Synchronizing local compiler cache.

.gitattributes

@@ -5980,3 +5980,49 @@ neuronxcc-2.22.12471.0+b4a00d10/MODULE_a0eeca96db30781301d6+519f203d/model.neff
 neuronxcc-2.22.12471.0+b4a00d10/MODULE_a4a4bab288905ceb01b4+84337dd9/model.neff filter=lfs diff=lfs merge=lfs -text
 neuronxcc-2.22.12471.0+b4a00d10/MODULE_a4a4bab288905ceb01b4+84337dd9/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
 neuronxcc-2.22.12471.0+b4a00d10/MODULE_c841826784a9a062f620+519f203d/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_0734f7a73b49a760847f+84337dd9/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_0734f7a73b49a760847f+84337dd9/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_09867c424bfa84dfcf5a+84337dd9/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_09867c424bfa84dfcf5a+84337dd9/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_0ab63364bcd05580a283+519f203d/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_0b667d06d984ae2415eb+84337dd9/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_0b667d06d984ae2415eb+84337dd9/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_0f8f90687a1080fc3678+0788fa03/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_1389221757af9c7b1df7+b7ba88c2/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_167ee952f5ffde447542+84337dd9/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_167ee952f5ffde447542+84337dd9/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_1a2557c4b251a50a2c8a+519f203d/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_29d1245367829925f390+84337dd9/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_29d1245367829925f390+84337dd9/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_3d4b43665f1b1f693abb+519f203d/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_4632a636780e68814286+84337dd9/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_4632a636780e68814286+84337dd9/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_509c1a379916e29760a3+519f203d/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_59198f7736511b94f6e7+e2a39c39/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_59198f7736511b94f6e7+e2a39c39/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_59b0dfe47acbb84fd629+519f203d/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_5cb29fd67ef93662a5a0+519f203d/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_5e143bd96278905e05e5+84337dd9/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_5e143bd96278905e05e5+84337dd9/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_75f7275becb21ddfde2a+84337dd9/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_75f7275becb21ddfde2a+84337dd9/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_842356c564eb183c4d65+84337dd9/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_842356c564eb183c4d65+84337dd9/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_92ac2881f9ca0f71d32e+84337dd9/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_92ac2881f9ca0f71d32e+84337dd9/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_97532c7ef5dc7e4f13a0+519f203d/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_a31af46e912d771cbdea+519f203d/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_a8c767775500acdd8e3a+519f203d/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_afaedb2859b9ad9b9fb6+519f203d/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_b2dc50dd0bb673c35a17+519f203d/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_bcae6a2eb9d69ea94ae6+84337dd9/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_bcae6a2eb9d69ea94ae6+84337dd9/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_c7960ef661d9d15cc8cc+ae2b3e2d/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_c7960ef661d9d15cc8cc+ae2b3e2d/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_c847e0791366b4c4b274+84337dd9/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_c847e0791366b4c4b274+84337dd9/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_ced952ae43941164aa1a+519f203d/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_e6164d1cf3b338ef4f0c+d1c00e0b/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_f44ea3026cb2ce9d2df7+519f203d/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_f84b3d15315ce03d158d+84337dd9/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.22.12471.0+b4a00d10/MODULE_f84b3d15315ce03d158d+84337dd9/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text

neuronxcc-2.22.12471.0+b4a00d10/0_REGISTRY/0.4.4.dev2/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/ab85492dab47a880ba57.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": "2.22.12471.0+b4a00d10",
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.4.4.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
+}

neuronxcc-2.22.12471.0+b4a00d10/0_REGISTRY/0.4.4.dev2/llama/llamafactory/tiny-random-Llama-3/5b1a29e1fc789f8196bc.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": "2.22.12471.0+b4a00d10",
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.4.4.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
+}

neuronxcc-2.22.12471.0+b4a00d10/0_REGISTRY/0.4.4.dev2/llama/unsloth/Llama-3.2-1B-Instruct/ef1d4a7a1638b14a3c76.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": "2.22.12471.0+b4a00d10",
+    "on_device_sampling": false,
+    "optimum_neuron_version": "0.4.4.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
+}

neuronxcc-2.22.12471.0+b4a00d10/0_REGISTRY/0.4.4.dev2/llama4_text/tiny-random/llama-4/ce86a511276f84327102.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": "2.22.12471.0+b4a00d10",
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.4.4.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
+}

