diff --git a/.gitattributes b/.gitattributes
index 67ccf8f670fcbe016bdd5cea850867fc779ec0a2..40c5197c656967ec1c987fd82702a0e2a7746a04 100644
--- a/.gitattributes
+++ b/.gitattributes
@@ -2933,3 +2933,12 @@ neuronxcc-2.17.194.0+d312836f/MODULE_47164c525d06157b6333+431f5505/model.neff fi
 neuronxcc-2.17.194.0+d312836f/MODULE_5dadd23af77b433b987a+bfe5714b/model.neff filter=lfs diff=lfs merge=lfs -text
 neuronxcc-2.17.194.0+d312836f/MODULE_616ec6966b1c927020d5+165e9558/model.neff filter=lfs diff=lfs merge=lfs -text
 neuronxcc-2.17.194.0+d312836f/MODULE_616ec6966b1c927020d5+165e9558/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_634439d56d5bb67e9812+165e9558/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_634439d56d5bb67e9812+165e9558/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_72281753d7f57952bfe4+165e9558/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_72281753d7f57952bfe4+165e9558/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_857bfa5db7b34c9a9421+165e9558/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_857bfa5db7b34c9a9421+165e9558/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_94ba84f59b341a98c066+bfe5714b/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_d3986b7b5efee5f065f4+bfe5714b/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_f84068611c70eb3c41a2+bfe5714b/model.neff filter=lfs diff=lfs merge=lfs -text
diff --git a/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/16eb552455637c961181.json b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/16eb552455637c961181.json
new file mode 100644
index 0000000000000000000000000000000000000000..d275807cdbfb870a0be53266909b4296b27a64a9
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/16eb552455637c961181.json
@@ -0,0 +1,51 @@
+{
+  "_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,
+  "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": "HloNeuronConfig",
+    "all_reduce_dtype": null,
+    "allow_flash_attention": true,
+    "attention_layout": "HSB",
+    "attn_output_transposed": false,
+    "auto_cast_type": "fp16",
+    "batch_size": 2,
+    "checkpoint_id": "hf-internal-testing/tiny-random-GraniteForCausalLM",
+    "checkpoint_revision": "c3074ebc0ac2fe545305f5e5f6cce2cc9b2aa0c5",
+    "collectives_layout": "HSB",
+    "continuous_batching": true,
+    "fuse_qkv": true,
+    "group_query_attention": null,
+    "log_softmax_scores": false,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "optimum_neuron_version": "0.3.0.dev1",
+    "output_all_logits": false,
+    "sequence_length": 100,
+    "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-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/98a5b36eff78463d521e.json b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/98a5b36eff78463d521e.json
new file mode 100644
index 0000000000000000000000000000000000000000..00eb3567442b2cfd634e41a17799aaecf7ea764c
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/98a5b36eff78463d521e.json
@@ -0,0 +1,51 @@
+{
+  "_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,
+  "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": "HloNeuronConfig",
+    "all_reduce_dtype": null,
+    "allow_flash_attention": true,
+    "attention_layout": "HSB",
+    "attn_output_transposed": false,
+    "auto_cast_type": "bf16",
+    "batch_size": 1,
+    "checkpoint_id": "hf-internal-testing/tiny-random-GraniteForCausalLM",
+    "checkpoint_revision": "c3074ebc0ac2fe545305f5e5f6cce2cc9b2aa0c5",
+    "collectives_layout": "HSB",
+    "continuous_batching": false,
+    "fuse_qkv": true,
+    "group_query_attention": null,
+    "log_softmax_scores": false,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "optimum_neuron_version": "0.3.0.dev1",
+    "output_all_logits": false,
+    "sequence_length": 100,
+    "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-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/b3f4b03f5c98af7258c7.json b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/b3f4b03f5c98af7258c7.json
new file mode 100644
index 0000000000000000000000000000000000000000..e570ab13e6bb7e702a8883bdb57fbd286231eeb9
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/granite/hf-internal-testing/tiny-random-GraniteForCausalLM/b3f4b03f5c98af7258c7.json
@@ -0,0 +1,51 @@
+{
+  "_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,
+  "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": "HloNeuronConfig",
+    "all_reduce_dtype": null,
+    "allow_flash_attention": true,
+    "attention_layout": "HSB",
+    "attn_output_transposed": false,
+    "auto_cast_type": "fp16",
+    "batch_size": 1,
+    "checkpoint_id": "hf-internal-testing/tiny-random-GraniteForCausalLM",
+    "checkpoint_revision": "c3074ebc0ac2fe545305f5e5f6cce2cc9b2aa0c5",
+    "collectives_layout": "HSB",
+    "continuous_batching": false,
+    "fuse_qkv": true,
+    "group_query_attention": null,
+    "log_softmax_scores": false,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "optimum_neuron_version": "0.3.0.dev1",
+    "output_all_logits": false,
+    "sequence_length": 100,
+    "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-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/llamafactory/tiny-random-Llama-3/23870c03582a624b981f.json b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/llamafactory/tiny-random-Llama-3/23870c03582a624b981f.json
new file mode 100644
index 0000000000000000000000000000000000000000..e2199f4bcd5afceef8b436b0604458d7ffa940d1
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/llamafactory/tiny-random-Llama-3/23870c03582a624b981f.json
@@ -0,0 +1,55 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "llamafactory/tiny-random-Llama-3",
+  "_task": "text-generation",
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "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": "HloNeuronConfig",
+    "all_reduce_dtype": null,
+    "allow_flash_attention": true,
+    "attention_layout": "BSH",
+    "attn_output_transposed": false,
+    "auto_cast_type": "fp16",
+    "batch_size": 1,
+    "checkpoint_id": "llamafactory/tiny-random-Llama-3",
+    "checkpoint_revision": "bf2a2e3bf199ad2ee96f02a3c00246c608db22a8",
+    "collectives_layout": "HSB",
+    "continuous_batching": false,
+    "fuse_qkv": true,
+    "group_query_attention": null,
+    "log_softmax_scores": false,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "optimum_neuron_version": "0.3.0.dev1",
+    "output_all_logits": false,
+    "sequence_length": 100,
+    "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-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/llamafactory/tiny-random-Llama-3/38c497769b1d1cbd7c0d.json b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/llamafactory/tiny-random-Llama-3/38c497769b1d1cbd7c0d.json
