diff --git a/.gitattributes b/.gitattributes index 57f93f4ae7a99b18a2f86fe971bf78674147e78a..618eb3a57ea7bf850bb31e13a14cdf11b1843a6e 100644 --- a/.gitattributes +++ b/.gitattributes @@ -16675,3 +16675,15 @@ neuronxcc-2.21.33363.0+82129205/MODULE_98dfce3ac3eb87801e0a+a02c3a36/model.neff neuronxcc-2.21.33363.0+82129205/MODULE_98dfce3ac3eb87801e0a+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text neuronxcc-2.21.33363.0+82129205/MODULE_cd7047dabeab8994abfe+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text neuronxcc-2.21.33363.0+82129205/MODULE_cd7047dabeab8994abfe+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_34d548ac7fd4d949f1fa+541be28a/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_34d548ac7fd4d949f1fa+541be28a/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_415b00061ed1b9f60492+1f479a5f/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_415b00061ed1b9f60492+1f479a5f/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_4ecd5c36fb60862f06e4+1dd6a3bf/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_4ecd5c36fb60862f06e4+1dd6a3bf/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_7ded0b9212ceac8698f4+477b7c62/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_7ded0b9212ceac8698f4+477b7c62/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_99a4c8ce2f07c82d25d4+76eb1f45/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_99a4c8ce2f07c82d25d4+76eb1f45/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_9cd1f2706a9a7386b4a1+edf8250f/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_9cd1f2706a9a7386b4a1+edf8250f/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/2da7a00f0478d50ae1e7f75f085c5b2773b5f355f427c61cf34cb6febd629d96/8522c8da70b6768237a5.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/2da7a00f0478d50ae1e7f75f085c5b2773b5f355f427c61cf34cb6febd629d96/8522c8da70b6768237a5.json new file mode 100644 index 0000000000000000000000000000000000000000..c800acf339d06789b0ccf320625f1b92f179c49b --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/2da7a00f0478d50ae1e7f75f085c5b2773b5f355f427c61cf34cb6febd629d96/8522c8da70b6768237a5.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": "0.4.6.dev1", + "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/0.4.6.dev1/441269935591cad8d370e512c0b93cdd2fce6247c40e5a4866d872ee5338b0de/0e89b203802f7fabc982.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/441269935591cad8d370e512c0b93cdd2fce6247c40e5a4866d872ee5338b0de/0e89b203802f7fabc982.json new file mode 100644 index 0000000000000000000000000000000000000000..610af520bbac634d047de47b210d2924ab6255ba --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/441269935591cad8d370e512c0b93cdd2fce6247c40e5a4866d872ee5338b0de/0e89b203802f7fabc982.json @@ -0,0 +1,81 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "unsloth/gemma-3-270m-it", + "_sliding_window_pattern": 6, + "_task": "text-generation", + "architectures": [ + "Gemma3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "attn_logit_softcapping": null, + "dtype": "bfloat16", + "final_logit_softcapping": null, + "head_dim": 256, + "hidden_activation": "gelu_pytorch_tanh", + "hidden_size": 640, + "initializer_range": 0.02, + "intermediate_size": 2048, + "layer_types": [ + "sliding_attention", + "sliding_attention", + "sliding_attention", + "sliding_attention", + "sliding_attention", + "full_attention", + "sliding_attention", + "sliding_attention", + "sliding_attention", + "sliding_attention", + "sliding_attention", + "full_attention", + "sliding_attention", + "sliding_attention", + "sliding_attention", + "sliding_attention", + "sliding_attention", + "full_attention" + ], + "max_position_embeddings": 32768, + "model_type": "gemma3_text", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 1, + "capacity_factor": null, + "checkpoint_id": "unsloth/gemma-3-270m-it", + "checkpoint_revision": "23cf460f6bb16954176b3ddcc8d4f250501458a9", + "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": "0.4.6.dev1", + "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": 18, + "num_key_value_heads": 1, + "query_pre_attn_scalar": 256, + "rms_norm_eps": 1e-06, + "rope_local_base_freq": 10000.0, + "rope_scaling": null, + "rope_theta": 1000000.0, + "sliding_window": 512, + "unsloth_fixed": true, + "use_bidirectional_attention": false, + "use_cache": true, + "vocab_size": 262144 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/4ab8140bc7eb4a553d95855c5c2be2cf8c0fbab21b823d76183b6f51e98b6fc5/9205d47535c8c8ab72fa.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/4ab8140bc7eb4a553d95855c5c2be2cf8c0fbab21b823d76183b6f51e98b6fc5/9205d47535c8c8ab72fa.json new file mode 100644 index 0000000000000000000000000000000000000000..9a62f49a750b9ed656cd78c522ac3241d70eb4e3 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/4ab8140bc7eb4a553d95855c5c2be2cf8c0fbab21b823d76183b6f51e98b6fc5/9205d47535c8c8ab72fa.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": "0.4.6.dev1", + "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/0.4.6.dev1/4cb7aff9e2a15c151396f2b684013e39d6739f0dec83e5c9dabbfe9d5fcf77b7/6f008bdec78b8ad51aee.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/4cb7aff9e2a15c151396f2b684013e39d6739f0dec83e5c9dabbfe9d5fcf77b7/6f008bdec78b8ad51aee.json new file mode 100644 index 0000000000000000000000000000000000000000..082c8cbe180d83f9944b013e2e37d1aa5257e7db --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/4cb7aff9e2a15c151396f2b684013e39d6739f0dec83e5c9dabbfe9d5fcf77b7/6f008bdec78b8ad51aee.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": "0.4.6.dev1", + "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/0.4.6.dev1/6454afdf3e9d66c7226c13a575b718845c25e53b0699600ba2bb4f883e9d841b/4725d0f98e9d733a5354.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/6454afdf3e9d66c7226c13a575b718845c25e53b0699600ba2bb4f883e9d841b/4725d0f98e9d733a5354.json new file mode 100644 index 0000000000000000000000000000000000000000..c904128e4983417ab36b72c597ddf69e8d5f666f --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/6454afdf3e9d66c7226c13a575b718845c25e53b0699600ba2bb4f883e9d841b/4725d0f98e9d733a5354.