-05, + "rope_scaling": null, + "rope_theta": 500000.0, + "router_aux_loss_coef": 0.001, + "router_jitter_noise": 0.0, + "tie_word_embeddings": true, + "use_cache": true, + "use_qk_norm": true, + "vocab_size": 202048 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/mixtral/dacorvo/Mixtral-tiny/1590241d4b08e264f0d2.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/mixtral/dacorvo/Mixtral-tiny/1590241d4b08e264f0d2.json new file mode 100644 index 0000000000000000000000000000000000000000..815c869988dcab1adaaf2124b9c51839b75cf3ac --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/mixtral/dacorvo/Mixtral-tiny/1590241d4b08e264f0d2.json @@ -0,0 +1,58 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "dacorvo/Mixtral-tiny", + "_task": "text-generation", + "architectures": [ + "MixtralForCausalLM" + ], + "attention_dropout": 0.0, + "dtype": "float16", + "head_dim": 32, + "hidden_act": "silu", + "hidden_size": 1024, + "initializer_range": 0.02, + "intermediate_size": 3584, + "max_position_embeddings": 1024, + "model_type": "mixtral", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 1, + "capacity_factor": null, + "checkpoint_id": "dacorvo/Mixtral-tiny", + "checkpoint_revision": "c557ba205ddff6ea911f4719e0d543d6c08356b6", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": false, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 1, + "max_context_length": 1024, + "max_topk": 256, + "n_active_tokens": 1024, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": false, + "optimum_neuron_version": "0.4.4.dev0", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 1024, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "float16", + "tp_degree": 2 + }, + "num_attention_heads": 32, + "num_experts_per_tok": 2, + "num_hidden_layers": 2, + "num_key_value_heads": 8, + "num_local_experts": 8, + "output_router_logits": false, + "rms_norm_eps": 1e-05, + "rope_theta": 10000.0, + "router_aux_loss_coef": 0.001, + "router_jitter_noise": 0.0, + "sliding_window": 4096, + "tie_word_embeddings": false, + "use_cache": true, + "vocab_size": 32000 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/phi3/microsoft/Phi-3-mini-4k-instruct/f1be233a4a51c058662d.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/phi3/microsoft/Phi-3-mini-4k-instruct/f1be233a4a51c058662d.json new file mode 100644 index 0000000000000000000000000000000000000000..c634b3f6be723d11cc770fb9da162d8324f8abc4 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/phi3/microsoft/Phi-3-mini-4k-instruct/f1be233a4a51c058662d.json @@ -0,0 +1,62 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "microsoft/Phi-3-mini-4k-instruct", + "_task": "text-generation", + "architectures": [ + "Phi3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "auto_map": { + "AutoConfig": "configuration_phi3.Phi3Config", + "AutoModelForCausalLM": "modeling_phi3.Phi3ForCausalLM" + }, + "dtype": "bfloat16", + "embd_pdrop": 0.0, + "hidden_act": "silu", + "hidden_size": 3072, + "initializer_range": 0.02, + "intermediate_size": 8192, + "max_position_embeddings": 4096, + "model_type": "phi3", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 4, + "capacity_factor": null, + "checkpoint_id": "microsoft/Phi-3-mini-4k-instruct", + "checkpoint_revision": "f39ac1d28e925b323eae81227eaba4464caced4e", + "continuous_batching": true, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 4, + "max_context_length": 4096, + "max_topk": 256, + "n_active_tokens": 4096, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.4.dev0", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 4096, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 32, + "num_hidden_layers": 32, + "num_key_value_heads": 32, + "original_max_position_embeddings": 4096, + "partial_rotary_factor": 1.0, + "resid_pdrop": 0.0, + "rms_norm_eps": 1e-05, + "rope_scaling": null, + "rope_theta": 10000.0, + "sliding_window": 2047, + "tie_word_embeddings": false, + "use_cache": true, + "vocab_size": 32064 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/phi3/yujiepan/phi-4-tiny-random/5d4426f635453972cc64.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/phi3/yujiepan/phi-4-tiny-random/5d4426f635453972cc64.json new file mode 100644 index 0000000000000000000000000000000000000000..2cce4007356ad5bec9e0cde5bea7dbd05db339c1 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/phi3/yujiepan/phi-4-tiny-random/5d4426f635453972cc64.json @@ -0,0 +1,59 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "yujiepan/phi-4-tiny-random", + "_task": "text-generation", + "architectures": [ + "Phi3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "auto_map": {}, + "dtype": "bfloat16", + "embd_pdrop": 0.0, + "hidden_act": "silu", + "hidden_size": 16, + "initializer_range": 0.02, + "intermediate_size": 32, + "max_position_embeddings": 16384, + "model_type": "phi3", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 1, + "capacity_factor": null, + "checkpoint_id": "yujiepan/phi-4-tiny-random", + "checkpoint_revision": "18a9a1168dc97ac6d128f811925670c275610f5a", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 1, + "max_context_length": 1024, + "max_topk": 256, + "n_active_tokens": 1024, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.4.dev0", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 1024, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 2, + "num_hidden_layers": 2, + "num_key_value_heads": 1, + "original_max_position_embeddings": 16384, + "partial_rotary_factor": 1.0, + "resid_pdrop": 0.0, + "rms_norm_eps": 1e-05, + "rope_scaling": null, + "rope_theta": 250000, + "sliding_window": null, + "tie_word_embeddings": false, + "use_cache": true, + "vocab_size": 100352 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/qwen2/Qwen/Qwen2.5-0.5B/a3841ef31133954fb76f.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/qwen2/Qwen/Qwen2.5-0.5B/a3841ef31133954fb76f.json new file mode 100644 index 0000000000000000000000000000000000000000..559a9cd764f74e9601368cd7a4c595dfd5771bb5 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/qwen2/Qwen/Qwen2.5-0.5B/a3841ef31133954fb76f.json @@ -0,0 +1,82 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "Qwen/Qwen2.5-0.5B", + "_task": "text-generation", + "architectures": [ + "Qwen2ForCausalLM" + ], + "attention_dropout": 0.0, + "dtype": "bfloat16", + "hidden_act": "silu", + "hidden_size": 896, + "initializer_range": 0.02, + "intermediate_size": 4864, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 32768, + "max_window_layers": 24, + "model_type": "qwen2", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 1, + "capacity_factor": null, + "checkpoint_id": "Qwen/Qwen2.5-0.5B", + "checkpoint_revision": "060db6499f32faf8b98477b0a26969ef7d8b9987", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": false, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 1, + "max_context_length": 4096, + "max_topk": 256, + "n_active_tokens": 4096, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.4.dev0", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 4096, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 14, + "num_hidden_layers": 24, + "num_key_value_heads": 2, + "rms_norm_eps": 1e-06, + "rope_scaling": null, + "rope_theta": 1000000.0, + "sliding_window": null, + "tie_word_embeddings": true, + "use_cache": true, + "use_mrope": false, + "use_sliding_window": false, + "vocab_size": 151936 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/qwen2/Qwen/Qwen2.5-0.5B/c1c6d556e7977f6007cd.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/qwen2/Qwen/Qwen2.5-0.5B/c1c6d556e7977f6007cd.json new file mode 100644 index 0000000000000000000000000000000000000000..d6e0a55ccd8504511d9746889b9f529e0f812c27 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/qwen2/Qwen/Qwen2.5-0.5B/c1c6d556e7977f6007cd.json @@ -0,0 +1,82 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "Qwen/Qwen2.5-0.5B", + "_task": "text-generation", + "architectures": [ + "Qwen2ForCausalLM" + ], + "attention_dropout": 0.0, + "dtype": "bfloat16", + "hidden_act": "silu", + "hidden_size": 896, + "initializer_range": 0.02, + "intermediate_size": 4864, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 32768, + "max_window_layers": 24, + "model_type": "qwen2", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 2, + "capacity_factor": null, + "checkpoint_id": "Qwen/Qwen2.5-0.5B", + "checkpoint_revision": "060db6499f32faf8b98477b0a26969ef7d8b9987", + "continuous_batching": true, + "ep_degree": 1, + "fused_qkv": false, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 2, + "max_context_length": 1024, + "max_topk": 256, + "n_active_tokens": 1024, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": false, + "optimum_neuron_version": "0.4.4.dev0", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 1024, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 14, + "num_hidden_layers": 24, + "num_key_value_heads": 2, + "rms_norm_eps": 1e-06, + "rope_scaling": null, + "rope_theta": 1000000.0, + "sliding_window": null, + "tie_word_embeddings": true, + "use_cache": true, + "use_mrope": false, + "use_sliding_window": false, + "vocab_size": 151936 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/qwen2/yujiepan/qwen2.5-128k-tiny-random/4059ceaa7cbe1c468099.