new file mode 100644
index 0000000000000000000000000000000000000000..d06d7489fd2f02e68ccc3db3bc42dc9aab412dab
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/llamafactory/tiny-random-Llama-3/38c497769b1d1cbd7c0d.json
@@ -0,0 +1,55 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "llamafactory/tiny-random-Llama-3",
+  "_task": "text-generation",
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "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": "HloNeuronConfig",
+    "all_reduce_dtype": null,
+    "allow_flash_attention": true,
+    "attention_layout": "BSH",
+    "attn_output_transposed": false,
+    "auto_cast_type": "bf16",
+    "batch_size": 1,
+    "checkpoint_id": "llamafactory/tiny-random-Llama-3",
+    "checkpoint_revision": "bf2a2e3bf199ad2ee96f02a3c00246c608db22a8",
+    "collectives_layout": "HSB",
+    "continuous_batching": false,
+    "fuse_qkv": true,
+    "group_query_attention": null,
+    "log_softmax_scores": false,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "optimum_neuron_version": "0.3.0.dev1",
+    "output_all_logits": false,
+    "sequence_length": 100,
+    "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-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/llamafactory/tiny-random-Llama-3/3f83ce0c2e5f27f6fa2d.json b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/llamafactory/tiny-random-Llama-3/3f83ce0c2e5f27f6fa2d.json
new file mode 100644
index 0000000000000000000000000000000000000000..01bdfd8810bccfc48298fe2d8748a300be7808b5
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/llamafactory/tiny-random-Llama-3/3f83ce0c2e5f27f6fa2d.json
@@ -0,0 +1,77 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "llamafactory/tiny-random-Llama-3",
+  "_task": "text-generation",
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "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",
+    "async_mode": false,
+    "attn_kernel_enabled": false,
+    "batch_size": 1,
+    "capacity_factor": null,
+    "cc_pipeline_tiling_factor": 2,
+    "checkpoint_id": "llamafactory/tiny-random-Llama-3",
+    "checkpoint_revision": "bf2a2e3bf199ad2ee96f02a3c00246c608db22a8",
+    "continuous_batching": false,
+    "enable_bucketing": false,
+    "ep_degree": 1,
+    "flash_decoding_enabled": false,
+    "fused_qkv": true,
+    "glu_mlp": true,
+    "is_chunked_prefill": false,
+    "local_ranks_size": 2,
+    "logical_nc_config": 1,
+    "max_batch_size": 1,
+    "max_context_length": 100,
+    "max_topk": 256,
+    "mlp_kernel_enabled": false,
+    "mlp_kernel_fuse_residual_add": false,
+    "n_active_tokens": 100,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "num_cores_per_group": 1,
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.3.0.dev1",
+    "output_logits": false,
+    "padding_side": "right",
+    "pp_degree": 1,
+    "qk_layernorm": false,
+    "qkv_kernel_enabled": false,
+    "rpl_reduce_dtype": "bfloat16",
+    "sequence_length": 100,
+    "sequence_parallel_enabled": false,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "torch_dtype": "bfloat16",
+    "tp_degree": 2,
+    "vocab_parallel": false
+  },
+  "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-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/llamafactory/tiny-random-Llama-3/8dcd6598dcebb27ef470.json b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/llamafactory/tiny-random-Llama-3/8dcd6598dcebb27ef470.json
new file mode 100644
index 0000000000000000000000000000000000000000..a264242c960e5343432f086a620fa49f1847e6ec
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/llamafactory/tiny-random-Llama-3/8dcd6598dcebb27ef470.json
@@ -0,0 +1,77 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "llamafactory/tiny-random-Llama-3",
+  "_task": "text-generation",
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "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",
+    "async_mode": false,
+    "attn_kernel_enabled": false,
+    "batch_size": 1,
+    "capacity_factor": null,
+    "cc_pipeline_tiling_factor": 2,
+    "checkpoint_id": "llamafactory/tiny-random-Llama-3",
+    "checkpoint_revision": "bf2a2e3bf199ad2ee96f02a3c00246c608db22a8",
+    "continuous_batching": false,
+    "enable_bucketing": false,
+    "ep_degree": 1,
+    "flash_decoding_enabled": false,
+    "fused_qkv": true,
+    "glu_mlp": true,
+    "is_chunked_prefill": false,
+    "local_ranks_size": 2,
+    "logical_nc_config": 1,
+    "max_batch_size": 1,
+    "max_context_length": 100,
+    "max_topk": 256,
+    "mlp_kernel_enabled": false,
+    "mlp_kernel_fuse_residual_add": false,
+    "n_active_tokens": 100,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "num_cores_per_group": 1,
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.3.0.dev1",
+    "output_logits": false,
+    "padding_side": "right",
+    "pp_degree": 1,
+    "qk_layernorm": false,
+    "qkv_kernel_enabled": false,
+    "rpl_reduce_dtype": "float16",
+    "sequence_length": 100,
+    "sequence_parallel_enabled": false,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "torch_dtype": "float16",
+    "tp_degree": 2,
+    "vocab_parallel": false
+  },
+  "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-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/llamafactory/tiny-random-Llama-3/b9624072379e00f37909.json b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/llamafactory/tiny-random-Llama-3/b9624072379e00f37909.json
new file mode 100644
index 0000000000000000000000000000000000000000..1a16723daefaa5f854b0ffaf693f91336bf4590e
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/llamafactory/tiny-random-Llama-3/b9624072379e00f37909.json
@@ -0,0 +1,55 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "llamafactory/tiny-random-Llama-3",
+  "_task": "text-generation",
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "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": "HloNeuronConfig",
+    "all_reduce_dtype": null,
+    "allow_flash_attention": true,
+    "attention_layout": "BSH",
+    "attn_output_transposed": false,
+    "auto_cast_type": "fp16",
+    "batch_size": 2,
+    "checkpoint_id": "llamafactory/tiny-random-Llama-3",
+    "checkpoint_revision": "bf2a2e3bf199ad2ee96f02a3c00246c608db22a8",
+    "collectives_layout": "HSB",
+    "continuous_batching": true,
+    "fuse_qkv": true,
+    "group_query_attention": null,
+    "log_softmax_scores": false,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "optimum_neuron_version": "0.3.0.dev1",
+    "output_all_logits": false,
+    "sequence_length": 100,
