Upload 17 files

config.json

@@ -0,0 +1,325 @@
+{
+  "_name_or_path": "RLinf/RLinf-OpenVLAOFT-GRPO-LIBERO-90",
+  "action_dim": 7,
+  "add_bias_linear": false,
+  "add_qkv_bias": true,
+  "arch_specifier": "no-align+fused-gelu-mlp",
+  "architectures": [
+    "OpenVLAOFTForRLActionPrediction"
+  ],
+  "attn_implementation": "flash_attention_2",
+  "auto_map": {
+    "AutoConfig": "configuration_prismatic.OpenVLAConfig",
+    "AutoModelForVision2Seq": "modeling_prismatic.OpenVLAForActionPrediction"
+  },
+  "center_crop": true,
+  "ckpt_path": "/mnt/public/mjwei/repo/RLinf-0915/logs/20250930-17:48:12/checkpoints/global_step_50/actor/model.pt",
+  "hf_llm_id": "meta-llama/Llama-2-7b-hf",
+  "hidden_size": 4096,
+  "image_resize_strategy": "resize-naive",
+  "image_size": [
+    224,
+    224
+  ],
+  "image_sizes": [
+    224,
+    224
+  ],
+  "is_lora": false,
+  "llm_backbone_id": "llama2-7b-pure",
+  "llm_max_length": 2048,
+  "lora_rank": 32,
+  "low_cpu_mem_usage": true,
+  "model_name": "openvla_oft",
+  "model_type": "openvla",
+  "n_action_bins": 256,
+  "norm_stats": {
+    "libero_10": {
+      "action": {
+        "mask": [
+          true,
+          true,
+          true,
+          true,
+          true,
+          true,
+          false
+        ],
+        "max": [
+          0.9375,
+          0.9375,
+          0.9375,
+          0.30000001192092896,
+          0.29357144236564636,
+          0.375,
+          1.0
+        ],
+        "mean": [
+          0.01820324920117855,
+          0.05858374014496803,
+          -0.05592384561896324,
+          0.004626928828656673,
+          0.00289608770981431,
+          -0.007673131301999092,
+          0.5457824468612671
+        ],
+        "min": [
+          -0.9375,
+          -0.9375,
+          -0.9375,
+          -0.23642857372760773,
+          -0.3053571283817291,
+          -0.3675000071525574,
+          0.0
+        ],
+        "q01": [
+          -0.6348214149475098,
+          -0.7741071581840515,
+          -0.7633928656578064,
+          -0.09749999642372131,
+          -0.14819999992847435,
+          -0.2742857038974762,
+          0.0
+        ],
+        "q99": [
+          0.7714285850524902,
+          0.8464285731315613,
+          0.9375,
+          0.13928571343421936,
+          0.15964286029338837,
+          0.3246428668498993,
+          1.0
+        ],
+        "std": [
+          0.2825464606285095,
+          0.35904666781425476,
+          0.3673802614212036,
+          0.03770702704787254,
+          0.05429719388484955,
+          0.08725254982709885,
+          0.49815231561660767
+        ]
+      },
+      "num_trajectories": 379,
+      "num_transitions": 101469,
+      "proprio": {
+        "max": [
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0
+        ],
+        "mean": [
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0
+        ],
+        "min": [
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0
+        ],
+        "q01": [
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0
+        ],
+        "q99": [
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0
+        ],
+        "std": [
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0
+        ]
+      }
+    },
+    "libero_90_no_noops_trajall": {
+      "action": {
+        "mask": [
+          true,
+          true,
+          true,
+          true,
+          true,
+          true,
+          false
+        ],
+        "max": [
+          0.9375,
+          0.9375,
+          0.9375,
+          0.375,
+          0.375,
+          0.375,
+          1.0
+        ],
+        "mean": [
+          0.04552333056926727,
+          0.037328675389289856,
+          -0.09673094749450684,
+          0.00501923356205225,
+          0.0022719360422343016,
+          -0.006229684222489595,
+          0.5282046794891357
+        ],
+        "min": [
+          -0.9375,
+          -0.9375,
+          -0.9375,
+          -0.3257142901420593,
+          -0.375,
+          -0.375,
+          0.0
+        ],
+        "q01": [
+          -0.6294642686843872,
+          -0.8705357313156128,
+          -0.8946428298950195,
+          -0.12321428209543228,
+          -0.1574999988079071,
+          -0.2775000035762787,
+          0.0
+        ],
+        "q99": [
+          0.8517857193946838,
+          0.8464285731315613,
+          0.9375,
+          0.1875,
+          0.1778571456670761,
+          0.3471428453922272,
+          1.0
+        ],
+        "std": [
+          0.2984432280063629,
+          0.3612210154533386,
+          0.40674421191215515,
+          0.04839186742901802,
+          0.05818791687488556,
+          0.08691468834877014,
+          0.4985364079475403
+        ]
+      },
+      "num_trajectories": 3954,
+      "num_transitions": 573965,
+      "proprio": {
+        "max": [
+          0.20274034142494202,
+          0.4884968400001526,
+          1.3584461212158203,
+          4.8432722091674805,
+          3.966320753097534,
+          2.4007365703582764,
+          0.04637677222490311,
+          0.0017036759527400136
+        ],
+        "mean": [
+          -0.08226173371076584,
+          0.010916306637227535,
+          0.9453046917915344,
+          2.9744503498077393,
+          -0.11405276507139206,
+          -0.09964624792337418,
+          0.026635831221938133,
+          -0.027010178193449974
+        ],
+        "min": [
+          -0.48259806632995605,
+          -0.3968846797943115,
+          0.4455491006374359,
+          -0.7501075863838196,
+          -4.363162040710449,
+          -3.2127554416656494,
+          -0.002592125441879034,
+          -0.04256961867213249
+        ],
+        "q01": [
+          -0.4019535529613495,
+          -0.2819894528388977,
+          0.458499813079834,
+          1.229066481590271,
+          -2.779330949783325,
+          -1.3500228834152221,
+          0.0016688233194872737,
+          -0.04004087835550308
+        ],
+        "q99": [
+          0.12681280374526968,
+          0.3188697147369384,
+          1.2563055849075317,
+          3.8263492584228516,
+          2.3427903938293455,
+          0.6062234616279595,
+          0.04022635221481323,
+          -0.0016752025950700054
+        ],
+        "std": [
+          0.11324016749858856,
+          0.14199486374855042,
+          0.23618334531784058,
+          0.43265825510025024,
+          0.9901652336120605,
+          0.32449689507484436,
+          0.014563564211130142,
+          0.01443739328533411
+        ]
+      }
+    }
+  },
+  "num_action_chunks": 8,
+  "num_images_in_input": 1,
+  "output_projector_states": false,
+  "pad_to_multiple_of": 64,
+  "pad_token_id": 32000,
+  "policy_setup": "widowx_bridge",
+  "precision": "bf16",
+  "text_config": {
+    "model_type": "llama",
+    "pad_token_id": 32000,
+    "torch_dtype": "bfloat16",
+    "vocab_size": 32064
+  },
+  "timm_model_ids": [
+    "vit_large_patch14_reg4_dinov2.lvd142m",
+    "vit_so400m_patch14_siglip_224"
+  ],
+  "timm_override_act_layers": [
+    null,
+    null
+  ],
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.40.1",
+  "trust_remote_code": true,
+  "unnorm_key": "libero_90_no_noops_trajall",
+  "use_fused_vision_backbone": true,
+  "use_proprio": false,
+  "value_type": "step_level",
+  "vh_mode": "a0",
+  "vision_backbone_id": "dinosiglip-vit-so-224px",
+  "vocab_size": 32000
+}