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/qwen2/yujiepan/qwen2.5-128k-tiny-random/4059ceaa7cbe1c468099.json new file mode 100644 index 0000000000000000000000000000000000000000..1886a648d781bfa95d8b6b0926a5085e038efebc --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/qwen2/yujiepan/qwen2.5-128k-tiny-random/4059ceaa7cbe1c468099.json @@ -0,0 +1,64 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "yujiepan/qwen2.5-128k-tiny-random", + "_task": "text-generation", + "architectures": [ + "Qwen2ForCausalLM" + ], + "attention_dropout": 0.0, + "dtype": "bfloat16", + "hidden_act": "silu", + "hidden_size": 8, + "initializer_range": 0.02, + "intermediate_size": 16, + "layer_types": [ + "full_attention", + "full_attention" + ], + "max_position_embeddings": 32768, + "max_window_layers": 1, + "model_type": "qwen2", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 1, + "capacity_factor": null, + "checkpoint_id": "yujiepan/qwen2.5-128k-tiny-random", + "checkpoint_revision": "c8296d4ca3f87782876d2382fbb6481d1beb8ef0", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": false, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 1, + "max_context_length": 1024, + "max_topk": 256, + "n_active_tokens": 1024, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.4.dev0", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 1024, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 4, + "num_hidden_layers": 2, + "num_key_value_heads": 2, + "rms_norm_eps": 1e-06, + "rope_scaling": { + "factor": 4.0, + "original_max_position_embeddings": 32768, + "rope_type": "yarn", + "type": "yarn" + }, + "rope_theta": 1000000.0, + "sliding_window": null, + "tie_word_embeddings": false, + "use_cache": true, + "use_sliding_window": false, + "vocab_size": 152064 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/qwen3/Qwen/Qwen3-Embedding-0.6B/1b96d2b17419448434ba.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/qwen3/Qwen/Qwen3-Embedding-0.6B/1b96d2b17419448434ba.json new file mode 100644 index 0000000000000000000000000000000000000000..5f70c135b9f21942ae3ddbd57f547c257873f4d8 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/qwen3/Qwen/Qwen3-Embedding-0.6B/1b96d2b17419448434ba.json @@ -0,0 +1,87 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "Qwen/Qwen3-Embedding-0.6B", + "_task": "text-generation", + "architectures": [ + "Qwen3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "dtype": "bfloat16", + "head_dim": 128, + "hidden_act": "silu", + "hidden_size": 1024, + "initializer_range": 0.02, + "intermediate_size": 3072, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 32768, + "max_window_layers": 28, + "model_type": "qwen3", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 6, + "capacity_factor": null, + "checkpoint_id": "Qwen/Qwen3-Embedding-0.6B", + "checkpoint_revision": "c54f2e6e80b2d7b7de06f51cec4959f6b3e03418", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 6, + "max_context_length": 1024, + "max_topk": 256, + "n_active_tokens": 1024, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": false, + "optimum_neuron_version": "0.4.4.dev0", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 1024, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 16, + "num_hidden_layers": 28, + "num_key_value_heads": 8, + "rms_norm_eps": 1e-06, + "rope_scaling": null, + "rope_theta": 1000000, + "sliding_window": null, + "tie_word_embeddings": true, + "use_cache": true, + "use_sliding_window": false, + "vocab_size": 151669 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/qwen3/Qwen/Qwen3-Embedding-0.6B/657aac4c188cc041c840.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/qwen3/Qwen/Qwen3-Embedding-0.6B/657aac4c188cc041c840.json new file mode 100644 index 0000000000000000000000000000000000000000..6b05a0757bedf01054397aa6baf7e70f54c71002 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/qwen3/Qwen/Qwen3-Embedding-0.6B/657aac4c188cc041c840.json @@ -0,0 +1,87 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "Qwen/Qwen3-Embedding-0.6B", + "_task": "text-generation", + "architectures": [ + "Qwen3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "dtype": "bfloat16", + "head_dim": 128, + "hidden_act": "silu", + "hidden_size": 1024, + "initializer_range": 0.02, + "intermediate_size": 3072, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 32768, + "max_window_layers": 28, + "model_type": "qwen3", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 4, + "capacity_factor": null, + "checkpoint_id": "Qwen/Qwen3-Embedding-0.6B", + "checkpoint_revision": "c54f2e6e80b2d7b7de06f51cec4959f6b3e03418", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 4, + "max_context_length": 1024, + "max_topk": 256, + "n_active_tokens": 1024, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": false, + "optimum_neuron_version": "0.4.4.dev0", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 1024, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 16, + "num_hidden_layers": 28, + "num_key_value_heads": 8, + "rms_norm_eps": 1e-06, + "rope_scaling": null, + "rope_theta": 1000000, + "sliding_window": null, + "tie_word_embeddings": true, + "use_cache": true, + "use_sliding_window": false, + "vocab_size": 151669 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/qwen3_moe/optimum-internal-testing/tiny-random-qwen3_moe/d7e521c77595672a2c6f.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/qwen3_moe/optimum-internal-testing/tiny-random-qwen3_moe/d7e521c77595672a2c6f.json new file mode 100644 index 0000000000000000000000000000000000000000..79450c5fcbaebc8dd4e81eaaa8cd2baf190ca8fb --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/qwen3_moe/optimum-internal-testing/tiny-random-qwen3_moe/d7e521c77595672a2c6f.json @@ -0,0 +1,65 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "optimum-internal-testing/tiny-random-qwen3_moe", + "_task": "text-generation", + "architectures": [ + "Qwen3MoeForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "decoder_sparse_step": 2, + "dtype": "float32", + "head_dim": 32, + "hidden_act": "silu", + "hidden_size": 64, + "initializer_range": 0.02, + "intermediate_size": 128, + "max_position_embeddings": 40960, + "max_window_layers": 1, + "mlp_only_layers": [], + "model_type": "qwen3_moe", + "moe_intermediate_size": 128, + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 1, + "capacity_factor": null, + "checkpoint_id": "optimum-internal-testing/tiny-random-qwen3_moe", + "checkpoint_revision": "e0230be2839556b44b7400a233c73c74b4abb7af", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": false, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 1, + "max_context_length": 1024, + "max_topk": 256, + "n_active_tokens": 1024, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.4.dev0", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 1024, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "float32", + "tp_degree": 2 + }, + "norm_topk_prob": true, + "num_attention_heads": 2, + "num_experts": 8, + "num_experts_per_tok": 2, + "num_hidden_layers": 2, + "num_key_value_heads": 1, + "output_router_logits": false, + "rms_norm_eps": 1e-06, + "rope_scaling": null, + "rope_theta": 1000000.0, + "router_aux_loss_coef": 0.001, + "sliding_window": null, + "tie_word_embeddings": true, + "use_cache": true, + "use_sliding_window": false, + "vocab_size": 151936 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/smollm3/HuggingFaceTB/SmolLM3-3B/a5c2f73dbe3905255d8f.