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": "0.4.6.dev1", + "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/0.4.6.dev1/6454afdf3e9d66c7226c13a575b718845c25e53b0699600ba2bb4f883e9d841b/79d1effc3dea92153467.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/6454afdf3e9d66c7226c13a575b718845c25e53b0699600ba2bb4f883e9d841b/79d1effc3dea92153467.json new file mode 100644 index 0000000000000000000000000000000000000000..50c446fb0e50ce8fb07edff591d82cc5cf792177 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/6454afdf3e9d66c7226c13a575b718845c25e53b0699600ba2bb4f883e9d841b/79d1effc3dea92153467.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": "0.4.6.dev1", + "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/0.4.6.dev1/6454afdf3e9d66c7226c13a575b718845c25e53b0699600ba2bb4f883e9d841b/dd5885d59f14f083843e.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/6454afdf3e9d66c7226c13a575b718845c25e53b0699600ba2bb4f883e9d841b/dd5885d59f14f083843e.json new file mode 100644 index 0000000000000000000000000000000000000000..6018dcd19916dc54ca17b6efc176b04983c0cffa --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/6454afdf3e9d66c7226c13a575b718845c25e53b0699600ba2bb4f883e9d841b/dd5885d59f14f083843e.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": "0.4.6.dev1", + "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/0.4.6.dev1/73707b485eab9008c7aba7f5dad0ce2384ac685318d5f888c12fa0d81ed90b19/597478f258749fdff84c.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/73707b485eab9008c7aba7f5dad0ce2384ac685318d5f888c12fa0d81ed90b19/597478f258749fdff84c.json new file mode 100644 index 0000000000000000000000000000000000000000..0282b53fc9ffe329d59ff479d067513c5fa8ca8a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/73707b485eab9008c7aba7f5dad0ce2384ac685318d5f888c12fa0d81ed90b19/597478f258749fdff84c.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": "0.4.6.dev1", + "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+82129205/0_REGISTRY/0.4.6.dev1/7518518c7e077820070186deda960d8cc49db068cdf0ac70664098fa2b6b698c/89ebc1898ab574188c6e.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/7518518c7e077820070186deda960d8cc49db068cdf0ac70664098fa2b6b698c/89ebc1898ab574188c6e.json new file mode 100644 index 0000000000000000000000000000000000000000..7a6e428330f2e72b2f8f52a7769a323383f97825 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/7518518c7e077820070186deda960d8cc49db068cdf0ac70664098fa2b6b698c/89ebc1898ab574188c6e.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": "0.4.6.dev1", + "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/0.4.6.dev1/920f44ce6d3e004d1ce547ae06644f7be262180644b04573153aa15d98742edc/3c1703b17314b02dcfb5.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/920f44ce6d3e004d1ce547ae06644f7be262180644b04573153aa15d98742edc/3c1703b17314b02dcfb5.json new file mode 100644 index 0000000000000000000000000000000000000000..acb53436bf0d5383875a3847b0260ef82087dc55 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/920f44ce6d3e004d1ce547ae06644f7be262180644b04573153aa15d98742edc/3c1703b17314b02dcfb5.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": "0.4.6.dev1", + "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/0.4.6.dev1/9c24a2080ec26e9c74d5bd90866f74131aa4d5c975415b96f5e6600f081d5647/357aab4e03a8b3b961c3.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/9c24a2080ec26e9c74d5bd90866f74131aa4d5c975415b96f5e6600f081d5647/357aab4e03a8b3b961c3.json new file mode 100644 index 0000000000000000000000000000000000000000..2defb3ec95a5dd1bb8b7814b8ca1c58af7101be4 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/9c24a2080ec26e9c74d5bd90866f74131aa4d5c975415b96f5e6600f081d5647/357aab4e03a8b3b961c3.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": "0.4.6.dev1", + "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/0.4.6.dev1/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/b02932ba1b54f85e3d84.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/b02932ba1b54f85e3d84.json new file mode 100644 index 0000000000000000000000000000000000000000..181653c6b8ffdf9be34b5b534610f2b7b3348156 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/b02932ba1b54f85e3d84.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": "0.4.6.dev1", + "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/0.4.6.dev1/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/d113152e0d34c2d365df.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/d113152e0d34c2d365df.json new file mode 100644 index 0000000000000000000000000000000000000000..9b84e0cd2402990e3e7c569df17d2d465bc0837a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/d113152e0d34c2d365df.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": "0.4.6.dev1", + "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/0.4.6.dev1/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/da6950729a5bb39e8b09.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/da6950729a5bb39e8b09.json new file mode 100644 index 0000000000000000000000000000000000000000..9394dacd4355f7db8d23bc51193db70dea7885d2 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/da6950729a5bb39e8b09.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": "5a8abab4a5d6f164389b1079fb721cfab8d7126c", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": true, + "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": "0.4.6.dev1", + "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/0.4.6.dev1/d139acf64685f15794bb983ff6eb881bdd31304bae88b0ce1ed20a54c21f2265/1fff0e9367994bde50df.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/d139acf64685f15794bb983ff6eb881bdd31304bae88b0ce1ed20a54c21f2265/1fff0e9367994bde50df.json new file mode 100644 index 0000000000000000000000000000000000000000..14355f8d17d1dda1d09d6066d5e681c9d60b4291 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/d139acf64685f15794bb983ff6eb881bdd31304bae88b0ce1ed20a54c21f2265/1fff0e9367994bde50df.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": "0.4.6.dev1", + "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/0.4.6.dev1/gemma3_text/unsloth/gemma-3-270m-it/0e89b203802f7fabc982.