+    "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-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/llamafactory/tiny-random-Llama-3/cfce0a36a7aad541df51.json b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/llamafactory/tiny-random-Llama-3/cfce0a36a7aad541df51.json
new file mode 100644
index 0000000000000000000000000000000000000000..985951ec31d99688d6a9f2f3088913292858a408
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/llamafactory/tiny-random-Llama-3/cfce0a36a7aad541df51.json
@@ -0,0 +1,77 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "llamafactory/tiny-random-Llama-3",
+  "_task": "text-generation",
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "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",
+    "async_mode": false,
+    "attn_kernel_enabled": false,
+    "batch_size": 2,
+    "capacity_factor": null,
+    "cc_pipeline_tiling_factor": 2,
+    "checkpoint_id": "llamafactory/tiny-random-Llama-3",
+    "checkpoint_revision": "bf2a2e3bf199ad2ee96f02a3c00246c608db22a8",
+    "continuous_batching": true,
+    "enable_bucketing": false,
+    "ep_degree": 1,
+    "flash_decoding_enabled": false,
+    "fused_qkv": true,
+    "glu_mlp": true,
+    "is_chunked_prefill": false,
+    "local_ranks_size": 2,
+    "logical_nc_config": 1,
+    "max_batch_size": 2,
+    "max_context_length": 100,
+    "max_topk": 256,
+    "mlp_kernel_enabled": false,
+    "mlp_kernel_fuse_residual_add": false,
+    "n_active_tokens": 100,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "num_cores_per_group": 1,
+    "on_device_sampling": false,
+    "optimum_neuron_version": "0.3.0.dev1",
+    "output_logits": false,
+    "padding_side": "right",
+    "pp_degree": 1,
+    "qk_layernorm": false,
+    "qkv_kernel_enabled": false,
+    "rpl_reduce_dtype": "float16",
+    "sequence_length": 100,
+    "sequence_parallel_enabled": false,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "torch_dtype": "float16",
+    "tp_degree": 2,
+    "vocab_parallel": false
+  },
+  "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-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/unsloth/Llama-3.2-1B-Instruct/bece693cb5ff2eaedc7d.json b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/unsloth/Llama-3.2-1B-Instruct/bece693cb5ff2eaedc7d.json
new file mode 100644
index 0000000000000000000000000000000000000000..b37c012b8de4311564d312e0db455b81aa1ce41c
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/llama/unsloth/Llama-3.2-1B-Instruct/bece693cb5ff2eaedc7d.json
@@ -0,0 +1,78 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "unsloth/Llama-3.2-1B-Instruct",
+  "_task": "text-generation",
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "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",
+    "async_mode": false,
+    "attn_kernel_enabled": false,
+    "batch_size": 4,
+    "capacity_factor": null,
+    "cc_pipeline_tiling_factor": 2,
+    "checkpoint_id": "unsloth/Llama-3.2-1B-Instruct",
+    "checkpoint_revision": "5a8abab4a5d6f164389b1079fb721cfab8d7126c",
+    "continuous_batching": true,
+    "enable_bucketing": false,
+    "ep_degree": 1,
+    "flash_decoding_enabled": false,
+    "fused_qkv": true,
+    "glu_mlp": true,
+    "is_chunked_prefill": false,
+    "local_ranks_size": 2,
+    "logical_nc_config": 1,
+    "max_batch_size": 4,
+    "max_context_length": 4096,
+    "max_topk": 256,
+    "mlp_kernel_enabled": false,
+    "mlp_kernel_fuse_residual_add": false,
+    "n_active_tokens": 4096,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "num_cores_per_group": 1,
+    "on_device_sampling": false,
+    "optimum_neuron_version": "0.3.0.dev1",
+    "output_logits": false,
+    "padding_side": "right",
+    "pp_degree": 1,
+    "qk_layernorm": false,
+    "qkv_kernel_enabled": false,
+    "rpl_reduce_dtype": "float16",
+    "sequence_length": 4096,
+    "sequence_parallel_enabled": false,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "torch_dtype": "float16",
+    "tp_degree": 2,
+    "vocab_parallel": false
+  },
+  "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-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/mixtral/dacorvo/Mixtral-tiny/1324c0afc0fb590822ad.json b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/mixtral/dacorvo/Mixtral-tiny/1324c0afc0fb590822ad.json
new file mode 100644
index 0000000000000000000000000000000000000000..4cef32eafadc72598528fee29ffc8bd1d3ed53e4
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/mixtral/dacorvo/Mixtral-tiny/1324c0afc0fb590822ad.json
@@ -0,0 +1,73 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "dacorvo/Mixtral-tiny",
+  "_task": "text-generation",
+  "architectures": [
+    "MixtralForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "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",
+    "async_mode": false,
+    "attn_kernel_enabled": false,
+    "batch_size": 1,
+    "capacity_factor": null,
+    "cc_pipeline_tiling_factor": 2,
+    "checkpoint_id": "dacorvo/Mixtral-tiny",
+    "checkpoint_revision": "c557ba205ddff6ea911f4719e0d543d6c08356b6",
+    "continuous_batching": false,
+    "enable_bucketing": false,
+    "ep_degree": 1,
+    "flash_decoding_enabled": false,
+    "fused_qkv": false,
+    "glu_mlp": true,
+    "is_chunked_prefill": false,
+    "local_ranks_size": 2,
+    "logical_nc_config": 1,
+    "max_batch_size": 1,
+    "max_context_length": 100,
+    "max_topk": 256,
+    "mlp_kernel_enabled": false,
+    "mlp_kernel_fuse_residual_add": false,
+    "n_active_tokens": 100,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "num_cores_per_group": 1,
+    "on_device_sampling": false,
+    "optimum_neuron_version": "0.3.0.dev1",
+    "output_logits": false,
+    "padding_side": "right",
+    "pp_degree": 1,
+    "qk_layernorm": false,
+    "qkv_kernel_enabled": false,
+    "rpl_reduce_dtype": "bfloat16",
+    "sequence_length": 100,
+    "sequence_parallel_enabled": false,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "torch_dtype": "bfloat16",
+    "tp_degree": 2,
+    "vocab_parallel": false
+  },
+  "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-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/mixtral/dacorvo/Mixtral-tiny/3c5f98b57fbf4eed7011.json b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/mixtral/dacorvo/Mixtral-tiny/3c5f98b57fbf4eed7011.json
new file mode 100644
index 0000000000000000000000000000000000000000..bc22200f7902327a730a4cdba88cd1463733255a
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/mixtral/dacorvo/Mixtral-tiny/3c5f98b57fbf4eed7011.json
@@ -0,0 +1,73 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "dacorvo/Mixtral-tiny",
+  "_task": "text-generation",
+  "architectures": [
+    "MixtralForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "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",