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/smollm3/HuggingFaceTB/SmolLM3-3B/a5c2f73dbe3905255d8f.json new file mode 100644 index 0000000000000000000000000000000000000000..b0943183db2a511fe233cc2da2457f5e7aba8b28 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/smollm3/HuggingFaceTB/SmolLM3-3B/a5c2f73dbe3905255d8f.json @@ -0,0 +1,134 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "HuggingFaceTB/SmolLM3-3B", + "_task": "text-generation", + "architectures": [ + "SmolLM3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "dtype": "bfloat16", + "hidden_act": "silu", + "hidden_size": 2048, + "initializer_range": 0.02, + "intermediate_size": 11008, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 65536, + "max_window_layers": 28, + "mlp_bias": false, + "model_type": "smollm3", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 4, + "capacity_factor": null, + "checkpoint_id": "HuggingFaceTB/SmolLM3-3B", + "checkpoint_revision": "a07cc9a04f16550a088caea529712d1d335b0ac1", + "continuous_batching": true, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 4, + "max_context_length": 4096, + "max_topk": 256, + "n_active_tokens": 4096, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.4.dev0", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 4096, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "no_rope_layer_interval": 4, + "no_rope_layers": [ + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0 + ], + "num_attention_heads": 16, + "num_hidden_layers": 36, + "num_key_value_heads": 4, + "pretraining_tp": 2, + "rms_norm_eps": 1e-06, + "rope_scaling": null, + "rope_theta": 5000000.0, + "sliding_window": null, + "use_cache": false, + "use_sliding_window": false, + "vocab_size": 128256 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/smollm3/HuggingFaceTB/SmolLM3-3B/a6ac0bc90de4088f4d5b.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/smollm3/HuggingFaceTB/SmolLM3-3B/a6ac0bc90de4088f4d5b.json new file mode 100644 index 0000000000000000000000000000000000000000..7491682c6b0db0dbb98b6904386f3ede7bfee1c1 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.4.dev0/smollm3/HuggingFaceTB/SmolLM3-3B/a6ac0bc90de4088f4d5b.json @@ -0,0 +1,134 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "HuggingFaceTB/SmolLM3-3B", + "_task": "text-generation", + "architectures": [ + "SmolLM3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "dtype": "bfloat16", + "hidden_act": "silu", + "hidden_size": 2048, + "initializer_range": 0.02, + "intermediate_size": 11008, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 65536, + "max_window_layers": 28, + "mlp_bias": false, + "model_type": "smollm3", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 1, + "capacity_factor": null, + "checkpoint_id": "HuggingFaceTB/SmolLM3-3B", + "checkpoint_revision": "a07cc9a04f16550a088caea529712d1d335b0ac1", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 1, + "max_context_length": 1024, + "max_topk": 256, + "n_active_tokens": 1024, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.4.dev0", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 1024, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "no_rope_layer_interval": 4, + "no_rope_layers": [ + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0 + ], + "num_attention_heads": 16, + "num_hidden_layers": 36, + "num_key_value_heads": 4, + "pretraining_tp": 2, + "rms_norm_eps": 1e-06, + "rope_scaling": null, + "rope_theta": 5000000.0, + "sliding_window": null, + "use_cache": false, + "use_sliding_window": false, + "vocab_size": 128256 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_10217061096959125489+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_10217061096959125489+e30acd3a/model.neff index 98b866c184304c99f142c729a56aa5ef1aa1ec30..a267697703f7308c629186e67a068ff52c5730c3 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_10217061096959125489+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_10217061096959125489+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_10244305442015770634+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_10244305442015770634+e30acd3a/model.neff index 094085c50482dd556f751455dd32beb509bf3d41..de4f1085605c1e338d4315d5d926f3bc96f0d453 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_10244305442015770634+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_10244305442015770634+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_10606948783918825529+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_10606948783918825529+e30acd3a/model.neff index 1232554e64271aff4cc80b7984e07c4fe1364918..3d5db322ceb1108b3737a95f663e64cc8c9f35fd 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_10606948783918825529+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_10606948783918825529+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_10645643398657092095+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_10645643398657092095+e30acd3a/model.neff index ac2dbc61001cb46a79f4539a3793017d07886e9d..01163c850cfa71cad6bd509125ec05b4fe7990af 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_10645643398657092095+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_10645643398657092095+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_10661660426924300837+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_10661660426924300837+e30acd3a/model.neff index 49ff6e1201ac11eeeebce47afee2a714cbccbb11..3ef9c66b6653ab1560c6e3c632c539da9e739645 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_10661660426924300837+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_10661660426924300837+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_10746122569655005679+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_10746122569655005679+e30acd3a/model.neff index 8cfb5750809a52e1181159ac1b50edfa78cb3a74..74699dc621badcd3d6004de19ce309b9e07c8d76 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_10746122569655005679+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_10746122569655005679+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_11086318750207148626+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_11086318750207148626+e30acd3a/model.neff index 31bcc85d66a816a182508223785312af4d394983..e53641d18f01449b2349ade0b9605739ee3716e8 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_11086318750207148626+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_11086318750207148626+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_11909310600244571805+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_11909310600244571805+e30acd3a/model.neff index 38b080719738b84cbd2f3a79acf9010cfb84a2f3..8bdbafcdfb51c4fb3b4ce1ce5cb4516525909220 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_11909310600244571805+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_11909310600244571805+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_11933818254123612383+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_11933818254123612383+e30acd3a/model.neff index 49e95f10c8d6c9c43ef7d6b0e95e83fabfbe2dde..8c79576ae245ff63cc53435f251502e0b48913b5 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_11933818254123612383+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_11933818254123612383+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_1210392327607194823+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_1210392327607194823+e30acd3a/model.neff index b8429cb847fd68c97464aa036e23fcdd21ae02a6..d1db24ca01efabd3b2a766b8239263e91ba19c9c 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_1210392327607194823+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_1210392327607194823+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_12487216553200321032+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_12487216553200321032+e30acd3a/model.neff index 3acddb8a6423fff8474f989613d6a9eab7f457a2..72fc078a08eec8591859ce946f76c7d947fdd071 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_12487216553200321032+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_12487216553200321032+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_13085549342645515693+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_13085549342645515693+e30acd3a/model.neff index 381041bdb9f4c84d769556360aaff95c82ddde33..75971b537b85c83a2cfdac99ac8ec4a6bc49a633 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_13085549342645515693+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_13085549342645515693+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_13102010590082783346+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_13102010590082783346+e30acd3a/model.neff index a5730c17f09446df8a556a9ac4b30352c23a74b9..b80abca726eb660ce361b1784f6bbbf586b18eca 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_13102010590082783346+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_13102010590082783346+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_13475273426270778455+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_13475273426270778455+e30acd3a/model.neff index fc4c08dc93815c54b65a5022dae0ebb26cfb86a8..499b8de623d017e2f92747b29f3ae42eee533216 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_13475273426270778455+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_13475273426270778455+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_13728813963059599796+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_13728813963059599796+e30acd3a/model.neff