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/gemma3_text/unsloth/gemma-3-270m-it/0e89b203802f7fabc982.json new file mode 100644 index 0000000000000000000000000000000000000000..610af520bbac634d047de47b210d2924ab6255ba --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/gemma3_text/unsloth/gemma-3-270m-it/0e89b203802f7fabc982.json @@ -0,0 +1,81 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "unsloth/gemma-3-270m-it", + "_sliding_window_pattern": 6, + "_task": "text-generation", + "architectures": [ + "Gemma3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "attn_logit_softcapping": null, + "dtype": "bfloat16", + "final_logit_softcapping": null, + "head_dim": 256, + "hidden_activation": "gelu_pytorch_tanh", + "hidden_size": 640, + "initializer_range": 0.02, + "intermediate_size": 2048, + "layer_types": [ + "sliding_attention", + "sliding_attention", + "sliding_attention", + "sliding_attention", + "sliding_attention", + "full_attention", + "sliding_attention", + "sliding_attention", + "sliding_attention", + "sliding_attention", + "sliding_attention", + "full_attention", + "sliding_attention", + "sliding_attention", + "sliding_attention", + "sliding_attention", + "sliding_attention", + "full_attention" + ], + "max_position_embeddings": 32768, + "model_type": "gemma3_text", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 1, + "capacity_factor": null, + "checkpoint_id": "unsloth/gemma-3-270m-it", + "checkpoint_revision": "23cf460f6bb16954176b3ddcc8d4f250501458a9", + "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": "0.4.6.dev1", + "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": 18, + "num_key_value_heads": 1, + "query_pre_attn_scalar": 256, + "rms_norm_eps": 1e-06, + "rope_local_base_freq": 10000.0, + "rope_scaling": null, + "rope_theta": 1000000.0, + "sliding_window": 512, + "unsloth_fixed": true, + "use_bidirectional_attention": false, + "use_cache": true, + "vocab_size": 262144 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/1fff0e9367994bde50df.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/1fff0e9367994bde50df.json new file mode 100644 index 0000000000000000000000000000000000000000..14355f8d17d1dda1d09d6066d5e681c9d60b4291 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/1fff0e9367994bde50df.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": "0.4.6.dev1", + "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/0.4.6.dev1/llama/llamafactory/tiny-random-Llama-3/79d1effc3dea92153467.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/llama/llamafactory/tiny-random-Llama-3/79d1effc3dea92153467.json new file mode 100644 index 0000000000000000000000000000000000000000..50c446fb0e50ce8fb07edff591d82cc5cf792177 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/llama/llamafactory/tiny-random-Llama-3/79d1effc3dea92153467.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": "0.4.6.dev1", + "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/0.4.6.dev1/llama/unsloth/Llama-3.2-1B-Instruct/b02932ba1b54f85e3d84.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/llama/unsloth/Llama-3.2-1B-Instruct/b02932ba1b54f85e3d84.json new file mode 100644 index 0000000000000000000000000000000000000000..181653c6b8ffdf9be34b5b534610f2b7b3348156 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/llama/unsloth/Llama-3.2-1B-Instruct/b02932ba1b54f85e3d84.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": "0.4.6.dev1", + "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/0.4.6.dev1/llama4_text/tiny-random/llama-4/357aab4e03a8b3b961c3.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/llama4_text/tiny-random/llama-4/357aab4e03a8b3b961c3.json new file mode 100644 index 0000000000000000000000000000000000000000..2defb3ec95a5dd1bb8b7814b8ca1c58af7101be4 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/llama4_text/tiny-random/llama-4/357aab4e03a8b3b961c3.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": "0.4.6.dev1", + "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/0.4.6.dev1/mixtral/dacorvo/Mixtral-tiny/9205d47535c8c8ab72fa.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/mixtral/dacorvo/Mixtral-tiny/9205d47535c8c8ab72fa.json new file mode 100644 index 0000000000000000000000000000000000000000..9a62f49a750b9ed656cd78c522ac3241d70eb4e3 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/mixtral/dacorvo/Mixtral-tiny/9205d47535c8c8ab72fa.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": "0.4.6.dev1", + "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/0.4.6.dev1/phi3/yujiepan/phi-4-tiny-random/8522c8da70b6768237a5.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/phi3/yujiepan/phi-4-tiny-random/8522c8da70b6768237a5.json new file mode 100644 index 0000000000000000000000000000000000000000..c800acf339d06789b0ccf320625f1b92f179c49b --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/phi3/yujiepan/phi-4-tiny-random/8522c8da70b6768237a5.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": "0.4.6.dev1", + "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/0.4.6.dev1/qwen2/Qwen/Qwen2.5-0.5B/6f008bdec78b8ad51aee.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/qwen2/Qwen/Qwen2.5-0.5B/6f008bdec78b8ad51aee.json new file mode 100644 index 0000000000000000000000000000000000000000..082c8cbe180d83f9944b013e2e37d1aa5257e7db --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/qwen2/Qwen/Qwen2.5-0.5B/6f008bdec78b8ad51aee.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": "0.4.6.dev1", + "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/0.4.6.dev1/qwen2/yujiepan/qwen2.5-128k-tiny-random/89ebc1898ab574188c6e.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/qwen2/yujiepan/qwen2.5-128k-tiny-random/89ebc1898ab574188c6e.json new file mode 100644 index 0000000000000000000000000000000000000000..7a6e428330f2e72b2f8f52a7769a323383f97825 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/qwen2/yujiepan/qwen2.5-128k-tiny-random/89ebc1898ab574188c6e.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": "0.4.6.dev1", + "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/0.4.6.dev1/qwen3_moe/optimum-internal-testing/tiny-random-qwen3_moe/3c1703b17314b02dcfb5.