+    "async_mode": false,
+    "attn_kernel_enabled": false,
+    "batch_size": 2,
+    "capacity_factor": null,
+    "cc_pipeline_tiling_factor": 2,
+    "checkpoint_id": "dacorvo/Mixtral-tiny",
+    "checkpoint_revision": "c557ba205ddff6ea911f4719e0d543d6c08356b6",
+    "continuous_batching": false,
+    "enable_bucketing": false,
+    "ep_degree": 1,
+    "flash_decoding_enabled": false,
+    "fused_qkv": false,
+    "glu_mlp": true,
+    "is_chunked_prefill": false,
+    "local_ranks_size": 2,
+    "logical_nc_config": 1,
+    "max_batch_size": 2,
+    "max_context_length": 100,
+    "max_topk": 256,
+    "mlp_kernel_enabled": false,
+    "mlp_kernel_fuse_residual_add": false,
+    "n_active_tokens": 100,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "num_cores_per_group": 1,
+    "on_device_sampling": false,
+    "optimum_neuron_version": "0.3.0.dev1",
+    "output_logits": false,
+    "padding_side": "right",
+    "pp_degree": 1,
+    "qk_layernorm": false,
+    "qkv_kernel_enabled": false,
+    "rpl_reduce_dtype": "float16",
+    "sequence_length": 100,
+    "sequence_parallel_enabled": false,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "torch_dtype": "float16",
+    "tp_degree": 2,
+    "vocab_parallel": false
+  },
+  "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-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/mixtral/dacorvo/Mixtral-tiny/e50ed7102c39809e27ac.json b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/mixtral/dacorvo/Mixtral-tiny/e50ed7102c39809e27ac.json
new file mode 100644
index 0000000000000000000000000000000000000000..5d8ded2e0cadb34394415461281d23f65bad2f91
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/mixtral/dacorvo/Mixtral-tiny/e50ed7102c39809e27ac.json
@@ -0,0 +1,73 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "dacorvo/Mixtral-tiny",
+  "_task": "text-generation",
+  "architectures": [
+    "MixtralForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "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",
+    "async_mode": false,
+    "attn_kernel_enabled": false,
+    "batch_size": 1,
+    "capacity_factor": null,
+    "cc_pipeline_tiling_factor": 2,
+    "checkpoint_id": "dacorvo/Mixtral-tiny",
+    "checkpoint_revision": "c557ba205ddff6ea911f4719e0d543d6c08356b6",
+    "continuous_batching": false,
+    "enable_bucketing": false,
+    "ep_degree": 1,
+    "flash_decoding_enabled": false,
+    "fused_qkv": false,
+    "glu_mlp": true,
+    "is_chunked_prefill": false,
+    "local_ranks_size": 2,
+    "logical_nc_config": 1,
+    "max_batch_size": 1,
+    "max_context_length": 100,
+    "max_topk": 256,
+    "mlp_kernel_enabled": false,
+    "mlp_kernel_fuse_residual_add": false,
+    "n_active_tokens": 100,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "num_cores_per_group": 1,
+    "on_device_sampling": false,
+    "optimum_neuron_version": "0.3.0.dev1",
+    "output_logits": false,
+    "padding_side": "right",
+    "pp_degree": 1,
+    "qk_layernorm": false,
+    "qkv_kernel_enabled": false,
+    "rpl_reduce_dtype": "float16",
+    "sequence_length": 100,
+    "sequence_parallel_enabled": false,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "torch_dtype": "float16",
+    "tp_degree": 2,
+    "vocab_parallel": false
+  },
+  "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-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/phi3/yujiepan/phi-4-tiny-random/2ae83bdd0abceabde586.json b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/phi3/yujiepan/phi-4-tiny-random/2ae83bdd0abceabde586.json
new file mode 100644
index 0000000000000000000000000000000000000000..51ecee0d74bf7f4205108309ea56c4e5a4ac9d3d
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/phi3/yujiepan/phi-4-tiny-random/2ae83bdd0abceabde586.json
@@ -0,0 +1,52 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "yujiepan/phi-4-tiny-random",
+  "_task": "text-generation",
+  "architectures": [
+    "Phi3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {},
+  "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": "HloNeuronConfig",
+    "all_reduce_dtype": null,
+    "allow_flash_attention": false,
+    "attention_layout": "HSB",
+    "attn_output_transposed": false,
+    "auto_cast_type": "bf16",
+    "batch_size": 1,
+    "checkpoint_id": "yujiepan/phi-4-tiny-random",
+    "checkpoint_revision": "18a9a1168dc97ac6d128f811925670c275610f5a",
+    "collectives_layout": "HSB",
+    "continuous_batching": false,
+    "fuse_qkv": true,
+    "group_query_attention": "replicated-heads",
+    "log_softmax_scores": false,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "optimum_neuron_version": "0.3.0.dev1",
+    "output_all_logits": false,
+    "sequence_length": 100,
+    "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-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/phi3/yujiepan/phi-4-tiny-random/3ed3625ef80163d27a4c.json b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/phi3/yujiepan/phi-4-tiny-random/3ed3625ef80163d27a4c.json
new file mode 100644
index 0000000000000000000000000000000000000000..0b39e09fdd981bd22b3a2220391784a1cc253c19
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/phi3/yujiepan/phi-4-tiny-random/3ed3625ef80163d27a4c.json
@@ -0,0 +1,52 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "yujiepan/phi-4-tiny-random",
+  "_task": "text-generation",
+  "architectures": [
+    "Phi3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {},
+  "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": "HloNeuronConfig",
+    "all_reduce_dtype": null,
+    "allow_flash_attention": false,
+    "attention_layout": "HSB",
+    "attn_output_transposed": false,
+    "auto_cast_type": "fp16",
+    "batch_size": 2,
+    "checkpoint_id": "yujiepan/phi-4-tiny-random",
+    "checkpoint_revision": "18a9a1168dc97ac6d128f811925670c275610f5a",
+    "collectives_layout": "HSB",
+    "continuous_batching": true,
+    "fuse_qkv": true,
+    "group_query_attention": "replicated-heads",
+    "log_softmax_scores": false,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "optimum_neuron_version": "0.3.0.dev1",
+    "output_all_logits": false,
+    "sequence_length": 100,
+    "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-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/phi3/yujiepan/phi-4-tiny-random/78bb146dc5773156a959.json b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/phi3/yujiepan/phi-4-tiny-random/78bb146dc5773156a959.json
new file mode 100644
index 0000000000000000000000000000000000000000..67faa73f3256763a9e6d2147fd2c2b98966f7c42
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/phi3/yujiepan/phi-4-tiny-random/78bb146dc5773156a959.json