index 55a365f5f45c375c33680e9d883e1da6274c43cc..6d9be9a83c681fc63f3e072230e323a666b7a919 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_13728813963059599796+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_13728813963059599796+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_13783872956654054643+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_13783872956654054643+e30acd3a/model.neff index 326bbe1c50b17cb9de30f018e11b4c7bccd68d35..80934dd0ed14adc8ace279c8bfe0e6a7a84e4bd1 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_13783872956654054643+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_13783872956654054643+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_14097776130612710282+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_14097776130612710282+e30acd3a/model.neff index d1d08144bc3ba539e9bcdd18c97011a22eec4ef4..5b8d7fb272275c3dd74a1bf09dbd126b5820ce7d 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_14097776130612710282+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_14097776130612710282+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_14155105943057125178+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_14155105943057125178+e30acd3a/model.neff index 796adabf13a93aeb07c0bde48280e63b7e4c1300..37ddbf31702432728f0ed311595be8dcdbdd8035 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_14155105943057125178+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_14155105943057125178+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_14173682194645188821+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_14173682194645188821+e30acd3a/model.neff index 448dcffb7fda3936b8ece415d1945d42c43f601a..bc1e1dbd50babbfefa64ba4ba152cc77125e3c72 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_14173682194645188821+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_14173682194645188821+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_14274833313744358348+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_14274833313744358348+e30acd3a/model.neff index 7e389beefbceb6370cb351ab32327ffee21acf64..c87ecc7697338a7d7ea306ec96d94ca897078030 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_14274833313744358348+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_14274833313744358348+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_14469716438804935215+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_14469716438804935215+e30acd3a/model.neff index 196577020592fd1ca14b70038e4d13abdf8396de..aeaf63b85d4f5771a7bf3d61015d1165a424b21b 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_14469716438804935215+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_14469716438804935215+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_14739672283771660808+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_14739672283771660808+e30acd3a/model.neff index 641fede92b006fd72aba9a1c3f1eae24ce761e13..f3b9e3164845a519f397de1b4d0d9349ffe0e7be 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_14739672283771660808+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_14739672283771660808+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_15442663025941492357+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_15442663025941492357+e30acd3a/model.neff index 2b5559647d455a7e96a420417c981e0a3cc1e0fd..f1204b3e997e72f0087f8249d5e7d3e98d68ef18 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_15442663025941492357+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_15442663025941492357+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_15557971880365771457+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_15557971880365771457+e30acd3a/model.neff index 6290c6a5303a9bc6bb60bbbe805190cb54cf75da..2a30b88f402f414bdd5b475fd65e5498a58bdb76 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_15557971880365771457+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_15557971880365771457+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_15564425602929126510+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_15564425602929126510+e30acd3a/model.neff index 84ba6792fa9c64213a47cb78a281571af2d43202..8ab595e68b6664900c0446a464a45434d3408a5c 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_15564425602929126510+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_15564425602929126510+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_15589838460896944293+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_15589838460896944293+e30acd3a/model.neff index 948ea2bbe796160e857f3c576adccee2b19a0440..57fb691542f27a31a0df7e9b5837b0d4dd13699c 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_15589838460896944293+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_15589838460896944293+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_15654572849171857535+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_15654572849171857535+e30acd3a/model.neff index 5811537ec2139f8ffbaec7b70e13c92f74fecaec..e798fe6b6418061d4d90b054786d9cfdcaa9ae89 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_15654572849171857535+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_15654572849171857535+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_15767383571209512795+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_15767383571209512795+e30acd3a/model.neff index 296c959558f86bde428e96499fe3be422244a13a..ac483edf15393c16739c5e3ca665527e6e089672 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_15767383571209512795+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_15767383571209512795+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_15888836045088309511+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_15888836045088309511+e30acd3a/model.neff index bf9f16e75f2978788f7b7b06cd72774480874bcd..75ef68277991a9e752e086026ba3f8e3ffd40d7c 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_15888836045088309511+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_15888836045088309511+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_15929451261464042997+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_15929451261464042997+e30acd3a/model.neff index 7ebef60f41edfc5598a09ebfa74abe6fbafe4d69..e55e392f1b162b5474eaac86d17f06bec7a5d58d 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_15929451261464042997+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_15929451261464042997+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_16016505958416521648+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_16016505958416521648+e30acd3a/model.neff index b16d46f82add590b0b46901ec1970b6dc0f7ab3c..db08da922a480c13d308aa233131d0cace68e07d 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_16016505958416521648+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_16016505958416521648+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_16053163479112702088+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_16053163479112702088+e30acd3a/model.neff index 9988c8f3cbdc6d98c4e6c41bfae069c30902ec7c..ea15d61a79eb9e1b71a9db3ca921b47e4c35708d 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_16053163479112702088+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_16053163479112702088+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_16204387852795926216+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_16204387852795926216+e30acd3a/model.neff index c44d18332ddca0cd08f618f3934837f405fbc054..1c086d79f117e4aa9c6fa11367e25e3359249309 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_16204387852795926216+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_16204387852795926216+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_1656719109221189948+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_1656719109221189948+e30acd3a/model.neff index 5f4eac4cc4514d0276dc758950e59b4762986608..dbf841f08cfaa5757a7e89c6febd1dc27d6b9460 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_1656719109221189948+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_1656719109221189948+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_16599571375348449904+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_16599571375348449904+e30acd3a/model.neff index edf7b270d229f79cf7758a2314fbe125abff1b6f..c3b6d10926658932e5d3fa8bb47647ace13a07de 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_16599571375348449904+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_16599571375348449904+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_1663757798483801648+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_1663757798483801648+e30acd3a/model.neff index 8830718bf03ce4df7a9bb5b9441e8b49f377542c..50703a2121e8c0163537d2f5a8ff6d01d39a81d1 