configuration_prismatic.py

@@ -0,0 +1,140 @@
+"""
+configuration_prismatic.py
+
+HuggingFace-style configuration definition for Prismatic VLMs, inheriting from `transformers.PretrainedConfig`.
+Default configuration specifies `siglip-224px+7b`.
+"""
+
+from typing import Any, Dict, List, Optional
+
+from transformers import PretrainedConfig
+from transformers.models.auto import CONFIG_MAPPING
+
+# === Utilities for Mapping Prismatic names to HF names ===
+# fmt: off
+VISION_BACKBONE_TO_RESOLUTION: Dict[str, List[int]] = {
+    "clip-vit-l": [224], "siglip-vit-so400m": [224], "dinov2-vit-l": [224], "in1k-vit-l": [224],
+
+    "clip-vit-l-336px": [336],
+    "siglip-vit-so400m-384px": [384],
+
+    "dinoclip-vit-l-336px": [336, 336],
+    "dinosiglip-vit-so-224px": [224, 224],
+    "dinosiglip-vit-so-384px": [384, 384],
+}
+VISION_BACKBONE_TO_TIMM_ID: Dict[str, List[str]] = {
+    "clip-vit-l": ["vit_large_patch14_clip_224.openai"],
+    "clip-vit-l-336px": ["vit_large_patch14_clip_336.openai"],
+
+    "dinov2-vit-l": ["vit_large_patch14_reg4_dinov2.lvd142m"],
+    "in1k-vit-l": ["vit_large_patch16_224.augreg_in21k_ft_in1k"],
+
+    "siglip-vit-so400m": ["vit_so400m_patch14_siglip_224"],
+    "siglip-vit-so400m-384px": ["vit_so400m_patch14_siglip_384"],
+
+    "dinoclip-vit-l-336px": ["vit_large_patch14_reg4_dinov2.lvd142m", "vit_large_patch14_clip_336.openai"],
+    "dinosiglip-vit-so-224px": ["vit_large_patch14_reg4_dinov2.lvd142m", "vit_so400m_patch14_siglip_224"],
+    "dinosiglip-vit-so-384px": ["vit_large_patch14_reg4_dinov2.lvd142m", "vit_so400m_patch14_siglip_384"],
+}
+TIMM_OVERRIDE_ACT_LAYER: Dict[str, List[Optional[str]]] = {
+    "clip-vit-l": ["quick_gelu"], "clip-vit-l-336px": ["quick_gelu"],
+    "dinov2-vit-l": [None], "in1k-vit-l": [None],
+    "siglip-vit-so400m": [None], "siglip-vit-so400m-384px": [None],
+    "dinoclip-vit-l-336px": [None, "quick_gelu"],
+    "dinosiglip-vit-so-224px": [None, None], "dinosiglip-vit-so-384px": [None, None]
+}
+
+LLM_BACKBONE_TO_HF_PATH = {
+    "llama2-7b-pure": "meta-llama/Llama-2-7b-hf", "llama2-13b-pure": "meta-llama/Llama-2-13b-hf",
+    "llama2-7b-chat": "meta-llama/Llama-2-7b-chat-hf", "llama2-13b-chat": "meta-llama/Llama-2-13b-chat-hf",
+
+    "vicuna-v15-7b": "lmsys/vicuna-7b-v1.5", "vicuna-v15-13b": "lmsys/vicuna-13b-v1.5",
+
+    "mistral-v0.1-7b-pure": "mistralai/Mistral-7B-v0.1",
+    "mistral-v0.1-7b-instruct": "mistralai/Mistral-7B-Instruct-v0.1",
+
+    "phi-2-3b": "microsoft/phi-2",
+}
+LLM_BACKBONE_TO_HF_METACLASS = {
+    "llama2-7b-pure": "llama", "llama2-13b-pure": "llama", "llama2-7b-chat": "llama", "llama2-13b-chat": "llama",
+    "vicuna-v15-7b": "llama", "vicuna-v15-13b": "llama",
+
+    "mistral-v0.1-7b-pure": "mistral", "mistral-v0.1-7b-instruct": "mistral",
+
+    "phi-2-3b": "phi",
+}
+
+VALID_VISION_BACKBONES = set(VISION_BACKBONE_TO_RESOLUTION.keys())
+VALID_LLM_BACKBONES = set(LLM_BACKBONE_TO_HF_PATH)
+# fmt: on
+
+
+class PrismaticConfig(PretrainedConfig):
+    model_type: str = "prismatic"
+    is_composition: bool = False
+
+    def __init__(
+        self,
+        vision_backbone_id: str = "siglip-vit-so400m",
+        llm_backbone_id: str = "vicuna-v15-7b",
+        arch_specifier: str = "no-align+gelu-mlp",
+        use_fused_vision_backbone: Optional[bool] = None,
+        image_resize_strategy: str = "letterbox",
+        text_config: Optional[Dict[str, Any]] = None,
+        llm_max_length: int = 2048,
+        pad_token_id: int = 32000,
+        pad_to_multiple_of: int = 64,
+        output_projector_states: bool = False,
+        **kwargs: str,
+    ) -> None:
+        if vision_backbone_id not in VALID_VISION_BACKBONES:
+            raise ValueError(f"Vision backbone `{vision_backbone_id}` not in {VALID_VISION_BACKBONES = }")
+
+        if llm_backbone_id not in VALID_LLM_BACKBONES:
+            raise ValueError(f"LLM backbone `{llm_backbone_id}` not in {VALID_LLM_BACKBONES = }")
+
+        # Set Prismatic Configuration Fields
+        self.vision_backbone_id = vision_backbone_id
+        self.llm_backbone_id = llm_backbone_id
+        self.arch_specifier = arch_specifier
+        self.output_projector_states = output_projector_states
+
+        # [Contract] All vision backbone parameters are lists =>> supports fused backbones with different preprocessing
+        self.use_fused_vision_backbone = (
+            use_fused_vision_backbone
+            if use_fused_vision_backbone is not None
+            else any(self.vision_backbone_id.startswith(v) for v in ["dinoclip", "dinosiglip"])
+        )
+
+        self.timm_model_ids = VISION_BACKBONE_TO_TIMM_ID[self.vision_backbone_id]
+        self.timm_override_act_layers = TIMM_OVERRIDE_ACT_LAYER[self.vision_backbone_id]
+        self.image_sizes = VISION_BACKBONE_TO_RESOLUTION[self.vision_backbone_id]
+        self.image_resize_strategy = image_resize_strategy
+
+        self.hf_llm_id = LLM_BACKBONE_TO_HF_PATH[self.llm_backbone_id]
+        self.llm_max_length = llm_max_length
+        self.pad_token_id, self.pad_to_multiple_of = pad_token_id, pad_to_multiple_of
+
+        # [IMPORTANT] HF Utilities actually look for a `text_config` field... we need to use that specific naming!
+        self.text_config = (
+            CONFIG_MAPPING[LLM_BACKBONE_TO_HF_METACLASS[self.llm_backbone_id]](**text_config)
+            if text_config is not None
+            else CONFIG_MAPPING[LLM_BACKBONE_TO_HF_METACLASS[self.llm_backbone_id]]()
+        )
+
+        # Dispatch **kwargs to super() =>> note that `pad_token_id` collides, so we pass it in here as well...
+        super().__init__(pad_token_id=pad_token_id, **kwargs)
+
+
+class OpenVLAConfig(PrismaticConfig):
+    model_type: str = "openvla"
+
+    def __init__(
+        self,
+        norm_stats: Optional[Dict[str, Dict[str, Dict[str, Dict[str, List[float]]]]]] = None,
+        n_action_bins: int = 256,
+        **kwargs: str,
+    ) -> None:
+        self.norm_stats, self.n_action_bins = norm_stats, n_action_bins
+
+        super().__init__(**kwargs)