@@ -0,0 +1,52 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "yujiepan/phi-4-tiny-random",
+  "_task": "text-generation",
+  "architectures": [
+    "Phi3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {},
+  "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": "HloNeuronConfig",
+    "all_reduce_dtype": null,
+    "allow_flash_attention": false,
+    "attention_layout": "HSB",
+    "attn_output_transposed": false,
+    "auto_cast_type": "fp16",
+    "batch_size": 1,
+    "checkpoint_id": "yujiepan/phi-4-tiny-random",
+    "checkpoint_revision": "18a9a1168dc97ac6d128f811925670c275610f5a",
+    "collectives_layout": "HSB",
+    "continuous_batching": false,
+    "fuse_qkv": true,
+    "group_query_attention": "replicated-heads",
+    "log_softmax_scores": false,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "optimum_neuron_version": "0.3.0.dev1",
+    "output_all_logits": false,
+    "sequence_length": 100,
+    "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-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/qwen2/yujiepan/qwen2.5-128k-tiny-random/0f369de663b01a949497.json b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/qwen2/yujiepan/qwen2.5-128k-tiny-random/0f369de663b01a949497.json
new file mode 100644
index 0000000000000000000000000000000000000000..625d08dca70ac1b81a75b7fef2f459399a6f6532
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/qwen2/yujiepan/qwen2.5-128k-tiny-random/0f369de663b01a949497.json
@@ -0,0 +1,75 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "yujiepan/qwen2.5-128k-tiny-random",
+  "_task": "text-generation",
+  "architectures": [
+    "Qwen2ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "hidden_act": "silu",
+  "hidden_size": 8,
+  "initializer_range": 0.02,
+  "intermediate_size": 16,
+  "max_position_embeddings": 32768,
+  "max_window_layers": 1,
+  "model_type": "qwen2",
+  "neuron": {
+    "_serialized_key": "NxDNeuronConfig",
+    "async_mode": false,
+    "attn_kernel_enabled": false,
+    "batch_size": 1,
+    "capacity_factor": null,
+    "cc_pipeline_tiling_factor": 2,
+    "checkpoint_id": "yujiepan/qwen2.5-128k-tiny-random",
+    "checkpoint_revision": "c8296d4ca3f87782876d2382fbb6481d1beb8ef0",
+    "continuous_batching": false,
+    "enable_bucketing": false,
+    "ep_degree": 1,
+    "flash_decoding_enabled": false,
+    "fused_qkv": false,
+    "glu_mlp": true,
+    "is_chunked_prefill": false,
+    "local_ranks_size": 2,
+    "logical_nc_config": 1,
+    "max_batch_size": 1,
+    "max_context_length": 100,
+    "max_topk": 256,
+    "mlp_kernel_enabled": false,
+    "mlp_kernel_fuse_residual_add": false,
+    "n_active_tokens": 100,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "num_cores_per_group": 1,
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.3.0.dev1",
+    "output_logits": false,
+    "padding_side": "right",
+    "pp_degree": 1,
+    "qk_layernorm": false,
+    "qkv_kernel_enabled": false,
+    "rpl_reduce_dtype": "float16",
+    "sequence_length": 100,
+    "sequence_parallel_enabled": false,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "torch_dtype": "float16",
+    "tp_degree": 2,
+    "vocab_parallel": false
+  },
+  "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": 131072,
+  "tie_word_embeddings": false,
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 152064
+}
\ No newline at end of file
diff --git a/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/qwen2/yujiepan/qwen2.5-128k-tiny-random/23dbff0523662bd7d6be.json b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/qwen2/yujiepan/qwen2.5-128k-tiny-random/23dbff0523662bd7d6be.json
new file mode 100644
index 0000000000000000000000000000000000000000..45276a6595cd3b61f50b7c8e45f6b662140c0c4d
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/qwen2/yujiepan/qwen2.5-128k-tiny-random/23dbff0523662bd7d6be.json
@@ -0,0 +1,75 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "yujiepan/qwen2.5-128k-tiny-random",
+  "_task": "text-generation",
+  "architectures": [
+    "Qwen2ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "hidden_act": "silu",
+  "hidden_size": 8,
+  "initializer_range": 0.02,
+  "intermediate_size": 16,
+  "max_position_embeddings": 32768,
+  "max_window_layers": 1,
+  "model_type": "qwen2",
+  "neuron": {
+    "_serialized_key": "NxDNeuronConfig",
+    "async_mode": false,
+    "attn_kernel_enabled": false,
+    "batch_size": 1,
+    "capacity_factor": null,
+    "cc_pipeline_tiling_factor": 2,
+    "checkpoint_id": "yujiepan/qwen2.5-128k-tiny-random",
+    "checkpoint_revision": "c8296d4ca3f87782876d2382fbb6481d1beb8ef0",
+    "continuous_batching": false,
+    "enable_bucketing": false,
+    "ep_degree": 1,
+    "flash_decoding_enabled": false,
+    "fused_qkv": false,
+    "glu_mlp": true,
+    "is_chunked_prefill": false,
+    "local_ranks_size": 2,
+    "logical_nc_config": 1,
+    "max_batch_size": 1,
+    "max_context_length": 100,
+    "max_topk": 256,
+    "mlp_kernel_enabled": false,
+    "mlp_kernel_fuse_residual_add": false,
+    "n_active_tokens": 100,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "num_cores_per_group": 1,
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.3.0.dev1",
+    "output_logits": false,
+    "padding_side": "right",
+    "pp_degree": 1,
+    "qk_layernorm": false,
+    "qkv_kernel_enabled": false,
+    "rpl_reduce_dtype": "bfloat16",
+    "sequence_length": 100,
+    "sequence_parallel_enabled": false,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "torch_dtype": "bfloat16",
+    "tp_degree": 2,
+    "vocab_parallel": false
+  },
+  "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": 131072,
+  "tie_word_embeddings": false,
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 152064
+}
\ No newline at end of file
diff --git a/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/qwen2/yujiepan/qwen2.5-128k-tiny-random/d8449f47ba76c9710cb1.json b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/qwen2/yujiepan/qwen2.5-128k-tiny-random/d8449f47ba76c9710cb1.json
new file mode 100644
index 0000000000000000000000000000000000000000..f8e969e38469674a00f91fb87a1df44c82526340
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev1/qwen2/yujiepan/qwen2.5-128k-tiny-random/d8449f47ba76c9710cb1.json
@@ -0,0 +1,75 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "yujiepan/qwen2.5-128k-tiny-random",
+  "_task": "text-generation",
+  "architectures": [
+    "Qwen2ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "hidden_act": "silu",
+  "hidden_size": 8,
+  "initializer_range": 0.02,
+  "intermediate_size": 16,
+  "max_position_embeddings": 32768,
+  "max_window_layers": 1,