neuronxcc-2.22.12471.0+b4a00d10/0_REGISTRY/0.4.4.dev2/mixtral/dacorvo/Mixtral-tiny/1825ae490ebaf0066f3d.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": "2.22.12471.0+b4a00d10",
+    "on_device_sampling": false,
+    "optimum_neuron_version": "0.4.4.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
+}

neuronxcc-2.22.12471.0+b4a00d10/0_REGISTRY/0.4.4.dev2/phi3/yujiepan/phi-4-tiny-random/0e9348e3d3a6c205f402.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": "2.22.12471.0+b4a00d10",
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.4.4.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
+}

neuronxcc-2.22.12471.0+b4a00d10/0_REGISTRY/0.4.4.dev2/qwen2/yujiepan/qwen2.5-128k-tiny-random/ee6c482035b6b3473cab.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": "2.22.12471.0+b4a00d10",
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.4.4.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
+}

neuronxcc-2.22.12471.0+b4a00d10/0_REGISTRY/0.4.4.dev2/qwen3_moe/optimum-internal-testing/tiny-random-qwen3_moe/a80d800541be4e880ee3.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": "2.22.12471.0+b4a00d10",
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.4.4.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
+}

neuronxcc-2.22.12471.0+b4a00d10/0_REGISTRY/0.4.4.dev2/smollm3/HuggingFaceTB/SmolLM3-3B/1032fefb59c006443ae4.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": "2.22.12471.0+b4a00d10",
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.4.4.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
+}

neuronxcc-2.22.12471.0+b4a00d10/MODULE_0734f7a73b49a760847f+84337dd9/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", "--verbose=35", "--logfile=/tmp/nxd_model/token_generation/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]

neuronxcc-2.22.12471.0+b4a00d10/MODULE_0734f7a73b49a760847f+84337dd9/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95d73215d4780160e939de28250ad42b94508d35c2ba646fad8a9285fcb222b6
+size 315826

neuronxcc-2.22.12471.0+b4a00d10/MODULE_0734f7a73b49a760847f+84337dd9/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4a9fd3cc53ad834cff00d8719e1d6445c5424c57969cf4d2d8bf778064a86e3
+size 308224

neuronxcc-2.22.12471.0+b4a00d10/MODULE_0734f7a73b49a760847f+84337dd9/wrapped_neff.hlo

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee357edc1ae39dde566b88e403a37b397dd8864e6ee6f99c0fe061fdc733293d
+size 319618

neuronxcc-2.22.12471.0+b4a00d10/MODULE_09867c424bfa84dfcf5a+84337dd9/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", "--verbose=35", "--logfile=/tmp/nxd_model/token_generation/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]

neuronxcc-2.22.12471.0+b4a00d10/MODULE_09867c424bfa84dfcf5a+84337dd9/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19d9b572dea8c7a7f1a70d2a2172b3b243b0a2c1cc1e3708705ba4805f245da5
+size 314544

neuronxcc-2.22.12471.0+b4a00d10/MODULE_09867c424bfa84dfcf5a+84337dd9/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0797d0342f401d9572e847b455b1f792e1005edad48d868e0d4b5003eed324f0
+size 308224

neuronxcc-2.22.12471.0+b4a00d10/MODULE_09867c424bfa84dfcf5a+84337dd9/wrapped_neff.hlo

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5fedb630b868e7a2ea54040e6e615724cc44eebe43c42e645603ea64753fb31d
+size 319618

neuronxcc-2.22.12471.0+b4a00d10/MODULE_0ab63364bcd05580a283+519f203d/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", "--verbose=35", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"]

neuronxcc-2.22.12471.0+b4a00d10/MODULE_0ab63364bcd05580a283+519f203d/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f173e475a11e2a7955ecbd2fa78e024bfbafa4c64ac64322268e10b476cc3b2
+size 244814

neuronxcc-2.22.12471.0+b4a00d10/MODULE_0ab63364bcd05580a283+519f203d/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dfe77e808f070091690bdaacfb27b7e4f8d94164ed16d0578b34f18e168b5a42
+size 287744

neuronxcc-2.22.12471.0+b4a00d10/MODULE_0b667d06d984ae2415eb+84337dd9/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", "--verbose=35", "--logfile=/tmp/nxd_model/token_generation/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]

neuronxcc-2.22.12471.0+b4a00d10/MODULE_0b667d06d984ae2415eb+84337dd9/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8069af4ea562006fc223816842ec8f47bfa1d48b033fed0f7ddc3fcc1138ba42
+size 1147291

neuronxcc-2.22.12471.0+b4a00d10/MODULE_0b667d06d984ae2415eb+84337dd9/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4dd76db8eda30b09a4d204962e945062469d08e2cc7cc7b7bb858adc107f6c9b
+size 594944

neuronxcc-2.22.12471.0+b4a00d10/MODULE_0b667d06d984ae2415eb+84337dd9/wrapped_neff.hlo

