diff --git a/.gitattributes b/.gitattributes index 93e65e1db1855872f697e2a94356db72a4708fb6..23ed554798b630422ca1d93548a6514804399f61 100644 --- a/.gitattributes +++ b/.gitattributes @@ -16710,3 +16710,85 @@ neuronxcc-2.21.33363.0+82129205/MODULE_518cfc3ed8323f4a8c4d+a32116a7/model.neff neuronxcc-2.21.33363.0+82129205/MODULE_4e8fb23a5a9a65cf1162+24129607/model.neff filter=lfs diff=lfs merge=lfs -text neuronxcc-2.21.33363.0+82129205/MODULE_9e05f620120bf1ec4c78+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text neuronxcc-2.21.33363.0+82129205/MODULE_9e05f620120bf1ec4c78+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_0007bf645cd8f44068db+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_0007bf645cd8f44068db+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_001a8c20b0b7f8f20fbc+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_001a8c20b0b7f8f20fbc+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_0680112a1970faff8f3a+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_0680112a1970faff8f3a+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_0a1bf3b6e86f854781c2+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_0a1bf3b6e86f854781c2+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_0c7914f92b283d2acea2+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_0c7914f92b283d2acea2+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_131d6c7c245c7131bd76+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_131d6c7c245c7131bd76+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_13ceff346d0e3479704a+6e4949b4/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_13f403bb4f61de68acf1+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_13f403bb4f61de68acf1+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_211a21d76f2e9aaf1477+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_211a21d76f2e9aaf1477+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_27b2c4720a6a1662f8b5+24129607/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_3b30a2906ed2f2261284+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_3b30a2906ed2f2261284+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_47c06cec950237374a0e+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_47c06cec950237374a0e+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_4da1b782d99a7304f0e2+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_4da1b782d99a7304f0e2+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_5ad2d5f306d13c506440+24129607/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_60f37ef2cb181ebeb434+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_60f37ef2cb181ebeb434+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_6ed16792f2699fac18e0+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_6ed16792f2699fac18e0+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_72c4957df590729a5c8d+ac10809c/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_72c4957df590729a5c8d+ac10809c/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_78ef14a5e66efce40c70+2dde74c7/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_78ef14a5e66efce40c70+2dde74c7/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_8051a8dfc97b2193171a+f68f772a/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_8051a8dfc97b2193171a+f68f772a/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_826bbed76f9f977f94c0+24129607/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_84656e81eaefaaeb593f+24129607/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_87be52aa7374785dde52+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_87be52aa7374785dde52+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_8f7c5fdf533e837a78bc+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_8f7c5fdf533e837a78bc+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_8fb7e0e33d5fb07b5f06+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_8fb7e0e33d5fb07b5f06+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_9a1b9a8b43d1665adae5+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_9a1b9a8b43d1665adae5+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_a1cd38756c74f5816b8b+24129607/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_b1f6368a309ea57ce627+3181e94a/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_b1f6368a309ea57ce627+3181e94a/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_b5e0b2d29beb17680855+a32116a7/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_b5eece93194f9ce71789+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_b5eece93194f9ce71789+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_b651bd725ae2d8726217+80d05c3f/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_b651bd725ae2d8726217+80d05c3f/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_b87961c02022113d30f4+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_b87961c02022113d30f4+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_bfde8ebf354636b209da+24129607/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_c3216c9356c729b39292+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_c3216c9356c729b39292+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_c6ac9f1b555d05bd7e39+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_c6ac9f1b555d05bd7e39+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_c75fd5e63a178ce4f3d1+677eeb9d/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_cd398d15ddca2b8db9a9+24129607/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_d28157993393c2fbf90f+1592a3e2/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_d28157993393c2fbf90f+1592a3e2/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_db3cfbc4f07f54d7ec28+24129607/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_e337acd66701db458735+24129607/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_e737e0fdb56bf5bd26fc+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_e737e0fdb56bf5bd26fc+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_ea4fa7da6b267fc2c5c7+4482fc6e/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_ea4fa7da6b267fc2c5c7+4482fc6e/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_eb5005f121a849e69eb7+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_eb5005f121a849e69eb7+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_ef6c48ae8ca983dc5ac7+4ef53020/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_ef6c48ae8ca983dc5ac7+4ef53020/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_ef93001ee65461791427+24129607/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_f33d66622d796b2877aa+283df001/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_fa29fc69337939622d95+24129607/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_fb598b8eba6e5a92e75f+24129607/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_fe491aea85c492a9d7ed+7e8048af/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_fe491aea85c492a9d7ed+7e8048af/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_fe6b8b40a8e1d622e71e+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text +neuronxcc-2.21.33363.0+82129205/MODULE_fe6b8b40a8e1d622e71e+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/0cc25526e2cfc37a8875a3752f33c4d7505d8a07b869d0f3f41915cf6e763b74/7ae1cb8d8d5ab5447a69.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/0cc25526e2cfc37a8875a3752f33c4d7505d8a07b869d0f3f41915cf6e763b74/7ae1cb8d8d5ab5447a69.json new file mode 100644 index 0000000000000000000000000000000000000000..0474ecd29f60def0567092ffed501ab0ac9fada7 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/0cc25526e2cfc37a8875a3752f33c4d7505d8a07b869d0f3f41915cf6e763b74/7ae1cb8d8d5ab5447a69.json @@ -0,0 +1,58 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "ibm-granite/granite-3.1-2b-instruct", + "_task": "text-generation", + "architectures": [ + "GraniteForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.1, + "attention_multiplier": 0.015625, + "dtype": "bfloat16", + "embedding_multiplier": 12.0, + "hidden_act": "silu", + "hidden_size": 2048, + "initializer_range": 0.02, + "intermediate_size": 8192, + "logits_scaling": 8.0, + "max_position_embeddings": 131072, + "mlp_bias": false, + "model_type": "granite", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 1, + "capacity_factor": null, + "checkpoint_id": "ibm-granite/granite-3.1-2b-instruct", + "checkpoint_revision": "bbc2aed595bd38bd770263dc3ab831db9794441d", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 1, + "max_context_length": 8192, + "max_topk": 256, + "n_active_tokens": 8192, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.6.dev2", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 8192, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 32, + "num_hidden_layers": 40, + "num_key_value_heads": 8, + "residual_multiplier": 0.22, + "rms_norm_eps": 1e-05, + "rope_scaling": null, + "rope_theta": 5000000.0, + "tie_word_embeddings": true, + "use_cache": true, + "vocab_size": 49155 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/0cc25526e2cfc37a8875a3752f33c4d7505d8a07b869d0f3f41915cf6e763b74/e30e0a23150fd668133c.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/0cc25526e2cfc37a8875a3752f33c4d7505d8a07b869d0f3f41915cf6e763b74/e30e0a23150fd668133c.json new file mode 100644 index 0000000000000000000000000000000000000000..9da2d2cd3632dd940ac81a725bf85e817c4e0244 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/0cc25526e2cfc37a8875a3752f33c4d7505d8a07b869d0f3f41915cf6e763b74/e30e0a23150fd668133c.json @@ -0,0 +1,58 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "ibm-granite/granite-3.1-2b-instruct", + "_task": "text-generation", + "architectures": [ + "GraniteForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.1, + "attention_multiplier": 0.015625, + "dtype": "bfloat16", + "embedding_multiplier": 12.0, + "hidden_act": "silu", + "hidden_size": 2048, + "initializer_range": 0.02, + "intermediate_size": 8192, + "logits_scaling": 8.0, + "max_position_embeddings": 131072, + "mlp_bias": false, + "model_type": "granite", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 4, + "capacity_factor": null, + "checkpoint_id": "ibm-granite/granite-3.1-2b-instruct", + "checkpoint_revision": "bbc2aed595bd38bd770263dc3ab831db9794441d", + "continuous_batching": true, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 4, + "max_context_length": 4096, + "max_topk": 256, + "n_active_tokens": 4096, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.6.dev2", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 4096, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 32, + "num_hidden_layers": 40, + "num_key_value_heads": 8, + "residual_multiplier": 0.22, + "rms_norm_eps": 1e-05, + "rope_scaling": null, + "rope_theta": 5000000.0, + "tie_word_embeddings": true, + "use_cache": true, + "vocab_size": 49155 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/2da7a00f0478d50ae1e7f75f085c5b2773b5f355f427c61cf34cb6febd629d96/d2900659f0c3438aa013.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/2da7a00f0478d50ae1e7f75f085c5b2773b5f355f427c61cf34cb6febd629d96/d2900659f0c3438aa013.json new file mode 100644 index 0000000000000000000000000000000000000000..4c75a90f39ef12dcbc55064c43db84818125d794 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/2da7a00f0478d50ae1e7f75f085c5b2773b5f355f427c61cf34cb6febd629d96/d2900659f0c3438aa013.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.dev2", + "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.dev2/441269935591cad8d370e512c0b93cdd2fce6247c40e5a4866d872ee5338b0de/04dbf3b30744c2b2a648.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/441269935591cad8d370e512c0b93cdd2fce6247c40e5a4866d872ee5338b0de/04dbf3b30744c2b2a648.json new file mode 100644 index 0000000000000000000000000000000000000000..ceaee3579382505ff3873e38e4db3bf3361b84c3 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/441269935591cad8d370e512c0b93cdd2fce6247c40e5a4866d872ee5338b0de/04dbf3b30744c2b2a648.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": 4, + "capacity_factor": null, + "checkpoint_id": "unsloth/gemma-3-270m-it", + "checkpoint_revision": "23cf460f6bb16954176b3ddcc8d4f250501458a9", + "continuous_batching": true, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 4, + "max_context_length": 4096, + "max_topk": 256, + "n_active_tokens": 4096, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.6.dev2", + "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": 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.dev2/441269935591cad8d370e512c0b93cdd2fce6247c40e5a4866d872ee5338b0de/7fde6971cc66d8965351.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/441269935591cad8d370e512c0b93cdd2fce6247c40e5a4866d872ee5338b0de/7fde6971cc66d8965351.json new file mode 100644 index 0000000000000000000000000000000000000000..fdad245a9f75b0a32510e2050a7bcbe357a35af4 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/441269935591cad8d370e512c0b93cdd2fce6247c40e5a4866d872ee5338b0de/7fde6971cc66d8965351.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.dev2", + "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.dev2/4ab8140bc7eb4a553d95855c5c2be2cf8c0fbab21b823d76183b6f51e98b6fc5/af2d330d6e4b2cf43703.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/4ab8140bc7eb4a553d95855c5c2be2cf8c0fbab21b823d76183b6f51e98b6fc5/af2d330d6e4b2cf43703.json new file mode 100644 index 0000000000000000000000000000000000000000..cfb064bfd57dbe0094a334e0e4f2390a2dccd1ef --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/4ab8140bc7eb4a553d95855c5c2be2cf8c0fbab21b823d76183b6f51e98b6fc5/af2d330d6e4b2cf43703.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.dev2", + "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.dev2/4cb7aff9e2a15c151396f2b684013e39d6739f0dec83e5c9dabbfe9d5fcf77b7/2e9d2a02987373a1d580.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/4cb7aff9e2a15c151396f2b684013e39d6739f0dec83e5c9dabbfe9d5fcf77b7/2e9d2a02987373a1d580.json new file mode 100644 index 0000000000000000000000000000000000000000..dad2e59357a029114ced4c00e3f7314288f9387f --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/4cb7aff9e2a15c151396f2b684013e39d6739f0dec83e5c9dabbfe9d5fcf77b7/2e9d2a02987373a1d580.json @@ -0,0 +1,82 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "Qwen/Qwen2.5-0.5B", + "_task": "text-generation", + "architectures": [ + "Qwen2ForCausalLM" + ], + "attention_dropout": 0.0, + "dtype": "bfloat16", + "hidden_act": "silu", + "hidden_size": 896, + "initializer_range": 0.02, + "intermediate_size": 4864, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 32768, + "max_window_layers": 24, + "model_type": "qwen2", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 1, + "capacity_factor": null, + "checkpoint_id": "Qwen/Qwen2.5-0.5B", + "checkpoint_revision": "060db6499f32faf8b98477b0a26969ef7d8b9987", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": false, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 1, + "max_context_length": 8192, + "max_topk": 256, + "n_active_tokens": 8192, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.6.dev2", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 8192, + "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.dev2/4cb7aff9e2a15c151396f2b684013e39d6739f0dec83e5c9dabbfe9d5fcf77b7/682142914fb2798c2038.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/4cb7aff9e2a15c151396f2b684013e39d6739f0dec83e5c9dabbfe9d5fcf77b7/682142914fb2798c2038.json new file mode 100644 index 0000000000000000000000000000000000000000..14571b33f4a306dfff256d6c83b12d6a14d3c933 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/4cb7aff9e2a15c151396f2b684013e39d6739f0dec83e5c9dabbfe9d5fcf77b7/682142914fb2798c2038.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.dev2", + "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.dev2/4cb7aff9e2a15c151396f2b684013e39d6739f0dec83e5c9dabbfe9d5fcf77b7/cb3cfc63fc49cc904f0d.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/4cb7aff9e2a15c151396f2b684013e39d6739f0dec83e5c9dabbfe9d5fcf77b7/cb3cfc63fc49cc904f0d.json new file mode 100644 index 0000000000000000000000000000000000000000..2cd4afb51e349c9350de1718e44537aedf18cf49 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/4cb7aff9e2a15c151396f2b684013e39d6739f0dec83e5c9dabbfe9d5fcf77b7/cb3cfc63fc49cc904f0d.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": 4, + "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": 4, + "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.dev2", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 4096, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 14, + "num_hidden_layers": 24, + "num_key_value_heads": 2, + "rms_norm_eps": 1e-06, + "rope_scaling": null, + "rope_theta": 1000000.0, + "sliding_window": null, + "tie_word_embeddings": true, + "use_cache": true, + "use_mrope": false, + "use_sliding_window": false, + "vocab_size": 151936 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/60feecaa0c4c075e2f3e46a3f55d9a273f0ddd75a0ecf64e4ae27352e0819506/7c0a10bc3ecf969f82ff.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/60feecaa0c4c075e2f3e46a3f55d9a273f0ddd75a0ecf64e4ae27352e0819506/7c0a10bc3ecf969f82ff.json new file mode 100644 index 0000000000000000000000000000000000000000..fc38ddcd499fa5b3aae4244c0eed35130af7d3f5 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/60feecaa0c4c075e2f3e46a3f55d9a273f0ddd75a0ecf64e4ae27352e0819506/7c0a10bc3ecf969f82ff.json @@ -0,0 +1,87 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "Qwen/Qwen3-Embedding-0.6B", + "_task": "feature-extraction", + "architectures": [ + "Qwen3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "dtype": "bfloat16", + "head_dim": 128, + "hidden_act": "silu", + "hidden_size": 1024, + "initializer_range": 0.02, + "intermediate_size": 3072, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 32768, + "max_window_layers": 28, + "model_type": "qwen3", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 6, + "capacity_factor": null, + "checkpoint_id": "Qwen/Qwen3-Embedding-0.6B", + "checkpoint_revision": "c54f2e6e80b2d7b7de06f51cec4959f6b3e03418", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 6, + "max_context_length": 8192, + "max_topk": 256, + "n_active_tokens": 8192, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": false, + "optimum_neuron_version": "0.4.6.dev2", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 8192, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 16, + "num_hidden_layers": 28, + "num_key_value_heads": 8, + "rms_norm_eps": 1e-06, + "rope_scaling": null, + "rope_theta": 1000000, + "sliding_window": null, + "tie_word_embeddings": true, + "use_cache": true, + "use_sliding_window": false, + "vocab_size": 151669 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/60feecaa0c4c075e2f3e46a3f55d9a273f0ddd75a0ecf64e4ae27352e0819506/7c0c5a6095e090c4f577.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/60feecaa0c4c075e2f3e46a3f55d9a273f0ddd75a0ecf64e4ae27352e0819506/7c0c5a6095e090c4f577.json new file mode 100644 index 0000000000000000000000000000000000000000..b478a799c5350dca3ac4e7e5862b7e9e12a434e3 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/60feecaa0c4c075e2f3e46a3f55d9a273f0ddd75a0ecf64e4ae27352e0819506/7c0c5a6095e090c4f577.json @@ -0,0 +1,87 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "Qwen/Qwen3-Embedding-0.6B", + "_task": "feature-extraction", + "architectures": [ + "Qwen3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "dtype": "bfloat16", + "head_dim": 128, + "hidden_act": "silu", + "hidden_size": 1024, + "initializer_range": 0.02, + "intermediate_size": 3072, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 32768, + "max_window_layers": 28, + "model_type": "qwen3", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 4, + "capacity_factor": null, + "checkpoint_id": "Qwen/Qwen3-Embedding-0.6B", + "checkpoint_revision": "c54f2e6e80b2d7b7de06f51cec4959f6b3e03418", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 4, + "max_context_length": 8192, + "max_topk": 256, + "n_active_tokens": 8192, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": false, + "optimum_neuron_version": "0.4.6.dev2", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 8192, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 16, + "num_hidden_layers": 28, + "num_key_value_heads": 8, + "rms_norm_eps": 1e-06, + "rope_scaling": null, + "rope_theta": 1000000, + "sliding_window": null, + "tie_word_embeddings": true, + "use_cache": true, + "use_sliding_window": false, + "vocab_size": 151669 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/6454afdf3e9d66c7226c13a575b718845c25e53b0699600ba2bb4f883e9d841b/77b737f7b0783b7b8363.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/6454afdf3e9d66c7226c13a575b718845c25e53b0699600ba2bb4f883e9d841b/77b737f7b0783b7b8363.json new file mode 100644 index 0000000000000000000000000000000000000000..b84debf9c0e5cf5ff43e12d6826d8c9151b01b40 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/6454afdf3e9d66c7226c13a575b718845c25e53b0699600ba2bb4f883e9d841b/77b737f7b0783b7b8363.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.dev2", + "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.dev2/6454afdf3e9d66c7226c13a575b718845c25e53b0699600ba2bb4f883e9d841b/9a1461765dbc24d49ac7.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/6454afdf3e9d66c7226c13a575b718845c25e53b0699600ba2bb4f883e9d841b/9a1461765dbc24d49ac7.json new file mode 100644 index 0000000000000000000000000000000000000000..c9ce2349cc64cf11d749feb64ef458f4677e6e5a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/6454afdf3e9d66c7226c13a575b718845c25e53b0699600ba2bb4f883e9d841b/9a1461765dbc24d49ac7.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.dev2", + "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.dev2/6454afdf3e9d66c7226c13a575b718845c25e53b0699600ba2bb4f883e9d841b/e52a6962dab5dfc81454.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/6454afdf3e9d66c7226c13a575b718845c25e53b0699600ba2bb4f883e9d841b/e52a6962dab5dfc81454.json new file mode 100644 index 0000000000000000000000000000000000000000..b708a3acaebf2719a3e32b07f08e40aa1b402e30 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/6454afdf3e9d66c7226c13a575b718845c25e53b0699600ba2bb4f883e9d841b/e52a6962dab5dfc81454.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.dev2", + "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.dev2/73707b485eab9008c7aba7f5dad0ce2384ac685318d5f888c12fa0d81ed90b19/8ac8c04582271fb515e7.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/73707b485eab9008c7aba7f5dad0ce2384ac685318d5f888c12fa0d81ed90b19/8ac8c04582271fb515e7.json new file mode 100644 index 0000000000000000000000000000000000000000..39dab72fe022b8817df56b39f8bd81574183546c --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/73707b485eab9008c7aba7f5dad0ce2384ac685318d5f888c12fa0d81ed90b19/8ac8c04582271fb515e7.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": 8192, + "max_topk": 256, + "n_active_tokens": 8192, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.6.dev2", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 8192, + "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.dev2/73707b485eab9008c7aba7f5dad0ce2384ac685318d5f888c12fa0d81ed90b19/8adf13bbd08271a796c8.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/73707b485eab9008c7aba7f5dad0ce2384ac685318d5f888c12fa0d81ed90b19/8adf13bbd08271a796c8.json new file mode 100644 index 0000000000000000000000000000000000000000..a132a2498b087f1c7ccca556a74a7d23482ac95f --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/73707b485eab9008c7aba7f5dad0ce2384ac685318d5f888c12fa0d81ed90b19/8adf13bbd08271a796c8.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.dev2", + "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.dev2/73707b485eab9008c7aba7f5dad0ce2384ac685318d5f888c12fa0d81ed90b19/a20cdb6c0f8b24266437.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/73707b485eab9008c7aba7f5dad0ce2384ac685318d5f888c12fa0d81ed90b19/a20cdb6c0f8b24266437.json new file mode 100644 index 0000000000000000000000000000000000000000..b1db6c6630c53cc7318685bdcc773549d644747f --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/73707b485eab9008c7aba7f5dad0ce2384ac685318d5f888c12fa0d81ed90b19/a20cdb6c0f8b24266437.json @@ -0,0 +1,134 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "HuggingFaceTB/SmolLM3-3B", + "_task": "text-generation", + "architectures": [ + "SmolLM3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "dtype": "bfloat16", + "hidden_act": "silu", + "hidden_size": 2048, + "initializer_range": 0.02, + "intermediate_size": 11008, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 65536, + "max_window_layers": 28, + "mlp_bias": false, + "model_type": "smollm3", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 4, + "capacity_factor": null, + "checkpoint_id": "HuggingFaceTB/SmolLM3-3B", + "checkpoint_revision": "a07cc9a04f16550a088caea529712d1d335b0ac1", + "continuous_batching": true, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 4, + "max_context_length": 4096, + "max_topk": 256, + "n_active_tokens": 4096, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.6.dev2", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 4096, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "no_rope_layer_interval": 4, + "no_rope_layers": [ + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0 + ], + "num_attention_heads": 16, + "num_hidden_layers": 36, + "num_key_value_heads": 4, + "pretraining_tp": 2, + "rms_norm_eps": 1e-06, + "rope_scaling": null, + "rope_theta": 5000000.0, + "sliding_window": null, + "use_cache": false, + "use_sliding_window": false, + "vocab_size": 128256 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/7518518c7e077820070186deda960d8cc49db068cdf0ac70664098fa2b6b698c/ee9ffbe1ac4a7ad2ee51.