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/qwen3_moe/optimum-internal-testing/tiny-random-qwen3_moe/3c1703b17314b02dcfb5.json new file mode 100644 index 0000000000000000000000000000000000000000..acb53436bf0d5383875a3847b0260ef82087dc55 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/qwen3_moe/optimum-internal-testing/tiny-random-qwen3_moe/3c1703b17314b02dcfb5.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": "0.4.6.dev1", + "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/0.4.6.dev1/smollm3/HuggingFaceTB/SmolLM3-3B/597478f258749fdff84c.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/smollm3/HuggingFaceTB/SmolLM3-3B/597478f258749fdff84c.json new file mode 100644 index 0000000000000000000000000000000000000000..0282b53fc9ffe329d59ff479d067513c5fa8ca8a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev1/smollm3/HuggingFaceTB/SmolLM3-3B/597478f258749fdff84c.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": "0.4.6.dev1", + "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 1d15e97dae9403c8c2b16727955bafe63273e8c0..435c2abdd651364499b1f0ac62751c9350712d79 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 b793b3849de631259b43da55ef427b13dce0585d..bde72ffaa2aa79846edaa5d93eb2f252e3972335 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 7da5d2bbffeadb277332cb3f642835b76542d6af..637b8accb2eeda5f6cb831487c8b0caea01467b3 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 dafa334d5b92fa581271b9f45767df550a92f7fe..2b7a4c287ebdb1656bfb7982d57981d0341d2575 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 6fd7358cb87141b2c7584e733f3edfb873b1664d..94051db7208bf785d762b655108b74cd460abe74 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 12476743ab9eb91763c6c100a167d31f5c6ff1fe..311c8ea2e19310e784b642eca461407e62d59c16 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 6cf9cd5b6b371ae1a3daf04356f4a7d578ac0f4c..9d37ef2d75514ffba611661c207b2e1eab5caa4a 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 d728f346f368b4c16230c0ed9666cf2016e8a4c5..1acc57c3bee517306af3a2abdc45e0dd517fee06 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 49109e084eb6d3c2f7ed1c56827362f93b367f10..36e8233e52394b154fc69ca9ccd224e2afca1fe6 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 fb5df1e3be290172affac9f342eaa1c9f9638d4c..3224d48163d743419823f06cf466285301238c9a 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 92fa74c728a63fd066d0327d8625c754882b7f2a..5a4f44d2b5f326b5d926a197f8ae748a949b5214 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 df4fbdc7f1dedefe1d8633ccacde4a632a0f6a82..c4833af468e77f1f559bc16a688887c01947d443 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 6bbd82256f9b82bea773d5d72952239745a67b0e..f8c64737ff4d2c2d5c8498fbd53b74edf7bfadb0 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 3067c1f4f4f6517ac1ea49560922b32d39bbf39c..a4aa658eeb8c2a78ab14c2ff1f81f99936d97645 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 04f0d500e9122bc74ce3230f7ac769d002c700b3..962f243914a38d52790567f1e66c8eecc721cf67 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 98e5fb6f03247d123c4c2d3c303d32b9e790cb33..7a0ed1d194e2440b728f85cfa0557248d5cba403 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 def3d79cd8f3d989428c590d9632000645ce23e8..2bcc3a174ca6a5224cd0b87d2920c3f4cefd6f5e 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 870e02c8b6c0c16349134e1f852fe7f87dfe115f..05c1e52235ab7443ab409ec9169ae0a67e6296ca 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 20b57633ea9c0e0e85550bf776f20ce7ceb75a30..29efb6f5b26cfcb9a8b2de24701df44966b62af3 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 f31c60adf539f97616529925c98cc5b1aa07604b..40b2801575fb52827895d30ad3ef06e8a300899b 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 89fd193855fb4197d234bd36531e3a857dfb375d..63641399c78893e9eef87e54e0b48c6b8af707ac 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 8a4a706ccb541727e4cba8f9b54e4b6c76b94787..7ff49e54b0ead1ab29eef887178269c401ad80b3 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 5978c85778f9ffac4367f16db7d0d99d5a89bc3c..6fb7ea45af29ec97c08f7d652cf2f51ba57e72a8 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 7d77982addd88b2443c35893eb4a9af3e8db307b..164524129c422bde0eef8df7f4b01f4c995748c2 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 0546ec075e2714a4ae1bc9aabc17a7c5eb376a21..754b4cfc257299803a760a2dd1749c0a4e0584bc 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 ebce233460f556ef8da0151da1889fa2379ac9a0..ceb869334a53b3fabad158f7c273aa9f68fb2bf4 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 4cd19ff5519541a9e57f4da0ffd996d28a2db74b..28f1350d29f3dca2345a49605a8b2c014f4218ae 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 c4fc22ccf16aa603ac46967068ad3cc61b633747..0de7ad99af42302f194956d0b6aeaa508972f197 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 b1586653270d554aebc6b6ea984d52dad05310f4..4ca930b1861afb43646b7cde930742a2f69b6f09 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 e21bbb9662dc1af2638135581cc55d11d7cd6610..0f6d7f7d458a415b8763dd785caeba448a847dc6 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 48dba9cdf2b8953946aa4b8883d7fda4c616a8b9..fa23a80bc46721fd1914a725c185d4acec199226 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 4a23b5b25f23db71b6fd00d63657b12dc920c625..47fbf97336fd751dfba553364f2f38aca61d61c9 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 64e8e0621dfe282197936d4e0417df6b546efb3a..6763a132fa1eec6635bae8c4a64937c18adb6a5a 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 652e3bc6bfcfda9b34560bc6043aa5e8688b8822..cfb01d83a442e0d7d002e492dca044cd87602f1f 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 25873f19030d1fb2dfb53785a4b02505c31a3f2d..f230c24d6866c3858250ca318f135c63078b468d 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 60994ae6e8161a0664d41d8ed5b6dffb19eb44e5..bba0860034032c6c17790a205e3fb3800a09b239 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 452190f4bc90064ce2c89472ce5f54ecbf7fe09e..cc4debad2a50d6ab5f7a1e3fa20f540c896c5de5 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 517c3d57e727e4d5d2e2a7267855f4816c0bc4b3..1e90367a3e0aac53aa277109fc5977ca94d462e5 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 391e396b116861967bc6ad1c8f12e91e42517f4b..1c410478789dd90f75e4d479250d0194ce588427 