dataset_statistics.json

@@ -0,0 +1,133 @@
+{
+  "libero_90_no_noops_trajall": {
+    "action": {
+      "mean": [
+        0.04552333056926727,
+        0.037328675389289856,
+        -0.09673094749450684,
+        0.00501923356205225,
+        0.0022719360422343016,
+        -0.006229684222489595,
+        0.5282046794891357
+      ],
+      "std": [
+        0.2984432280063629,
+        0.3612210154533386,
+        0.40674421191215515,
+        0.04839186742901802,
+        0.05818791687488556,
+        0.08691468834877014,
+        0.4985364079475403
+      ],
+      "max": [
+        0.9375,
+        0.9375,
+        0.9375,
+        0.375,
+        0.375,
+        0.375,
+        1.0
+      ],
+      "min": [
+        -0.9375,
+        -0.9375,
+        -0.9375,
+        -0.3257142901420593,
+        -0.375,
+        -0.375,
+        0.0
+      ],
+      "q01": [
+        -0.6294642686843872,
+        -0.8705357313156128,
+        -0.8946428298950195,
+        -0.12321428209543228,
+        -0.1574999988079071,
+        -0.2775000035762787,
+        0.0
+      ],
+      "q99": [
+        0.8517857193946838,
+        0.8464285731315613,
+        0.9375,
+        0.1875,
+        0.1778571456670761,
+        0.3471428453922272,
+        1.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        false
+      ]
+    },
+    "proprio": {
+      "mean": [
+        -0.08226173371076584,
+        0.010916306637227535,
+        0.9453046917915344,
+        2.9744503498077393,
+        -0.11405276507139206,
+        -0.09964624792337418,
+        0.026635831221938133,
+        -0.027010178193449974
+      ],
+      "std": [
+        0.11324016749858856,
+        0.14199486374855042,
+        0.23618334531784058,
+        0.43265825510025024,
+        0.9901652336120605,
+        0.32449689507484436,
+        0.014563564211130142,
+        0.01443739328533411
+      ],
+      "max": [
+        0.20274034142494202,
+        0.4884968400001526,
+        1.3584461212158203,
+        4.8432722091674805,
+        3.966320753097534,
+        2.4007365703582764,
+        0.04637677222490311,
+        0.0017036759527400136
+      ],
+      "min": [
+        -0.48259806632995605,
+        -0.3968846797943115,
+        0.4455491006374359,
+        -0.7501075863838196,
+        -4.363162040710449,
+        -3.2127554416656494,
+        -0.002592125441879034,
+        -0.04256961867213249
+      ],
+      "q01": [
+        -0.4019535529613495,
+        -0.2819894528388977,
+        0.458499813079834,
+        1.229066481590271,
+        -2.779330949783325,
+        -1.3500228834152221,
+        0.0016688233194872737,
+        -0.04004087835550308
+      ],
+      "q99": [
+        0.12681280374526968,
+        0.3188697147369384,
+        1.2563055849075317,
+        3.8263492584228516,
+        2.3427903938293455,
+        0.6062234616279595,
+        0.04022635221481323,
+        -0.0016752025950700054
+      ]
+    },
+    "num_transitions": 573965,
+    "num_trajectories": 3954
+  }
+}

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "pad_token_id": 32000,
+  "transformers_version": "4.40.1"
+}

model-00001-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f3e26bb9c109bf8f909daf13f0b08a662184036164b0cecd7b8b20812ec90b5
+size 4925122448

model-00002-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:49c51285a710d9dac42759fbce1a4cfebd5fbf6e226d0893614edb12e6fd43db
+size 4947392496

model-00003-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4e35784711a9ccf5b12164757d3466114726064ba208241d03f4fa3ed6465fbe
+size 4947417456

model-00004-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e21b4736ecd3c0993b7614a54085d47588ceb876044e7537c3a2795169694981
+size 266997624