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/7518518c7e077820070186deda960d8cc49db068cdf0ac70664098fa2b6b698c/ee9ffbe1ac4a7ad2ee51.json new file mode 100644 index 0000000000000000000000000000000000000000..05b1c9e977e9ce897fe5a98011d3b19f37bb938b --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/7518518c7e077820070186deda960d8cc49db068cdf0ac70664098fa2b6b698c/ee9ffbe1ac4a7ad2ee51.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.dev2", + "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.dev2/8c90ac2593ed0b7f1ecb60e82cb184fb11f2ea640befa1cc7b10766a5c02525d/2375ae0a6703b35f26db.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/8c90ac2593ed0b7f1ecb60e82cb184fb11f2ea640befa1cc7b10766a5c02525d/2375ae0a6703b35f26db.json new file mode 100644 index 0000000000000000000000000000000000000000..d5cb0b967a02f6f4b309e91a5965d98d2485d453 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/8c90ac2593ed0b7f1ecb60e82cb184fb11f2ea640befa1cc7b10766a5c02525d/2375ae0a6703b35f26db.json @@ -0,0 +1,164 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "microsoft/Phi-3.5-mini-instruct", + "_task": "text-generation", + "architectures": [ + "Phi3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "auto_map": { + "AutoConfig": "configuration_phi3.Phi3Config", + "AutoModelForCausalLM": "modeling_phi3.Phi3ForCausalLM" + }, + "dtype": "bfloat16", + "embd_pdrop": 0.0, + "hidden_act": "silu", + "hidden_size": 3072, + "initializer_range": 0.02, + "intermediate_size": 8192, + "max_position_embeddings": 131072, + "model_type": "phi3", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 4, + "capacity_factor": null, + "checkpoint_id": "microsoft/Phi-3.5-mini-instruct", + "checkpoint_revision": "2fe192450127e6a83f7441aef6e3ca586c338b77", + "continuous_batching": true, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 4, + "max_context_length": 4096, + "max_topk": 256, + "n_active_tokens": 4096, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.6.dev2", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 4096, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 32, + "num_hidden_layers": 32, + "num_key_value_heads": 32, + "original_max_position_embeddings": 4096, + "partial_rotary_factor": 1.0, + "resid_pdrop": 0.0, + "rms_norm_eps": 1e-05, + "rope_scaling": { + "long_factor": [ + 1.0800000429153442, + 1.1100000143051147, + 1.1399999856948853, + 1.340000033378601, + 1.5899999141693115, + 1.600000023841858, + 1.6200000047683716, + 2.620000123977661, + 3.2300000190734863, + 3.2300000190734863, + 4.789999961853027, + 7.400000095367432, + 7.700000286102295, + 9.09000015258789, + 12.199999809265137, + 17.670000076293945, + 24.46000099182129, + 28.57000160217285, + 30.420001983642578, + 30.840002059936523, + 32.590003967285156, + 32.93000411987305, + 42.320003509521484, + 44.96000289916992, + 50.340003967285156, + 50.45000457763672, + 57.55000305175781, + 57.93000411987305, + 58.21000289916992, + 60.1400032043457, + 62.61000442504883, + 62.62000274658203, + 62.71000289916992, + 63.1400032043457, + 63.1400032043457, + 63.77000427246094, + 63.93000411987305, + 63.96000289916992, + 63.970001220703125, + 64.02999877929688, + 64.06999969482422, + 64.08000183105469, + 64.12000274658203, + 64.41000366210938, + 64.4800033569336, + 64.51000213623047, + 64.52999877929688, + 64.83999633789062 + ], + "short_factor": [ + 1.0, + 1.0199999809265137, + 1.0299999713897705, + 1.0299999713897705, + 1.0499999523162842, + 1.0499999523162842, + 1.0499999523162842, + 1.0499999523162842, + 1.0499999523162842, + 1.0699999332427979, + 1.0999999046325684, + 1.1099998950958252, + 1.1599998474121094, + 1.1599998474121094, + 1.1699998378753662, + 1.2899998426437378, + 1.339999794960022, + 1.679999828338623, + 1.7899998426437378, + 1.8199998140335083, + 1.8499997854232788, + 1.8799997568130493, + 1.9099997282028198, + 1.9399996995925903, + 1.9899996519088745, + 2.0199997425079346, + 2.0199997425079346, + 2.0199997425079346, + 2.0199997425079346, + 2.0199997425079346, + 2.0199997425079346, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0799996852874756, + 2.0899996757507324, + 2.189999580383301, + 2.2199995517730713, + 2.5899994373321533, + 2.729999542236328, + 2.749999523162842, + 2.8399994373321533 + ], + "type": "longrope" + }, + "rope_theta": 10000.0, + "sliding_window": 262144, + "tie_word_embeddings": false, + "use_cache": true, + "vocab_size": 32064 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/8c90ac2593ed0b7f1ecb60e82cb184fb11f2ea640befa1cc7b10766a5c02525d/3c2aa3ff232ee546fbda.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/8c90ac2593ed0b7f1ecb60e82cb184fb11f2ea640befa1cc7b10766a5c02525d/3c2aa3ff232ee546fbda.json new file mode 100644 index 0000000000000000000000000000000000000000..e26ce8a8536eb67702b5c50f9f64a6d2e4883c38 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/8c90ac2593ed0b7f1ecb60e82cb184fb11f2ea640befa1cc7b10766a5c02525d/3c2aa3ff232ee546fbda.json @@ -0,0 +1,164 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "microsoft/Phi-3.5-mini-instruct", + "_task": "text-generation", + "architectures": [ + "Phi3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "auto_map": { + "AutoConfig": "configuration_phi3.Phi3Config", + "AutoModelForCausalLM": "modeling_phi3.Phi3ForCausalLM" + }, + "dtype": "bfloat16", + "embd_pdrop": 0.0, + "hidden_act": "silu", + "hidden_size": 3072, + "initializer_range": 0.02, + "intermediate_size": 8192, + "max_position_embeddings": 131072, + "model_type": "phi3", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 1, + "capacity_factor": null, + "checkpoint_id": "microsoft/Phi-3.5-mini-instruct", + "checkpoint_revision": "2fe192450127e6a83f7441aef6e3ca586c338b77", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 1, + "max_context_length": 8192, + "max_topk": 256, + "n_active_tokens": 8192, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.6.dev2", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 8192, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 32, + "num_hidden_layers": 32, + "num_key_value_heads": 32, + "original_max_position_embeddings": 4096, + "partial_rotary_factor": 1.0, + "resid_pdrop": 0.0, + "rms_norm_eps": 1e-05, + "rope_scaling": { + "long_factor": [ + 1.0800000429153442, + 1.1100000143051147, + 1.1399999856948853, + 1.340000033378601, + 1.5899999141693115, + 1.600000023841858, + 1.6200000047683716, + 2.620000123977661, + 3.2300000190734863, + 3.2300000190734863, + 4.789999961853027, + 7.400000095367432, + 7.700000286102295, + 9.09000015258789, + 12.199999809265137, + 17.670000076293945, + 24.46000099182129, + 28.57000160217285, + 30.420001983642578, + 30.840002059936523, + 32.590003967285156, + 32.93000411987305, + 42.320003509521484, + 44.96000289916992, + 50.340003967285156, + 50.45000457763672, + 57.55000305175781, + 57.93000411987305, + 58.21000289916992, + 60.1400032043457, + 62.61000442504883, + 62.62000274658203, + 62.71000289916992, + 63.1400032043457, + 63.1400032043457, + 63.77000427246094, + 63.93000411987305, + 63.96000289916992, + 63.970001220703125, + 64.02999877929688, + 64.06999969482422, + 64.08000183105469, + 64.12000274658203, + 64.41000366210938, + 64.4800033569336, + 64.51000213623047, + 64.52999877929688, + 64.83999633789062 + ], + "short_factor": [ + 1.0, + 1.0199999809265137, + 1.0299999713897705, + 1.0299999713897705, + 1.0499999523162842, + 1.0499999523162842, + 1.0499999523162842, + 1.0499999523162842, + 1.0499999523162842, + 1.0699999332427979, + 1.0999999046325684, + 1.1099998950958252, + 1.1599998474121094, + 1.1599998474121094, + 1.1699998378753662, + 1.2899998426437378, + 1.339999794960022, + 1.679999828338623, + 1.7899998426437378, + 1.8199998140335083, + 1.8499997854232788, + 1.8799997568130493, + 1.9099997282028198, + 1.9399996995925903, + 1.9899996519088745, + 2.0199997425079346, + 2.0199997425079346, + 2.0199997425079346, + 2.0199997425079346, + 2.0199997425079346, + 2.0199997425079346, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0799996852874756, + 2.0899996757507324, + 2.189999580383301, + 2.2199995517730713, + 2.5899994373321533, + 2.729999542236328, + 2.749999523162842, + 2.8399994373321533 + ], + "type": "longrope" + }, + "rope_theta": 10000.0, + "sliding_window": 262144, + "tie_word_embeddings": false, + "use_cache": true, + "vocab_size": 32064 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/920f44ce6d3e004d1ce547ae06644f7be262180644b04573153aa15d98742edc/a4bf640c57a2ab4c037e.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/920f44ce6d3e004d1ce547ae06644f7be262180644b04573153aa15d98742edc/a4bf640c57a2ab4c037e.json new file mode 100644 index 0000000000000000000000000000000000000000..4bf6d5f0262aec2537fdd0ed394528e794a49135 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/920f44ce6d3e004d1ce547ae06644f7be262180644b04573153aa15d98742edc/a4bf640c57a2ab4c037e.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.dev2", + "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.dev2/929b02754a13cbfdf657d863c3fc6f3bce672879bc6ae48ab45be21e881e9ec2/15af6878da0b86d6586a.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/929b02754a13cbfdf657d863c3fc6f3bce672879bc6ae48ab45be21e881e9ec2/15af6878da0b86d6586a.json new file mode 100644 index 0000000000000000000000000000000000000000..3d7ea8538f80747d89895c77886c094bf04b2f69 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/929b02754a13cbfdf657d863c3fc6f3bce672879bc6ae48ab45be21e881e9ec2/15af6878da0b86d6586a.json @@ -0,0 +1,87 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "Qwen/Qwen3-0.6B", + "_task": "text-generation", + "architectures": [ + "Qwen3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "dtype": "bfloat16", + "head_dim": 128, + "hidden_act": "silu", + "hidden_size": 1024, + "initializer_range": 0.02, + "intermediate_size": 3072, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 40960, + "max_window_layers": 28, + "model_type": "qwen3", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 4, + "capacity_factor": null, + "checkpoint_id": "Qwen/Qwen3-0.6B", + "checkpoint_revision": "c1899de289a04d12100db370d81485cdf75e47ca", + "continuous_batching": true, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 4, + "max_context_length": 4096, + "max_topk": 256, + "n_active_tokens": 4096, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": false, + "optimum_neuron_version": "0.4.6.dev2", + "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": 16, + "num_hidden_layers": 28, + "num_key_value_heads": 8, + "rms_norm_eps": 1e-06, + "rope_scaling": null, + "rope_theta": 1000000, + "sliding_window": null, + "tie_word_embeddings": true, + "use_cache": true, + "use_sliding_window": false, + "vocab_size": 151936 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/929b02754a13cbfdf657d863c3fc6f3bce672879bc6ae48ab45be21e881e9ec2/532d1f3cc32c693450dc.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/929b02754a13cbfdf657d863c3fc6f3bce672879bc6ae48ab45be21e881e9ec2/532d1f3cc32c693450dc.json new file mode 100644 index 0000000000000000000000000000000000000000..bac4a5833499d97f94bfc495a1283f2f7ad8f076 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/929b02754a13cbfdf657d863c3fc6f3bce672879bc6ae48ab45be21e881e9ec2/532d1f3cc32c693450dc.json @@ -0,0 +1,87 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "Qwen/Qwen3-0.6B", + "_task": "text-generation", + "architectures": [ + "Qwen3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "dtype": "bfloat16", + "head_dim": 128, + "hidden_act": "silu", + "hidden_size": 1024, + "initializer_range": 0.02, + "intermediate_size": 3072, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 40960, + "max_window_layers": 28, + "model_type": "qwen3", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 1, + "capacity_factor": null, + "checkpoint_id": "Qwen/Qwen3-0.6B", + "checkpoint_revision": "c1899de289a04d12100db370d81485cdf75e47ca", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 1, + "max_context_length": 8192, + "max_topk": 256, + "n_active_tokens": 8192, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.6.dev2", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 8192, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 16, + "num_hidden_layers": 28, + "num_key_value_heads": 8, + "rms_norm_eps": 1e-06, + "rope_scaling": null, + "rope_theta": 1000000, + "sliding_window": null, + "tie_word_embeddings": true, + "use_cache": true, + "use_sliding_window": false, + "vocab_size": 151936 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/9c24a2080ec26e9c74d5bd90866f74131aa4d5c975415b96f5e6600f081d5647/16b8e47a972c18b057c1.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/9c24a2080ec26e9c74d5bd90866f74131aa4d5c975415b96f5e6600f081d5647/16b8e47a972c18b057c1.json new file mode 100644 index 0000000000000000000000000000000000000000..3408e49ff28bf73d02316047a752a2665c2f0917 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/9c24a2080ec26e9c74d5bd90866f74131aa4d5c975415b96f5e6600f081d5647/16b8e47a972c18b057c1.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.dev2", + "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.dev2/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/201d4c1f716698f7f279.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/201d4c1f716698f7f279.json new file mode 100644 index 0000000000000000000000000000000000000000..77d0728884813aed637c2bf1ece7eb4518000161 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/201d4c1f716698f7f279.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.dev2", + "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.dev2/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/237f3427f0ffd5bd2890.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/237f3427f0ffd5bd2890.json new file mode 100644 index 0000000000000000000000000000000000000000..88dc361547a68cfb0cfaec81c2bbac00f7c0240c --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/237f3427f0ffd5bd2890.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.dev2", + "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.dev2/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/4d1906e093e7da0b8a2f.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/4d1906e093e7da0b8a2f.json new file mode 100644 index 0000000000000000000000000000000000000000..854a8dcab2a473dc026ba079b83a408efc972d00 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/4d1906e093e7da0b8a2f.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": 4, + "capacity_factor": null, + "checkpoint_id": "unsloth/Llama-3.2-1B-Instruct", + "checkpoint_revision": "5a8abab4a5d6f164389b1079fb721cfab8d7126c", + "continuous_batching": true, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 4, + "max_context_length": 4096, + "max_topk": 256, + "n_active_tokens": 4096, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.6.dev2", + "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.dev2/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/b52ec561f1c1fa96cdd0.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/b52ec561f1c1fa96cdd0.json new file mode 100644 index 0000000000000000000000000000000000000000..3ee69b335e80ddc18519ffb89bdd5b184835f702 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/cf6b9a360dcf294104671106bae2adbd9fd291823bb60a351883163684073231/b52ec561f1c1fa96cdd0.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.dev2", + "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.dev2/d139acf64685f15794bb983ff6eb881bdd31304bae88b0ce1ed20a54c21f2265/646e661265ff77f78ecc.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/d139acf64685f15794bb983ff6eb881bdd31304bae88b0ce1ed20a54c21f2265/646e661265ff77f78ecc.json new file mode 100644 index 0000000000000000000000000000000000000000..27b0a4d2187e411f24170d67551dc4c7c08dfbc0 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/d139acf64685f15794bb983ff6eb881bdd31304bae88b0ce1ed20a54c21f2265/646e661265ff77f78ecc.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.dev2", + "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.dev2/gemma3_text/unsloth/gemma-3-270m-it/04dbf3b30744c2b2a648.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/gemma3_text/unsloth/gemma-3-270m-it/04dbf3b30744c2b2a648.json new file mode 100644 index 0000000000000000000000000000000000000000..ceaee3579382505ff3873e38e4db3bf3361b84c3 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/gemma3_text/unsloth/gemma-3-270m-it/04dbf3b30744c2b2a648.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": 4, + "capacity_factor": null, + "checkpoint_id": "unsloth/gemma-3-270m-it", + "checkpoint_revision": "23cf460f6bb16954176b3ddcc8d4f250501458a9", + "continuous_batching": true, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 4, + "max_context_length": 4096, + "max_topk": 256, + "n_active_tokens": 4096, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.6.dev2", + "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": 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.dev2/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/646e661265ff77f78ecc.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/646e661265ff77f78ecc.json new file mode 100644 index 0000000000000000000000000000000000000000..27b0a4d2187e411f24170d67551dc4c7c08dfbc0 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/646e661265ff77f78ecc.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.dev2", + "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.dev2/granite/ibm-granite/granite-3.1-2b-instruct/e30e0a23150fd668133c.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/granite/ibm-granite/granite-3.1-2b-instruct/e30e0a23150fd668133c.json new file mode 100644 index 0000000000000000000000000000000000000000..9da2d2cd3632dd940ac81a725bf85e817c4e0244 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/granite/ibm-granite/granite-3.1-2b-instruct/e30e0a23150fd668133c.json @@ -0,0 +1,58 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "ibm-granite/granite-3.1-2b-instruct", + "_task": "text-generation", + "architectures": [ + "GraniteForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.1, + "attention_multiplier": 0.015625, + "dtype": "bfloat16", + "embedding_multiplier": 12.0, + "hidden_act": "silu", + "hidden_size": 2048, + "initializer_range": 0.02, + "intermediate_size": 8192, + "logits_scaling": 8.0, + "max_position_embeddings": 131072, + "mlp_bias": false, + "model_type": "granite", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 4, + "capacity_factor": null, + "checkpoint_id": "ibm-granite/granite-3.1-2b-instruct", + "checkpoint_revision": "bbc2aed595bd38bd770263dc3ab831db9794441d", + "continuous_batching": true, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 4, + "max_context_length": 4096, + "max_topk": 256, + "n_active_tokens": 4096, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.6.dev2", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 4096, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 32, + "num_hidden_layers": 40, + "num_key_value_heads": 8, + "residual_multiplier": 0.22, + "rms_norm_eps": 1e-05, + "rope_scaling": null, + "rope_theta": 5000000.0, + "tie_word_embeddings": true, + "use_cache": true, + "vocab_size": 49155 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/llama/llamafactory/tiny-random-Llama-3/9a1461765dbc24d49ac7.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/llama/llamafactory/tiny-random-Llama-3/9a1461765dbc24d49ac7.json new file mode 100644 index 0000000000000000000000000000000000000000..c9ce2349cc64cf11d749feb64ef458f4677e6e5a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/llama/llamafactory/tiny-random-Llama-3/9a1461765dbc24d49ac7.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.dev2", + "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.dev2/llama/unsloth/Llama-3.2-1B-Instruct/4d1906e093e7da0b8a2f.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/llama/unsloth/Llama-3.2-1B-Instruct/4d1906e093e7da0b8a2f.json new file mode 100644 index 0000000000000000000000000000000000000000..854a8dcab2a473dc026ba079b83a408efc972d00 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/llama/unsloth/Llama-3.2-1B-Instruct/4d1906e093e7da0b8a2f.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": 4, + "capacity_factor": null, + "checkpoint_id": "unsloth/Llama-3.2-1B-Instruct", + "checkpoint_revision": "5a8abab4a5d6f164389b1079fb721cfab8d7126c", + "continuous_batching": true, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 4, + "max_context_length": 4096, + "max_topk": 256, + "n_active_tokens": 4096, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.6.dev2", + "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.dev2/llama4_text/tiny-random/llama-4/16b8e47a972c18b057c1.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/llama4_text/tiny-random/llama-4/16b8e47a972c18b057c1.json new file mode 100644 index 0000000000000000000000000000000000000000..3408e49ff28bf73d02316047a752a2665c2f0917 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/llama4_text/tiny-random/llama-4/16b8e47a972c18b057c1.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.dev2", + "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.dev2/mixtral/dacorvo/Mixtral-tiny/af2d330d6e4b2cf43703.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/mixtral/dacorvo/Mixtral-tiny/af2d330d6e4b2cf43703.json new file mode 100644 index 0000000000000000000000000000000000000000..cfb064bfd57dbe0094a334e0e4f2390a2dccd1ef --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/mixtral/dacorvo/Mixtral-tiny/af2d330d6e4b2cf43703.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.dev2", + "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.dev2/phi3/microsoft/Phi-3.5-mini-instruct/2375ae0a6703b35f26db.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/phi3/microsoft/Phi-3.5-mini-instruct/2375ae0a6703b35f26db.json new file mode 100644 index 0000000000000000000000000000000000000000..d5cb0b967a02f6f4b309e91a5965d98d2485d453 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/phi3/microsoft/Phi-3.5-mini-instruct/2375ae0a6703b35f26db.json @@ -0,0 +1,164 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "microsoft/Phi-3.5-mini-instruct", + "_task": "text-generation", + "architectures": [ + "Phi3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "auto_map": { + "AutoConfig": "configuration_phi3.Phi3Config", + "AutoModelForCausalLM": "modeling_phi3.Phi3ForCausalLM" + }, + "dtype": "bfloat16", + "embd_pdrop": 0.0, + "hidden_act": "silu", + "hidden_size": 3072, + "initializer_range": 0.02, + "intermediate_size": 8192, + "max_position_embeddings": 131072, + "model_type": "phi3", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 4, + "capacity_factor": null, + "checkpoint_id": "microsoft/Phi-3.5-mini-instruct", + "checkpoint_revision": "2fe192450127e6a83f7441aef6e3ca586c338b77", + "continuous_batching": true, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 4, + "max_context_length": 4096, + "max_topk": 256, + "n_active_tokens": 4096, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.6.dev2", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 4096, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 32, + "num_hidden_layers": 32, + "num_key_value_heads": 32, + "original_max_position_embeddings": 4096, + "partial_rotary_factor": 1.0, + "resid_pdrop": 0.0, + "rms_norm_eps": 1e-05, + "rope_scaling": { + "long_factor": [ + 1.0800000429153442, + 1.1100000143051147, + 1.1399999856948853, + 1.340000033378601, + 1.5899999141693115, + 1.600000023841858, + 1.6200000047683716, + 2.620000123977661, + 3.2300000190734863, + 3.2300000190734863, + 4.789999961853027, + 7.400000095367432, + 7.700000286102295, + 9.09000015258789, + 12.199999809265137, + 17.670000076293945, + 24.46000099182129, + 28.57000160217285, + 30.420001983642578, + 30.840002059936523, + 32.590003967285156, + 32.93000411987305, + 42.320003509521484, + 44.96000289916992, + 50.340003967285156, + 50.45000457763672, + 57.55000305175781, + 57.93000411987305, + 58.21000289916992, + 60.1400032043457, + 62.61000442504883, + 62.62000274658203, + 62.71000289916992, + 63.1400032043457, + 63.1400032043457, + 63.77000427246094, + 63.93000411987305, + 63.96000289916992, + 63.970001220703125, + 64.02999877929688, + 64.06999969482422, + 64.08000183105469, + 64.12000274658203, + 64.41000366210938, + 64.4800033569336, + 64.51000213623047, + 64.52999877929688, + 64.83999633789062 + ], + "short_factor": [ + 1.0, + 1.0199999809265137, + 1.0299999713897705, + 1.0299999713897705, + 1.0499999523162842, + 1.0499999523162842, + 1.0499999523162842, + 1.0499999523162842, + 1.0499999523162842, + 1.0699999332427979, + 1.0999999046325684, + 1.1099998950958252, + 1.1599998474121094, + 1.1599998474121094, + 1.1699998378753662, + 1.2899998426437378, + 1.339999794960022, + 1.679999828338623, + 1.7899998426437378, + 1.8199998140335083, + 1.8499997854232788, + 1.8799997568130493, + 1.9099997282028198, + 1.9399996995925903, + 1.9899996519088745, + 2.0199997425079346, + 2.0199997425079346, + 2.0199997425079346, + 2.0199997425079346, + 2.0199997425079346, + 2.0199997425079346, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0299997329711914, + 2.0799996852874756, + 2.0899996757507324, + 2.189999580383301, + 2.2199995517730713, + 2.5899994373321533, + 2.729999542236328, + 2.749999523162842, + 2.8399994373321533 + ], + "type": "longrope" + }, + "rope_theta": 10000.0, + "sliding_window": 262144, + "tie_word_embeddings": false, + "use_cache": true, + "vocab_size": 32064 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/phi3/yujiepan/phi-4-tiny-random/d2900659f0c3438aa013.