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 3779798b7262813f64fd72cf8c114a4a4e9f5c0d..4dbb02688ef9ae3ae2bc3eb7b6486f4c54bc3d28 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 a44e82d5ec7c97c1f3815b77e84ab3736c3bac2a..3b40fd9bef1a1e504e18b6477961a755f9531b58 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 e3a9bc07931121be026e634ebf48fe4fe86523fa..8ef409e722d647a00cea116d9bdfc8c67fc795ed 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 fe8b1b79a1e2849c2ff80e5cfb51df1854173092..026fb1e3dd805e496b11bc29455df08f173b9ec0 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 61a6fb9160f857db7c615222738b9ef1a06f3b7e..2d297c98ec43c00b3161f8d105320470ac0463d1 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 c088e645684793d7f536fb15df3255dcc23166d3..b570d3c0319a32f03c576c845f735ae3cd04b983 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 a195c7a28cfae11a59c81cc15fccf15b348ee821..a463c0011a8ece70b1eac21776bf8f19855889ab 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 25a5e4cefe1b3761bdb6829144cff453fb43e02d..0a20a239827b3f8d5d43590b20e10ce8cccf8a68 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 8eb43ee5de1dc2a0a5aa091557970a22a62f0b74..e6497d8d5582abf6ca693933e63563ff4d5ed6a8 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 2574a1ae887d502923bdaf1b2a76efdb4119f469..309ce8c85e109b8d9e48b57ce6b4726c2313f95f 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 908d1c3058940bb965dff8bcb222756ce1107336..8ffe66d286445bac1e54c8cf41e641e975f80ab8 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 afb89d9f282f314500893504f66b89b05b71305e..7dab360da833f650a69d299ee87f72ce22eac321 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 0c419d78bcc689840dc10c064bfa4b9bbfc6b30a..dd70bd622e3c9c25642a5f3d25657ff29535657d 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 eb3d3202e68806f6599848af4c158321bce0b301..7ef66754f77c43e7ab116ca92d23a07f00d1b256 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 c33c387bb1ce5b291ba88c37e084284df3699622..5f6764beb108725e0cbc9d6a89e3edf1f585db7c 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 361e2d7bed78906473329dc8feb0f7d4057bc100..1553c4008b17197b895c5a27d6785f16210ed8ab 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 f114a0daa91590af213b9c1e5af0b80c3b168527..760252b18942ef703b3e5c97f523510f4aed3ef5 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_2ce3d3065088729c7d7f+24129607/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_2ce3d3065088729c7d7f+24129607/model.hlo_module.pb index 1b8352e1ee9da452fd713c1a756c86482df2737e..5b31a795fa051fe6382001027af513aab21cff12 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_2ce3d3065088729c7d7f+24129607/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_2ce3d3065088729c7d7f+24129607/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e635e3fd90e62f2f22ffdf53f78d7bdd4fa50499755a35b844ef028afd5c49ab +oid sha256:9a587d5ddb67a42533df39b1b05311d022b9cb500122dfc7b382e9f459a2b148 size 593552 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_2ce3d3065088729c7d7f+24129607/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_2ce3d3065088729c7d7f+24129607/model.neff index bc1eb276cb4434077e8c16083567de69fb0bbfbd..19d57ab036df5a355c0b32336d85a582ca5ff866 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_2ce3d3065088729c7d7f+24129607/model.neff +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_2ce3d3065088729c7d7f+24129607/model.neff @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:fca3a2584696bc4c79d8367cae210bbf0552db0af7bdcb595aec8fc002685b6f +oid sha256:a9a1689ed7d2b0223c8cb4efa62045beb080793c0aece10a790120c2720e031a size 3769344 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 2aa08d946f12f338fbef91dec2206d0d9f1b6520..3be737032057c462d0b982aecdc05addec8ef013 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_34d548ac7fd4d949f1fa+541be28a/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_34d548ac7fd4d949f1fa+541be28a/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..c0b30a27bf6ab423b5030c1bdf590787761d7440 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_34d548ac7fd4d949f1fa+541be28a/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_0d189f0c-2a32-472c-924a-4a4b608e338d/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_34d548ac7fd4d949f1fa+541be28a/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_34d548ac7fd4d949f1fa+541be28a/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_34d548ac7fd4d949f1fa+541be28a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_34d548ac7fd4d949f1fa+541be28a/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..e6084ce6d845ff91b01c0cbfd758d24019f38ef9 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_34d548ac7fd4d949f1fa+541be28a/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_34d548ac7fd4d949f1fa+541be28a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_34d548ac7fd4d949f1fa+541be28a/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..e1c76bf6c79ff9e393e27d48d773828336f46bc8 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_34d548ac7fd4d949f1fa+541be28a/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f5a7623704c025f08b92088742a809ac00d183aa94cccb9656a8c27cfe8268f +size 134144 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_34d548ac7fd4d949f1fa+541be28a/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_34d548ac7fd4d949f1fa+541be28a/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..d8b9ea808a3a2b8be107b1723d9f6cdac630727f --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_34d548ac7fd4d949f1fa+541be28a/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:148bb32c719de8dedc7d3614686ee06bf86a140235254e9b1a5ba226eea60061 +size 136222 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 a5aff362dad3bee7945a424133d78da4e3e05424..e5e8557c70eb56e9fb34eb0bf26d44221d8d89de 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 b8341f3dda6c31e0b8be65719b4e3d2600f32df3..7c82691617d7c5fa30942e34fbddec88e112ad36 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_415b00061ed1b9f60492+1f479a5f/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_415b00061ed1b9f60492+1f479a5f/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..eaae4a7aaf53c2bf4be5ce378945e4075ba36578 