model.safetensors.index.json

@@ -0,0 +1,994 @@
+{
+  "metadata": {
+    "total_size": 15086803072
+  },
+  "weight_map": {
+    "language_model.lm_head.weight": "model-00004-of-00004.safetensors",
+    "language_model.model.embed_tokens.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.0.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.0.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.0.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.0.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.0.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.0.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.0.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.0.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.0.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.1.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.1.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.1.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.1.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.1.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.1.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.1.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.1.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.1.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.10.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.10.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.10.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.10.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.10.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.10.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.10.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.10.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.10.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.11.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.11.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.11.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.11.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.11.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.11.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.11.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.11.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.11.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.12.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.12.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.12.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.12.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.12.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.12.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.12.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.12.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.12.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.13.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.13.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.13.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.13.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.13.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.13.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.13.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.13.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.13.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.14.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.14.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.14.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.14.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.14.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.14.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.14.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.14.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.14.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.15.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.15.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.15.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.15.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.15.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.15.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.15.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.15.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.15.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.16.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.16.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.16.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.16.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.16.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.16.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.16.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.16.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.16.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.17.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.17.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.17.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.17.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.17.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.17.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.17.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.17.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.17.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.18.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.18.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.18.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.18.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.18.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.18.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.18.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.18.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.18.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.19.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.19.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.19.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.19.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.19.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.19.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.19.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.19.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.19.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.2.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.2.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.2.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.2.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.2.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.2.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.2.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.2.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.2.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.20.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.20.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.20.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.20.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.20.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.20.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.20.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.20.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.20.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.21.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.21.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.21.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.21.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.21.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.21.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.21.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.21.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.21.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.22.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.22.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.22.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.22.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.22.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.22.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.22.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.22.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.22.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.23.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.23.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.23.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.23.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.23.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.23.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.23.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.23.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.23.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.24.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.24.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.24.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.24.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.24.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.24.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.24.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.24.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.24.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.25.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.25.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.25.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.25.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.25.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.25.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.25.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.25.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.25.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.26.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.26.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.26.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.26.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.26.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.26.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.26.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.26.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.26.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.27.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.27.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.27.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.27.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.27.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.27.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.27.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.27.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.27.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.28.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.28.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.28.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.28.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.28.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.28.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.28.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.28.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.28.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.29.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.29.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.29.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.29.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.29.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.29.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.29.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.29.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.29.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.3.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.3.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.3.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.3.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.3.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.3.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.3.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.3.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.3.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.30.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.30.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.30.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.30.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.30.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.30.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.30.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.30.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.30.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.31.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.31.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.31.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.31.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.31.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.31.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.31.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.31.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.31.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "language_model.model.layers.4.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.4.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.4.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.4.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.4.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.4.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.4.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.4.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.4.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.5.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.5.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.5.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.5.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.5.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.5.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.5.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.5.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.5.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.6.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.6.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.6.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.6.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.6.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.6.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.6.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.6.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.6.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.7.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.7.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.7.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.7.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.7.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.7.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.7.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.7.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.7.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "language_model.model.layers.8.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.8.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.8.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.8.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.8.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.8.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.8.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.8.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.8.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.9.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.9.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.9.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.9.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.9.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.9.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.9.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.9.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.layers.9.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "language_model.model.norm.weight": "model-00003-of-00004.safetensors",
+    "projector.fc1.bias": "model-00001-of-00004.safetensors",
+    "projector.fc1.weight": "model-00001-of-00004.safetensors",
+    "projector.fc2.bias": "model-00001-of-00004.safetensors",
+    "projector.fc2.weight": "model-00001-of-00004.safetensors",
+    "projector.fc3.bias": "model-00001-of-00004.safetensors",
+    "projector.fc3.weight": "model-00001-of-00004.safetensors",
+    "value_head.head_l1.bias": "model-00004-of-00004.safetensors",
+    "value_head.head_l1.weight": "model-00004-of-00004.safetensors",
+    "value_head.head_l2.bias": "model-00004-of-00004.safetensors",
+    "value_head.head_l2.weight": "model-00004-of-00004.safetensors",
+    "value_head.head_l3.weight": "model-00004-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.0.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.0.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.0.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.0.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.0.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.0.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.0.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.0.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.0.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.0.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.0.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.0.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.0.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.0.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.1.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.1.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.1.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.1.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.1.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.1.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.1.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.1.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.1.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.1.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.1.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.1.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.1.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.1.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.10.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.10.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.10.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.10.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.10.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.10.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.10.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.10.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.10.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.10.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.10.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.10.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.10.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.10.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.11.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.11.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.11.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.11.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.11.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.11.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.11.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.11.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.11.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.11.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.11.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.11.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.11.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.11.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.12.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.12.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.12.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.12.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.12.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.12.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.12.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.12.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.12.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.12.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.12.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.12.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.12.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.12.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.13.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.13.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.13.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.13.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.13.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.13.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.13.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.13.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.13.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.13.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.13.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.13.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.13.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.13.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.14.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.14.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.14.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.14.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.14.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.14.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.14.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.14.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.14.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.14.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.14.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.14.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.14.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.14.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.15.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.15.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.15.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.15.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.15.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.15.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.15.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.15.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.15.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.15.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.15.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.15.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.15.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.15.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.16.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.16.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.16.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.16.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.16.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.16.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.16.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.16.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.16.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.16.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.16.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.16.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.16.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.16.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.17.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.17.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.17.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.17.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.17.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.17.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.17.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.17.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.17.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.17.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.17.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.17.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.17.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.17.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.18.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.18.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.18.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.18.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.18.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.18.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.18.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.18.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.18.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.18.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.18.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.18.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.18.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.18.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.19.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.19.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.19.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.19.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.19.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.19.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.19.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.19.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.19.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.19.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.19.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.19.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.19.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.19.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.2.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.2.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.2.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.2.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.2.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.2.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.2.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.2.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.2.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.2.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.2.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.2.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.2.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.2.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.20.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.20.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.20.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.20.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.20.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.20.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.20.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.20.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.20.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.20.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.20.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.20.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.20.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.20.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.21.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.21.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.21.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.21.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.21.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.21.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.21.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.21.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.21.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.21.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.21.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.21.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.21.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.21.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.22.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.22.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.22.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.22.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.22.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.22.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.22.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.22.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.22.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.22.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.22.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.22.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.22.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.22.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.23.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.23.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.23.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.23.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.23.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.23.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.23.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.23.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.23.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.23.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.23.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.23.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.23.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.23.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.3.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.3.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.3.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.3.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.3.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.3.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.3.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.3.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.3.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.3.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.3.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.3.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.3.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.3.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.4.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.4.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.4.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.4.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.4.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.4.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.4.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.4.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.4.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.4.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.4.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.4.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.4.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.4.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.5.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.5.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.5.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.5.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.5.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.5.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.5.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.5.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.5.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.5.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.5.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.5.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.5.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.5.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.6.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.6.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.6.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.6.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.6.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.6.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.6.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.6.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.6.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.6.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.6.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.6.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.6.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.6.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.7.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.7.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.7.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.7.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.7.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.7.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.7.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.7.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.7.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.7.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.7.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.7.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.7.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.7.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.8.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.8.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.8.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.8.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.8.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.8.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.8.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.8.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.8.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.8.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.8.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.8.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.8.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.8.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.9.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.9.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.9.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.9.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.9.ls1.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.9.ls2.scale_factor": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.9.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.9.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.9.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.9.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.9.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.9.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.9.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.blocks.9.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.cls_token": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.norm.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.norm.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.patch_embed.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.patch_embed.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.pos_embed": "model-00001-of-00004.safetensors",
+    "vision_backbone.featurizer.reg_token": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.attn_pool.kv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.attn_pool.kv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.attn_pool.latent": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.attn_pool.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.attn_pool.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.attn_pool.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.attn_pool.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.attn_pool.norm.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.attn_pool.norm.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.attn_pool.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.attn_pool.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.attn_pool.q.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.attn_pool.q.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.0.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.0.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.0.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.0.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.0.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.0.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.0.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.0.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.0.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.0.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.0.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.0.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.1.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.1.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.1.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.1.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.1.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.1.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.1.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.1.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.1.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.1.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.1.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.1.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.10.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.10.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.10.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.10.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.10.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.10.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.10.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.10.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.10.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.10.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.10.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.10.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.11.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.11.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.11.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.11.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.11.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.11.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.11.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.11.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.11.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.11.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.11.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.11.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.12.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.12.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.12.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.12.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.12.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.12.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.12.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.12.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.12.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.12.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.12.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.12.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.13.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.13.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.13.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.13.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.13.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.13.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.13.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.13.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.13.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.13.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.13.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.13.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.14.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.14.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.14.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.14.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.14.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.14.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.14.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.14.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.14.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.14.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.14.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.14.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.15.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.15.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.15.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.15.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.15.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.15.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.15.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.15.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.15.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.15.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.15.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.15.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.16.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.16.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.16.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.16.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.16.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.16.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.16.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.16.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.16.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.16.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.16.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.16.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.17.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.17.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.17.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.17.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.17.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.17.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.17.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.17.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.17.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.17.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.17.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.17.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.18.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.18.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.18.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.18.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.18.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.18.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.18.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.18.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.18.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.18.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.18.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.18.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.19.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.19.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.19.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.19.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.19.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.19.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.19.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.19.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.19.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.19.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.19.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.19.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.2.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.2.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.2.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.2.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.2.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.2.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.2.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.2.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.2.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.2.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.2.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.2.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.20.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.20.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.20.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.20.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.20.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.20.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.20.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.20.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.20.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.20.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.20.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.20.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.21.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.21.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.21.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.21.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.21.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.21.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.21.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.21.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.21.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.21.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.21.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.21.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.22.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.22.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.22.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.22.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.22.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.22.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.22.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.22.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.22.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.22.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.22.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.22.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.23.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.23.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.23.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.23.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.23.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.23.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.23.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.23.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.23.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.23.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.23.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.23.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.24.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.24.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.24.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.24.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.24.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.24.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.24.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.24.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.24.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.24.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.24.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.24.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.25.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.25.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.25.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.25.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.25.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.25.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.25.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.25.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.25.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.25.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.25.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.25.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.26.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.26.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.26.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.26.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.26.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.26.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.26.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.26.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.26.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.26.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.26.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.26.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.3.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.3.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.3.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.3.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.3.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.3.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.3.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.3.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.3.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.3.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.3.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.3.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.4.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.4.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.4.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.4.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.4.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.4.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.4.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.4.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.4.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.4.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.4.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.4.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.5.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.5.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.5.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.5.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.5.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.5.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.5.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.5.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.5.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.5.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.5.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.5.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.6.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.6.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.6.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.6.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.6.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.6.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.6.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.6.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.6.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.6.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.6.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.6.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.7.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.7.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.7.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.7.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.7.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.7.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.7.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.7.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.7.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.7.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.7.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.7.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.8.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.8.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.8.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.8.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.8.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.8.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.8.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.8.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.8.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.8.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.8.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.8.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.9.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.9.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.9.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.9.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.9.mlp.fc1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.9.mlp.fc1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.9.mlp.fc2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.9.mlp.fc2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.9.norm1.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.9.norm1.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.9.norm2.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.blocks.9.norm2.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.norm.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.norm.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.patch_embed.proj.bias": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.patch_embed.proj.weight": "model-00001-of-00004.safetensors",
+    "vision_backbone.fused_featurizer.pos_embed": "model-00001-of-00004.safetensors"
+  }
+}