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_1663757798483801648+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_1663757798483801648+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_16738296820980389103+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_16738296820980389103+e30acd3a/model.neff index 9dd82fbd81310ec01187bd1fe7cd306b621f4057..1804fb6f4b7fba90f81509fc8e29fc671b8bfe69 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_16738296820980389103+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_16738296820980389103+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_1696825468766062114+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_1696825468766062114+e30acd3a/model.neff index aae162c144464e27370625381a78dcc5300b908c..7f65e63fb1ed7f85a0618dfd52e923827d5df766 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_1696825468766062114+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_1696825468766062114+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_17476209562158013765+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_17476209562158013765+e30acd3a/model.neff index c14b17c9e9cc96e87cf370612f25df58f8f8a95c..a06641c3d164c710a4303867ee266240e3b0691d 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_17476209562158013765+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_17476209562158013765+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_17781596253725927902+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_17781596253725927902+e30acd3a/model.neff index 1fd6ee583e24a046385b18b4a94294444995cc54..0ea9038cc66ff9aba2882bc2a2f8b0914a6fb809 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_17781596253725927902+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_17781596253725927902+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_1795740353831178306+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_1795740353831178306+e30acd3a/model.neff index 51926b5e0eb0392a7502cfb09977ed2c855014a6..92cb4f8d85c9edd6633c2e557c1f84f0ec02e31a 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_1795740353831178306+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_1795740353831178306+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_1800832390737682969+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_1800832390737682969+e30acd3a/model.neff index 22b77389e1b8e2a31161a05f553f5f0bfab77aeb..66b9eb0a426b2676db21cf86f79b9d180e58639c 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_1800832390737682969+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_1800832390737682969+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_1805481651134498710+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_1805481651134498710+e30acd3a/model.neff index d5820ac39645073ab7b6436c787ee2c99381534d..69c09df2543d6c9129897e93e811b7df2c00b33b 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_1805481651134498710+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_1805481651134498710+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_1805728312057401221+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_1805728312057401221+e30acd3a/model.neff index 04594896fe9aa866522b705d04928c43c62abe34..088e6805af1651ef25864bc55da4a32071b9ec9a 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_1805728312057401221+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_1805728312057401221+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_18070749384531238695+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_18070749384531238695+e30acd3a/model.neff index 2ef5378aa478bd274c48d5dc1401aa24b9a434bc..c2f307ee237883e5f3d6838167e6ee9731ce71a3 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_18070749384531238695+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_18070749384531238695+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_18134517905922687964+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_18134517905922687964+e30acd3a/model.neff index 96660078deb07d2f4585dfa1b93f93615de724da..d6b066b1084fa02e626ec744dc1a0edc14ea5970 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_18134517905922687964+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_18134517905922687964+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_18207475553554060390+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_18207475553554060390+e30acd3a/model.neff index f91ae3ce32b017d4a8452b64f9f669e5f31c04e7..537ca5d92244cec375a473d2718adc339e7aa8f6 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_18207475553554060390+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_18207475553554060390+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_2157613318347839507+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_2157613318347839507+e30acd3a/model.neff index 58decc7694375039dd18030d237da252912e759c..2ee2232a5f43e7f9bb51e77d17ca267ed1370419 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_2157613318347839507+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_2157613318347839507+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_2203200257461345827+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_2203200257461345827+e30acd3a/model.neff index 72b7335cfb0dc62d786765f59fe18bed89f3b40d..976c95644ba178414d864eb650b62a0ade844b5e 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_2203200257461345827+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_2203200257461345827+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_2410426589418994260+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_2410426589418994260+e30acd3a/model.neff index 7e34f66d3ce93fc13d3325a95339eb081ecded0c..a75736b15e3fd10d46f8849b7d139eb39a27f310 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_2410426589418994260+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_2410426589418994260+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_2426414314187505427+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_2426414314187505427+e30acd3a/model.neff index 6f6aa2a9419eaa9b1a35c7c16adcec423f00125f..0bfe7a2cbda1e4b3695b97ea7c9261f731fec634 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_2426414314187505427+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_2426414314187505427+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_2576759111807165188+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_2576759111807165188+e30acd3a/model.neff index 8ea550d1587d39a18b1ba354549380bd900b5df0..84cee4077f6d4377a2ef316aea3be163488be33a 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_2576759111807165188+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_2576759111807165188+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_2650447136112456251+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_2650447136112456251+e30acd3a/model.neff index 26124d783c87a8b8e5176a33a805974660da560f..05b9a361cb6ea56bcdd0101be269235b4346a9ed 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_2650447136112456251+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_2650447136112456251+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_2682953264353234433+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_2682953264353234433+e30acd3a/model.neff index 05be5aa39f4176645e89bd75f4b2f295b8a4a1c5..b85efc4a60f84313283cbc1b2f9627319eee2507 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_2682953264353234433+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_2682953264353234433+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_2699280712748688265+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_2699280712748688265+e30acd3a/model.neff index e0ff14b5f46f4bb4a9c90df64d1eca6fc47f26ba..87346ff9785398c1b8e6e31b929baabd1fabc840 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_2699280712748688265+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_2699280712748688265+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_2811495285170804454+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_2811495285170804454+e30acd3a/model.neff index 9598c884815164ce992a9050bd41c3c2883b8d37..6a92060783f3d8d763c4fe0e5ae4a8941cf53ae9 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_2811495285170804454+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_2811495285170804454+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_302950992539679003+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_302950992539679003+e30acd3a/model.neff index f665d08d4fd6704cf29aa8a4a440edf58d1bc378..4fbd9f43206dab31e378027ae0e81a63c3158c9a 