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_415b00061ed1b9f60492+1f479a5f/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_78a61d54-fb7a-4519-ab05-760d552e5f67/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_415b00061ed1b9f60492+1f479a5f/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_415b00061ed1b9f60492+1f479a5f/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_415b00061ed1b9f60492+1f479a5f/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_415b00061ed1b9f60492+1f479a5f/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..be5670c6258a72b3253c262c04b0a3c084a1810d --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_415b00061ed1b9f60492+1f479a5f/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_415b00061ed1b9f60492+1f479a5f/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_415b00061ed1b9f60492+1f479a5f/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..dd2f7c81c5cf7e44b9c09edccf7cddcefbd9aaea --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_415b00061ed1b9f60492+1f479a5f/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8c812869008257ea83cf2905d62f637cdda0ad72cc0b05f24d3b046616ca68c +size 328704 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_415b00061ed1b9f60492+1f479a5f/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_415b00061ed1b9f60492+1f479a5f/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..03967eba011b19c58c9fcf3aa4fd26c0b74ee91c --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_415b00061ed1b9f60492+1f479a5f/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd9501806ac4243c96adde981fa37340f45f2397045c7e6cfd0dfb946a076400 +size 334452 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 63f47511ddfb5de1039aecc37b3891444fe94169..abaa4018e5c85fd073d456d8ebc85c486e914e00 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 42a7e3dfc94ca3011527166f8193e94a614c3dfd..d0f3f98f3ecbc0bbaf7c24a79f501ac46a5f6eeb 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 d34af3de493181f90980c169c490c59d58ef8d44..85146c5536f79b29558ae47c0e556acb8e160658 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 cdb0c5812c22df8d1abd5dd97c954b498c717b0a..6612f2f2d5e4211efda76ba1bd85948f6f5b600f 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 e24e81eb309b6823ce1d23c5da955e8e0a7653fc..e397e677b38621e4c4847b8ddcc19df5bda1a4a5 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 b097c11a6263fdccb6fe5b3c0a0d94171156a9c6..1dcaeeb9c80a58560d4d8dcc00880f104e091a86 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 51b4ddfe4f79bfe58224ea10a4f431b0db46330f..17d5eefff9e308359854f2d29fb64ef42205bafe 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 4c336353c0f4366a11d63d1940d175738aa8b0b6..04955733352a3926ebfd33d41f3bf9fe15fc2d70 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_4ecd5c36fb60862f06e4+1dd6a3bf/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_4ecd5c36fb60862f06e4+1dd6a3bf/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..0dcc6976338681cfbd90a5b90c4bb4e83ddcb939 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_4ecd5c36fb60862f06e4+1dd6a3bf/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_685f28d0-1689-4b9b-b368-79c411bfec4b/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_4ecd5c36fb60862f06e4+1dd6a3bf/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_4ecd5c36fb60862f06e4+1dd6a3bf/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_4ecd5c36fb60862f06e4+1dd6a3bf/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_4ecd5c36fb60862f06e4+1dd6a3bf/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..a11cc062167102b0f4faa9fdf16388d72d9e0015 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_4ecd5c36fb60862f06e4+1dd6a3bf/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_4ecd5c36fb60862f06e4+1dd6a3bf/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_4ecd5c36fb60862f06e4+1dd6a3bf/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..f10b75802071f38e5ee194cae3202f8c572eb04a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_4ecd5c36fb60862f06e4+1dd6a3bf/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcd9ff777822eeafc874fbf5311f767bf2b013e49d43108bd2db21a5ec8571b1 +size 1444864 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_4ecd5c36fb60862f06e4+1dd6a3bf/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_4ecd5c36fb60862f06e4+1dd6a3bf/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..efac5124901ee1d6558b8f1775cec67f6fd4260b --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_4ecd5c36fb60862f06e4+1dd6a3bf/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3385d2cbfa9bbe19492e6e8954dc7287eb88ccb4f01d18a24ce2fc0d0ef1e79 +size 1447845 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 5354b643c15823f0b77a82c016402992c63ea290..7fde6cff05ee74e0fdf85e814d301a8d5ba16241 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_5724544066416780383+e30acd3a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_5724544066416780383+e30acd3a/model.neff index 7905d9314429bcbc8d5fb67794d3bb887a48694d..508930a68cb468573b210d3daf83e55b489cf8bf 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 bc138bde352783ad4d837e2290b53318d8e5485f..23811d085c782a0f434b18f4af8e81860338cbbe 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 92ee1d38883193dbc86a0add4c54fbadcc13ad93..5f3d71376dddf4c0d9223d029e32a15f5b45261a 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 5acb25b506b831d0ef9bb3d7ea783121ea5020d4..b945bedf0b95a3c105e4c2f38554ec79810312fc 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 69b87f002013d6667273c88c52280a6d6066cc47..04fa848502b723e268e14c32f7e5c94d090a22c6 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 5dd3c9079cb8f0dd11850475e950d347aded996b..840088dfe9afb15b81d0e8c53b71c3a83c436518 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 f204ed39826246e16f8f1ecb35a1c219b9dd3c45..e54f39b0b72dc0ea9fb135e44242a00e935bf67a 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 fd88f24880e04e1ead99e720ce896cad4993a5ac..872e340fd0fc40da6e9d7d163003107cad7d1f7e 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 fdbd6316a68b572b79f423fc5326a09917d9b0be..7b1376525cc384529b32c647e86acf03681373cb 