modeling_prismatic.py

@@ -0,0 +1,1085 @@
+"""
+modeling_prismatic.py
+
+Core HuggingFace-style PrismaticPreTrainedModel and PrismaticForConditionalGeneration class definitions.
+Inherits from the default `transformers.PretrainedModel`. Meant to be standalone and self-contained,
+but exactly replicate the logic in `prismatic.models.vlms.prismatic.py`.
+"""
+
+import logging
+from dataclasses import dataclass
+from functools import partial
+from typing import Any, Callable, ClassVar, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import timm
+import tokenizers
+import torch
+import torch.nn as nn
+import transformers
+from timm.models.vision_transformer import LayerScale
+from transformers import AutoModelForCausalLM, PretrainedConfig, PreTrainedModel
+from transformers.modeling_outputs import ModelOutput
+
+from prismatic.training.train_utils import (
+    get_current_action_mask,
+    get_next_actions_mask,
+)
+from prismatic.vla.constants import (
+    ACTION_DIM,
+    ACTION_PROPRIO_NORMALIZATION_TYPE,
+    ACTION_TOKEN_BEGIN_IDX,
+    IGNORE_INDEX,
+    NUM_ACTIONS_CHUNK,
+    STOP_INDEX,
+    NormalizationType,
+)
+
+from .configuration_prismatic import OpenVLAConfig, PrismaticConfig
+
+# Set up logger
+logger = logging.getLogger(__name__)
+
+
+# === Utility Functions for Monkey-Patching ===
+def unpack_tuple(fn: Callable[[Any], Tuple[Any]]) -> Callable[[Any], Any]:
+    def wrapper(*args: Any, **kwargs: Any) -> Any:
+        result = fn(*args, **kwargs)
+        return result[0] if isinstance(result, tuple) else result
+
+    return wrapper
+
+
+# HF Transformers overwrites parameters with names containing `gamma`; we're going to patch VisionBackbone.LayerScale.
+#   =>> TIMM :: https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/vision_transformer.py#L109
+#   =>> Transformers :: https://github.com/huggingface/transformers/blob/main/src/transformers/modeling_utils.py#L3960
+def _ls_new_forward(self, x: torch.Tensor) -> torch.Tensor:
+    return x.mul_(self.scale_factor) if self.inplace else x * self.scale_factor
+
+
+def ls_apply_patch(ls_module: LayerScale):
+    ls_module.scale_factor = nn.Parameter(ls_module.gamma.clone())
+    ls_module.forward = _ls_new_forward.__get__(ls_module, LayerScale)
+    del ls_module.gamma
+
+
+# === Prismatic Vision Backbone (nn.Module) Definitions (w/ Fused Backbone Support) ===
+class PrismaticVisionBackbone(nn.Module):
+    """
+    Vision backbone for Prismatic models that handles image feature extraction.
+
+    Supports both single backbone (e.g., SigLIP) and fused backbone (e.g., SigLIP + DINOv2) configurations.
+    For fused backbones, features from both models are concatenated along the feature dimension.
+    """
+
+    def __init__(
+        self,
+        use_fused_vision_backbone: bool,
+        image_sizes: List[int],
+        timm_model_ids: List[str],
+        timm_override_act_layers: List[Optional[str]],
+    ) -> None:
+        """
+        Initialize the vision backbone.
+
+        Args:
+            use_fused_vision_backbone: Whether to use two backbones and fuse their features
+            image_sizes: List of image sizes for each backbone
+            timm_model_ids: List of TIMM model IDs to use for each backbone
+            timm_override_act_layers: List of activation layer overrides for each backbone
+        """
+        super().__init__()
+        self.use_fused_vision_backbone = use_fused_vision_backbone
+        self.num_images_in_input = 1  # Default value, can be overridden later
+
+        # Validate number of (fused) vision backbones
+        if len(timm_model_ids) > 2:
+            raise ValueError("Prismatic models only support up to 2 (fused) vision backbones!")
+
+        # Create primary featurizer
+        self.featurizer = self._create_featurizer(
+            model_id=timm_model_ids[0], img_size=image_sizes[0], act_layer=timm_override_act_layers[0]
+        )
+        self.embed_dim = self.featurizer.embed_dim
+
+        # Create secondary featurizer if using fused backbone
+        if self.use_fused_vision_backbone:
+            self.fused_featurizer = self._create_featurizer(
+                model_id=timm_model_ids[1], img_size=image_sizes[1], act_layer=timm_override_act_layers[1]
+            )
+            self.embed_dim += self.fused_featurizer.embed_dim
+
+        # Patch LayerScale modules for HF compatibility
+        self._patch_layer_scales()
+
+    def _create_featurizer(self, model_id: str, img_size: int, act_layer: Optional[str]) -> nn.Module:
+        """
+        Create a TIMM-based featurizer model with appropriate configurations.
+
+        Args:
+            model_id: The TIMM model ID to load
+            img_size: Input image size for the model
+            act_layer: Override for the activation layer type
+
+        Returns:
+            A configured featurizer model
+        """
+        featurizer = timm.create_model(
+            model_id,
+            pretrained=False,
+            num_classes=0,
+            img_size=img_size,
+            act_layer=act_layer,
+        )
+
+        # Monkey-patch the forward function to extract the second-to-last layer features
+        num_blocks = len(featurizer.blocks)
+        featurizer.forward = unpack_tuple(partial(featurizer.get_intermediate_layers, n={num_blocks - 2}))
+
+        return featurizer
+
+    def _patch_layer_scales(self) -> None:
+        """
+        Patch all LayerScale modules to be compatible with HF's parameter naming.
+
+        HF Transformers overwrites parameters with names containing 'gamma',
+        so we need to rename and modify the forward method.
+        """
+        # Patch primary featurizer
+        for module in self.featurizer.modules():
+            if isinstance(module, LayerScale):
+                ls_apply_patch(module)
+
+        # Patch secondary featurizer if it exists
+        if self.use_fused_vision_backbone:
+            for module in self.fused_featurizer.modules():
+                if isinstance(module, LayerScale):
+                    ls_apply_patch(module)
+
+    def get_num_patches(self) -> int:
+        """
+        Returns the number of vision patches output by the vision backbone.
+
+        Returns:
+            Number of patches per image
+        """
+        return self.featurizer.patch_embed.num_patches
+
+    def get_num_images_in_input(self) -> int:
+        """
+        Returns the number of input images for the vision backbone.
+
+        Returns:
+            Number of images expected in the input
+        """
+        return self.num_images_in_input
+
+    def set_num_images_in_input(self, num_images_in_input: int) -> None:
+        """
+        Sets the number of input images for the vision backbone.
+
+        Args:
+            num_images_in_input: Number of images to expect in the input
+        """
+        self.num_images_in_input = num_images_in_input
+
+    def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
+        """