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_302950992539679003+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_302950992539679003+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_3b4b3de0fa211055bd74+47866c66/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_3b4b3de0fa211055bd74+47866c66/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..d610711cec2ed09aa9d1c82596a3669ff45ad3ac --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_3b4b3de0fa211055bd74+47866c66/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_e782f974-431c-47fc-85de-67bacaffd1e2/compiler_workdir/NeuronRMSNorm/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_3b4b3de0fa211055bd74+47866c66/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_3b4b3de0fa211055bd74+47866c66/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_3b4b3de0fa211055bd74+47866c66/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_3b4b3de0fa211055bd74+47866c66/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..6a2752c25c9f7e1092e2b55c010694e168610ce3 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_3b4b3de0fa211055bd74+47866c66/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c774c6626fc91684e924b843c1fe807dadbbac6d8a8e19443663b43c13eac35 +size 1165 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_3b4b3de0fa211055bd74+47866c66/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_3b4b3de0fa211055bd74+47866c66/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..af17c600879a4702c412a62d3c282438a861694c --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_3b4b3de0fa211055bd74+47866c66/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11ae4468e1db37708f9b7934b5a6286ddd0850127bff7d219413318a55240e2b +size 103424 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_3b4b3de0fa211055bd74+47866c66/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_3b4b3de0fa211055bd74+47866c66/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..60e96dff130f20bb0c6d248c33fd5b8b2e601d44 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_3b4b3de0fa211055bd74+47866c66/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dad6b9bbeef7b385cb625f7f55027bdf7c2d856696f045c28f868a768e037e45 +size 104320 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_4013984518400530928+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_4013984518400530928+e30acd3a/model.neff index f23312368c25a6c7cec0424ff55df835ff8a4737..063315f759e85fba74664f4f01dc83a503ffa8b5 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_4013984518400530928+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_4013984518400530928+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_4052264821257342969+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_4052264821257342969+e30acd3a/model.neff index 6c1bcef816a242bc6630e81fc070efa4009b29d9..f85364ed112b88dbadecc007502ff642ca19eb2e 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_4052264821257342969+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_4052264821257342969+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_4166695227062860792+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_4166695227062860792+e30acd3a/model.neff index 5caea38d2ba3711dc046ad2ee5aa4506bb4a9a9a..91fc32fda0493dd5583cacb31ff7cf2917bd07aa 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_4166695227062860792+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_4166695227062860792+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_4232557591140652245+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_4232557591140652245+e30acd3a/model.neff index 07d6df74d4134149f56d279a3b531d62d0b19778..0aeaee8c383fcd4e3d19ba272f1437bb9ca4f38d 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_4232557591140652245+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_4232557591140652245+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_4606491920565381523+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_4606491920565381523+e30acd3a/model.neff index c91f3daf321bb31a740e6df8cbbcb31aec4dc187..97bbce01938d6f33c91bf7ecde988e4bba0e4265 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_4606491920565381523+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_4606491920565381523+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_4608333438381222049+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_4608333438381222049+e30acd3a/model.neff index cb2f07f5f8e7a8fbc17c0c1ef22fe1e5264be5bd..4a63e09bcb2093fac2a2d068b4c102681636345f 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_4608333438381222049+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_4608333438381222049+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_4683016291147007110+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_4683016291147007110+e30acd3a/model.neff index 70950b229a1af6567960b87778ebe4a9b9dadbd3..64be36f57e886e83c0a58ccd05369325cfcd099b 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_4683016291147007110+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_4683016291147007110+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_4816422433712790690+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_4816422433712790690+e30acd3a/model.neff index 06f165d667d1e26e59a3353211edf491a9a61b75..d1a577cb04b2e9145ded440a538d705815e7950c 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_4816422433712790690+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_4816422433712790690+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_4842952601010372434+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_4842952601010372434+e30acd3a/model.neff index 5d9e815d1703eeea6988811a748b9e5603c7799c..0d70a8510ef729b68876432749bac4d9c5bcd5d1 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_4842952601010372434+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_4842952601010372434+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_4939721357859779936+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_4939721357859779936+e30acd3a/model.neff index 938d599a939fd55ab41661a216eaf94871664694..174fb7bfc13fc501f0b249158626493ad913e6cb 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_4939721357859779936+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_4939721357859779936+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_5397223456911199516+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_5397223456911199516+e30acd3a/model.neff index 2031c38a5b11a4827b087504d74ed7aae53a8707..32d6347e6aead993056cbfc3cc4558424d9e8c05 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_5397223456911199516+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_5397223456911199516+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_542c0bd110414da7bca1+f931dc9f/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_542c0bd110414da7bca1+f931dc9f/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..ca1ab14cb6d579a778a202466fc8ce8e05f7fd57 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_542c0bd110414da7bca1+f931dc9f/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_3a24e800-a920-4789-b290-28e4d09f6ffd/compiler_workdir/AttentionModelWrapper/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_542c0bd110414da7bca1+f931dc9f/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_542c0bd110414da7bca1+f931dc9f/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_542c0bd110414da7bca1+f931dc9f/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_542c0bd110414da7bca1+f931dc9f/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..8496781848c02ecd308b5b2c5a3c342b702425e7 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_542c0bd110414da7bca1+f931dc9f/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51e79a5cc325f9a317d158d931d75ec73ae1cfb16a4d83343a2593edd95d0044 +size 8979 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_542c0bd110414da7bca1+f931dc9f/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_542c0bd110414da7bca1+f931dc9f/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..804da2143a6562b3f67c83e63eec07df392d7403 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_542c0bd110414da7bca1+f931dc9f/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd37484c9b1eaf5f5068590ca15723a05bbe17fc92d2bb05c2a580a60671f1c3 +size 246784 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_542c0bd110414da7bca1+f931dc9f/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_542c0bd110414da7bca1+f931dc9f/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..0be24b7a608f18b7ec2edb1dd152948faaaa7085 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_542c0bd110414da7bca1+f931dc9f/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb2304444fbe1afe703f4542f5739c976aff0ce3c346e959ac91483fa333ea11 +size 249608 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_5724544066416780383+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_5724544066416780383+e30acd3a/model.neff