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 79af444aedc6cd092c36550fb0795b85522b5c3e..7cc9d35f53b0c1067c0c21b6b5f9865bed6a3bcb 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_7ded0b9212ceac8698f4+477b7c62/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_7ded0b9212ceac8698f4+477b7c62/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..6213d4cf166eda39ba44cca81da6653f2bc81bd9 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_7ded0b9212ceac8698f4+477b7c62/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_fb8b7054-9975-40b9-8607-6d51ecebaad9/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_7ded0b9212ceac8698f4+477b7c62/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_7ded0b9212ceac8698f4+477b7c62/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_7ded0b9212ceac8698f4+477b7c62/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_7ded0b9212ceac8698f4+477b7c62/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..07bf62460bc7813e4a95f208598e2fa4ae551d41 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_7ded0b9212ceac8698f4+477b7c62/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_7ded0b9212ceac8698f4+477b7c62/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_7ded0b9212ceac8698f4+477b7c62/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..06d0a9fd1e5175039b6d8c034ff032f118c197ff --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_7ded0b9212ceac8698f4+477b7c62/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c5d352673f0c03922d6ab106d507ffa1ead2647ef0cc1096e1b7232c67cc82 +size 267264 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_7ded0b9212ceac8698f4+477b7c62/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_7ded0b9212ceac8698f4+477b7c62/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..89ee84ec9532dea7e8e47cd33f744aa9c2773336 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_7ded0b9212ceac8698f4+477b7c62/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:813f6f8a1688ec44f7b53d83ef4faf5d195ff2721f8967be4cf01974023dc6ad +size 272962 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 69a28620b4f076e47dcc4eb325de713c35e3ceca..de31f3165d9264d681c52b5f5ce10651aa56fa33 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 07811ca3d62ded215f4fdd8066619de335098aeb..330bf4f60ffe2f50ecba95399305154cb85c071a 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 0c047b8509658bacac3f705f4220db4ded05c9be..e754db87565601ad17a045d878b7b1c4dfdaa6bb 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 c8670fdfc6ef9c0035322a4e40c85a41972ee7b6..8f125c2be65638a2692a23d1a4da1cef0c9eb82b 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 c70c1389ac98beeec3a63728623f9a9e1cbe4588..0a3160c8a0c5b8647e0a564ae2f8700c694b347d 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 f8f649e43b36354b38c55b91df85fcdbc70dfc48..21423c8824e63fca810dbf7ba5c44158f43defd5 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 a0d81f2e60b927b779f156904049f4180c4c3889..e31837372b6afd5677c207a71b996b3360978735 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 901484c9387cb1a86fe5693ddc3234de4ba81946..9dfa11caf2c6754b995e0990b038f5f5570f8cd4 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 448392b33246097b16ac967e881c940ea879462d..0823d6c1877f724d047ae7ea776c5d6bb41b8e07 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_99a4c8ce2f07c82d25d4+76eb1f45/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_99a4c8ce2f07c82d25d4+76eb1f45/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..ceec222a66946b45e9cac50614aef4f465473131 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_99a4c8ce2f07c82d25d4+76eb1f45/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_da5128bb-8554-44c9-a4a7-3a7f19e0e806/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_99a4c8ce2f07c82d25d4+76eb1f45/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_99a4c8ce2f07c82d25d4+76eb1f45/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_99a4c8ce2f07c82d25d4+76eb1f45/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_99a4c8ce2f07c82d25d4+76eb1f45/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..6a2752c25c9f7e1092e2b55c010694e168610ce3 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_99a4c8ce2f07c82d25d4+76eb1f45/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_99a4c8ce2f07c82d25d4+76eb1f45/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_99a4c8ce2f07c82d25d4+76eb1f45/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..ba1c98d81956be5615a95e140a14b6ce65bbf4aa --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_99a4c8ce2f07c82d25d4+76eb1f45/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c4da957391a2c478b55c710d2e0b19172fd84b474d887b1c230e0fe6f7b4358 +size 103424 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_99a4c8ce2f07c82d25d4+76eb1f45/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_99a4c8ce2f07c82d25d4+76eb1f45/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..47b9b994f5b709962d239a9efc010b81eb2cbbe4 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_99a4c8ce2f07c82d25d4+76eb1f45/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fdde416702a5c1065a5032e7cb5e0936bf9494c659c9d972e763b7afb33d94ff +size 104320 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_9cd1f2706a9a7386b4a1+edf8250f/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_9cd1f2706a9a7386b4a1+edf8250f/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5ffc01fa1eeaa33fb85f07db90149d3678081f6b --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_9cd1f2706a9a7386b4a1+edf8250f/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_448badab-1089-404e-9447-38313b46d758/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_9cd1f2706a9a7386b4a1+edf8250f/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_9cd1f2706a9a7386b4a1+edf8250f/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_9cd1f2706a9a7386b4a1+edf8250f/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_9cd1f2706a9a7386b4a1+edf8250f/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..8496781848c02ecd308b5b2c5a3c342b702425e7 