+        Implements the forward pass for the vision backbone.
+
+        If `self.use_fused_vision_backbone == True`, uses both SigLIP and DINOv2 transformers to extract visual features
+        (otherwise uses SigLIP only). Allows multi-image inputs (but only for fused vision backbone).
+
+        Args:
+            pixel_values (torch.Tensor): Pixels for input image(s), (B, C, H, W).
+        """
+        if self.num_images_in_input == 1:
+            if not self.use_fused_vision_backbone:
+                return self.featurizer(pixel_values)
+
+            # Split `pixel_values :: [bsz, 2 * 3, resolution, resolution]` =>> featurize =>> channel stack
+            img, img_fused = torch.split(pixel_values, [3, 3], dim=1)
+            patches, patches_fused = self.featurizer(img), self.fused_featurizer(img_fused)
+
+            return torch.cat([patches, patches_fused], dim=2)
+
+        else:
+            assert self.use_fused_vision_backbone, "Multi-image inputs require using fused backbone!"
+
+            # Split `pixel_values` into individual images (each with 6 channels: 3 for SigLIP + 3 for DINOv2)
+            images = torch.split(pixel_values, [6] * self.num_images_in_input, dim=1)
+
+            # Process each image and collect patches
+            all_patches = []
+            for img in images:
+                # Split each image further into two stacks of channels (each with 3 channels)
+                img_regular, img_fused = torch.split(img, [3, 3], dim=1)
+
+                # Get patches from both SigLIP and DINOv2 vision transformers
+                patches = self.featurizer(img_regular)
+                patches_fused = self.fused_featurizer(img_fused)
+
+                # Concatenate SigLIP and DINOv2 patches along the hidden dimension
+                combined_patches = torch.cat([patches, patches_fused], dim=2)
+                all_patches.append(combined_patches)
+
+            # Concatenate all patches along the patch dimension
+            return torch.cat(all_patches, dim=1)
+
+
+# === Prismatic Projector (nn.Module) Definitions ===
+class PrismaticProjector(nn.Module):
+    def __init__(self, use_fused_vision_backbone: bool, vision_dim: int, llm_dim: int) -> None:
+        super().__init__()
+        self.use_fused_vision_backbone = use_fused_vision_backbone
+        self.vision_dim, self.llm_dim = vision_dim, llm_dim
+
+        # Switch on `use_fused_vision_backbone` =>> use slightly different MLPs and projection factors!
+        if not self.use_fused_vision_backbone:
+            self.fc1 = nn.Linear(self.vision_dim, self.llm_dim, bias=True)
+            self.fc2 = nn.Linear(self.llm_dim, self.llm_dim, bias=True)
+            self.act_fn1 = nn.GELU()
+        else:
+            initial_projection_dim = 4 * vision_dim
+            self.fc1 = nn.Linear(self.vision_dim, initial_projection_dim, bias=True)
+            self.fc2 = nn.Linear(initial_projection_dim, self.llm_dim, bias=True)
+            self.fc3 = nn.Linear(self.llm_dim, self.llm_dim, bias=True)
+            self.act_fn1 = nn.GELU()
+            self.act_fn2 = nn.GELU()
+
+    def forward(self, img_patches: torch.Tensor) -> torch.Tensor:
+        if not self.use_fused_vision_backbone:
+            projected_features = self.fc1(img_patches)
+            projected_features = self.act_fn1(projected_features)
+            projected_features = self.fc2(projected_features)
+        else:
+            projected_features = self.fc1(img_patches)
+            projected_features = self.act_fn1(projected_features)
+            projected_features = self.fc2(projected_features)
+            projected_features = self.act_fn2(projected_features)
+            projected_features = self.fc3(projected_features)
+
+        return projected_features
+
+
+# === Main HF Class Definitions ===
+@dataclass
+class PrismaticCausalLMOutputWithPast(ModelOutput):
+    """Base class for Prismatic casual (visually-conditioned) language model outputs; also exposes visual features."""
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+    # Additions for VLMs
+    projector_features: Optional[torch.FloatTensor] = None
+
+
+class PrismaticPreTrainedModel(PreTrainedModel):
+    config_class: PretrainedConfig = PrismaticConfig
+    base_model_prefix: str = "model"
+    supports_gradient_checkpointing: bool = True
+
+    _no_split_modules: ClassVar[List[str]] = ["PrismaticProjector"]
+    _skip_keys_device_placement: str = "past_key_values"
+    _supports_flash_attn_2: bool = True
+
+    def _init_weights(self, module: nn.Module) -> None:
+        # Important :: this HF ported version is *not* meant for training from scratch; only inference and fine-tuning!
+        #   => As such, this init_weights code is not correct; if training VLMs from scratch, use the main codebase at
+        #      https://github.com/TRI-ML/prismatic-vlms
+        std = (
+            self.config.initializer_range
+            if hasattr(self.config, "initializer_range")
+            else self.config.text_config.initializer_range
+        )
+
+        if hasattr(module, "class_embedding"):
+            module.class_embedding.data.normal_(mean=0.0, std=std)
+
+        if isinstance(module, (nn.Linear, nn.Conv2d)):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    @property
+    def _supports_sdpa(self) -> bool:
+        """Check LLM supports SDPA Attention"""
+        return self.language_model._supports_sdpa
+
+
+class PrismaticForConditionalGeneration(PrismaticPreTrainedModel):
+    def __init__(self, config: PrismaticConfig) -> None:
+        super().__init__(config)
+
+        # [Validation] Lightweight Validate on `config` Fields + Dependency Versions
+        if config.use_fused_vision_backbone is None:
+            raise ValueError("Missing config field `use_fused_vision_backbone`")
+
+        if timm.__version__ not in {"0.9.10", "0.9.11", "0.9.12", "0.9.16"}:
+            raise NotImplementedError(
+                "TIMM Version must be >= 0.9.10 and < 1.0.0 (breaking); please raise a GitHub Issue "
+                "if you urgently need support for latest TIMM versions."
+            )
+
+        if (transformers.__version__ != "4.40.1") or (tokenizers.__version__ != "0.19.1"):
+            logger.warning(
+                f"Expected `transformers==4.40.1` and `tokenizers==0.19.1` but got "
+                f"`transformers=={transformers.__version__}` and `tokenizers=={tokenizers.__version__}`; "
+                f"there might be inference-time regressions due to dependency changes. If in doubt, please"
+                f"use the above versions."
+            )
+
+        # Instantiate PrismaticVisionBackbone (w/ Potential Fused Backbone)
+        self.vision_backbone = PrismaticVisionBackbone(
+            config.use_fused_vision_backbone, config.image_sizes, config.timm_model_ids, config.timm_override_act_layers
+        )
+
+        # Create Multimodal Projector
+        self.projector = PrismaticProjector(
+            config.use_fused_vision_backbone,