+  "model_type": "qwen2",
+  "neuron": {
+    "_serialized_key": "NxDNeuronConfig",
+    "async_mode": false,
+    "attn_kernel_enabled": false,
+    "batch_size": 2,
+    "capacity_factor": null,
+    "cc_pipeline_tiling_factor": 2,
+    "checkpoint_id": "yujiepan/qwen2.5-128k-tiny-random",
+    "checkpoint_revision": "c8296d4ca3f87782876d2382fbb6481d1beb8ef0",
+    "continuous_batching": true,
+    "enable_bucketing": false,
+    "ep_degree": 1,
+    "flash_decoding_enabled": false,
+    "fused_qkv": false,
+    "glu_mlp": true,
+    "is_chunked_prefill": false,
+    "local_ranks_size": 2,
+    "logical_nc_config": 1,
+    "max_batch_size": 2,
+    "max_context_length": 100,
+    "max_topk": 256,
+    "mlp_kernel_enabled": false,
+    "mlp_kernel_fuse_residual_add": false,
+    "n_active_tokens": 100,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "num_cores_per_group": 1,
+    "on_device_sampling": false,
+    "optimum_neuron_version": "0.3.0.dev1",
+    "output_logits": false,
+    "padding_side": "right",
+    "pp_degree": 1,
+    "qk_layernorm": false,
+    "qkv_kernel_enabled": false,
+    "rpl_reduce_dtype": "float16",
+    "sequence_length": 100,
+    "sequence_parallel_enabled": false,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "torch_dtype": "float16",
+    "tp_degree": 2,
+    "vocab_parallel": false
+  },
+  "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": 131072,
+  "tie_word_embeddings": false,
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 152064
+}
\ No newline at end of file
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_4c948ba1d275cea9b124+431f5505/compile_flags.json b/neuronxcc-2.17.194.0+d312836f/MODULE_4c948ba1d275cea9b124+431f5505/compile_flags.json
new file mode 100644
index 0000000000000000000000000000000000000000..9932c70d27d759d781be88485b74b17b49b4c04f
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_4c948ba1d275cea9b124+431f5505/compile_flags.json
@@ -0,0 +1 @@
+"--model-type=transformer -O1 --lnc=1 --internal-hlo2tensorizer-options=--experimental-unsafe-fp8e4m3fn-as-fp8e4m3 --logfile=/tmp/nxd_model/layout_opt/log-neuron-cc.txt"
\ No newline at end of file
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_4c948ba1d275cea9b124+431f5505/model.done b/neuronxcc-2.17.194.0+d312836f/MODULE_4c948ba1d275cea9b124+431f5505/model.done
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_4c948ba1d275cea9b124+431f5505/model.hlo_module.pb b/neuronxcc-2.17.194.0+d312836f/MODULE_4c948ba1d275cea9b124+431f5505/model.hlo_module.pb
new file mode 100644
index 0000000000000000000000000000000000000000..2cb9abfb8e1ee1c3a50d91035b5819017ac4f072
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_4c948ba1d275cea9b124+431f5505/model.hlo_module.pb
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9cce13c2c8d6e621050c4f6d2a392fb09e41dd7942671696a8fe6a6738dfe181
+size 10362
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_4c948ba1d275cea9b124+431f5505/model.neff b/neuronxcc-2.17.194.0+d312836f/MODULE_4c948ba1d275cea9b124+431f5505/model.neff
new file mode 100644
index 0000000000000000000000000000000000000000..8a131482b172f4d5bcdd500e7c733b18c4dab230
Binary files /dev/null and b/neuronxcc-2.17.194.0+d312836f/MODULE_4c948ba1d275cea9b124+431f5505/model.neff differ
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_634439d56d5bb67e9812+165e9558/compile_flags.json b/neuronxcc-2.17.194.0+d312836f/MODULE_634439d56d5bb67e9812+165e9558/compile_flags.json
new file mode 100644
index 0000000000000000000000000000000000000000..e1dd82452b8e16fe8d25db04b67b50fefc4d0215
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_634439d56d5bb67e9812+165e9558/compile_flags.json
@@ -0,0 +1 @@
+"--auto-cast=none --model-type=transformer --tensorizer-options='--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-dge-dma --vectorize-strided-dma ' -O2  --internal-num-neuroncores-per-sengine=1 --logfile=/tmp/nxd_model/token_generation_model/_tp0_bk0/log-neuron-cc.txt --enable-internal-neff-wrapper"
\ No newline at end of file
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_634439d56d5bb67e9812+165e9558/model.done b/neuronxcc-2.17.194.0+d312836f/MODULE_634439d56d5bb67e9812+165e9558/model.done
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_634439d56d5bb67e9812+165e9558/model.hlo_module.pb b/neuronxcc-2.17.194.0+d312836f/MODULE_634439d56d5bb67e9812+165e9558/model.hlo_module.pb
new file mode 100644
index 0000000000000000000000000000000000000000..c1fb56e7082598b8366ecc1e26e3508fca8b84f7
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_634439d56d5bb67e9812+165e9558/model.hlo_module.pb
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ac7b363ad14391ea6e90f95eff1d88a5146cc28af5699413964e7fc4febe0b2
+size 55929
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_634439d56d5bb67e9812+165e9558/model.neff b/neuronxcc-2.17.194.0+d312836f/MODULE_634439d56d5bb67e9812+165e9558/model.neff
new file mode 100644
index 0000000000000000000000000000000000000000..26c75f37364b760d300715130cda08963aed7d99
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_634439d56d5bb67e9812+165e9558/model.neff
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:82e85e7c33a12ba1aded0285eb00c9631d44713a2e29c7e17268dcf5109dce34
+size 2366464
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_634439d56d5bb67e9812+165e9558/wrapped_neff.hlo b/neuronxcc-2.17.194.0+d312836f/MODULE_634439d56d5bb67e9812+165e9558/wrapped_neff.hlo
new file mode 100644
index 0000000000000000000000000000000000000000..f4e89d7da3b61b35b25092ea0c2d9f009c71afa8
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_634439d56d5bb67e9812+165e9558/wrapped_neff.hlo
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4acd67a4467ef3a0cb39e5f4ff6d7234573efbd930e665c2631b471c1a2ab93
+size 2376747
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_72281753d7f57952bfe4+165e9558/compile_flags.json b/neuronxcc-2.17.194.0+d312836f/MODULE_72281753d7f57952bfe4+165e9558/compile_flags.json
new file mode 100644
index 0000000000000000000000000000000000000000..e1dd82452b8e16fe8d25db04b67b50fefc4d0215
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_72281753d7f57952bfe4+165e9558/compile_flags.json
@@ -0,0 +1 @@
+"--auto-cast=none --model-type=transformer --tensorizer-options='--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-dge-dma --vectorize-strided-dma ' -O2  --internal-num-neuroncores-per-sengine=1 --logfile=/tmp/nxd_model/token_generation_model/_tp0_bk0/log-neuron-cc.txt --enable-internal-neff-wrapper"
\ No newline at end of file
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_72281753d7f57952bfe4+165e9558/model.done b/neuronxcc-2.17.194.0+d312836f/MODULE_72281753d7f57952bfe4+165e9558/model.done