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f84761431b16ca78119e709070be533518005e2d0f623da2c1ba7846b33ddfd5
+size 614445

neuronxcc-2.22.12471.0+b4a00d10/MODULE_0f8f90687a1080fc3678+0788fa03/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", "--verbose=35", "--logfile=/tmp/nxd_model/speculation_model/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]

neuronxcc-2.22.12471.0+b4a00d10/MODULE_0f8f90687a1080fc3678+0788fa03/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8c78b43d7d485cdedc31104fdfd885a0a12552c14d2d2450f3f9375770c1073
+size 1783977

neuronxcc-2.22.12471.0+b4a00d10/MODULE_0f8f90687a1080fc3678+0788fa03/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:436ba50f0621abf64ea88297c63b219c0925b4b73cef49f22a78f97290016c7c
+size 4035584

neuronxcc-2.22.12471.0+b4a00d10/MODULE_1389221757af9c7b1df7+b7ba88c2/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", "--verbose=35", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"]

neuronxcc-2.22.12471.0+b4a00d10/MODULE_1389221757af9c7b1df7+b7ba88c2/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c57ef43047a70bae4b6a7176f61bcfadaa9beded533dba23db140e66505f077
+size 322952

neuronxcc-2.22.12471.0+b4a00d10/MODULE_1389221757af9c7b1df7+b7ba88c2/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:15e0794c58deaa6bd3422b54f6085f439093af87ad2a9d631667e03518c8783a
+size 441344

neuronxcc-2.22.12471.0+b4a00d10/MODULE_167ee952f5ffde447542+84337dd9/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", "--verbose=35", "--logfile=/tmp/nxd_model/token_generation/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]

neuronxcc-2.22.12471.0+b4a00d10/MODULE_167ee952f5ffde447542+84337dd9/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc48ca1b2611b5ad051b02e8ed3b44dec07322e3b2300e6a2f995bd18f7ce921
+size 272727

neuronxcc-2.22.12471.0+b4a00d10/MODULE_167ee952f5ffde447542+84337dd9/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6e0296fbc87d52070801873ffd3be4cbbbd8a4600637354fcfe5b34110ee733
+size 308224

neuronxcc-2.22.12471.0+b4a00d10/MODULE_167ee952f5ffde447542+84337dd9/wrapped_neff.hlo

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4eab23b70063e6899f29dceee0d08b468e05508bdca3b3484723fb735fbb53e5
+size 316437

neuronxcc-2.22.12471.0+b4a00d10/MODULE_1a2557c4b251a50a2c8a+519f203d/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", "--verbose=35", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"]

neuronxcc-2.22.12471.0+b4a00d10/MODULE_1a2557c4b251a50a2c8a+519f203d/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96f8ba3e488ab72785cb4ddebbeeb2b227880b7f4be178aa092174ff8d0bf6dd
+size 292043

neuronxcc-2.22.12471.0+b4a00d10/MODULE_1a2557c4b251a50a2c8a+519f203d/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53f7d975a1fa7926a7451e2a3cf83029efe1da26991b07aad45808b4c3a5fa03
+size 390144

neuronxcc-2.22.12471.0+b4a00d10/MODULE_29d1245367829925f390+84337dd9/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", "--verbose=35", "--logfile=/tmp/nxd_model/token_generation/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]

neuronxcc-2.22.12471.0+b4a00d10/MODULE_29d1245367829925f390+84337dd9/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b35fdb2007733da53c843e8e47518ba0b805f4480c4457363003876acf9470d
+size 270365

neuronxcc-2.22.12471.0+b4a00d10/MODULE_29d1245367829925f390+84337dd9/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:828f1c6ad55f4992b0b53c0539040115cbf737a50e93220f9042adb7a3031d82
+size 277504