+            vision_dim=self.vision_backbone.embed_dim,
+            llm_dim=config.text_config.hidden_size,
+        )
+
+        # Instantiate LLM Backbone
+        self.language_model = AutoModelForCausalLM.from_config(
+            config.text_config, attn_implementation=config._attn_implementation
+        )
+        self.vocab_size = config.text_config.vocab_size
+        self.pad_token_id = config.pad_token_id
+        self.llm_dim = config.text_config.hidden_size
+
+        # HF Boilerplate =>> initializes weights via `_init_weights()` and sets gradient checkpointing
+        self.post_init()
+
+    # === `PreTrainedModel` Boilerplate ===
+    def get_input_embeddings(self) -> nn.Module:
+        return self.language_model.get_input_embeddings()
+
+    def set_input_embeddings(self, value: nn.Module) -> None:
+        self.language_model.set_input_embeddings(value)
+
+    def get_output_embeddings(self) -> nn.Module:
+        return self.language_model.get_output_embeddings()
+
+    def set_output_embeddings(self, new_embeddings: nn.Module) -> None:
+        self.language_model.set_output_embeddings(new_embeddings)
+
+    def get_decoder(self) -> nn.Module:
+        return self.language_model.get_decoder()
+
+    def set_decoder(self, decoder: nn.Module) -> None:
+        self.language_model.set_decoder(decoder)
+
+    def tie_weights(self) -> None:
+        self.language_model.tie_weights()  # Note: `Llama-2` and `Mistral` don't tie weights (no-op)
+
+    def resize_token_embeddings(
+        self, new_num_tokens: Optional[int] = None, pad_to_multiple_of: Optional[int] = None
+    ) -> nn.Embedding:
+        updated_embeddings = self.language_model.resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
+
+        # Update config/instance variables
+        self.config.text_config.vocab_size = updated_embeddings.num_embeddings
+        self.vocab_size = updated_embeddings.num_embeddings
+
+        return updated_embeddings
+
+    def _replace_input_embeddings(self, input_embeddings, all_actions_mask, noisy_action_features):
+        """
+        Replace embeddings in input_embeddings at positions where all_actions_mask is True
+        with embeddings from noisy_action_features, using vectorized operations.
+
+        Args:
+            input_embeddings: Tensor of shape (B, S, D)
+            all_actions_mask: Boolean tensor of shape (B, S)
+            noisy_action_features: Tensor of shape (B, K, D) where K is the number of True values in mask per sample
+
+        Returns:
+            Modified input_embeddings tensor
+        """
+        # Clone input to avoid modifying the original tensor
+        new_input_embeddings = input_embeddings.clone()
+
+        # Create a tensor with the same shape of input_embeddings to hold the noisy action features
+        repositioned_noisy_action_features = torch.zeros_like(input_embeddings)
+
+        # Create batch indices for splicing
+        batch_indices = torch.arange(input_embeddings.shape[0], device=input_embeddings.device)
+        batch_indices = batch_indices.unsqueeze(1).expand(-1, noisy_action_features.shape[1])
+
+        # Get indices where mask is True for each sample
+        masked_indices = torch.stack([torch.where(mask)[0] for mask in all_actions_mask])
+
+        # Move the noisy action features into their correct positions
+        repositioned_noisy_action_features[batch_indices, masked_indices] = noisy_action_features
+
+        # Combine original input embeddings and noisy action embeddings using the mask
+        new_input_embeddings = torch.where(
+            all_actions_mask.unsqueeze(-1), repositioned_noisy_action_features, new_input_embeddings
+        )
+
+        return new_input_embeddings
+
+    def _process_action_masks(self, labels):
+        """Helper to get action masks from labels"""
+        current_action_mask = get_current_action_mask(labels)
+        next_actions_mask = get_next_actions_mask(labels)
+        all_actions_mask = current_action_mask | next_actions_mask  # (B, seq_len)
+        return all_actions_mask
+
+    def _process_vision_features(self, pixel_values, language_embeddings=None, use_film=False):
+        """Process vision features with optional FiLM conditioning"""
+        if use_film:
+            # FiLM: Infuse language inputs into visual features
+            patch_features = self.vision_backbone(pixel_values, language_embeddings)  # (bsz, 256 * num_images, D)
+        else:
+            patch_features = self.vision_backbone(pixel_values)  # (bsz, 256 * num_images, D)
+
+        # Project patch embeddings into language embedding space
+        return self.projector(patch_features)
+
+    def _process_proprio_features(self, projected_patch_embeddings, proprio, proprio_projector):
+        """Process proprioceptive features and append to vision features"""
+        if proprio_projector is not None and proprio is not None:
+            # projected_patch_embeddings: (bsz, num_patches * num_images, llm_dim)
+            # proprio: (bsz, proprio_dim) or (propro_dim,)
+            proprio = proprio.reshape(projected_patch_embeddings.shape[0], -1)  # (bsz, proprio_dim)
+            proprio_features = proprio_projector(proprio)  # (bsz, llm_dim)
+            proprio_features = proprio_features.unsqueeze(dim=1)  # (bsz, 1, llm_dim)
+            # For simplicity, just append proprio token to the end of projected vision patch tokens
+            return torch.cat((projected_patch_embeddings, proprio_features), dim=1)
+        return projected_patch_embeddings
+
+    def _build_multimodal_attention(self, input_embeddings, projected_patch_embeddings, attention_mask):
+        """Build multimodal embeddings and attention mask"""
+        # Update attention mask
+        projected_patch_attention_mask = None
+        if attention_mask is not None:
+            projected_patch_attention_mask = torch.full(
+                (projected_patch_embeddings.shape[0], projected_patch_embeddings.shape[1]),
+                fill_value=True,
+                dtype=attention_mask.dtype,
+                device=attention_mask.device,
+            )
+
+        # Build multimodal embeddings & attention mask; insert embeddings after <BOS> token (1:)
+        multimodal_embeddings = torch.cat(
+            [input_embeddings[:, :1, :], projected_patch_embeddings, input_embeddings[:, 1:, :]], dim=1
+        )
+
+        multimodal_attention_mask = None
+        if attention_mask is not None:
+            multimodal_attention_mask = torch.cat(
+                [attention_mask[:, :1], projected_patch_attention_mask, attention_mask[:, 1:]], dim=1
+            )
+
+        return multimodal_embeddings, multimodal_attention_mask
+
+    def _build_multimodal_labels(self, labels, projected_patch_embeddings):
+        """Build multimodal labels with IGNORE_INDEX for patch embeddings"""
+        if labels is not None:
+            projected_patch_labels = torch.full(