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/phi3/yujiepan/phi-4-tiny-random/d2900659f0c3438aa013.json new file mode 100644 index 0000000000000000000000000000000000000000..4c75a90f39ef12dcbc55064c43db84818125d794 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/phi3/yujiepan/phi-4-tiny-random/d2900659f0c3438aa013.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.dev2", + "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.dev2/qwen2/Qwen/Qwen2.5-0.5B/cb3cfc63fc49cc904f0d.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/qwen2/Qwen/Qwen2.5-0.5B/cb3cfc63fc49cc904f0d.json new file mode 100644 index 0000000000000000000000000000000000000000..2cd4afb51e349c9350de1718e44537aedf18cf49 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/qwen2/Qwen/Qwen2.5-0.5B/cb3cfc63fc49cc904f0d.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": 4, + "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": 4, + "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.dev2", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 4096, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 14, + "num_hidden_layers": 24, + "num_key_value_heads": 2, + "rms_norm_eps": 1e-06, + "rope_scaling": null, + "rope_theta": 1000000.0, + "sliding_window": null, + "tie_word_embeddings": true, + "use_cache": true, + "use_mrope": false, + "use_sliding_window": false, + "vocab_size": 151936 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/qwen2/yujiepan/qwen2.5-128k-tiny-random/ee9ffbe1ac4a7ad2ee51.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/qwen2/yujiepan/qwen2.5-128k-tiny-random/ee9ffbe1ac4a7ad2ee51.json new file mode 100644 index 0000000000000000000000000000000000000000..05b1c9e977e9ce897fe5a98011d3b19f37bb938b --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/qwen2/yujiepan/qwen2.5-128k-tiny-random/ee9ffbe1ac4a7ad2ee51.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.dev2", + "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.dev2/qwen3/Qwen/Qwen3-0.6B/15af6878da0b86d6586a.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/qwen3/Qwen/Qwen3-0.6B/15af6878da0b86d6586a.json new file mode 100644 index 0000000000000000000000000000000000000000..3d7ea8538f80747d89895c77886c094bf04b2f69 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/qwen3/Qwen/Qwen3-0.6B/15af6878da0b86d6586a.json @@ -0,0 +1,87 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "Qwen/Qwen3-0.6B", + "_task": "text-generation", + "architectures": [ + "Qwen3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "dtype": "bfloat16", + "head_dim": 128, + "hidden_act": "silu", + "hidden_size": 1024, + "initializer_range": 0.02, + "intermediate_size": 3072, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 40960, + "max_window_layers": 28, + "model_type": "qwen3", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 4, + "capacity_factor": null, + "checkpoint_id": "Qwen/Qwen3-0.6B", + "checkpoint_revision": "c1899de289a04d12100db370d81485cdf75e47ca", + "continuous_batching": true, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 4, + "max_context_length": 4096, + "max_topk": 256, + "n_active_tokens": 4096, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": false, + "optimum_neuron_version": "0.4.6.dev2", + "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": 16, + "num_hidden_layers": 28, + "num_key_value_heads": 8, + "rms_norm_eps": 1e-06, + "rope_scaling": null, + "rope_theta": 1000000, + "sliding_window": null, + "tie_word_embeddings": true, + "use_cache": true, + "use_sliding_window": false, + "vocab_size": 151936 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/qwen3/Qwen/Qwen3-Embedding-0.6B/7c0c5a6095e090c4f577.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/qwen3/Qwen/Qwen3-Embedding-0.6B/7c0c5a6095e090c4f577.json new file mode 100644 index 0000000000000000000000000000000000000000..b478a799c5350dca3ac4e7e5862b7e9e12a434e3 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/qwen3/Qwen/Qwen3-Embedding-0.6B/7c0c5a6095e090c4f577.json @@ -0,0 +1,87 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "Qwen/Qwen3-Embedding-0.6B", + "_task": "feature-extraction", + "architectures": [ + "Qwen3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "dtype": "bfloat16", + "head_dim": 128, + "hidden_act": "silu", + "hidden_size": 1024, + "initializer_range": 0.02, + "intermediate_size": 3072, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 32768, + "max_window_layers": 28, + "model_type": "qwen3", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 4, + "capacity_factor": null, + "checkpoint_id": "Qwen/Qwen3-Embedding-0.6B", + "checkpoint_revision": "c54f2e6e80b2d7b7de06f51cec4959f6b3e03418", + "continuous_batching": false, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 4, + "max_context_length": 8192, + "max_topk": 256, + "n_active_tokens": 8192, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": false, + "optimum_neuron_version": "0.4.6.dev2", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 8192, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "num_attention_heads": 16, + "num_hidden_layers": 28, + "num_key_value_heads": 8, + "rms_norm_eps": 1e-06, + "rope_scaling": null, + "rope_theta": 1000000, + "sliding_window": null, + "tie_word_embeddings": true, + "use_cache": true, + "use_sliding_window": false, + "vocab_size": 151669 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/qwen3_moe/optimum-internal-testing/tiny-random-qwen3_moe/a4bf640c57a2ab4c037e.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/qwen3_moe/optimum-internal-testing/tiny-random-qwen3_moe/a4bf640c57a2ab4c037e.json new file mode 100644 index 0000000000000000000000000000000000000000..4bf6d5f0262aec2537fdd0ed394528e794a49135 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/qwen3_moe/optimum-internal-testing/tiny-random-qwen3_moe/a4bf640c57a2ab4c037e.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.dev2", + "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.dev2/smollm3/HuggingFaceTB/SmolLM3-3B/a20cdb6c0f8b24266437.json b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/smollm3/HuggingFaceTB/SmolLM3-3B/a20cdb6c0f8b24266437.json new file mode 100644 index 0000000000000000000000000000000000000000..b1db6c6630c53cc7318685bdcc773549d644747f --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/smollm3/HuggingFaceTB/SmolLM3-3B/a20cdb6c0f8b24266437.json @@ -0,0 +1,134 @@ +{ + "_entry_class": "SingleModelCacheEntry", + "_model_id": "HuggingFaceTB/SmolLM3-3B", + "_task": "text-generation", + "architectures": [ + "SmolLM3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "dtype": "bfloat16", + "hidden_act": "silu", + "hidden_size": 2048, + "initializer_range": 0.02, + "intermediate_size": 11008, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 65536, + "max_window_layers": 28, + "mlp_bias": false, + "model_type": "smollm3", + "neuron": { + "_serialized_key": "NxDNeuronConfig", + "batch_size": 4, + "capacity_factor": null, + "checkpoint_id": "HuggingFaceTB/SmolLM3-3B", + "checkpoint_revision": "a07cc9a04f16550a088caea529712d1d335b0ac1", + "continuous_batching": true, + "ep_degree": 1, + "fused_qkv": true, + "glu_mlp": true, + "local_ranks_size": 2, + "max_batch_size": 4, + "max_context_length": 4096, + "max_topk": 256, + "n_active_tokens": 4096, + "neuronxcc_version": "2.21.33363.0+82129205", + "on_device_sampling": true, + "optimum_neuron_version": "0.4.6.dev2", + "output_logits": false, + "pp_degree": 1, + "sequence_length": 4096, + "speculation_length": 0, + "start_rank_id": 0, + "target": "trn1", + "torch_dtype": "bfloat16", + "tp_degree": 2 + }, + "no_rope_layer_interval": 4, + "no_rope_layers": [ + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0, + 1, + 1, + 1, + 0 + ], + "num_attention_heads": 16, + "num_hidden_layers": 36, + "num_key_value_heads": 4, + "pretraining_tp": 2, + "rms_norm_eps": 1e-06, + "rope_scaling": null, + "rope_theta": 5000000.0, + "sliding_window": null, + "use_cache": false, + "use_sliding_window": false, + "vocab_size": 128256 +} \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_0007bf645cd8f44068db+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_0007bf645cd8f44068db+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_0007bf645cd8f44068db+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_0007bf645cd8f44068db+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_0007bf645cd8f44068db+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_0007bf645cd8f44068db+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_0007bf645cd8f44068db+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..e738d22a04cf8409ac46e869018434bfde1f43c0 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_0007bf645cd8f44068db+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f605a954b114104fa81403712a2efbe24d99be08d5bcfe37bb5b9a565639ad02 +size 443127 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_0007bf645cd8f44068db+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_0007bf645cd8f44068db+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..08c5d494fff98a1fe6675a4c7b9a04ed72892d4a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_0007bf645cd8f44068db+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4959586c29d7dbe597ae03fad6b1c0ffc0f7bc0a0b574257c2a43c418ff8469c +size 2171904 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_0007bf645cd8f44068db+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_0007bf645cd8f44068db+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..6b53af04d13787b43ec25244ec25142f6709c473 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_0007bf645cd8f44068db+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82d721b25142da6db296f0e7a48219cfb200220310600e0214afb4872bff2ad5 +size 2245991 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_001a8c20b0b7f8f20fbc+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_001a8c20b0b7f8f20fbc+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_001a8c20b0b7f8f20fbc+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_001a8c20b0b7f8f20fbc+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_001a8c20b0b7f8f20fbc+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_001a8c20b0b7f8f20fbc+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_001a8c20b0b7f8f20fbc+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..3f1b7222ffef38421e6ec4b8d77b8b21eb144aa6 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_001a8c20b0b7f8f20fbc+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0c09aa6c6ce30cf490abc4d08daa94efa87fdf35526d5dab840ab7d65edc245 +size 788292 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_001a8c20b0b7f8f20fbc+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_001a8c20b0b7f8f20fbc+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..c68a806e165d16b84f837961a7e7a5a0f5970cfb --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_001a8c20b0b7f8f20fbc+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50181c53053b5860557d2b2c973ba3f2d6cb1669515bf66233848f12bfe0d880 +size 2571264 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_001a8c20b0b7f8f20fbc+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_001a8c20b0b7f8f20fbc+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..21eafc6f70a169d91a34563c2c53a2e2cec63c35 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_001a8c20b0b7f8f20fbc+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b27b0a5c00f9712024400bbbb02191386282a0c2b2f55097b7e26e056844d59a +size 2708432 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_0680112a1970faff8f3a+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_0680112a1970faff8f3a+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_0680112a1970faff8f3a+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_0680112a1970faff8f3a+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_0680112a1970faff8f3a+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_0680112a1970faff8f3a+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_0680112a1970faff8f3a+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..19c16cfc86a6f1d72b1fe130742a1de40c66e10a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_0680112a1970faff8f3a+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9aede2c8d2d65ee80658dcddb37304acb08ec589dbb2707962c654478719d7fd +size 904137 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_0680112a1970faff8f3a+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_0680112a1970faff8f3a+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..005cd09a4effa1e19f38cbd37316cdaf53cdfc79 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_0680112a1970faff8f3a+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82cfc3dfcdb273c5075fd53c7cf5a1969e63dbeecdcd045794264fde9742392c +size 4998144 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_0680112a1970faff8f3a+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_0680112a1970faff8f3a+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..c96dfb52d1a0e3f3f832894cb90d003637ed7431 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_0680112a1970faff8f3a+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a58280421d3ed9a1e0893521b9dfa1c92a6db24cd230c60c0687bfefc18e04c +size 5164668 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_0a1bf3b6e86f854781c2+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_0a1bf3b6e86f854781c2+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_0a1bf3b6e86f854781c2+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_0a1bf3b6e86f854781c2+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_0a1bf3b6e86f854781c2+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_0a1bf3b6e86f854781c2+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_0a1bf3b6e86f854781c2+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..717d36ea4a225299bdf0d0646c14ae0f158981f1 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_0a1bf3b6e86f854781c2+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47a649ed18d8cf22148cf1727477a6e0cee892573790834d9455503d83038ec7 +size 85208 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_0a1bf3b6e86f854781c2+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_0a1bf3b6e86f854781c2+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..afcec2dd0eb4063ac5dd28f4b2f1e21783d25470 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_0a1bf3b6e86f854781c2+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:546630b25a8444602e63c1aea3a8363e60f2c8c09e9681284c833f6e05eb395d +size 257024 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_0a1bf3b6e86f854781c2+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_0a1bf3b6e86f854781c2+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..463c09c2ea6fe1c3d86c4a02dfc5702fd71b47ef --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_0a1bf3b6e86f854781c2+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70e826c53104b3589c79de96bde775712a147396c7b4b681ed3453a7227629ba +size 265207 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_0c7914f92b283d2acea2+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_0c7914f92b283d2acea2+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_0c7914f92b283d2acea2+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_0c7914f92b283d2acea2+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_0c7914f92b283d2acea2+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_0c7914f92b283d2acea2+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_0c7914f92b283d2acea2+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..ab01b86bd6ff7f378a52350f5d0635741acbbabf --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_0c7914f92b283d2acea2+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4eca84f3fe1c63f66a254fd36737cd3ccef69e39730abe1047b4d74798cbddd +size 1131253 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_0c7914f92b283d2acea2+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_0c7914f92b283d2acea2+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..114e7049cc32a2ffb61b306ae61f77025dab052f --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_0c7914f92b283d2acea2+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e09a422912c4c46309de2e9aa22a341e56c1b5e55d6e022f7e190759be32a96 +size 5305344 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_0c7914f92b283d2acea2+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_0c7914f92b283d2acea2+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..57c9b3e9cacc93cbaed26cd3340964a90a4b4c5e --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_0c7914f92b283d2acea2+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a440b46964b33cf8b9efd38e73241796b21ba255a6a996e5112f13b8d6c3fd02 +size 5489523 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_10244305442015770634+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_10244305442015770634+e30acd3a/model.hlo_module.pb index 2537e60527205941d984fdd93ad6b7a0cd6b0f15..3991e08562a61b997d7b7f88f577de9d7425592d 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_10244305442015770634+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_10244305442015770634+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0113611968c8ed0cbeaea5ff3bba2ce4eecd5fc989e2592c6b8ac5fdf09c91c2 +oid sha256:e8857bc71bce3a47e270445722e6c35ea87adad4a29c0ed7da5c42bcb9d4a235 size 1562 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 a8606cef4ecf1e11bde21fd15e59610ad4a47f83..bfcbc3ccd8e1d818fbb50305539bd5e0b66e40d0 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_10606948783918825529+e30acd3a/model.hlo_module.pb index eb7582b902bd2edb54b480408228e6707405eacd..38f880d95305e20bd388ec54b0ba721a62fefdb7 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_10606948783918825529+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_10606948783918825529+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a35335d3e96ee86956cdf93862bf42ae6f83017e9b331b77531ad6942a62366d +oid sha256:a986626ba85932cdfa72d7fd785a84a7841a1030e9093b06b094c36b7cd00dd6 size 1269 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 dc993c479ba2e8eba852d67ec4921eecf2866327..abe3c971ff2aa835efd5b0fdc325a5aedc179ee6 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_10645643398657092095+e30acd3a/model.hlo_module.pb index 3ad9c3bfd79a132579dd5f782b30383d991bfe54..004f76640104ee3696144319d7049feb952a2af0 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_10645643398657092095+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_10645643398657092095+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:153fdddade7e1b284f8d78df84523bd332eccc7ddfc7a0698ad8653d560c55ca +oid sha256:5809f339914003355e65777d2fc04033f1e440f0ccaf17193f8c82d7e8239456 size 1124 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 8b0ed41d93842fb649a771f51cc3c67a50527add..16f669d2d1aed2363bebe3a5941b50e0073a8f49 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_10661660426924300837+e30acd3a/model.hlo_module.pb index 933103060a7f2312d11c5ae375b27203b96a5d79..0638ae422d49522c240818f52a1c11ffd4093c85 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_10661660426924300837+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_10661660426924300837+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7434cf1ede5dc6759a8ad5cdc472fb8e1668b1a8bd8dd7e7e08b5c09401b8065 +oid sha256:43a22567c4e7f4f23930e75eceafe9c3e24fc04f874fa613d73e420a262403a9 size 1562 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 5eeabe500332635f3de3b9a219afa23cbf345c8c..5e9f52cee0a437f078d59914338fe6640276c4d0 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_10746122569655005679+e30acd3a/model.hlo_module.pb index ec05bf8d51fc25e29356d9107de4ebb162ce96d2..75de0c11a24f0e0142a6de5bfe869669cff39eff 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_10746122569655005679+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_10746122569655005679+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e7000c9195e6c3c9abfcbac09928468ef4c204488094a046ec7abfd1b7ea8657 +oid sha256:6239ad1a4e7efb29d39953b88372ea15911f12ac0ea8d9656c92d7863a0b9a41 size 1563 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 655c7df498a0eefe66e15357b3ff908c9970962c..3c72391cfcd5f766a8dcd51b2cf809e36752652f 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_11086318750207148626+e30acd3a/model.hlo_module.pb index 2e03205f1e361ab764610e16e461b9b00a8c03ca..9dbb8340ab1390b17de1f1cd68cb378ba59e0c2c 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_11086318750207148626+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_11086318750207148626+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2fcfd4f1816e23bee263009b76a032d3de2454e4937f73423fade31c7a0f0278 +oid sha256:c4e4fca2e2d0f999693dccb3c833fde50612b85f5f12fe91ae8dc419a8ebadf4 size 1562 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 dfc620176ff019c1743b5f2e7609f6281d673d8d..8cf9fed9ff69c70e389940c3ff4115d76242d03d 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_11909310600244571805+e30acd3a/model.hlo_module.pb index 2f82cf1c07efabf7f4ab0f220b1b9575678a758b..605b11b03d8afaf0637505d4e60e2461bb4d56b1 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_11909310600244571805+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_11909310600244571805+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:701e22032f5900dff59bf252dc2804e14345ee4b52f9587d7305edd2a497770f +oid sha256:43e4ee6ab08d3241ae1b0776c0e1e14bd38eece7694e57cf40b9f2065f671b16 size 1269 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 77a913fc6af039c549045000394facddd423edd2..3af09fe76800c47d9367a2b13077c47ceae51587 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_11933818254123612383+e30acd3a/model.hlo_module.pb index 7dc0ea46bd2b69b761f6695a7fa5949ce0b2d1a5..bd8b9cb34ad0e2a284044cf957f11976718e7e2a 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_11933818254123612383+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_11933818254123612383+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a3ba973ff38c3c4be9f7ceda4e02dd7d04d9799bf7507445264e5a513b8bcc4e +oid sha256:e3d9897d12071a12ca2e14456c514db2c211bf8d57937bdfcc7a9e85acbc7aed size 1124 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 febd76225c7aed9f27a371387f682b56272aed23..2519c3a4680d2530c6872488e60fedc251f6a979 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_1210392327607194823+e30acd3a/model.hlo_module.pb index af5b3a3e420298db1ca7eb70b24cc993252e2df2..f20f4f0fcdcee9fec770958041e8ef6fc20a5ccc 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_1210392327607194823+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_1210392327607194823+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1d49de9ffa9ae82afa026ae1eb0f4b7596894c67fa760474a7363dc25657e8e7 +oid sha256:498c87cc6dedd16291dc3889c1c4ad2c55a99cc9e751509422dd2e4d39fe2a6b size 1562 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 d4a3ee581bbf32ecbeaad6bd02576cd642901d7a..9e26357dbe09ef3f80bb1da3a3b6a32c96b2fbe7 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_13102010590082783346+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_13102010590082783346+e30acd3a/model.hlo_module.pb index 43bf5107ff083bb19e2ebb030dd638fb6e003264..035933557cf5c7dfc807b086fe88103183d5ad64 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_13102010590082783346+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_13102010590082783346+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:57f0928a33533f819a379b0ad9ff0ae0c1c1b54e6ae13339d55ea0610a9e2389 +oid sha256:86c949e40344d781a7a27174b5ea4350f7dba5c8778007ac8865390c8cb672b0 size 1562 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 bb21a30124bf2b6e2a420394e9acd193e720e383..59a3fed7f86039cfb9e3b1c314e7ed9e5e0b38cd 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_131d6c7c245c7131bd76+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_131d6c7c245c7131bd76+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_131d6c7c245c7131bd76+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_131d6c7c245c7131bd76+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_131d6c7c245c7131bd76+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_131d6c7c245c7131bd76+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_131d6c7c245c7131bd76+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..99428203f594a84d1ee466e0f301526cf8f42f73 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_131d6c7c245c7131bd76+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:194247f6c266078d61032d45dc11cfdc31dcce5959d33c6337fa7d27454171da +size 87126 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_131d6c7c245c7131bd76+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_131d6c7c245c7131bd76+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..5c0ffc1a4112e38961156e7522ea00898cee93e5 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_131d6c7c245c7131bd76+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b94924e9a5902f3716cb75da310be67a1a3c4f14f2c1dbfb7999fdfe0f1dadee +size 287744 