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_72281753d7f57952bfe4+165e9558/model.hlo_module.pb b/neuronxcc-2.17.194.0+d312836f/MODULE_72281753d7f57952bfe4+165e9558/model.hlo_module.pb
new file mode 100644
index 0000000000000000000000000000000000000000..cee68abee60c0300cdc0f7363612aae53fb136c5
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_72281753d7f57952bfe4+165e9558/model.hlo_module.pb
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3de47ddfe9c2dd2dd851b262ed8e72f81d7bd5088bd4c866e33eccb1235c731
+size 88686
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_72281753d7f57952bfe4+165e9558/model.neff b/neuronxcc-2.17.194.0+d312836f/MODULE_72281753d7f57952bfe4+165e9558/model.neff
new file mode 100644
index 0000000000000000000000000000000000000000..d714cc233ea87d3373db30b3e4342fb76028267e
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_72281753d7f57952bfe4+165e9558/model.neff
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:670eb5f6e43393459448f206825777f42128ddb80a24d3859cee614b17f37054
+size 236544
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_72281753d7f57952bfe4+165e9558/wrapped_neff.hlo b/neuronxcc-2.17.194.0+d312836f/MODULE_72281753d7f57952bfe4+165e9558/wrapped_neff.hlo
new file mode 100644
index 0000000000000000000000000000000000000000..2d298b31a325d9db281b468fc86d636891ed5277
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_72281753d7f57952bfe4+165e9558/wrapped_neff.hlo
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e79163920afe05ec7bb4285c4a6cc84559fbbf3b4fde58c15eccb5c8001b019c
+size 247153
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_857bfa5db7b34c9a9421+165e9558/compile_flags.json b/neuronxcc-2.17.194.0+d312836f/MODULE_857bfa5db7b34c9a9421+165e9558/compile_flags.json
new file mode 100644
index 0000000000000000000000000000000000000000..e1dd82452b8e16fe8d25db04b67b50fefc4d0215
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_857bfa5db7b34c9a9421+165e9558/compile_flags.json
@@ -0,0 +1 @@
+"--auto-cast=none --model-type=transformer --tensorizer-options='--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-dge-dma --vectorize-strided-dma ' -O2  --internal-num-neuroncores-per-sengine=1 --logfile=/tmp/nxd_model/token_generation_model/_tp0_bk0/log-neuron-cc.txt --enable-internal-neff-wrapper"
\ No newline at end of file
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_857bfa5db7b34c9a9421+165e9558/model.done b/neuronxcc-2.17.194.0+d312836f/MODULE_857bfa5db7b34c9a9421+165e9558/model.done
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_857bfa5db7b34c9a9421+165e9558/model.hlo_module.pb b/neuronxcc-2.17.194.0+d312836f/MODULE_857bfa5db7b34c9a9421+165e9558/model.hlo_module.pb
new file mode 100644
index 0000000000000000000000000000000000000000..3ca90971c7cb3ce9400fca9bcac1cb70f8bc0979
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_857bfa5db7b34c9a9421+165e9558/model.hlo_module.pb
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ef04c56dfa88ba3744238690dd27414699ab135700c99f5f366574da1704460
+size 88686
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_857bfa5db7b34c9a9421+165e9558/model.neff b/neuronxcc-2.17.194.0+d312836f/MODULE_857bfa5db7b34c9a9421+165e9558/model.neff
new file mode 100644
index 0000000000000000000000000000000000000000..07fa907695665cdeecb55ac97dbf2ce42d5b2634
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_857bfa5db7b34c9a9421+165e9558/model.neff
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b96ecdcfc5ff6f5c8bb2c6c748d4dc4fea4c4013f221ee4cbce75c1367605cc
+size 236544
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_857bfa5db7b34c9a9421+165e9558/wrapped_neff.hlo b/neuronxcc-2.17.194.0+d312836f/MODULE_857bfa5db7b34c9a9421+165e9558/wrapped_neff.hlo
new file mode 100644
index 0000000000000000000000000000000000000000..3e7bd1f1570a3261c3e5a4f83bfbb316c9dd5358
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_857bfa5db7b34c9a9421+165e9558/wrapped_neff.hlo
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a96e164632fabab7abd4e40b82d9ffddbb90bbb3261ddde437c1e7bebddcb01
+size 247153
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_94ba84f59b341a98c066+bfe5714b/compile_flags.json b/neuronxcc-2.17.194.0+d312836f/MODULE_94ba84f59b341a98c066+bfe5714b/compile_flags.json
new file mode 100644
index 0000000000000000000000000000000000000000..dcde6c3978ed8b79ae83ebd29a4874956f871e22
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_94ba84f59b341a98c066+bfe5714b/compile_flags.json
@@ -0,0 +1 @@
+"--auto-cast=none --model-type=transformer --tensorizer-options='--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-dge-dma --vectorize-strided-dma ' -O2  --internal-num-neuroncores-per-sengine=1 --logfile=/tmp/nxd_model/context_encoding_model/_tp0_bk0/log-neuron-cc.txt"
\ No newline at end of file
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_94ba84f59b341a98c066+bfe5714b/model.done b/neuronxcc-2.17.194.0+d312836f/MODULE_94ba84f59b341a98c066+bfe5714b/model.done
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_94ba84f59b341a98c066+bfe5714b/model.hlo_module.pb b/neuronxcc-2.17.194.0+d312836f/MODULE_94ba84f59b341a98c066+bfe5714b/model.hlo_module.pb
new file mode 100644
index 0000000000000000000000000000000000000000..d71f3b29f54c7c04856a9c56a09ce9e0b18d7373
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_94ba84f59b341a98c066+bfe5714b/model.hlo_module.pb
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ba65d5335646539f3aa28ede06c54054cea89d33efa6900bb95e22d5e9aeb0f
+size 53519
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_94ba84f59b341a98c066+bfe5714b/model.neff b/neuronxcc-2.17.194.0+d312836f/MODULE_94ba84f59b341a98c066+bfe5714b/model.neff
new file mode 100644
index 0000000000000000000000000000000000000000..61226b4a0a42b21eb979000d4f07fd00ffa9323b
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_94ba84f59b341a98c066+bfe5714b/model.neff
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39b3c90fdb13ea980144897c0236d2d8432d49c31695db60316c9665c66bcd39
+size 164864
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_9c37025554da49ae454e+431f5505/compile_flags.json b/neuronxcc-2.17.194.0+d312836f/MODULE_9c37025554da49ae454e+431f5505/compile_flags.json
new file mode 100644
index 0000000000000000000000000000000000000000..9932c70d27d759d781be88485b74b17b49b4c04f
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_9c37025554da49ae454e+431f5505/compile_flags.json
@@ -0,0 +1 @@
+"--model-type=transformer -O1 --lnc=1 --internal-hlo2tensorizer-options=--experimental-unsafe-fp8e4m3fn-as-fp8e4m3 --logfile=/tmp/nxd_model/layout_opt/log-neuron-cc.txt"
\ No newline at end of file