index 0c1842f53122ba03b81cf37bbe390cf73bb26372..7dc9c022b969b48b3d6d76bdda46cc6b2d1e7dd5 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_5724544066416780383+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_5724544066416780383+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_5765916484866650909+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_5765916484866650909+e30acd3a/model.neff index 53aad5efdb38fe5c6e89de37a5817d030f24fb7b..04e2506bea84126a9af5f1d59a3bc0c17fe40332 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_5765916484866650909+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_5765916484866650909+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_6368409506294683105+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_6368409506294683105+e30acd3a/model.neff index 2130e4defe4e779d5a8a22aeebf0118aac61b7fb..afe16c666b3e5934a2ad4200906ac9b136681b90 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_6368409506294683105+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_6368409506294683105+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_6375780938374445148+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_6375780938374445148+e30acd3a/model.neff index 15770df6f860566349fa78cc0ce784e2d70391b5..41efb39e1e508d3c84596f1e36ab67416b102036 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_6375780938374445148+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_6375780938374445148+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_6417297157123190451+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_6417297157123190451+e30acd3a/model.neff index 352f5e99ed4cc81a17214f60b40fe6db092495f7..32e8160340e7bb6b5011c8b297bb10d1614c1539 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_6417297157123190451+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_6417297157123190451+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_665145310648895012+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_665145310648895012+e30acd3a/model.neff index d3335e5995dcf1da795ca6d76c6c2bb72c9281a4..59a6ab1b495f8b5985134904c3277779a016fee3 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_665145310648895012+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_665145310648895012+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_6711765444274762085+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_6711765444274762085+e30acd3a/model.neff index f3944ebfa7c8f1ab4556768264e3bb40ede3cdb7..cf02da59611bcd042c3976d3e1336cd0e4a2d30e 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_6711765444274762085+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_6711765444274762085+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_7049343467335223052+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_7049343467335223052+e30acd3a/model.neff index 6c673dd7b44a4b5451fce39a7b8b93e2e35f3cde..521b66514f74d235e0627a1907aa29a49273fc93 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_7049343467335223052+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_7049343467335223052+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_7877866862268499659+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_7877866862268499659+e30acd3a/model.neff index 3da2dd19f68d1e2541d50674dc10565461ffaec4..fca86dc84e56d261f502f3c52f380e1d35ca0eb0 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_7877866862268499659+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_7877866862268499659+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_793722132614349680+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_793722132614349680+e30acd3a/model.neff index a805ff511f6301c1a722a36326524101b91a9aa4..6d77f4c897d03dbbce8c863fc1f6c3993ca6a1cc 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_793722132614349680+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_793722132614349680+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_8142158350842512240+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_8142158350842512240+e30acd3a/model.neff index a901bf9020a5486143b63176193bf238ca0214b1..e2e786d5b16831b3c092fea58fb2732cb012095c 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_8142158350842512240+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_8142158350842512240+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_8146958338478347620+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_8146958338478347620+e30acd3a/model.neff index 99fe98cceb9e4e2e20b7ab956262ac432f1b2c1b..f2e46cbd8fe9f6ceb00cb9e0dd5102e5202adb20 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_8146958338478347620+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_8146958338478347620+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_8245165830758578911+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_8245165830758578911+e30acd3a/model.neff index 5ed851719764a561d656fd828244b97b46b98ef6..da6f66fc0d0f07d9aaf78678fc3a7e80c05fa359 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_8245165830758578911+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_8245165830758578911+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_8679096817551502409+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_8679096817551502409+e30acd3a/model.neff index 4766d01f1aa78ff13ddad5e9614c633250db6cb4..f43a637adf9be06a6fa109aa9fd398c85b3a0458 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_8679096817551502409+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_8679096817551502409+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_8882592820567938515+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_8882592820567938515+e30acd3a/model.neff index d1d6d98cdf3647bb3c9b306415593dc6b1e9a520..fd5ac926cc28f9f1f957b000b93eeeda250852b1 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_8882592820567938515+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_8882592820567938515+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_9168843780052024308+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_9168843780052024308+e30acd3a/model.neff index 0cd17c1abc830548049d54d4edf7ea205065c77d..de35c6a90c1e6c8ba266682d154869d7bb8664c4 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_9168843780052024308+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_9168843780052024308+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_9527558113976496538+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_9527558113976496538+e30acd3a/model.neff index 794faecff6a2bfd4c1b472ac19481b6a2ba0e061..6f6583433dfd361bd47e784439fc347a55602458 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_9527558113976496538+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_9527558113976496538+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_9533302140457976101+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_9533302140457976101+e30acd3a/model.neff index 9bafc434fe6cf32c5beb407c18432dec0e9a2f3a..51840d0a7d69a8a717edb101c25637d700ef907b 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_9533302140457976101+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_9533302140457976101+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_9770544877851564228+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_9770544877851564228+e30acd3a/model.neff index e7b128983e702d3c6b124c358bfcd08cf16a701a..409349fb881c641f1ee42583c851a713b53fcd21 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_9770544877851564228+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_9770544877851564228+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b2bf569c84c54e75eabd+ee54247e/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_b2bf569c84c54e75eabd+ee54247e/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..ddb74a748c2eb1c52f2126808b44bf7cb5cd17c4 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b2bf569c84c54e75eabd+ee54247e/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_56ba1234-480d-4898-a75f-eb7e3d31ca25/compiler_workdir/NeuronLlamaMLP/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b2bf569c84c54e75eabd+ee54247e/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_b2bf569c84c54e75eabd+ee54247e/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b2bf569c84c54e75eabd+ee54247e/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_b2bf569c84c54e75eabd+ee54247e/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..e6084ce6d845ff91b01c0cbfd758d24019f38ef9 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b2bf569c84c54e75eabd+ee54247e/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a99355de7043422bc03a8df795afecf586999ec744407bc4fbffe63f002ba2df +size 1931 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b2bf569c84c54e75eabd+ee54247e/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_b2bf569c84c54e75eabd+ee54247e/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..5febababecb8a0d3a85412844ef9e1f7f0d78cb6 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b2bf569c84c54e75eabd+ee54247e/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d4440c6cb59304b4bfc8ed0c3e9d15b204e3bb0006a38f670c7005fbe036d22f +size 134144 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b2bf569c84c54e75eabd+ee54247e/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_b2bf569c84c54e75eabd+ee54247e/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..2a4ef4afb76be7ab29f6dc7d20dd7cc3ba3b5468 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b2bf569c84c54e75eabd+ee54247e/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8b00d3f8847a6064ba8158b75b52ee8ec2811b23117878e654665f3be3d55e4 +size 136222 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_c6050ed2a2db99f0f447+45a194b0/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_c6050ed2a2db99f0f447+45a194b0/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..8d90dcd122435f6034214ad3e89dbdf2d2a548ce --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_c6050ed2a2db99f0f447+45a194b0/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_1164f85d-e63a-4327-932d-334a32225048/compiler_workdir/DecoderLayerWrapper/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_c6050ed2a2db99f0f447+45a194b0/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_c6050ed2a2db99f0f447+45a194b0/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_c6050ed2a2db99f0f447+45a194b0/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_c6050ed2a2db99f0f447+45a194b0/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..07bf62460bc7813e4a95f208598e2fa4ae551d41 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_c6050ed2a2db99f0f447+45a194b0/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76dc525043c8d0fc6b46382b17659e6a2f2e981788c635021a8a61aa832d2866 +size 14480 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_c6050ed2a2db99f0f447+45a194b0/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_c6050ed2a2db99f0f447+45a194b0/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..b89a72b1b2ac12d183ed7be8d8c0fad2b63a92fb --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_c6050ed2a2db99f0f447+45a194b0/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c7a544d28722fc3b6081a9321168b8351e4c17290a9d0d241492898f3faf391 +size 267264 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_c6050ed2a2db99f0f447+45a194b0/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_c6050ed2a2db99f0f447+45a194b0/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..317eea8ce92964296dbde97bcca430d8399f5b17 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_c6050ed2a2db99f0f447+45a194b0/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74cb8895f1d6812690d4418c95ccda83af93b16e0d0424aa95f5ee87ee400ff2 +size 272962 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_daebaecf4ac674417399+c0fe283e/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_daebaecf4ac674417399+c0fe283e/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..2a20ac1cf64c3ccc812b3f30d2d31631e5a0dfe2 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_daebaecf4ac674417399+c0fe283e/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_d1724f78-aa29-449e-b82a-1d22b985cfba/compiler_workdir/Llama3RotaryEmbedding/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_daebaecf4ac674417399+c0fe283e/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_daebaecf4ac674417399+c0fe283e/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_daebaecf4ac674417399+c0fe283e/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_daebaecf4ac674417399+c0fe283e/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..f1e2473c54cffac949ecf31d3b6ec636c290545b --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_daebaecf4ac674417399+c0fe283e/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4a29d4cf27aec9af90e53324eff9ee69e033679482f4016a9dd3e5416c7c68a +size 1584 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_daebaecf4ac674417399+c0fe283e/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_daebaecf4ac674417399+c0fe283e/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..49c9cc0987c41a87745bb2d0a2f11a0f5efb3297 Binary files /dev/null and b/neuronxcc-2.21.33363.0+82129205/MODULE_daebaecf4ac674417399+c0fe283e/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_df7cef7d783cd9aa888e+2528ca2d/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_df7cef7d783cd9aa888e+2528ca2d/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..0d38cc6263811e70c3dcd57624477be52b39a6c4 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_df7cef7d783cd9aa888e+2528ca2d/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_02901cb8-de75-490b-ba24-7ce34c03dbc1/compiler_workdir/AttentionModelWrapper/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_df7cef7d783cd9aa888e+2528ca2d/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_df7cef7d783cd9aa888e+2528ca2d/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_df7cef7d783cd9aa888e+2528ca2d/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_df7cef7d783cd9aa888e+2528ca2d/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..a11cc062167102b0f4faa9fdf16388d72d9e0015 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_df7cef7d783cd9aa888e+2528ca2d/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4590b82eb167fd46963abdbb717cc18cb9511961a3393d100eeacdbdc45b0243 +size 11280 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_df7cef7d783cd9aa888e+2528ca2d/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_df7cef7d783cd9aa888e+2528ca2d/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..9a8d2d7808114e9093cd2aa0acd5368f663d819e --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_df7cef7d783cd9aa888e+2528ca2d/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8fe86eb29892ba2983d947156de541dec93aa0bc653e88ed5b15f202a24bf250 +size 1444864 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_df7cef7d783cd9aa888e+2528ca2d/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_df7cef7d783cd9aa888e+2528ca2d/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..ff24399aeb2ed977f5500f3b7905739efcb81aaa --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_df7cef7d783cd9aa888e+2528ca2d/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:532965bbef28aabda0275dd19d416add880688155d8f360334ae5278a8ac1984 +size 1447845 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fb29e0ab34c0bf999b52+1745662a/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_fb29e0ab34c0bf999b52+1745662a/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..64e4730ac5b58007560f53b3433dd2c611569ad3 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fb29e0ab34c0bf999b52+1745662a/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_ef296a04-24d8-4d43-adc3-66cb0a17270d/compiler_workdir/DecoderLayerWrapper/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fb29e0ab34c0bf999b52+1745662a/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_fb29e0ab34c0bf999b52+1745662a/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fb29e0ab34c0bf999b52+1745662a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_fb29e0ab34c0bf999b52+1745662a/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..be5670c6258a72b3253c262c04b0a3c084a1810d --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fb29e0ab34c0bf999b52+1745662a/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3dc99fd8fa13fb76a08ae04c8275958cd87878cfbf86f63ae613d7efcf775bfc +size 29412 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fb29e0ab34c0bf999b52+1745662a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_fb29e0ab34c0bf999b52+1745662a/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..a30afcb8819390f4474387873ba4974ec2bfff62 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fb29e0ab34c0bf999b52+1745662a/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49d36548358b2fc59c1fb40e6619319abd958c87a7d12040d01b10d65b31ce50 +size 328704 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fb29e0ab34c0bf999b52+1745662a/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_fb29e0ab34c0bf999b52+1745662a/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..2f0073c336bd1e9178787bfca5800aae3a93446f --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fb29e0ab34c0bf999b52+1745662a/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a7d893c472d17192fcd135bc84b7ad408d35b5a4c7095093c14f54a78f10f18 +size 334452