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_131d6c7c245c7131bd76+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_131d6c7c245c7131bd76+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..3dc8a3ae418dfc53b4ba9c08ab266cfe13f577c4 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_131d6c7c245c7131bd76+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7489edb28de75b5a233411b7723a46e902615262b35c0ac6ac9887eec09236d +size 295957 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_13475273426270778455+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_13475273426270778455+e30acd3a/model.hlo_module.pb index 1aae8517a655f99b06a2cdca212dec6a3faeba15..50bddc07741991a22c8217953f323ce320328831 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_13475273426270778455+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_13475273426270778455+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a591de74b26ea85b3c61bc1be494c47f9c1fa8923b3b9f6180e955eabaef6657 +oid sha256:b16bb3d9322f5555a9b2acf982cd22461cfc5c092ff7e50e3abbabfa6abb3563 size 2839 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 14b8890877311c150270cfc660cdd68167e6ed86..248622b5c78f31ce4db62da5329d2706a37e2416 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_13728813963059599796+e30acd3a/model.hlo_module.pb index bcd627e4879d274ec124e61fa05ae10f66bbfbf0..0587e621428b08da9948f4496b1bbf25fba78eed 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_13728813963059599796+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_13728813963059599796+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e26c0bd2ba77576e5498a156696cfcfeac1be86021368feb0908e79755eed533 +oid sha256:780133bcf4e604fe47a0ea0f2d719a8f44f049443e99322953116328624cea16 size 1124 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 c301e14ad2ca998cfb8b2dee9a5204b926ec12ba..5c9ed2356a65031b8e42dfd36aef8e3a8d6498a6 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_13783872956654054643+e30acd3a/model.hlo_module.pb index 32ab6d67f3b6497f57a2c0736c753f5d903db802..5e262f77bf4379a85356260d33740a0c5ffc7cbe 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_13783872956654054643+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_13783872956654054643+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:199ec7f00113f98677fd447807a97ab72737d90617fda54a1af4afd18c9b33d3 +oid sha256:c72c1a25e0886e24915548c61784a7b21394458e2ba3b7f856f425695bbd85a4 size 1562 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 0ce737b31871905ff51341b98b0c15ad9915a77e..b2a67f8011119ba007e8db9553816492fda61499 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_13ceff346d0e3479704a+6e4949b4/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_13ceff346d0e3479704a+6e4949b4/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..2ad69cd01383a682715fed6c7faaf92a72533ca6 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_13ceff346d0e3479704a+6e4949b4/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--enable-mixed-precision-accumulation", "--model-type", "transformer", "-O1", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2", "--auto-cast=none", "--internal-enable-dge-levels", "vector_dynamic_offsets", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_13ceff346d0e3479704a+6e4949b4/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_13ceff346d0e3479704a+6e4949b4/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_13ceff346d0e3479704a+6e4949b4/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_13ceff346d0e3479704a+6e4949b4/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..9dfa4b248bf7c86f476d45bc072d65d054c81302 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_13ceff346d0e3479704a+6e4949b4/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c235fd154884dae0bb0408b3066c676d3279b38350a7637b8a4248b6691c936f +size 84807 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_13ceff346d0e3479704a+6e4949b4/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_13ceff346d0e3479704a+6e4949b4/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..f83f5aad0ddf31f8ee67358ed0be55d0bfff2566 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_13ceff346d0e3479704a+6e4949b4/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10b906ddbae4fb13528dc91ceda5cb21c3da0d86586f0cf463ec33be1141dcd2 +size 646144 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_13f403bb4f61de68acf1+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_13f403bb4f61de68acf1+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_13f403bb4f61de68acf1+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_13f403bb4f61de68acf1+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_13f403bb4f61de68acf1+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_13f403bb4f61de68acf1+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_13f403bb4f61de68acf1+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..6b2cf6d140ae61b3d93781e11c51fd697a7ac18d --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_13f403bb4f61de68acf1+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05a160b90c94a7d78c01e57e3bda99556c41b92cbf5abfa643b3f139cdaf7cd8 +size 668949 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_13f403bb4f61de68acf1+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_13f403bb4f61de68acf1+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..e5d7f91e68ad48959e85fd620494b3fc6ab514ea --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_13f403bb4f61de68acf1+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1e3834afeb2101bd0a28a456e2facf6076cc0cb50d1616940b6104b6c873486 +size 1885184 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_13f403bb4f61de68acf1+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_13f403bb4f61de68acf1+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..ed88cd7b46bc461f5407de935a97aaa232b430d7 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_13f403bb4f61de68acf1+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:938e9910fa13c306e5a0f6f4b2e79abb8063a1190b7c0c32c58444cda5e68695 +size 2041490 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_14097776130612710282+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_14097776130612710282+e30acd3a/model.hlo_module.pb index 557c2105986ba3f85e1286f9ede7c19f0a9c64ac..61cb38b394c1c8144276d47603e482af5a92267a 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_14097776130612710282+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_14097776130612710282+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:edf55ecff7d3028bae2cbf5f2fbc03167c198e64aad7e464445ac6a5e2424538 +oid sha256:83f8282952bf7f4a3c90dbff8d986e98374ffbce31e3ff79a7673e8fe6bf2ffe size 1562 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 00f597960b6520225b1b5700cd01b1b220819b4f..7ca6b0420617baa20ac2cda07585942164422c2e 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_14155105943057125178+e30acd3a/model.hlo_module.pb index a63dd40ae39f47f53c81ec17460d40a392273c2e..3f33cce06eeaf23c4466425d62f4a6e0868852d7 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_14155105943057125178+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_14155105943057125178+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:10e54a0a2ec0667067a1dfc61da1f33cb3266c4f6da202d67a9669c2fcffe960 +oid sha256:c1762ba9af8abcdbcfcc241679596a9d954b072ae9166c4bddd7c266e56d354b size 1560 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 0b09bedd8655394baf2df2eb3224a0db93996a5d..70d4d1fd66a5de63443934e861a2bd050f50db50 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_14173682194645188821+e30acd3a/model.hlo_module.pb index 80d4cda0fe6fc602da5ceadc769d2dc3c37e08e6..cb67d9c734706aabb303d76d59c91dd5f7600cce 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_14173682194645188821+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_14173682194645188821+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9419e130aa8fd79326ba8327b8064d560c0f35f608be7c38677a89ba67b9ba53 +oid sha256:cccdc5b5dc0f627c390a1d8ea6493eba66bb431235eb7cc4267a383051e1de43 size 1269 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 faf81f22d0a607a35e35a655f448d08d0a99f477..3379a621242b7e62d2de83f11e4c189dbd333d3a 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_14274833313744358348+e30acd3a/model.hlo_module.pb index 3440a59e17d5181dc3b9cdd45fe1d22f07f406ed..8e1c267e6662242bfff0d337675d60e661040db3 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_14274833313744358348+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_14274833313744358348+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c2ce50fd35b56002a320a0e665a0140adb13b78d610af21a12809da39de7890c +oid sha256:05983ba390bd275e95b015e98e80b34cb3840f84d7ec485271a4d6e5d72abb89 size 1563 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 8e179b820ff158382253ada7846dbf3a09832f61..1506f8ac0816a8f5e5a670daf5aa45ac6b7f0954 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_14469716438804935215+e30acd3a/model.hlo_module.pb index 97051668f16154c2861ada42d8b38fedc1942184..e5cd6c2940bc5a8b119e5d737e6d5f4f260f7897 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_14469716438804935215+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_14469716438804935215+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ce8eca7f514016ca7ddf18c32df5168016ffca7741edb33f4b602c1efa072fcb +oid sha256:4b98ceff411119e55712d559f77ef621ffdf1a225a67c89570e985cc3e3c3895 size 1122 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 9783f73ca61cd64e76591a01487616e7f7f39435..f76c9d31c7ef31e723d2d3a7259aa0480a9ba24b 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_14739672283771660808+e30acd3a/model.hlo_module.pb index f7450f5ec21a1ec9e73bdfa0902db77908f05f6e..c4be878f8afe2cf4ba5d5de64e9032c1af0d8cd3 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_14739672283771660808+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_14739672283771660808+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2781f132c46e031208d39895974a19362c478bc055747d8ad4afb390abd652cb +oid sha256:f2b97aaa8498e010551a8ef5d5aa65bc3a6912f7cc3cf7f55edd572a1a55ba45 size 1563 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 e5afcaf440bbd9e9e317f95f0c77e5c2deb7c86e..b94b86f03e8f9e51c0b07622bb4d92d042eb551a 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_15557971880365771457+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_15557971880365771457+e30acd3a/model.hlo_module.pb index 0b0ba3bd87b1cf10544ede814622d1313a99a3f9..b92c2e06d431bf199e1bebbdce18ad3371e26e0c 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_15557971880365771457+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_15557971880365771457+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4bec3e6e32c2946a75cf366caffcf035c59b4ed66669998a7d2eb9e92746a235 +oid sha256:0779c164d4f3ee1f86fb275ebb86471435fc7ca0a67b85210429e444044a5cb9 size 1269 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 90a90f5a20c75fde7eaf08370f6589aaad169bcb..922215f016e0a129740249043478fb650257c6f0 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_15564425602929126510+e30acd3a/model.hlo_module.pb index 93a3b21ba9b4d60786523ff171fc51adfc06c7b0..a889f695bac1c6219da3e873fca5cf07b7ced9b7 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_15564425602929126510+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_15564425602929126510+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:97ecef61d0a9738171f16b5fc8a68eddab4ce2d2084a47dd73f3c848138efae2 +oid sha256:294f9e30c896594920919185505c9d5c35dd47fc63c064a39b50b21b30b88228 size 1563 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 5050507c22afed7e641b2ca1aab3f3e8293fb043..5136f7a0f856691f23eb6c35b6d4243d8cb4768d 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_15589838460896944293+e30acd3a/model.hlo_module.pb index 6016fc36a1ef3b3cb5bf0588ff450d9ec89546ef..fe15cb485f61e5d442fac81d047f9b7936ec8925 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_15589838460896944293+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_15589838460896944293+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:45584387bcadd48cad47007b2515871b4d8501b6cf5e10a63bcbb0293e3f8f2c +oid sha256:cb28270f0d514a05f868fe32b59caa34b646516cc261af7622fd42d15673d752 size 1562 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 df0aea2de78b69bb39926ef966f754e7109b3ca5..434ef6a021394b8fa4944cc4f3bb62b2469c54a6 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_15654572849171857535+e30acd3a/model.hlo_module.pb index 41bfb6c0d020dd122d3f8c6396d17a3b2b193072..bd4d97bae44d2b9c90d825051b6b2e37941b6789 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_15654572849171857535+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_15654572849171857535+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7a430bf143c7170afa093bd06643be2f181456b01992615d3e06ad9b9c58c732 +oid sha256:e5811848f0aaf579ca1135035bf264660958c8bf38f24218776c413ece972af4 size 2839 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 f8b195f534933983bb9126f78d65f31053fb3287..b72822a2d5fb3a047b8e5113ff6699950c4a9e28 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_15767383571209512795+e30acd3a/model.hlo_module.pb index 4846661d33abe144a46e8a8511a4453f273a7583..003a32339e91c6c9ed2a90e93d655c8696104551 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_15767383571209512795+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_15767383571209512795+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9320e8826ce4d627cb8c5441eb549e929da7a656ca17b9840b64f02fb6012e96 +oid sha256:8f1ee95a4ba7491ed2054e6d00f99d98f5ca6f7558594ce80e60d69a18b8ec72 size 1562 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 a8272a8985c95a56f455ccf64776ed3f69324c38..10aedcba5fc1b117596c1137e8b1c6d0a1c89103 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_15888836045088309511+e30acd3a/model.hlo_module.pb index f9b49ab6317064eb1ab2274498c53daa6eca05bf..5a03e68079e66f3b2c2ea18caab26b8a131bc766 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_15888836045088309511+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_15888836045088309511+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:14c9b1d5fc37de8bf294a5f8ff61d7f7db479ba96b0467007d77f370caf64468 +oid sha256:f33e06b0e9f27a318f5ab923c7b5738ddab4b8832dc0348f747717b6dd46c54f size 1562 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 4a79766a563798c53b1043a8c99c861c62c9f42c..3c1547a7134ec1168d64a16a9616c2c425de6345 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_15929451261464042997+e30acd3a/model.hlo_module.pb index 2b1365fdf9429caed333c19b1170509e0e786735..f648848049ed68d540f6167c394aa55ad9b370ac 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_15929451261464042997+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_15929451261464042997+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7758ff75875978a913c9926f5e19c1e4a4a478250069b1e432e104536d6acdef +oid sha256:6d3c8b83e7f3bb71945abb0b7ce3c09a9ac717f01fb7f431a7a4a175035efa5f size 1563 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 f8d8679292217de71fb8e743b63bd62b91445e77..f247be16bd504b1534536d62da0292ccb386001f 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_16016505958416521648+e30acd3a/model.hlo_module.pb index 60030f6ff406f8a5644055f6b4b2ccff16ffad32..0b5374d68b4ba8cf3e1f6d319df2f6c55e8c3cb4 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_16016505958416521648+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_16016505958416521648+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8f100e3073b2d7739e1fd872b5bb2f099fe1128a7c805a326ac6ae8c044d4b03 +oid sha256:01770bf8156da3c9d6bbd2984a9362a5047a7b1746a80e2782e233361bce8661 size 1124 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 5c3284f425688eb15a40bb03137e7317b495080e..803a7a1ff16344cabebf0154f5ea3c38bc4ed75b 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_16053163479112702088+e30acd3a/model.hlo_module.pb index 471a4ed1a3838d951011e8210a48c87277ff0d5b..835c5bba3e4b6db9ec3e598f42a25fcc4f77b789 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_16053163479112702088+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_16053163479112702088+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ab9a6c15381056cc07bf3a7da82329fb72e5357cd91a45c2a57c37df4c714a82 +oid sha256:a0282f0dff1f473788c962dbcb8cfc0a8b4a4968d92c647c79686f548fab7473 size 1124 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 c6c03386f27cdbc0c34a4d99d851211c6e3cb35b..01a5fd7aa74b85507dae40c42963df0bd3c05f0d 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_16204387852795926216+e30acd3a/model.hlo_module.pb index a6376da61b349e1afc7fb0c4d50d69b0e1cda5e9..a724562b5977a73441d0be507747ec38fbb715c6 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_16204387852795926216+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_16204387852795926216+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:3d0ff42338f985ea765bd39144baa05fddcf04e2e0c87f9a579dc81dec070344 +oid sha256:f2f0264b4c4f1c0c0c99a684331301325c85f750727255f2aa6836b7881d1d0e size 1562 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 29442f1d86d515a26b5e4a2a37d7ea7334187853..d2abc01343cf4bfdf88db5a8dc3b26c2610d009f 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_1656719109221189948+e30acd3a/model.hlo_module.pb index edea36e568de0ab02cc13346ebd9d5e0cfd8d7c3..93abb067a029131691e26543db4032bb9c7fdba1 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_1656719109221189948+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_1656719109221189948+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:10871e9557189f265c8590afe0579930d47d192d2388ab7c54129a14d25447ac +oid sha256:108dc50d1709d33f0b153631eddf1954a12bc58d9d3bff8c9c5f426f54418336 size 1562 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 52e198ed647f516462b33c7490b81d45b98bb98e..47e86c42107ef769e7af35e05a6bab8400eaa9ae 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_16599571375348449904+e30acd3a/model.hlo_module.pb index c0d044272042a3ca314b1393ce4dbb89cafac36d..1eb26a4c103b780f5d2a7533fe4308c99c843bd2 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_16599571375348449904+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_16599571375348449904+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:be771db5b4c31679b8e509e6af1c059ab3aa543f3e241c53836c5bd228edb83f +oid sha256:0ceab3e58b37f7243dfd77f6beb88a3536963c89b0fb06d088b1fa73ac32baf0 size 1562 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 18302b6ef279113bfb9beb4c920a35192d09d00f..ce3fe7262a6b13e0090e8ce863606a23aa3ed090 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_1663757798483801648+e30acd3a/model.hlo_module.pb index e6792d53f0585b86fc564270f53adb548756d061..87d5d9965782ddca81873fe40c8b0c4a58ea6b3f 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_1663757798483801648+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_1663757798483801648+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1ae51e2ca45b600b16b5ca4db1d0eb0b08c3f6b52af11722ea501893a8ed3d23 +oid sha256:d34643956946d3f9f55aa284d64073a6dcaa8a657ea8fc779ec4f2ace1b06b68 size 1124 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 a99bb1b22d0585c87ea3fd0290b51131e5070f02..18fc2b633ab31494213ab00f24d76b6f20794938 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_16738296820980389103+e30acd3a/model.hlo_module.pb index 8294a60306648ef6ce885c400e5a609634060f6b..71f8f335edf59e356441d9dbbab26df2e73b22ab 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_16738296820980389103+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_16738296820980389103+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9164c961d569ed6684d61269aa435d636d22d94ae8310d04257b7a8e3f5fd470 +oid sha256:42ff1bbb4f463240a2af990ec3883cd9ada599176b8cf62338a1ba60b264a38d size 1562 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 8bd3ed758f9175d2e7386f4f6f492bf319c460a1..1da4ede3b6e21636a93172015663033368365324 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_1696825468766062114+e30acd3a/model.hlo_module.pb index 525eda02fb8262c9abb5b8a5f4320ec4e8ae224a..5fe0b70cd26f98717997a1f5d29ef6b4c58adccb 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_1696825468766062114+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_1696825468766062114+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:31eb937eceb5dcdcc63bb5542e67ca3f8ef80566b085221cf2fd6ea66518226c +oid sha256:f7ecefd23ec64fde9afec3c5a206f1f5585916572dae7b9d09c609d1ca9ac526 size 1562 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 718b80e0c715865b723818041e5469de6c78f8d5..2699a7c27ef9a4ce751acd7de85d370827fd7d6b 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 fecb8387df20f7344a970375e68d18741dc3e665..b6ab11a8d52d9dabec946c3b6fd37764e29e5fa5 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_17781596253725927902+e30acd3a/model.hlo_module.pb index 88a4ef02099481b9d946662e5b2c32baf3a9bc9a..bf5e490478be89f0be23564e930e0318a7e498a7 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_17781596253725927902+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_17781596253725927902+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:bbe570c56e2c771a29adda9deb46e617811178af3af758a6556f2e75e20d6132 +oid sha256:1fb234503b6c46273f7c20d09b2d14b6d06bd1b297fd1692cc33a24fa6c56e35 size 1562 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 7640b7e55db9882284deb3ba2ff1d101bbfd7ae9..e253e16a99226277e02d451b3cb53ffa306e2b06 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_1795740353831178306+e30acd3a/model.hlo_module.pb index 7d5a85d9e03058c4faa66be14d13ae8882927e1b..8ac2233005ef94cc72b46f727806e53f64c52d1f 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_1795740353831178306+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_1795740353831178306+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7ba0b6c2332db3c53fe4edb740ffed1437ef4a6762bff77e81adb105d00e0824 +oid sha256:afc1c6b71db75bc1b5d43b93bb581bcda3568f07ca8702efd0512db8df405b58 size 1562 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 ef025fea120ea674f640202c14f25e9daf7aa661..76f525de2db142f26501a57f4937346843350946 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_1800832390737682969+e30acd3a/model.hlo_module.pb index c024c1179d30e4e7ac77bd3bbfe44f5f362273c9..77ca369fd4f13f57ce95db553027e13d98bb2c62 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_1800832390737682969+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_1800832390737682969+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8ca9ed7c09a1fcd89ad987455681dc5857f4d26abc0322491579cfdddd6a436d +oid sha256:1d998336582669e8875d2de1439ae0c68c849e3933fe4e4adea2d8a773d810a4 size 1562 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 301e52bd81e5211c8e12e9e38329433dc3242734..d79b8b8e3e357edeca6df71571dc9873f3ec49dc 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_1805481651134498710+e30acd3a/model.hlo_module.pb index d1bd00c58e0bb946c2cb9091a2d31a05056199a5..09988c4b4fd44999e241077a94f2c62d88d6d545 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_1805481651134498710+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_1805481651134498710+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:447362e58f9444e9409f472a58a1dad4aa005006b5f794ecc5dd1d062aff6ccc +oid sha256:3dae310d5d05091f8e95e21b0a06b3f53a570b95c07ddc188b5325331c822980 size 1124 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 5608fd418b1e67ff917846bcae9ebd6fc4d7d3b1..cf3be1ee441885d98104deb8a316b4c41e258c15 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_1805728312057401221+e30acd3a/model.hlo_module.pb index c3ab68ad465d47cba7f3790fb0bf33d6a273dbb9..e381da5642f7bab603abd043f22d1523b2835c99 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_1805728312057401221+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_1805728312057401221+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2d64f9c086e86830f9b63c0b8c32b103b1f6b4bcb0a56f0525bc97c8eff7a1b1 +oid sha256:c963b833a6133f7fd55e21ce19b3bf390c5943e719996a9fe00b32d5088434c2 size 1562 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 be91fa26932c76e729f5e6240c9ce5ec282202d4..8aa580e5eb6bd21fcf71f9da0a96dd8e366c05fb 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_18070749384531238695+e30acd3a/model.hlo_module.pb index fc5131986d8a90bfbe832dec7daa6df2e7277e9b..c34c6c655f66e46cd828321a08a795e0131925c5 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_18070749384531238695+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_18070749384531238695+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8f417c3d3891a4aa551e8c388ee608a70eb0070fffbf03c1baa9e0c8f89268da +oid sha256:ea3301fb7c0370a91e3ede747c224b22a45803c0c7a14d669a2ed426962259f9 size 1269 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 3ddffa7f395bfbfffe9b7f26ab6fe128575e4f3b..fdc6035cd75463601749b7b7bb62cb656559892b 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_18134517905922687964+e30acd3a/model.hlo_module.pb index 86ae55e8d48a76d99062a94fff36ceb5b595c7d6..af811d588985647769cf44e7c5451e717fe72b61 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_18134517905922687964+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_18134517905922687964+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b63a01e64fc6b527b267564c81355d0a269278589ee0440913f16d2ca259d7b8 +oid sha256:5124d6aaa66a1e6dc0091101fa9f0fbd01160ec384c86b9ac4352e156aa512d2 size 1562 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 