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_9cd1f2706a9a7386b4a1+edf8250f/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_9cd1f2706a9a7386b4a1+edf8250f/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_9cd1f2706a9a7386b4a1+edf8250f/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..899003b059523fb75750fffcf262875f60b98d8c --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_9cd1f2706a9a7386b4a1+edf8250f/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f555eaadacf8ad3b8241ee9383db91f43c69a04dc35d211f3ab342d8ed09480c +size 246784 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_9cd1f2706a9a7386b4a1+edf8250f/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_9cd1f2706a9a7386b4a1+edf8250f/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..de47d585f4a9db0c8eef3b995f0060a91f144058 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_9cd1f2706a9a7386b4a1+edf8250f/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e7a39e27535087c6438ba48c227518b8ef87fa363dc71f91424b6253d221a77 +size 249608 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_ebb2335ea86c893fd330+cd6024cb/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_ebb2335ea86c893fd330+cd6024cb/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..347f8663853a48e3c4d0ce60bfd91cb3ad960406 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_ebb2335ea86c893fd330+cd6024cb/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_debc978d-c3b1-4a89-99d7-c0b0811ce966/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_ebb2335ea86c893fd330+cd6024cb/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_ebb2335ea86c893fd330+cd6024cb/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_ebb2335ea86c893fd330+cd6024cb/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_ebb2335ea86c893fd330+cd6024cb/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..f1e2473c54cffac949ecf31d3b6ec636c290545b --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_ebb2335ea86c893fd330+cd6024cb/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_ebb2335ea86c893fd330+cd6024cb/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_ebb2335ea86c893fd330+cd6024cb/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..5b30e40450732b5f2004c4b6d1820831a82d0f51 Binary files /dev/null and b/neuronxcc-2.21.33363.0+82129205/MODULE_ebb2335ea86c893fd330+cd6024cb/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_f19ecf2e2bcc3a4b1bd9+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_f19ecf2e2bcc3a4b1bd9+a02c3a36/model.hlo_module.pb index 087adad3be38716000a37f13a671422143669eca..db90fa54202e67f9c2b63a9f960524c316db768c 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_f19ecf2e2bcc3a4b1bd9+a02c3a36/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_f19ecf2e2bcc3a4b1bd9+a02c3a36/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:18a82b3beb86fc5ad8db17a90281a3d8d1e499f2379ca89b897d6d0517a715c4 +oid sha256:8ed2e09754ffce7b5957680d04d07b9b18166a7809582694d1e73f3af4b21434 size 590808 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_f19ecf2e2bcc3a4b1bd9+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_f19ecf2e2bcc3a4b1bd9+a02c3a36/model.neff index b7df2ab105b2c4d178400f74a1ef85eba5c80720..1c4d36523ab361879edddbeb9386f0255d2ef751 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_f19ecf2e2bcc3a4b1bd9+a02c3a36/model.neff +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_f19ecf2e2bcc3a4b1bd9+a02c3a36/model.neff @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b23b4a2a766db8ea02925c38c042a8596ad883b1314ba3bad71006bc2e503a11 +oid sha256:8a4d1820de61a7d69c692dda846a2beefab202121ece6bf3b2f3d3f57b6615ef size 1547264 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_f19ecf2e2bcc3a4b1bd9+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_f19ecf2e2bcc3a4b1bd9+a02c3a36/wrapped_neff.hlo index 0a68dbcc5151289820b637cbe80b48e0e0ca74fb..6bd39e951a18efffa4c7afec9d14b8dc12b2564c 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_f19ecf2e2bcc3a4b1bd9+a02c3a36/wrapped_neff.hlo +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_f19ecf2e2bcc3a4b1bd9+a02c3a36/wrapped_neff.hlo @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:3f213a8293dec0ed4180bff71d6a19ed22229a32087da31e728ccea8b8b87c9f +oid sha256:09068bc76e1d7f12bfc2a2a0a894345a7b43e7ba0364c7d8ca52c83125a85fce size 1703455 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fde2b1111ad10bd7504e+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_fde2b1111ad10bd7504e+a02c3a36/model.hlo_module.pb index 47dc969631220fdb40e5fe10f7bfae7a5454d613..1f4ba21c9f53777d279e086808276ab93c3b0615 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_fde2b1111ad10bd7504e+a02c3a36/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fde2b1111ad10bd7504e+a02c3a36/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:018a62e9a6d7b944baff314113a3d60b4bee472c3df99216573823afea25c4f2 +oid sha256:6661ed9cdcff742b06a6c5ab009981de926925468498a831dec14ba213178f65 size 426072 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fde2b1111ad10bd7504e+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_fde2b1111ad10bd7504e+a02c3a36/model.neff index 9fb49c749b593ac819cfd7c0cb2471431e188a83..2d9f1255e7f21cac515745f0bf3e5d2619eb74a6 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_fde2b1111ad10bd7504e+a02c3a36/model.neff +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fde2b1111ad10bd7504e+a02c3a36/model.neff @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4d90e44e9c444635c0ee1460fc00130b2c11d3ebcdc17e01db03772a3bd23b47 +oid sha256:78a45a412b6e0562fe26aed1091353d940bb875ab9302dc24eb0c24d6e52029c size 2100224 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fde2b1111ad10bd7504e+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_fde2b1111ad10bd7504e+a02c3a36/wrapped_neff.hlo index 0acb3ba1a67798c90b430c2921b8775f761ec859..d245795d2f67279f3fbe1eafc884ef5e4daa8773 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_fde2b1111ad10bd7504e+a02c3a36/wrapped_neff.hlo +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fde2b1111ad10bd7504e+a02c3a36/wrapped_neff.hlo @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:aab386728f7b078571fb53e43a40ff52622f677a44fa11b4b900702de01ab81b +oid sha256:9a1dee73cc6b9bb4a4a4dc8dc072f92e7f134800fa3f17b1317b073dfb40bd24 size 2174311