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_9c37025554da49ae454e+431f5505/model.done b/neuronxcc-2.17.194.0+d312836f/MODULE_9c37025554da49ae454e+431f5505/model.done
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_9c37025554da49ae454e+431f5505/model.hlo_module.pb b/neuronxcc-2.17.194.0+d312836f/MODULE_9c37025554da49ae454e+431f5505/model.hlo_module.pb
new file mode 100644
index 0000000000000000000000000000000000000000..170e46b40ac0b56c809e91059c3be6e81be09164
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_9c37025554da49ae454e+431f5505/model.hlo_module.pb
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0116aa51c873801bce62eab5b30eedbeb02a12112baaecad8ff8960d50e370f
+size 10013
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_9c37025554da49ae454e+431f5505/model.neff b/neuronxcc-2.17.194.0+d312836f/MODULE_9c37025554da49ae454e+431f5505/model.neff
new file mode 100644
index 0000000000000000000000000000000000000000..02f2a5b21b9a8d34f2f484c9e3b6eadd815239f4
Binary files /dev/null and b/neuronxcc-2.17.194.0+d312836f/MODULE_9c37025554da49ae454e+431f5505/model.neff differ
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_cbd29f12a931a6034fdc+431f5505/compile_flags.json b/neuronxcc-2.17.194.0+d312836f/MODULE_cbd29f12a931a6034fdc+431f5505/compile_flags.json
new file mode 100644
index 0000000000000000000000000000000000000000..9932c70d27d759d781be88485b74b17b49b4c04f
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_cbd29f12a931a6034fdc+431f5505/compile_flags.json
@@ -0,0 +1 @@
+"--model-type=transformer -O1 --lnc=1 --internal-hlo2tensorizer-options=--experimental-unsafe-fp8e4m3fn-as-fp8e4m3 --logfile=/tmp/nxd_model/layout_opt/log-neuron-cc.txt"
\ No newline at end of file
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_cbd29f12a931a6034fdc+431f5505/model.done b/neuronxcc-2.17.194.0+d312836f/MODULE_cbd29f12a931a6034fdc+431f5505/model.done
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_cbd29f12a931a6034fdc+431f5505/model.hlo_module.pb b/neuronxcc-2.17.194.0+d312836f/MODULE_cbd29f12a931a6034fdc+431f5505/model.hlo_module.pb
new file mode 100644
index 0000000000000000000000000000000000000000..22dcf0f8c79ea418cd5ca2568a4ae6f6315ccc56
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_cbd29f12a931a6034fdc+431f5505/model.hlo_module.pb
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8919938a040060bd5077eef2dfa9d7707e8483b155a3cc2449f75b6e5ccf99f7
+size 10362
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_cbd29f12a931a6034fdc+431f5505/model.neff b/neuronxcc-2.17.194.0+d312836f/MODULE_cbd29f12a931a6034fdc+431f5505/model.neff
new file mode 100644
index 0000000000000000000000000000000000000000..73a459f32b950bf6b4ddb88c31cdf7d77951cc00
Binary files /dev/null and b/neuronxcc-2.17.194.0+d312836f/MODULE_cbd29f12a931a6034fdc+431f5505/model.neff differ
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_d3986b7b5efee5f065f4+bfe5714b/compile_flags.json b/neuronxcc-2.17.194.0+d312836f/MODULE_d3986b7b5efee5f065f4+bfe5714b/compile_flags.json
new file mode 100644
index 0000000000000000000000000000000000000000..dcde6c3978ed8b79ae83ebd29a4874956f871e22
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_d3986b7b5efee5f065f4+bfe5714b/compile_flags.json
@@ -0,0 +1 @@
+"--auto-cast=none --model-type=transformer --tensorizer-options='--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-dge-dma --vectorize-strided-dma ' -O2  --internal-num-neuroncores-per-sengine=1 --logfile=/tmp/nxd_model/context_encoding_model/_tp0_bk0/log-neuron-cc.txt"
\ No newline at end of file
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_d3986b7b5efee5f065f4+bfe5714b/model.done b/neuronxcc-2.17.194.0+d312836f/MODULE_d3986b7b5efee5f065f4+bfe5714b/model.done
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_d3986b7b5efee5f065f4+bfe5714b/model.hlo_module.pb b/neuronxcc-2.17.194.0+d312836f/MODULE_d3986b7b5efee5f065f4+bfe5714b/model.hlo_module.pb
new file mode 100644
index 0000000000000000000000000000000000000000..61f4fa8fa48ba12cb1cb5e77fb6d5ce569a19424
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_d3986b7b5efee5f065f4+bfe5714b/model.hlo_module.pb
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2d3490ad3f154cf17e30d467065e0d249bfda6d6ad064598f06cefb0c4ca4072
+size 87293
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_d3986b7b5efee5f065f4+bfe5714b/model.neff b/neuronxcc-2.17.194.0+d312836f/MODULE_d3986b7b5efee5f065f4+bfe5714b/model.neff
new file mode 100644
index 0000000000000000000000000000000000000000..a19bcbbb0e4b6ec579b7ed7dc97048bff8e8b442
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_d3986b7b5efee5f065f4+bfe5714b/model.neff
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1276d19ea2012e1c74fbd66b9b7fb2c4f116c5d75822dc9365bda22b1d6bac7e
+size 246784
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_f84068611c70eb3c41a2+bfe5714b/compile_flags.json b/neuronxcc-2.17.194.0+d312836f/MODULE_f84068611c70eb3c41a2+bfe5714b/compile_flags.json
new file mode 100644
index 0000000000000000000000000000000000000000..dcde6c3978ed8b79ae83ebd29a4874956f871e22
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_f84068611c70eb3c41a2+bfe5714b/compile_flags.json
@@ -0,0 +1 @@
+"--auto-cast=none --model-type=transformer --tensorizer-options='--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-dge-dma --vectorize-strided-dma ' -O2  --internal-num-neuroncores-per-sengine=1 --logfile=/tmp/nxd_model/context_encoding_model/_tp0_bk0/log-neuron-cc.txt"
\ No newline at end of file
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_f84068611c70eb3c41a2+bfe5714b/model.done b/neuronxcc-2.17.194.0+d312836f/MODULE_f84068611c70eb3c41a2+bfe5714b/model.done
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_f84068611c70eb3c41a2+bfe5714b/model.hlo_module.pb b/neuronxcc-2.17.194.0+d312836f/MODULE_f84068611c70eb3c41a2+bfe5714b/model.hlo_module.pb
new file mode 100644
index 0000000000000000000000000000000000000000..351d209f0561d87d2ac91522986a3bc10e886f94
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_f84068611c70eb3c41a2+bfe5714b/model.hlo_module.pb
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b4aeec679a69511a8128b7836f1aa52b284c380397507732dcacb0e8d540385
+size 87293
diff --git a/neuronxcc-2.17.194.0+d312836f/MODULE_f84068611c70eb3c41a2+bfe5714b/model.neff b/neuronxcc-2.17.194.0+d312836f/MODULE_f84068611c70eb3c41a2+bfe5714b/model.neff
new file mode 100644
index 0000000000000000000000000000000000000000..203864cade35b10c1fcd74c241aa4e5960548946
--- /dev/null
+++ b/neuronxcc-2.17.194.0+d312836f/MODULE_f84068611c70eb3c41a2+bfe5714b/model.neff
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:753350d9678f5d9a3a7a205fed3e4aa3f6a0cb0273310fc36b07edf138c09299
+size 246784