1fefb3958aa04fcd6c795ba2a4a5dde1510ae5f4..1b2f743c41077580c42514fa2fd30561b2d112c3 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_18207475553554060390+e30acd3a/model.hlo_module.pb index d36520a190fe8cfe40bb90f520d93262c51cfe8f..f3194429afdf45cb1511aece042afa35b34b53e4 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_18207475553554060390+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_18207475553554060390+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:bf16d08f74844be992a2146623298b5f30c43b624c0be7355d8c526783e870a5 +oid sha256:998f578376c945e645fa034e1ffd3f70192bb857c307bb7bb76c38415f9449c1 size 1124 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 c750d6730d68e7dec17ba7ea898f853e707b1a9f..cc26aaa8e8d2fd1b5446f7bdc3e44607ad8556e2 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_211a21d76f2e9aaf1477+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_211a21d76f2e9aaf1477+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_211a21d76f2e9aaf1477+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_211a21d76f2e9aaf1477+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_211a21d76f2e9aaf1477+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_211a21d76f2e9aaf1477+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_211a21d76f2e9aaf1477+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..6616094d803cbaf159c68991933dd526aedd176b --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_211a21d76f2e9aaf1477+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11a0ad1bafcde952d02ad784e9416dc772da8b47bd2b06a9fee64722cb69530e +size 741972 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_211a21d76f2e9aaf1477+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_211a21d76f2e9aaf1477+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..bec4c9a11784325bc761545e60aef20865a04c5c --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_211a21d76f2e9aaf1477+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd91d6b9e4d6b3581034d387bbc20b952b294d2a15ae3bf7b9c6ad8158a14e58 +size 8909824 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_211a21d76f2e9aaf1477+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_211a21d76f2e9aaf1477+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..0a0d36f38ad0ebaddda1386b63d657cda44e3a1c --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_211a21d76f2e9aaf1477+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a11817c86f872f84da78b5ef734af96ea2831357c7fa18ca20dc4ad422afacd +size 9056840 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_21ca87b128dcdd0120be+d4317e28/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_21ca87b128dcdd0120be+d4317e28/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..75896f9b68fa67a429f9ddd0951768450cee8980 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_21ca87b128dcdd0120be+d4317e28/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_f425ac8a-fd61-467c-a8b5-fce12e80a1a8/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_21ca87b128dcdd0120be+d4317e28/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_21ca87b128dcdd0120be+d4317e28/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_21ca87b128dcdd0120be+d4317e28/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_21ca87b128dcdd0120be+d4317e28/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..f1e2473c54cffac949ecf31d3b6ec636c290545b --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_21ca87b128dcdd0120be+d4317e28/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_21ca87b128dcdd0120be+d4317e28/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_21ca87b128dcdd0120be+d4317e28/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..5aaf015f2632046619b393b7d6e8a91edb8e6b32 Binary files /dev/null and b/neuronxcc-2.21.33363.0+82129205/MODULE_21ca87b128dcdd0120be+d4317e28/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_2203200257461345827+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_2203200257461345827+e30acd3a/model.hlo_module.pb index 9ce7e3e0488985578abdde828c9b1ea9f1c0fb45..e8599a0cf960756b0e20680af605013607a7f061 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_2203200257461345827+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_2203200257461345827+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:45456b37b53b0b6a06b610216e14a515d98135ca5224f322df706ea17c09ec78 +oid sha256:d9d529e4ec6ef2fc62c3e90424817369d59b44565743d37db01ba973fb959d8e size 1124 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 059d1f56012551cd227f0a6b3f2db12880ab72d5..1a78610d164a39a0f94e67ff0cb7100baa054126 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_2410426589418994260+e30acd3a/model.hlo_module.pb index 608bec449a43a29ca3449653f863d5ac22dd9c2e..cb33bb459dcd04987d118a0b2bbe1727baef70f2 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_2410426589418994260+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_2410426589418994260+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:87a0fb4e60f0b8dab31549071774496d1084aa943d60bae75fd03ba975f51e40 +oid sha256:6c68d03612557a870a352972685d3d3eb790173c26910f7144fdf82d65c8a244 size 1269 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 c61876140be1db595e9269ee4dd48fc71e6741a3..4e8ff27dc6949833d96dde7fd783772dc9fa0e92 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_2426414314187505427+e30acd3a/model.hlo_module.pb index b2b53e161795e6f8fa9731ef8b474f52f20759b2..bdf1a76847d954bd7381212893a88100b84869bd 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_2426414314187505427+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_2426414314187505427+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:568a9462c809b77387087e485e9f6b26fcc1fdfb399610a6dc2a51706cce4476 +oid sha256:5f216a6a1622450851010e635d9b77b50d65735e050dd577ca75cf1347bb1a08 size 1562 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 944fc5b0ee99fb8494ff40bff11c2ae57577d5a5..95330b5892086f8cb95ec4f844aef412c28a682a 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_2576759111807165188+e30acd3a/model.hlo_module.pb index 65704d9768c03aac666774dda2cb66b947e691d8..a135660d0f25b9402d28b3d7bffed8a62e18d6fc 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_2576759111807165188+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_2576759111807165188+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:eea0159b626c550a13a8905089c534debb265105c7c809d918473fb6c6d39b28 +oid sha256:3ffa8cabf4d297b1042009232ccab674116ed1615fb5077f233e666ceb753348 size 1563 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 8775a02bee0b92d7bf2cf852a8c931190490ae45..f698e07ef6b226aaf08d4f9f3cf4810e2e86de14 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_2650447136112456251+e30acd3a/model.hlo_module.pb index 15143a1e694fdaf77d7d75ecef0f5d256d4c7a4f..1a597abbc0dc0a2eca544a4d0c1785ebc7018718 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_2650447136112456251+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_2650447136112456251+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:3d61fa5a96412b9a7c6faa3450178901ac4ce51bfd0c93a474da54ea3aeb18f0 +oid sha256:ca6d0e25778efb5cf6894ddb58f5566717f422fd85b88b7c20f6c0699e0f4375 size 1562 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 7a76658f39b15320e7fbf2bcfe9710f2873f4d24..3b14eeaa0123729cdad89c69075cfaacb0241433 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_2682953264353234433+e30acd3a/model.hlo_module.pb index 1bc201d0e1f32e0d24e1cd010c60ade707a56a0a..f34b1b1f6546bb8e2cc772d4eba03c5a6acc5b82 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_2682953264353234433+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_2682953264353234433+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ce6baeaf5dade8a8f217484d76835daf8e2d046fea864ab1af575806cac89fa3 +oid sha256:bbd8a577f7bfb0c83ba6e46dc08ad9db8f43e0676227087b7fd64ad9c08faab7 size 1562 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 c7bc4734d8281fac5d0b17faf5403c5c2184f966..2a38e4b34cff50499a19d083c88557b5c704c0d4 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_2699280712748688265+e30acd3a/model.hlo_module.pb index 98e535a37499e77fa0155dccc1e3b230813c7cf0..203974d7a7b13689d3ce343c8c29b29b8aeabee6 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_2699280712748688265+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_2699280712748688265+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:dcf05601e9eeb39a55dc3f524c47b8a6a1c12c7d0709369bbbfc58798e96410f +oid sha256:f2418661ffc63d5aa090bb6b361ac20c0aad1482ca7e469f46814f5fe159df7e size 1124 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 af7071a0d83cfb70378fdd4ab64194161c5a5536..842a944b1c9f9b6f6ca873d5a46e3b32635c532c 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_27b2c4720a6a1662f8b5+24129607/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_27b2c4720a6a1662f8b5+24129607/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..836724f44545ce0dedda1521fd4c623a6ea8ec72 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_27b2c4720a6a1662f8b5+24129607/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_27b2c4720a6a1662f8b5+24129607/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_27b2c4720a6a1662f8b5+24129607/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_27b2c4720a6a1662f8b5+24129607/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_27b2c4720a6a1662f8b5+24129607/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..90a9be2fe1a681994bf75de85c03215f147cd2b6 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_27b2c4720a6a1662f8b5+24129607/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1dddf91c6a86546b55890eda4dd0026330a7ea59d853f39f8ac2758b366729a6 +size 961456 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_27b2c4720a6a1662f8b5+24129607/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_27b2c4720a6a1662f8b5+24129607/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..6f664ff148e06931f35dea1e5aa23319736c9efe --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_27b2c4720a6a1662f8b5+24129607/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c9b303d5ffb9dbe60b8a13f4198948947628044e3564750dd748a1831dcfceb +size 14654464 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_2811495285170804454+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_2811495285170804454+e30acd3a/model.hlo_module.pb index f1b267c75dcb0cc9e08fffddaaac180e64bc50a2..42f88809f0b6fd5a6bc277c134049c0b79dfea19 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_2811495285170804454+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_2811495285170804454+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:fb403f8238eef0d57655a1d2f7c4852b4dc960caaeea77dcfaea4ff0fc2e4297 +oid sha256:58eda6d436cf71dba223ded27d3c57151d7becaceac10631eac6216d79766f9f size 2839 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 79e55d75b5413395d65e984c6f7e35895c5323ce..51aaab92608a64db7b39b1b1541718b76db329ef 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_2811495285170804454+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_2811495285170804454+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_302950992539679003+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_302950992539679003+e30acd3a/model.hlo_module.pb index 49d05a8cb91848c922cc659a2e42ab6df25c04d9..7181b41514260ba24d8db52b981864bb6b24be62 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_302950992539679003+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_302950992539679003+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:bbd9819fbbc2940d3d17a4acca136bcfe23177cdf48e41dee38682ff0f141560 +oid sha256:5751ad072a31cfc2a9a291e4f022d3f0fb7b4ea034c0745c6694cc0ea2538940 size 1269 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 2d962c9ac1967b53965c850071f9fe9005faac90..9c58b0c42e48239d1756260c4d63733e61014b96 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_3b30a2906ed2f2261284+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_3b30a2906ed2f2261284+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_3b30a2906ed2f2261284+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_3b30a2906ed2f2261284+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_3b30a2906ed2f2261284+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_3b30a2906ed2f2261284+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_3b30a2906ed2f2261284+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..5e953f00279d7914979edbc1e4e712cc67f7cbef --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_3b30a2906ed2f2261284+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7a2fcc74eae6f630646a73ad794ef48d8241b552b60b4a064b7cf4af67acfdf +size 443823 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_3b30a2906ed2f2261284+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_3b30a2906ed2f2261284+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..0f4b7af66bc30a958f2c3ba11ecaa186776af11a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_3b30a2906ed2f2261284+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85de0a89cc5f5de0bea00f7e465f0214c79d3636318193d72c99c11ffdef5fb1 +size 3134464 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_3b30a2906ed2f2261284+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_3b30a2906ed2f2261284+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..6b522cef16fb9660b98eb239a9715194c38e915d --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_3b30a2906ed2f2261284+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12c9a536b6ef9531988320a4f3510fb42d2cb1acf8ff8306ddede4677ccd33bc +size 3208565 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_4013984518400530928+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_4013984518400530928+e30acd3a/model.hlo_module.pb index ead377b1c20b000831d42100c43368c555f6cb96..a9879b6ca1a4d731a59cb770edd0f49056d70f7f 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_4013984518400530928+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_4013984518400530928+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:efff66a8c2f78f8a8b6017b4f1658730b0acf6b16231a6cc09d88c4d1787e80a +oid sha256:35ef60e1167a2431df546bb2488a90c23886592b1651cf073918fa01aabcac5e size 1562 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 889ff47931ea3eb54475e077bd14ac4801f56d85..dd75a92177ee20e0f9a39eacf7aca68f03dd2450 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_4052264821257342969+e30acd3a/model.hlo_module.pb index 228e7397cb7429b492ca0ecf3e17d3000f437681..f82f279129dfea9f69c4aa491ea1ce1318cedd1b 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_4052264821257342969+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_4052264821257342969+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5706f45c70599fd5f452d6cbb6551dc245a40ac81c7fb2e33ccfa9c51ed7575f +oid sha256:1055ac7426465f79318ce28b635258e34def38605bf2596401a253cbd5594a9b size 1562 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 cd7d33786c3f3d8371a6df600cf344b876b1ae2f..1dff34b733d0a9a794c6ee78a99671c45d70a775 100644 Binary files a/neuronxcc-2.21.33363.0+82129205/MODULE_4052264821257342969+e30acd3a/model.neff and b/neuronxcc-2.21.33363.0+82129205/MODULE_4052264821257342969+e30acd3a/model.neff differ diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_4166695227062860792+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_4166695227062860792+e30acd3a/model.hlo_module.pb index 7018c545b777bbe09056ae44c14acd58dece6fc2..3891e1dacd3cdaedb9086e334983228b4ab1e4a6 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_4166695227062860792+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_4166695227062860792+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9e93e3084a1b7e8f11120529d2e0d2de375c80e38c9e7dcd4bf8fc05a8da51fe +oid sha256:fe9e9881afb03f2b11ecf5db18b868f5a53fa26bea6ee05df46d423417be7352 size 1562 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 6dafd667c879807e2cf6aeb5066c174f99d9393e..010b6f0c7897786fac04ee29cbd92f38cec9eeae 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_4232557591140652245+e30acd3a/model.hlo_module.pb index ddae64ca3b1ad4b0838a3af191a64be1b874bc7f..e8ed0985d5b0a62f118015b8447087d13bcd80b2 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_4232557591140652245+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_4232557591140652245+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d702e49d6eabd97584c918a76c0872c349a8330fff8222dab5980cc566b47d02 +oid sha256:4d80d9906664bd7fe12754d0eeb711207b1cad5e533a61231a03399afb42b3ea size 1269 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 be060b96a512165362fbadc53d5b35844681565b..da56cd1a3485b91b6ea9727c3c636b646c050e18 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_4606491920565381523+e30acd3a/model.hlo_module.pb index ae07adbb5bd9f95359a6b4b4321e79da69789526..73e2c21790174288c3f4a75542bda8ca0cdda4bd 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_4606491920565381523+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_4606491920565381523+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:46fe8fdb1fb41ba01b5dace63a4834e8dd897529f0ca17b449734819dc0ba2dc +oid sha256:e1e1c235d325dae24efd11e30622a63a03c6f5930d2a0775ffe4ff8452c1cd2e size 1562 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 e6f1f78d55f28b7fb3aa5c11346b15901927d11f..9c2245ec0f234b89cd4a6ff46eef4df57fa4b78d 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_4608333438381222049+e30acd3a/model.hlo_module.pb index 6ecaa624db0502603c5d5de8c76a80ad78facf46..81b8e3f0cebf479492eda8b344d95989638d6957 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_4608333438381222049+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_4608333438381222049+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:31fcf22de5701ef0965f731595d85f24a2a33c39999a280c4337a0f8d002c27c +oid sha256:1f2c20d5461ea36987b59cbdab555d34b02a6bad9fa104a0009788b6d2a3a050 size 1269 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 fe9ac2e6a56a2c7729bc626fac24f95aec30c4ea..d4e120a60f8cb119304873ac3e7dd0625395a590 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_4683016291147007110+e30acd3a/model.hlo_module.pb index 25fcd81ae3aa1b49c12b5eb5964611bd71bcfca9..888b0db9d617b6446515801bc301f6d16cebc756 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_4683016291147007110+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_4683016291147007110+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:832dbdefae1c623549fdc85f4cf110f5672eb9193a64711a4c432dd21cca1e2c +oid sha256:f5e74e436b038570a84bc734899893b77c4f464426555acb3e9217ce4ade7de4 size 1124 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 5b2d5620a4e6abc960d1e3123dbf430e1459010f..f146e382ca53c9bc5a4f5ea77bef3616f4b51b42 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_47c06cec950237374a0e+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_47c06cec950237374a0e+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_47c06cec950237374a0e+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_47c06cec950237374a0e+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_47c06cec950237374a0e+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_47c06cec950237374a0e+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_47c06cec950237374a0e+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..6d675be7cf7972edf83eaff87f47240aa17d3353 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_47c06cec950237374a0e+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d8edecf22b4185a0fd12d029e8a0a761ac76ef60c9d6054b180594a1128a00f +size 82629 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_47c06cec950237374a0e+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_47c06cec950237374a0e+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..132e3ad43249b32c2e73121da32f745168f5f481 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_47c06cec950237374a0e+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10cf0d8e6b6f0b09c9759911af6482a82df2f590f6394c3ebf7e274704c63274 +size 287744 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_47c06cec950237374a0e+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_47c06cec950237374a0e+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..63ded5b190bce02a50395edf7242aada72e1eca4 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_47c06cec950237374a0e+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83869159f0d69e24395c4ca4c8ead35c5600e4a26de86d06e5424156fbaa0353 +size 295957 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_4816422433712790690+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_4816422433712790690+e30acd3a/model.hlo_module.pb index 533dbac3facd87ce2a77338b9c4f1dab5e6b6c28..5e3848c767d2924f4882d14ed6e2037f5f37f14b 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_4816422433712790690+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_4816422433712790690+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9ff06073fbb3bf24c4bc85add37e0048b4b7804d30f6352c8b597823822821f2 +oid sha256:6d45d472b00f172e67b7b6218ba4be61acad18677059ac9036f3c2f6afaa5d81 size 1562 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 23fc07456c16b151494f1baddd3b102db899de92..972bd8d9a2e4e02930f80437ace6f40305bab803 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_4842952601010372434+e30acd3a/model.hlo_module.pb index f7fa6db0106c5c747f1a128b9ddb0d13c602607b..ab521c5a71a8d95edb64b8e8d7d4e83932c902f7 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_4842952601010372434+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_4842952601010372434+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:dbe5c13f92458747422f9d0c68f9d956dcd6910209befb270000db6bd1c0cf23 +oid sha256:861376f75f7f5671344381b7a343bf16e7ae0bcf5ac396b966f8d06342230bb8 size 1562 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 b959bdaa4737a6f82626fd9db436671a9fc03d77..7827746010fa44d8ae370245e53b7fd0c890c7f9 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_4939721357859779936+e30acd3a/model.hlo_module.pb index 58f2c38ca89f9aa2f340552fdd9d5de7ca4b7cee..a9526676936bd9b687481b3b2df5d3a048abe38b 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_4939721357859779936+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_4939721357859779936+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:3e0b6f210a6acdf12812b8b6fe0a9542363d340090387d393ccf52c918cb1ab2 +oid sha256:e99504064d6a79172db169b6fce4a0ae7f7041d24934555f6d9d66e8be6f94f9 size 1124 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 b1088e4f8dd5c035952501497798be4f69f4ec11..4e40a1d71eab1e6a9b6b19689f9b6989f415ec1a 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_4da1b782d99a7304f0e2+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_4da1b782d99a7304f0e2+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_4da1b782d99a7304f0e2+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_4da1b782d99a7304f0e2+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_4da1b782d99a7304f0e2+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_4da1b782d99a7304f0e2+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_4da1b782d99a7304f0e2+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..8d0e1832f36371c947458374734b53549fa617bc --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_4da1b782d99a7304f0e2+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b4fd29746283c3bf9e8a296c0a0c1a4aba33eef54f20dd7eed545f4dff63e335 +size 680975 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_4da1b782d99a7304f0e2+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_4da1b782d99a7304f0e2+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..5f39bf0019b5d6c11fe25d2bcc44309f8d573485 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_4da1b782d99a7304f0e2+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5eb4395f507db8e19479ec3f9af8984ef0ed77b7d702202468b3634faf02f67a +size 543744 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_4da1b782d99a7304f0e2+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_4da1b782d99a7304f0e2+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..2309d403dbc2b9397870779f4e6c659d86826696 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_4da1b782d99a7304f0e2+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a764fed6315873da19c5fe7a67ef0ba9527e6597d0dc04c2829e2f32f8a0594c +size 563245 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_5397223456911199516+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_5397223456911199516+e30acd3a/model.hlo_module.pb index 3d2a403c5d8bfb71a4eb5d4fa9ad83dea2bdc001..c1000de8e5a9c60bbd7022715e745c9146d11431 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_5397223456911199516+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_5397223456911199516+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:14e48c5427bae1afb972e0db0cd400c86b50dde6366cfbeeb4a9e2ff34d4c69f +oid sha256:ee423330b8cf67333a7590859ceb649498e636f1bc7893b2ee1343e91906316a size 1562 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 ae5ebdcec64ac037537823cc6689c00cfd2ff6f5..00f1f0639fbebe8a7150ff0b97c34e7e5addcd9f 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_5724544066416780383+e30acd3a/model.hlo_module.pb index f6063a063fd797828fb19096905b5b325f91c9e6..122c935c4e41adb4425e854ba75203afb5af4750 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_5724544066416780383+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_5724544066416780383+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c4d53710ce19d1c4554fc4d7c52ac3be95c5032f3a853b056f221b868a598272 -size 1562 +oid sha256:9e9fd1c704bfb5ac3bc4c3bfa7ab7af8c55f973d2c5171a166faa96581674ff3 +size 1563 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 256504c4a16627419f30ac189ab927cf09666417..a736068c4a6254e4503e553e9e5961ed8988eb23 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_5765916484866650909+e30acd3a/model.hlo_module.pb index 859cbe32eb34843c3da5a231bb1696a767accc52..4a36c07c03050e7f25c396f72f6fba50e7a8325e 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_5765916484866650909+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_5765916484866650909+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e46ea7735f4aadde8f2d382dcb96da002de17fcd16fea8ea7f4535f5d3b3b9b9 +oid sha256:6b1f448dc7baa25a25d40e72219230256e3f6cbed4e3bb2e452401b8d9203532 size 1269 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 b421ed90f76bf316af1b9af21405ff347437f721..5a5fb06724609ce47bd336b719ff41812c8a19c7 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_5ad2d5f306d13c506440+24129607/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_5ad2d5f306d13c506440+24129607/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..836724f44545ce0dedda1521fd4c623a6ea8ec72 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_5ad2d5f306d13c506440+24129607/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_5ad2d5f306d13c506440+24129607/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_5ad2d5f306d13c506440+24129607/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_5ad2d5f306d13c506440+24129607/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_5ad2d5f306d13c506440+24129607/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..8d6fa80089c130a6372bef6c30adb443327ef40d --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_5ad2d5f306d13c506440+24129607/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:546055e83ecab4869934d006e22de9b0749a7ba92647736a2fd3717385302934 +size 81516 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_5ad2d5f306d13c506440+24129607/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_5ad2d5f306d13c506440+24129607/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..1743386812db96c1985d42ca2a640fc29f489ec2 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_5ad2d5f306d13c506440+24129607/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f30a96faa2c3d7d53564b4faa524529849c0bafebb2a7cd2a2d3565199b4afab +size 267264 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_60f37ef2cb181ebeb434+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_60f37ef2cb181ebeb434+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_60f37ef2cb181ebeb434+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_60f37ef2cb181ebeb434+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_60f37ef2cb181ebeb434+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_60f37ef2cb181ebeb434+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_60f37ef2cb181ebeb434+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..008a42b511d0bd31b09eabbf90595cc60b356662 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_60f37ef2cb181ebeb434+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f29d1c45ce60b946ac91556bafe5d8ef1f50ef99b89954c5529d18122e6c8b05 +size 581355 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_60f37ef2cb181ebeb434+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_60f37ef2cb181ebeb434+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..44ccec38914b4b2e4f03e4c81dc3c89b056edefd --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_60f37ef2cb181ebeb434+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1fe4ebd73ffc01c6a9f0e1b592025b6316c1795432ee900302da206df867cc79 +size 1772544 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_60f37ef2cb181ebeb434+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_60f37ef2cb181ebeb434+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..13973c534ac0df98a8fedb4f0d9457bc40ad1970 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_60f37ef2cb181ebeb434+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ebfa95057eb5713aa05862ae992aaa77ce4045ea5f806dcc3726efca85a605a +size 1894933 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_6368409506294683105+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_6368409506294683105+e30acd3a/model.hlo_module.pb index 8b0e12e2580cc2fc4aa66eeee756599d121ae261..74705b8318a445e0648f9fde3508fab8b16caaa8 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_6368409506294683105+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_6368409506294683105+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e89fa3fc9d44fa90d6a6771776e18597d1fb8561f9238c51c9dba3b2a483a73f +oid sha256:d4a35d9a793aed572792342e4c392fbbe2c251c7ae404d8d6c73b5108461e966 size 1269 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 e4c2de99367d40927501238f5c734b05669e5e02..533b7f56cc8e6ed243ce641a37dc65175cd9db32 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_6375780938374445148+e30acd3a/model.hlo_module.pb index 98578d52b81febec85234521f26ee32203bf2a01..e769e7df6df4474e6214c4572870b42bfe6a848b 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_6375780938374445148+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_6375780938374445148+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:398244ad15ff0322d06d35f20b198a98f1d59c2fdce69b15b70c802cb587c418 +oid sha256:10fa8b54610c56299358af63f88cd00ae19bc45aac12a419e5ac7b9a05f2462e size 1562 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 d6a72bba1fcd1c111557ac5296748ec99f8f74a2..f02958dfc476dc60c2ec88152d7ed0ddb94b2590 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_6417297157123190451+e30acd3a/model.hlo_module.pb index 96c1648c163b507ee68c5774a5eea4b5b4194d59..e4027cb6403e9c70ab57bb06cfda86943cd3e650 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_6417297157123190451+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_6417297157123190451+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:666a24fc2cd759709aab566d5768ef5cb3fbdbf3a86ec63f31e255d04fc3d07d +oid sha256:6bd84a0d4c66e18f634658f2bee166ac0b55e45f6eab1c25527a1908c1aeea68 size 1269 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 f91f85fce31559653db0e9c293c4e9f20ac42982..dafc640f0df56ff3328ebeb7f7a3c7e9d42eb691 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_665145310648895012+e30acd3a/model.hlo_module.pb index d448d12f7eb638799b8ff9a5a083058cb0d6ef3e..f405a9a3bdb7031cf076e09427863ee33fcac9b4 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_665145310648895012+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_665145310648895012+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ed684c5d08ebfc375d5043c02039d7175de51c5f523157a5f9df6e2446b13344 +oid sha256:b6809c25dac91c3f32f0a8e8c347f1de44e8a10937cdbb15aae5e2f322fd4c1c size 1124 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 c90715bb823d824ca962467996ebfbf41be4bae0..f647c10adebf43cb05abb3c12f8641d932e09127 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_6711765444274762085+e30acd3a/model.hlo_module.pb index 2259e64f6d87ea65d39153666872c71ae4909811..5c95528727e970cbdf6d54305992e3ec5683ac3e 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_6711765444274762085+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_6711765444274762085+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0f626908ffced16ec5986bd5c3b22ddf75aacc96c5bea3115fb7593bb3b7d35b +oid sha256:2ec9e882d8b50c69128f4a6ac05177fa0ec6f24797487635015012581dff35b2 size 1562 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 0b9b626509a3173430c6ffaf4a0bdcda026c46a4..c73dd070f846d92ec3370193042cb38236967013 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_6ed16792f2699fac18e0+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_6ed16792f2699fac18e0+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_6ed16792f2699fac18e0+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_6ed16792f2699fac18e0+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_6ed16792f2699fac18e0+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_6ed16792f2699fac18e0+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_6ed16792f2699fac18e0+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..8441d02f09fde5e8d54ee7cd9e8a2f94997042a3 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_6ed16792f2699fac18e0+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ddfc4058c88c14f1f3999a3f3aa5c3fb167e3eff80a639290325506909c6b2f +size 95557 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_6ed16792f2699fac18e0+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_6ed16792f2699fac18e0+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..b850e04ea7ac2f8c86c905152a36d73edb4ad162 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_6ed16792f2699fac18e0+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:386711bc7e3781a301911fef0d205883713fcc35449c319fae73e7fc0dee9ce4 +size 277504 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_6ed16792f2699fac18e0+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_6ed16792f2699fac18e0+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..7716bcd00894211b0f8aa6aebd1ec5f982f43be9 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_6ed16792f2699fac18e0+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d018e14fafeaea830330e04210880e735034466d75b6f8afb853aaa10e8828d +size 288898 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_7049343467335223052+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_7049343467335223052+e30acd3a/model.hlo_module.pb index c7a344e0407bfd870825ed62394cf1506135d0ce..3dadac4c17fce0493ef4a16208fc9cf98e765f69 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_7049343467335223052+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_7049343467335223052+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2827fec44c2160f1c31d2ec711b12e971c0439fb4305cfba357249c12959a67c +oid sha256:fd0debe7e160da4c0a17cf7d9ff41210aa464503608de3313b5594af94694bc9 size 1563 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 a29ff43eadf9716ce39ea68d6728c0f8096e8e05..1e800743557603b210a70c173ed82fb63ebf3fb5 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_72c4957df590729a5c8d+ac10809c/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_72c4957df590729a5c8d+ac10809c/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..7d1678bf25666438bbebf65702a2233ae75c43c9 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_72c4957df590729a5c8d+ac10809c/compile_flags.json @@ -0,0 +1 @@ +["--target=trn2", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=2", "--logfile=/tmp/nxd_model/token_generation/_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_72c4957df590729a5c8d+ac10809c/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_72c4957df590729a5c8d+ac10809c/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_72c4957df590729a5c8d+ac10809c/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_72c4957df590729a5c8d+ac10809c/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..0a4a3f7ef189330c8d8479fa8685d93ed15e26ac --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_72c4957df590729a5c8d+ac10809c/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3f6d632a3bb5166336f1b7739b8d06b318be62f8969d580293dedad698b235d +size 1602245 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_72c4957df590729a5c8d+ac10809c/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_72c4957df590729a5c8d+ac10809c/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..a9ba2e7f1d87a12e244e7707fea5fc0cc27275c0 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_72c4957df590729a5c8d+ac10809c/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ec368c7abc981c8d623476626479619d24a7310d190a56e52b416db9dc758af +size 769024 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_72c4957df590729a5c8d+ac10809c/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_72c4957df590729a5c8d+ac10809c/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..081fbe43020ec0416b02f33511b9948417ea206b --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_72c4957df590729a5c8d+ac10809c/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f895b8d167b4e14c7d919acef147eab1831c6e385df44cfa5575c78b9bac8f1b +size 776820 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_7877866862268499659+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_7877866862268499659+e30acd3a/model.hlo_module.pb index 6b91b1bff11aaac8a2018fab2cb2f351f8bbede3..77faaf00a0a76021c729e053d24dfe9e62da3b31 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_7877866862268499659+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_7877866862268499659+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e8488b9ff78cf28711fae06c1840e772100935c15b3c7ea13dc83dd72977d738 +oid sha256:bd11e94da1ca033b60037108346eb911fa84bd731e210d980977d346f57419a5 size 1562 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 e28ba3d8d0f6f49c3ce88d65008043d16d7536f9..c060f8f2f8382c111c6e1a9b98395d59116c23e6 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_78ef14a5e66efce40c70+2dde74c7/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_78ef14a5e66efce40c70+2dde74c7/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..a7a40c022faf4beb0fb20246d9c1f9c8f32c6900 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_78ef14a5e66efce40c70+2dde74c7/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--enable-mixed-precision-accumulation", "--model-type", "transformer", "-O1", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2", "--auto-cast=none", "--internal-enable-dge-levels", "vector_dynamic_offsets", "--internal-hlo2tensorizer-options=--verify-hlo=true", "--logfile=/tmp/nxd_model/token_generation/_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_78ef14a5e66efce40c70+2dde74c7/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_78ef14a5e66efce40c70+2dde74c7/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_78ef14a5e66efce40c70+2dde74c7/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_78ef14a5e66efce40c70+2dde74c7/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..8c89a0fe3bb68fde36fb6e7b7289304585e2b0a1 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_78ef14a5e66efce40c70+2dde74c7/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b99ecac1d2fba0ada5643f1cdf386f9d8501ebe4ed6c14f367a826c19efce49 +size 92677 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_78ef14a5e66efce40c70+2dde74c7/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_78ef14a5e66efce40c70+2dde74c7/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..e0cf2c4a68422a7a924e641b9a049a1159d86fab --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_78ef14a5e66efce40c70+2dde74c7/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c4be434d2ef7d79586a1f957fdaa1048bfdc869de59c03b5b2a87826f7806ea +size 369664 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_78ef14a5e66efce40c70+2dde74c7/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_78ef14a5e66efce40c70+2dde74c7/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..8aeca084c40ce9870a875d47ab518cddfda21657 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_78ef14a5e66efce40c70+2dde74c7/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a47c16359d170b75e756c9f1f95b3331b4c6b52d8b6d628f039725d14acbe3ba +size 379225 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_793722132614349680+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_793722132614349680+e30acd3a/model.hlo_module.pb index f4f452a0f7a5599f79d8580e2003d1c0c65964a0..fec8157124654f258800ec77a0875ec8633778d2 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_793722132614349680+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_793722132614349680+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1ce53467f03398096635507713bb04d5f6307e7f4d5193d0caf3e04c2d36646c +oid sha256:2082adaf3aa8f8276ec6dc85132d42c996c4bcb2433ef9da2d03fbe8b01e89e0 size 1562 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 0f8c130ca3b5ee8f3366ba6179cb35299f449795..6d4067940c4b5e1a556955bcf49870a69c4d5eff 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_8051a8dfc97b2193171a+f68f772a/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_8051a8dfc97b2193171a+f68f772a/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..acb13661fc3dcdced172b53027a0b54dac8c5e94 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_8051a8dfc97b2193171a+f68f772a/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_03910494-dc4b-40ae-b072-b1d36dcb9ce3/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_8051a8dfc97b2193171a+f68f772a/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_8051a8dfc97b2193171a+f68f772a/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_8051a8dfc97b2193171a+f68f772a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_8051a8dfc97b2193171a+f68f772a/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..6a2752c25c9f7e1092e2b55c010694e168610ce3 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_8051a8dfc97b2193171a+f68f772a/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_8051a8dfc97b2193171a+f68f772a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_8051a8dfc97b2193171a+f68f772a/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..cf153cf93efeaab87664330d0447cc6972e274e2 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_8051a8dfc97b2193171a+f68f772a/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:254b1d40b84d8ae8519f26abe8b482ac741a24a5d5505c908bf4fc90a16fb3f6 +size 103424 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_8051a8dfc97b2193171a+f68f772a/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_8051a8dfc97b2193171a+f68f772a/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..4e13eb1a31cda513dc7b411353c1df7786ce26e9 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_8051a8dfc97b2193171a+f68f772a/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c7a59b16cce087c81dbb8733c43645db65759f0e67492df226e48932b4bead8 +size 104320 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_8142158350842512240+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_8142158350842512240+e30acd3a/model.hlo_module.pb index b503bf05f37ca24492c099c4ed600f9a35b080d5..958512e101420326bed6804ba7a8ee5286121040 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_8142158350842512240+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_8142158350842512240+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:77ee8e9cc8fadbed653dc442c1fe0ffbabf12e69fc72302103989ac86480243b +oid sha256:b2f2e1b1a25410b4acbec66a482f2cfbb378e6493f8b0d20cec1e8b01178ef1f size 1562 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 95b2d284a76dcdc59b903d9d12a330e993475b96..a2d13aa53df8f7c48d07efa8e8c03d92ecc67d7b 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_8146958338478347620+e30acd3a/model.hlo_module.pb index a183f592fb2998384e9cf59b6455cdc90b0f51ec..498516752dbf2ff4c10ed273bd2d272d0f10dc37 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_8146958338478347620+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_8146958338478347620+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f9dd8cffcffb0a341d84ff1d1cb2dca0875149e5263a7c68f0a5935fea891493 +oid sha256:98cb3615d85bd62c84407901de29163a77650abb4f11b1ded37a2894613d1f6f size 1562 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 df3b471fb0d0249aef0cfbaf534ac8ba7aeb1e44..816317bc74b938efe735e78a9ee98b7dad4219eb 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_8245165830758578911+e30acd3a/model.hlo_module.pb index a13f38f8b376e4eb8b072ac6e4625a08c5e99349..c307e99ce17396ff22e762c2ec047b769082a268 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_8245165830758578911+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_8245165830758578911+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9b675bd71d57acc8f43816e11cca613820bc22be1299c49f9426655bfff683e0 +oid sha256:7d4b3b1a52bc98ba0eb4f88712dbda67ebde0aec6d5defc2c6210f4a45f81515 size 1124 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 fa527a501809e483bdb3cd2eb00188481e0ea792..ed89f6db55b757d3b7beaa7fb43e0627c0ec42e8 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_826bbed76f9f977f94c0+24129607/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_826bbed76f9f977f94c0+24129607/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..836724f44545ce0dedda1521fd4c623a6ea8ec72 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_826bbed76f9f977f94c0+24129607/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_826bbed76f9f977f94c0+24129607/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_826bbed76f9f977f94c0+24129607/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_826bbed76f9f977f94c0+24129607/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_826bbed76f9f977f94c0+24129607/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..82b08d5da7d400509d623760df8cfc407b2ca577 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_826bbed76f9f977f94c0+24129607/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ebdbd2f771cf3c318754304a5f5f110207bf5f7735e07a5254607e3abca36f63 +size 79552 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_826bbed76f9f977f94c0+24129607/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_826bbed76f9f977f94c0+24129607/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..b8163ceb503f28d31b508e4263924118ba117f72 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_826bbed76f9f977f94c0+24129607/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe59a5f39ca2b14c71aabb12a156bf84eb6bf1b914abd12026ba8db1de52315f +size 338944 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_84656e81eaefaaeb593f+24129607/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_84656e81eaefaaeb593f+24129607/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..836724f44545ce0dedda1521fd4c623a6ea8ec72 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_84656e81eaefaaeb593f+24129607/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_84656e81eaefaaeb593f+24129607/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_84656e81eaefaaeb593f+24129607/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_84656e81eaefaaeb593f+24129607/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_84656e81eaefaaeb593f+24129607/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..6e81e985cdadd827f927a352ef6c7c27a9e24f6b --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_84656e81eaefaaeb593f+24129607/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:226520757a94ced9214c615bc2496045526dab8114f2ab79beb78c7b064930a9 +size 694128 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_84656e81eaefaaeb593f+24129607/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_84656e81eaefaaeb593f+24129607/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..319c4bf390c3a1ac817681965edbb43a5b6ec79a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_84656e81eaefaaeb593f+24129607/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eeb89ba7755fc74f16da3c26c1219d52f5266f08f97e876517318fe8314a213c +size 625664 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_8679096817551502409+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_8679096817551502409+e30acd3a/model.hlo_module.pb index ac85e995e77342b635918c9cdf254f1e3c681d23..c1c96a2914fc611c392fe743274c4ee2cf6576ef 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_8679096817551502409+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_8679096817551502409+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7abd3f10f212616e20c47a6f1bcb6621cff1ec887b005c23ac7ccfbacde88589 +oid sha256:e164f767add2c2e56774b342a7be20c6f077df6725c3c40b8850f43abc1a6839 size 1562 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 bd8c3bb88ee5b5aab163dbc04a5c3610db467cc2..f4848a5086bb3dea176f8e62d01a0e7bfe9fab87 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_87be52aa7374785dde52+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_87be52aa7374785dde52+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_87be52aa7374785dde52+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_87be52aa7374785dde52+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_87be52aa7374785dde52+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_87be52aa7374785dde52+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_87be52aa7374785dde52+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..2e7979062bc24ca089ad8ec31590eb1ff2df2829 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_87be52aa7374785dde52+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:86d470dbc20e1941daaabcd28fd83da8923f9db6c7856319a719121c3b650e3b +size 616389 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_87be52aa7374785dde52+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_87be52aa7374785dde52+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..024da56e4068f7ce5549c1257ad2a91c252ee4a6 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_87be52aa7374785dde52+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a84281cc2b00a1279b9892f08842d1b8f43842b4b52bf0a43c3cc87f231fd47e +size 1711104 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_87be52aa7374785dde52+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_87be52aa7374785dde52+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..90905b0377d74b9fd0bf26cb8ca8c67d7d6b65c1 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_87be52aa7374785dde52+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c9427afe09c7e67f12b988814bc3b937df36a2e75411a2672361f07e0d19c6a5 +size 1867295 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_8882592820567938515+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_8882592820567938515+e30acd3a/model.hlo_module.pb index d784ce72e29b5099335c478ff0919bb33371aa83..5ff3deec582d676db69466412f2351f5ba741ee4 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_8882592820567938515+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_8882592820567938515+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a6f91e1a75519c46114a666484125f9f268121b7be45de24d074290695d05b20 +oid sha256:900b8bbfaadddbc319860881ba88f3502c42c068fc997d745ef3ce6e8a7d2614 size 1269 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 8ea6713900adb01d1d3265719a1fc8a64d11538e..6334dc7d75c8e89d0c2430ad44ff013b752250ff 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_8f7c5fdf533e837a78bc+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_8f7c5fdf533e837a78bc+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_8f7c5fdf533e837a78bc+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_8f7c5fdf533e837a78bc+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_8f7c5fdf533e837a78bc+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_8f7c5fdf533e837a78bc+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_8f7c5fdf533e837a78bc+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..7164162bccb52523502408bb69e04c0c5fd74742 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_8f7c5fdf533e837a78bc+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b41b6d511cdabc37cece8376c22ab5b2fc18d412c69364097970b82fe0aae427 +size 741676 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_8f7c5fdf533e837a78bc+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_8f7c5fdf533e837a78bc+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..c343268b2ca59bbc93dbaf762919f6145877cc98 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_8f7c5fdf533e837a78bc+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f34f42180a2803fcf54d028d7be8f06bffe2717e834c2227a73818cd8c4069a +size 3861504 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_8f7c5fdf533e837a78bc+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_8f7c5fdf533e837a78bc+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..2e7bdb0ccb14e632fa7348ee08583cf985fb47ad --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_8f7c5fdf533e837a78bc+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70c7bd02c7105e0294dce5ff400527dabf49327a312ef20b3f13d6baab928849 +size 3998561 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_8fb7e0e33d5fb07b5f06+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_8fb7e0e33d5fb07b5f06+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_8fb7e0e33d5fb07b5f06+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_8fb7e0e33d5fb07b5f06+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_8fb7e0e33d5fb07b5f06+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_8fb7e0e33d5fb07b5f06+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_8fb7e0e33d5fb07b5f06+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..1b3f65920585a67c8f847fbe9826c3ce4036b2d7 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_8fb7e0e33d5fb07b5f06+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1918ca2b4c869d1b73ba6c3a28affbf724dd37e2ebcd52e868010f6d09bfddd0 +size 622533 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_8fb7e0e33d5fb07b5f06+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_8fb7e0e33d5fb07b5f06+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..bd0e5c3a60a4252e02e3b8e5d76a538e74d06c0b --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_8fb7e0e33d5fb07b5f06+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb2da831a56f8bf86a7f998101e94dd364cfe128469a7e4bcddce08d372fd264 +size 2192384 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_8fb7e0e33d5fb07b5f06+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_8fb7e0e33d5fb07b5f06+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..5432c76afa489cfaf8cbfaaa5edd9f0071c8915d --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_8fb7e0e33d5fb07b5f06+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d0fa5908858b198712472c30a92c21d3a41a4b5d9ef49e57613f1cb749f3e6d +size 2348578 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_9168843780052024308+e30acd3a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_9168843780052024308+e30acd3a/model.hlo_module.pb index 7b5711a620a1ad72517d3a49034a2498e7562b5a..06d0cf2950223ccb53b4ab60582c72e2f0967d9a 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_9168843780052024308+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_9168843780052024308+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d64b3b25c7de683aed8adf0312bc92adbfa521b917df86023ce20a7bf2976aa0 +oid sha256:2c4f1f754db274298e66afda889ec497217e80e8e54e832118c2acd3bf16f899 size 1562 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 6e958451c8b323ad8c50c1bf22ac29e8158aaa93..6e1098b2c0f087113be20f3ba6c8b19c2bbd3247 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_9527558113976496538+e30acd3a/model.hlo_module.pb index 8ae4330fda6101e86cdee0bdf738f7268b0268fe..e256a0583bbdb6ff237bbc686be787aa13fd1528 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_9527558113976496538+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_9527558113976496538+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:95f4dfeb50732360e1caef0cb517568d513cc2dbf1560aa21ccc645b2f915493 +oid sha256:6945702dcfdf82f86f9e312b0f1c015bf24f2f23011de3977e7ea5070d9ecde9 size 1269 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 a41129081a0562e3a821abeb3b91b3185d82875f..62745f5a288ebf3cda18b545debe1f8a99d5b7a2 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_9533302140457976101+e30acd3a/model.hlo_module.pb index ae76c64a7626d16b0029bd150ef450fd0831607a..85e31072761d0a395de06bf55fbfb08e8dc9a4ae 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_9533302140457976101+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_9533302140457976101+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ae7d3adba443e1506864fd8954c89e898991e5e755ead7d2c8fd9d53a082da74 +oid sha256:318f6e2ba764c7f7f194b85c1cad48eaa2961f02affeecea5ef0a43b5e88fe20 size 1563 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 f6e84dcc6098fbe31c3b86fa4161e84068112257..914affbbd359fee54029336064c037f6be352f6e 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.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_9770544877851564228+e30acd3a/model.hlo_module.pb index 3b67b0517df5c01b8e466026c5f859b3f21efac6..c75532b950656d686ad0f996577c79273318cace 100644 --- a/neuronxcc-2.21.33363.0+82129205/MODULE_9770544877851564228+e30acd3a/model.hlo_module.pb +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_9770544877851564228+e30acd3a/model.hlo_module.pb @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:edb3950f99ea6bc3ae5a4a2afb866f714971a583eaa2b3df46d7d7b2179ba75d +oid sha256:36a800003b93aac06331a8f082eceb6b5701281197a053ccf7beea9d93cc45fd size 1562 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 c336ecde1ed1f9bffc64336349d675b8e69bdb32..d5ba2258273b863734690bfdf2eff7a99b72b312 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_9a1b9a8b43d1665adae5+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_9a1b9a8b43d1665adae5+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_9a1b9a8b43d1665adae5+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_9a1b9a8b43d1665adae5+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_9a1b9a8b43d1665adae5+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_9a1b9a8b43d1665adae5+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_9a1b9a8b43d1665adae5+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..a7c33dc82b0d1d5fa61c5a5c012a84bb5a8b8bfc --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_9a1b9a8b43d1665adae5+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc407d20d296d772d6a91a2b7c75c17a97766e94444ef6950771e65d8db11fe6 +size 910911 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_9a1b9a8b43d1665adae5+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_9a1b9a8b43d1665adae5+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..32a31ef07e2ddd674cefc8d8e32cf70a8a656d23 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_9a1b9a8b43d1665adae5+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ebd63e98cbc834e70e68e7e8713f06e25476da1f703eec434396deb2f1eab8a +size 6964224 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_9a1b9a8b43d1665adae5+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_9a1b9a8b43d1665adae5+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..cb5df4423c9968efe33e670d6e9737a99eafb815 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_9a1b9a8b43d1665adae5+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d11db6a94447d4517f5a18ac9b0de64e94aabbb9792a2f5f16f6e92ff6965911 +size 7130748 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_a1cd38756c74f5816b8b+24129607/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_a1cd38756c74f5816b8b+24129607/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..836724f44545ce0dedda1521fd4c623a6ea8ec72 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_a1cd38756c74f5816b8b+24129607/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_a1cd38756c74f5816b8b+24129607/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_a1cd38756c74f5816b8b+24129607/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_a1cd38756c74f5816b8b+24129607/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_a1cd38756c74f5816b8b+24129607/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..8b3b83651c6fadd4d8f7355f90c20f0dea795ac0 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_a1cd38756c74f5816b8b+24129607/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5afb2944ba0711a4441c137bd39995accbbe79dc510f590407ae77b49a33df0d +size 1050679 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_a1cd38756c74f5816b8b+24129607/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_a1cd38756c74f5816b8b+24129607/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..a5edbee3f81b7fc208c0f966cbdc13c9c02e8262 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_a1cd38756c74f5816b8b+24129607/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:652897628aae608aa5ee58fc2ab10b8599eba028f43996e8519fe056c5c402f8 +size 5827584 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b1f6368a309ea57ce627+3181e94a/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_b1f6368a309ea57ce627+3181e94a/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..e35ab266d9f51949c9347f951479a2ff314da1e4 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b1f6368a309ea57ce627+3181e94a/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_9d35d101-ee67-4383-96a9-60332ea3d668/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_b1f6368a309ea57ce627+3181e94a/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_b1f6368a309ea57ce627+3181e94a/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b1f6368a309ea57ce627+3181e94a/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_b1f6368a309ea57ce627+3181e94a/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..50604677fbf25d7b519098dc89a8c2738e04cf1a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b1f6368a309ea57ce627+3181e94a/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9cab74238b621677b38d0f2660428e2499dca903faab0b9597c2f828d7feb87 +size 29532 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b1f6368a309ea57ce627+3181e94a/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_b1f6368a309ea57ce627+3181e94a/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..60e7dd49020f00c2cd4014a9b8455c25475bad43 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b1f6368a309ea57ce627+3181e94a/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44c350a10b53a13f67e90f9d371b28d4626ead63d0ab95dd38df6625595ddd1e +size 328704 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b1f6368a309ea57ce627+3181e94a/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_b1f6368a309ea57ce627+3181e94a/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..acd3fb1ff9c09a3d3e56f1d3529c8e94fd771262 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b1f6368a309ea57ce627+3181e94a/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dcf519338c7f7df2d2b0c7a8c244b9d249f579cb1aa09d8b1bbc9a0f3cf344b0 +size 334452 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b5e0b2d29beb17680855+a32116a7/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_b5e0b2d29beb17680855+a32116a7/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..3c18214665ab208ee16402ace66a8dd081416d1a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b5e0b2d29beb17680855+a32116a7/compile_flags.json @@ -0,0 +1 @@ +["--target=trn2", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=2", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b5e0b2d29beb17680855+a32116a7/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_b5e0b2d29beb17680855+a32116a7/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b5e0b2d29beb17680855+a32116a7/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_b5e0b2d29beb17680855+a32116a7/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..193dd51af6cc41c744c400ae69ad2db882a60375 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b5e0b2d29beb17680855+a32116a7/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:714a6830f526c63be97ababe983fe227fa2a7fecb5cab1f1a4045a005e41473d +size 1314616 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b5e0b2d29beb17680855+a32116a7/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_b5e0b2d29beb17680855+a32116a7/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..7485b42b86df7be1d54b495602ae6ea490ad41d0 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b5e0b2d29beb17680855+a32116a7/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78742ae2ad6f32b589bf0c135e8e76c44feaea1984d793f6b59965fe0bf89af7 +size 728064 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b5eece93194f9ce71789+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_b5eece93194f9ce71789+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b5eece93194f9ce71789+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_b5eece93194f9ce71789+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_b5eece93194f9ce71789+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b5eece93194f9ce71789+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_b5eece93194f9ce71789+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..79a0f0c84e8b7d06758405d9712710e9e81f6d6f --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b5eece93194f9ce71789+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f439ba657f29daeb38932d93cc76281739af44d62554f8bedc65a2360627f79d +size 734453 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b5eece93194f9ce71789+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_b5eece93194f9ce71789+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..e8922bbb8ee57e6b717607efa85887b88b52370a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b5eece93194f9ce71789+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84c1c8df51a53f62b1d9f94b73a7f1ee7e18cf6ab16e900e721c095af5ed24d4 +size 12329984 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b5eece93194f9ce71789+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_b5eece93194f9ce71789+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..fc2b4f054964386aab594c48753f8deb34194168 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b5eece93194f9ce71789+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60023ba4fd81922718fadca76d0faea3d7b32ab7c86f24bcf4f071a410b555c0 +size 12477000 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b651bd725ae2d8726217+80d05c3f/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_b651bd725ae2d8726217+80d05c3f/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..3d7d85cd3d975162683053faa369dcc6259604af --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b651bd725ae2d8726217+80d05c3f/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--enable-mixed-precision-accumulation", "--model-type", "transformer", "-O1", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2", "--auto-cast=none", "--internal-enable-dge-levels", "vector_dynamic_offsets", "--logfile=/tmp/nxd_model/token_generation/_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_b651bd725ae2d8726217+80d05c3f/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_b651bd725ae2d8726217+80d05c3f/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b651bd725ae2d8726217+80d05c3f/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_b651bd725ae2d8726217+80d05c3f/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..3c1d3f8d82d6f45827207e27bc89760cea90909c --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b651bd725ae2d8726217+80d05c3f/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25d19f95bd2d2d22716d50dd7ff25b13a26b0bd88df29dca923442e3a52de242 +size 74684 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b651bd725ae2d8726217+80d05c3f/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_b651bd725ae2d8726217+80d05c3f/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..3a1829605302fdb9da4ca4e86c66bf09f359ae4b --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b651bd725ae2d8726217+80d05c3f/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ede9d4f2677bed7427f874e5f503791ef1cc8be4af7d4e6bd1a7aa6bb9ee467d +size 277504 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b651bd725ae2d8726217+80d05c3f/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_b651bd725ae2d8726217+80d05c3f/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..f18cf246f89228d80f42a35b5d855f2109134a05 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b651bd725ae2d8726217+80d05c3f/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ac4db0e4cb996e91e41b41f3e9aaf8204fb1bbd888dc69bb4606c019bf83fdc +size 289434 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b87961c02022113d30f4+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_b87961c02022113d30f4+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b87961c02022113d30f4+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_b87961c02022113d30f4+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_b87961c02022113d30f4+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b87961c02022113d30f4+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_b87961c02022113d30f4+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..bcff2cce1144329bac7c6cb34cdcaabd69311e8a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b87961c02022113d30f4+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22087adc7a012a994ce8911f4eda9a541258e32cbea09a62369de2492d056ab8 +size 405481 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b87961c02022113d30f4+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_b87961c02022113d30f4+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..2db58043032e46a9ff844aed4b060e91282dca63 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b87961c02022113d30f4+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9f81c3f224e8d6a6d6143de80e7fee956d55148d61f7ba601bf225dfaea01a0 +size 1997824 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_b87961c02022113d30f4+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_b87961c02022113d30f4+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..4ad774f6b1be8a69e605a360eb0490ea4502e49d --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_b87961c02022113d30f4+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:767dccde03e8e5cfbfdabc2ad4c4a7e35b803036f5df49dee8b0e106aca1a752 +size 2090737 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_bfde8ebf354636b209da+24129607/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_bfde8ebf354636b209da+24129607/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..836724f44545ce0dedda1521fd4c623a6ea8ec72 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_bfde8ebf354636b209da+24129607/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_bfde8ebf354636b209da+24129607/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_bfde8ebf354636b209da+24129607/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_bfde8ebf354636b209da+24129607/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_bfde8ebf354636b209da+24129607/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..6a1343f8d1c22c108ec81c0e96453226d1b6dcad --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_bfde8ebf354636b209da+24129607/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90e059fee7359da59d6c908ee73e7eeeb1886c53da6a1d5e1b498fac4c6908ed +size 83504 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_bfde8ebf354636b209da+24129607/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_bfde8ebf354636b209da+24129607/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..2c51c608536892d96eec0d13bf65a0ab644b6834 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_bfde8ebf354636b209da+24129607/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60f3d973be675fca98ebaccd4b2dbcdb5526cfae37c4d87ec659ba0322c49684 +size 287744 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_c3216c9356c729b39292+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_c3216c9356c729b39292+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_c3216c9356c729b39292+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_c3216c9356c729b39292+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_c3216c9356c729b39292+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_c3216c9356c729b39292+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_c3216c9356c729b39292+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..90dce993d1c27da71e05ba619baa352560aa307f --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_c3216c9356c729b39292+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9522be72d7bef5233cbcd5f03292acf81d1d8dc6f7a0ccb3ff1c25c373e7da6 +size 1123832 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_c3216c9356c729b39292+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_c3216c9356c729b39292+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..862793202fe2d4a9910fd27c3667bcd371f9c5bb --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_c3216c9356c729b39292+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec65def4eaa0a20b35f5a345458cb0e4d1a841b0ecc4e1e3542ad7d6a25bcb6a +size 7005184 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_c3216c9356c729b39292+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_c3216c9356c729b39292+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..057b9edf3a7d4053f3af27e8659e829fe50f5b29 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_c3216c9356c729b39292+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c241a84921c14baea9c54a7d105922da3285efe8ab386d3505ba7db75f66a1b0 +size 7189377 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_c6ac9f1b555d05bd7e39+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_c6ac9f1b555d05bd7e39+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_c6ac9f1b555d05bd7e39+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_c6ac9f1b555d05bd7e39+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_c6ac9f1b555d05bd7e39+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_c6ac9f1b555d05bd7e39+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_c6ac9f1b555d05bd7e39+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..3467a28f4d54fb66e965ba13a331395f7be50279 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_c6ac9f1b555d05bd7e39+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aea4e9cc13ab00a8e2e6f922b982507ed418e56b8f1b5c6c8e3d72c9dfdd8d44 +size 574388 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_c6ac9f1b555d05bd7e39+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_c6ac9f1b555d05bd7e39+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..16b067f0e2339bcb9dec283c5994380923e5dc9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_c6ac9f1b555d05bd7e39+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b32525c928c868fca4bff354f75d0607cbe9c470aaafaba8e34a9da5a37cf8ba +size 1086464 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_c6ac9f1b555d05bd7e39+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_c6ac9f1b555d05bd7e39+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..bb463f6d6686c4cee216899638237371322ff207 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_c6ac9f1b555d05bd7e39+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa3f91118d993e6dceb74b20a7733d83daf3b5589cca7da9638151b8a6efb76a +size 1208853 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_c75fd5e63a178ce4f3d1+677eeb9d/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_c75fd5e63a178ce4f3d1+677eeb9d/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..ae7e0c6790082c43cd145aae8cc7e3cc89e15ca8 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_c75fd5e63a178ce4f3d1+677eeb9d/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/speculation_model/_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_c75fd5e63a178ce4f3d1+677eeb9d/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_c75fd5e63a178ce4f3d1+677eeb9d/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_c75fd5e63a178ce4f3d1+677eeb9d/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_c75fd5e63a178ce4f3d1+677eeb9d/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..75e0d55ae0c6061c1e8b08f4f2195fa17f80e49a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_c75fd5e63a178ce4f3d1+677eeb9d/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed708e361a873a0f5a0f54a37a60ceaefda28896ee334ed372676c770f162697 +size 455594 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_c75fd5e63a178ce4f3d1+677eeb9d/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_c75fd5e63a178ce4f3d1+677eeb9d/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..9b891c7aadd149e5d575bce29ab60584948cfb40 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_c75fd5e63a178ce4f3d1+677eeb9d/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18051bbe5941a0521ceed1401548d145062cc95c2f1bb7a9c886e052cbf7a121 +size 3922944 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_cd398d15ddca2b8db9a9+24129607/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_cd398d15ddca2b8db9a9+24129607/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..836724f44545ce0dedda1521fd4c623a6ea8ec72 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_cd398d15ddca2b8db9a9+24129607/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_cd398d15ddca2b8db9a9+24129607/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_cd398d15ddca2b8db9a9+24129607/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_cd398d15ddca2b8db9a9+24129607/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_cd398d15ddca2b8db9a9+24129607/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..6e4701bc0c10a25a0a06cba3aaa67a9a5f6e4d7b --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_cd398d15ddca2b8db9a9+24129607/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ada1a8cdd1e2563985ad4f6b5da6f07dfb3891689064800b5ae54f325d71200 +size 83504 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_cd398d15ddca2b8db9a9+24129607/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_cd398d15ddca2b8db9a9+24129607/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..abc83a2b1cd62217075810c4627fb6d1e190f44a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_cd398d15ddca2b8db9a9+24129607/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5149332130fc9012cb373e61bd59cf414015efae208f0432a94f8067650f33e +size 328704 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_d28157993393c2fbf90f+1592a3e2/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_d28157993393c2fbf90f+1592a3e2/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..d97e4d0133f070c2dd7275b28c5d410e5119fee1 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_d28157993393c2fbf90f+1592a3e2/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_19528f74-a51d-46a3-a297-4609ebe7dcc9/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_d28157993393c2fbf90f+1592a3e2/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_d28157993393c2fbf90f+1592a3e2/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_d28157993393c2fbf90f+1592a3e2/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_d28157993393c2fbf90f+1592a3e2/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..e6084ce6d845ff91b01c0cbfd758d24019f38ef9 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_d28157993393c2fbf90f+1592a3e2/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_d28157993393c2fbf90f+1592a3e2/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_d28157993393c2fbf90f+1592a3e2/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..a43d728c9903f7378f66ae6ba5e884009341b466 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_d28157993393c2fbf90f+1592a3e2/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a23f5e7f07922bc80bb87db284b38bdf893d46072d47382548833df5dc62249 +size 134144 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_d28157993393c2fbf90f+1592a3e2/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_d28157993393c2fbf90f+1592a3e2/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..d20e1e801fa097c2a92128439a829c16934b2555 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_d28157993393c2fbf90f+1592a3e2/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e71ea0d0a351279393c9e8a8f76bccdd57d3109b72bd8bcaf75ff51933012261 +size 136222 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_db3cfbc4f07f54d7ec28+24129607/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_db3cfbc4f07f54d7ec28+24129607/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..836724f44545ce0dedda1521fd4c623a6ea8ec72 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_db3cfbc4f07f54d7ec28+24129607/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_db3cfbc4f07f54d7ec28+24129607/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_db3cfbc4f07f54d7ec28+24129607/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_db3cfbc4f07f54d7ec28+24129607/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_db3cfbc4f07f54d7ec28+24129607/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..da297c0af91ac37297be3a20e620cd26c71620cd --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_db3cfbc4f07f54d7ec28+24129607/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:919ec4a413b1cf92bee992795b9851958ebf6262d6e0a4e5f84848017747b1ad +size 593552 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_db3cfbc4f07f54d7ec28+24129607/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_db3cfbc4f07f54d7ec28+24129607/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..3b88b6ac48f53bbd26432d1cc5e33e99c025ec6b --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_db3cfbc4f07f54d7ec28+24129607/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:06a3c4224750067c6cbd1a7eb69271c58ac7c3580cdb0bb24cc8b5b4ee9f3ce9 +size 3769344 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_e337acd66701db458735+24129607/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_e337acd66701db458735+24129607/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..836724f44545ce0dedda1521fd4c623a6ea8ec72 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_e337acd66701db458735+24129607/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_e337acd66701db458735+24129607/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_e337acd66701db458735+24129607/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_e337acd66701db458735+24129607/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_e337acd66701db458735+24129607/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..fe83deb92860f016e77e8058cfa8ef9da53dd57c --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_e337acd66701db458735+24129607/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8577f49ec00439c40e1058ddef82ca8a142655c3cc99be5bdd45d4a97b4ec763 +size 532500 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_e337acd66701db458735+24129607/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_e337acd66701db458735+24129607/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..1ee3e0bb830c12d9f8f7eee399942590125b4d95 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_e337acd66701db458735+24129607/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:043a5ae6643a58a7ff6a5bf07db42c29e3482ade0f0b523dd5b4151f3413cfdd +size 1813504 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_e737e0fdb56bf5bd26fc+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_e737e0fdb56bf5bd26fc+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_e737e0fdb56bf5bd26fc+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_e737e0fdb56bf5bd26fc+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_e737e0fdb56bf5bd26fc+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_e737e0fdb56bf5bd26fc+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_e737e0fdb56bf5bd26fc+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..8f6a9f6860d90ce21700383cf34274f7eff95c39 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_e737e0fdb56bf5bd26fc+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:126886d674179a37c2e080667a65d212841def7e512e4014e61d74962dcd0132 +size 86102 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_e737e0fdb56bf5bd26fc+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_e737e0fdb56bf5bd26fc+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..efff402ec2afbde2078fa7c193fc6492fddc31e7 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_e737e0fdb56bf5bd26fc+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:45516796e50db563f43b7e4cf8e2f1b3a146c338eaa63e0ae44ca540d1836128 +size 277504 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_e737e0fdb56bf5bd26fc+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_e737e0fdb56bf5bd26fc+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..f2639522ae6bda456c07245e87a9ed57f1189c73 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_e737e0fdb56bf5bd26fc+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93f1c519ba2d7fd91265c7b0ba4446976e68a1d32d0bca2cc5a686186209ff8a +size 285717 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_ea4fa7da6b267fc2c5c7+4482fc6e/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_ea4fa7da6b267fc2c5c7+4482fc6e/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..72e8649e886b59ce51bb1e03a61d18d8137e2970 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_ea4fa7da6b267fc2c5c7+4482fc6e/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_d2bec72a-fc7f-4c4d-98be-dadade1ec7fd/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_ea4fa7da6b267fc2c5c7+4482fc6e/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_ea4fa7da6b267fc2c5c7+4482fc6e/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_ea4fa7da6b267fc2c5c7+4482fc6e/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_ea4fa7da6b267fc2c5c7+4482fc6e/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..307f90c8e8aa2efa59b5a7f04944e13f8a7e1b5f --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_ea4fa7da6b267fc2c5c7+4482fc6e/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc96f7fb7fbf64514a818a779dd383411b8c502b328094192490c454e498a825 +size 14480 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_ea4fa7da6b267fc2c5c7+4482fc6e/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_ea4fa7da6b267fc2c5c7+4482fc6e/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..1a18c66969bb07136f3d420525dfd4ec32b5f75a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_ea4fa7da6b267fc2c5c7+4482fc6e/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47fa19cc7313831f44f01915dfbc7b0fa6dcca398b58fecc9bec3ac6ec6f52bc +size 267264 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_ea4fa7da6b267fc2c5c7+4482fc6e/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_ea4fa7da6b267fc2c5c7+4482fc6e/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..cb3655aba91f99d02d4893394d910da4485eb15d --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_ea4fa7da6b267fc2c5c7+4482fc6e/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1088a73a4725da3e83d72f06ef3cc3ebfcfb515910f51644e84aba28bc3b8d4d +size 272962 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_eb5005f121a849e69eb7+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_eb5005f121a849e69eb7+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_eb5005f121a849e69eb7+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_eb5005f121a849e69eb7+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_eb5005f121a849e69eb7+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_eb5005f121a849e69eb7+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_eb5005f121a849e69eb7+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..1a3606dd60158f628ee4dba36476e80af3ef54fc --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_eb5005f121a849e69eb7+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8be4eff62c7d95ba6141d3b9629ad6ba857b158185449a221229a5e829f32b56 +size 918473 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_eb5005f121a849e69eb7+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_eb5005f121a849e69eb7+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..e2c68cff408215bf0bf876f72c6fbab8f3610f9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_eb5005f121a849e69eb7+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e50abbc9b548b192b540a1dead3bf4253640fd811ceb9dabe4e1770bcccb2d4f +size 5704704 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_eb5005f121a849e69eb7+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_eb5005f121a849e69eb7+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..809ad95def3668ec80b131d94cd210ffebb569e1 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_eb5005f121a849e69eb7+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9049a432cd117654c62bae8fa876d45da969e25f7a100b4c7ae8c4cea7c58890 +size 5871228 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_ef6c48ae8ca983dc5ac7+4ef53020/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_ef6c48ae8ca983dc5ac7+4ef53020/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..399698a9ce7d5df427c8b3892c2a25045432717f --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_ef6c48ae8ca983dc5ac7+4ef53020/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_e770e1f1-a135-4d46-9c7b-c250b5d64bad/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_ef6c48ae8ca983dc5ac7+4ef53020/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_ef6c48ae8ca983dc5ac7+4ef53020/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_ef6c48ae8ca983dc5ac7+4ef53020/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_ef6c48ae8ca983dc5ac7+4ef53020/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..89eb2339905a3e3d5d05af1d33aeb632bbb1436e --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_ef6c48ae8ca983dc5ac7+4ef53020/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14a5d6816ffd369be6c10cb45f27b95ad073d2be256d587642764fb87bce2d53 +size 8979 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_ef6c48ae8ca983dc5ac7+4ef53020/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_ef6c48ae8ca983dc5ac7+4ef53020/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..2c6bfdc34fc83047e22ca034f199d3505bc0def7 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_ef6c48ae8ca983dc5ac7+4ef53020/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efe201b0eda30c79ee1b59a9974bd0cfc299c2ce0a02a17814fd6c8d3497d4c0 +size 246784 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_ef6c48ae8ca983dc5ac7+4ef53020/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_ef6c48ae8ca983dc5ac7+4ef53020/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..07a59c141b4137a3efe2789ab74162a546a1397f --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_ef6c48ae8ca983dc5ac7+4ef53020/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af278bae97ddad99468786aab17f26d3be3642c314d3f4738c0c708b0650dc45 +size 249608 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_ef93001ee65461791427+24129607/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_ef93001ee65461791427+24129607/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..836724f44545ce0dedda1521fd4c623a6ea8ec72 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_ef93001ee65461791427+24129607/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_ef93001ee65461791427+24129607/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_ef93001ee65461791427+24129607/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_ef93001ee65461791427+24129607/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_ef93001ee65461791427+24129607/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..0b3c8f147b8bac5783832babfa2b8dbc7477e6be --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_ef93001ee65461791427+24129607/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df9f6cc64dd38c523fa33828cc0cca5eb3e0edb9c47024b7f6c748e59c22cb54 +size 82772 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_ef93001ee65461791427+24129607/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_ef93001ee65461791427+24129607/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..9bb814d32491348c038ac36048cba43b28ac2a5b --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_ef93001ee65461791427+24129607/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03f6e5ccf3b0e4969397c1b114409aa2e1cd46189425201864551b0b64a89b19 +size 267264 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_f33d66622d796b2877aa+283df001/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_f33d66622d796b2877aa+283df001/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..58fbf6b9366d9e28d7f19321e85acd4cd96c0614 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_f33d66622d796b2877aa+283df001/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--enable-mixed-precision-accumulation", "--model-type", "transformer", "-O1", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2", "--auto-cast=none", "--internal-enable-dge-levels", "vector_dynamic_offsets", "--internal-hlo2tensorizer-options=--verify-hlo=true", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_f33d66622d796b2877aa+283df001/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_f33d66622d796b2877aa+283df001/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_f33d66622d796b2877aa+283df001/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_f33d66622d796b2877aa+283df001/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..999c3d79c7a98a57ce36d6cf123c7b54ad0588b1 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_f33d66622d796b2877aa+283df001/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:335c133575936870fb0fae94881672481e07292ccd05798a21f93552aef3b7fd +size 97794 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_f33d66622d796b2877aa+283df001/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_f33d66622d796b2877aa+283df001/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..826509d6eb373a6ed7f50bb84e2358525d9f7443 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_f33d66622d796b2877aa+283df001/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4bd0c5e8dcc29d47f0b544761ad02481e325f7bc70226ed229787ae27520b274 +size 410624 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fa29fc69337939622d95+24129607/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_fa29fc69337939622d95+24129607/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..836724f44545ce0dedda1521fd4c623a6ea8ec72 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fa29fc69337939622d95+24129607/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fa29fc69337939622d95+24129607/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_fa29fc69337939622d95+24129607/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fa29fc69337939622d95+24129607/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_fa29fc69337939622d95+24129607/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..7fb4c9bbc41a898533b0531c218b32b9039aa882 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fa29fc69337939622d95+24129607/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5758f104aa59185e2d260ad19edb597a812c714aa81dc87048736add612f9418 +size 90382 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fa29fc69337939622d95+24129607/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_fa29fc69337939622d95+24129607/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..99d6d6953d3fa39fe7e0a14b49b5eee452c54cfb --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fa29fc69337939622d95+24129607/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40efb18e323706a3b13f75af55a66201e078f38c2ed010d562521d6f9812aa7e +size 359424 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fb598b8eba6e5a92e75f+24129607/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_fb598b8eba6e5a92e75f+24129607/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..836724f44545ce0dedda1521fd4c623a6ea8ec72 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fb598b8eba6e5a92e75f+24129607/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"] \ No newline at end of file diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fb598b8eba6e5a92e75f+24129607/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_fb598b8eba6e5a92e75f+24129607/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fb598b8eba6e5a92e75f+24129607/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_fb598b8eba6e5a92e75f+24129607/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..063bbf04ebf17e3e2c055b77dc1fd871e19ac9b6 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fb598b8eba6e5a92e75f+24129607/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a447ade2a2d2fbae6cb8aba2246b5ec0bf2861cbf08f1ed4c1bf22b7625db059 +size 761066 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fb598b8eba6e5a92e75f+24129607/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_fb598b8eba6e5a92e75f+24129607/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..92b18666ed672f0ab53eb50254d30de327134b35 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fb598b8eba6e5a92e75f+24129607/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1577e7107e5832f9db26e7a3f4aa9c24355fb28f03595c396a8ebcf1283bd46 +size 12493824 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fe491aea85c492a9d7ed+7e8048af/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_fe491aea85c492a9d7ed+7e8048af/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..1d4f95d5cdc5a81187948551c55a411b62e31605 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fe491aea85c492a9d7ed+7e8048af/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--enable-saturate-infinity", "--auto-cast=none", "--model-type=transformer", "-O1", "--logfile=/tmp/nxdi_test_07152ca8-2992-4834-a6f1-979ad40f60f6/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_fe491aea85c492a9d7ed+7e8048af/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_fe491aea85c492a9d7ed+7e8048af/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fe491aea85c492a9d7ed+7e8048af/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_fe491aea85c492a9d7ed+7e8048af/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..8203a5aa5acd37ae7e6d6839b35c7438ee9e48fd --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fe491aea85c492a9d7ed+7e8048af/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0eb674f0ac78f8d22ded7c755c983de3edb92563308d60b734d19e694f489568 +size 11280 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fe491aea85c492a9d7ed+7e8048af/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_fe491aea85c492a9d7ed+7e8048af/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..2d0ed7338c02bd5c87b662e67e626e3906aa40a3 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fe491aea85c492a9d7ed+7e8048af/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1e0410a9e2e263363b44a6fe093fcefd6e207da131bf1acefa977eb8c325499 +size 1444864 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fe491aea85c492a9d7ed+7e8048af/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_fe491aea85c492a9d7ed+7e8048af/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..2c719838ed41ef44a48de0cbc654252c96b57f30 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fe491aea85c492a9d7ed+7e8048af/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55e514cfdbc85b8cd9bea9a412a8f50a6b27f21f81a2df79d5274764dcca94aa +size 1447845 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fe6b8b40a8e1d622e71e+a02c3a36/compile_flags.json b/neuronxcc-2.21.33363.0+82129205/MODULE_fe6b8b40a8e1d622e71e+a02c3a36/compile_flags.json new file mode 100644 index 0000000000000000000000000000000000000000..5726abc7d1d8c52fa95bc7919439a23a23fe3b9a --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fe6b8b40a8e1d622e71e+a02c3a36/compile_flags.json @@ -0,0 +1 @@ +["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_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_fe6b8b40a8e1d622e71e+a02c3a36/model.done b/neuronxcc-2.21.33363.0+82129205/MODULE_fe6b8b40a8e1d622e71e+a02c3a36/model.done new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fe6b8b40a8e1d622e71e+a02c3a36/model.hlo_module.pb b/neuronxcc-2.21.33363.0+82129205/MODULE_fe6b8b40a8e1d622e71e+a02c3a36/model.hlo_module.pb new file mode 100644 index 0000000000000000000000000000000000000000..90b29d24d1deec0ccffe07084324f7b6bc10d4f4 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fe6b8b40a8e1d622e71e+a02c3a36/model.hlo_module.pb @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7bca3bd92e92ac125defbadf7e843c65755895cb44e5cdcbc92712a880e49ab1 +size 84493 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fe6b8b40a8e1d622e71e+a02c3a36/model.neff b/neuronxcc-2.21.33363.0+82129205/MODULE_fe6b8b40a8e1d622e71e+a02c3a36/model.neff new file mode 100644 index 0000000000000000000000000000000000000000..5c2f0da8a8b70ab020f0d59a673754c2042e775d --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fe6b8b40a8e1d622e71e+a02c3a36/model.neff @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3aced29a9eab1a81387206b2a9a5f5a3716aafec360b3350f677fd860cb5d8a5 +size 216064 diff --git a/neuronxcc-2.21.33363.0+82129205/MODULE_fe6b8b40a8e1d622e71e+a02c3a36/wrapped_neff.hlo b/neuronxcc-2.21.33363.0+82129205/MODULE_fe6b8b40a8e1d622e71e+a02c3a36/wrapped_neff.hlo new file mode 100644 index 0000000000000000000000000000000000000000..8f8c5600163c84a747bde4c53e4cf27eeaa69539 --- /dev/null +++ b/neuronxcc-2.21.33363.0+82129205/MODULE_fe6b8b40a8e1d622e71e+a02c3a36/wrapped_neff.hlo @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:539c539df8ca90271f687afdf1aa3a80a65aa6dedea21b0bfd01eadbac362261 +size 224275