+                (projected_patch_embeddings.shape[0], projected_patch_embeddings.shape[1]),
+                fill_value=IGNORE_INDEX,
+                dtype=labels.dtype,
+                device=labels.device,
+            )
+            return torch.cat([labels[:, :1], projected_patch_labels, labels[:, 1:]], dim=1)
+        return None
+
+    # === Core Prismatic VLM `forward()` Logic ===
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_projector_features: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        proprio=None,
+        proprio_projector=None,
+        noisy_actions=None,
+        noisy_action_projector=None,
+        diffusion_timestep_embeddings=None,
+        use_film: bool = False,
+    ) -> Union[Tuple, PrismaticCausalLMOutputWithPast]:
+        """Run a forward pass through the VLM, returning a PrismaticCausalLMOutputWithPast instance."""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        output_projector_features = output_projector_features if output_projector_features is not None else False
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # Respect `use_cache` only if not training (even if `gradient_checkpointing` is off)
+        use_cache = use_cache and not self.training
+
+        # Instantiate Placeholder for Projector Features
+        projected_patch_embeddings = None
+
+        # === Handle Generation with Cache (`input_ids.shape[1] == 1`) =>> requires `past_keys_values` ===
+        if input_ids.shape[1] == 1:
+            assert input_ids.shape[0] == 1, "Generation is only currently supported for batch size of 1!"
+            assert past_key_values is not None, "You must provide `past_key_values` during cached generation!"
+            assert labels is None, "Unexpected key `labels` provided during cached generation!"
+
+            language_model_output = self.language_model(
+                input_ids=input_ids,
+                attention_mask=None,
+                position_ids=None,
+                past_key_values=past_key_values,
+                inputs_embeds=None,
+                labels=None,
+                use_cache=use_cache,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+
+        # === Handle Unimodal Forward ===
+        elif pixel_values is None:
+            assert (input_ids is not None) and (inputs_embeds is None), "Missing `input_ids` in language-only forward!"
+            assert past_key_values is None, "Unexpected key `past_key_values` provided during language-only forward!"
+
+            language_model_output = self.language_model(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                position_ids=None,
+                past_key_values=None,
+                inputs_embeds=None,
+                labels=labels,
+                use_cache=use_cache,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+
+        # === Handle Multimodal Forward ===
+        elif (input_ids.shape[0] == pixel_values.shape[0]) or (inputs_embeds.shape[0] == pixel_values.shape[0]):
+            assert past_key_values is None, "Unexpected key `past_key_values` provided during multimodal forward!"
+
+            # Get input embeddings (from language model embeddings)
+            input_embeddings = self.get_input_embeddings()(input_ids)  # (B, seq_len, D)
+
+            # Extract action masks
+            all_actions_mask = self._process_action_masks(labels)
+
+            # Extract the language portion of the input embeddings (i.e. remove the action tokens portion)
+            language_embeddings = input_embeddings[~all_actions_mask].reshape(
+                input_embeddings.shape[0], -1, input_embeddings.shape[2]
+            )  # (B, lang_seq_len, llm_dim)
+
+            # Get visual features
+            projected_patch_embeddings = self._process_vision_features(pixel_values, language_embeddings, use_film)
+
+            # Add proprioceptive state if provided
+            projected_patch_embeddings = self._process_proprio_features(
+                projected_patch_embeddings, proprio, proprio_projector
+            )
+
+            # [Diffusion] Add diffusion timestep embedding if provided
+            if diffusion_timestep_embeddings is not None:
+                # For simplicity, just append diffusion timestep embedding to the end of projected vision patch tokens
+                projected_patch_embeddings = torch.cat(
+                    (projected_patch_embeddings, diffusion_timestep_embeddings), dim=1
+                )
+
+            # Process action embeddings
+            if noisy_actions is not None:
+                # Get mask corresponding to all action tokens
+                all_actions_mask = self._process_action_masks(labels)
+
+                # Reshape noisy actions into individual action tokens
+                # noisy_actions: (B, chunk_len, action_dim) -> (B, chunk_len * action_dim, 1)
+                B = noisy_actions.shape[0]
+                noisy_actions = noisy_actions.reshape(B, -1).unsqueeze(-1)
+
+                # Project noisy action tokens into language model embedding space
+                noisy_action_features = noisy_action_projector(noisy_actions)  # (B, chunk_len * action_dim, llm_dim)
+
+                # Replace embeddings of the action tokens with noisy action embeddings
+                input_embeddings = self._replace_input_embeddings(
+                    input_embeddings, all_actions_mask, noisy_action_features
+                )
+            else:
+                # Replace the embeddings of the action tokens with zeros
+                # (Later on, the positional embeddings will be added to them)
+                all_actions_mask = all_actions_mask.unsqueeze(-1)  # (B, seq_len, 1)
+                input_embeddings = input_embeddings * ~all_actions_mask
+
+            # Build multimodal embeddings & attention mask
+            multimodal_embeddings, multimodal_attention_mask = self._build_multimodal_attention(
+                input_embeddings, projected_patch_embeddings, attention_mask
+            )
+
+            # Build labels for multimodal sequence if needed
+            multimodal_labels = self._build_multimodal_labels(labels, projected_patch_embeddings)
+
+            # Dispatch to language model
+            language_model_output = self.language_model(
+                input_ids=None,
+                attention_mask=multimodal_attention_mask,
+                position_ids=None,
+                past_key_values=None,
+                inputs_embeds=multimodal_embeddings,
+                labels=multimodal_labels,
+                use_cache=use_cache,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+
+        # === Otherwise =>> Assume Invalid! ===
+        elif (input_ids.shape[0] != pixel_values.shape[0]) or (inputs_embeds.shape[0] != pixel_values.shape[0]):
+            raise ValueError("Non-homogenous batch of (text, image) input -- forward() does not support mixed batches!")
+
+        else:
+            raise ValueError(
+                "Invalid PrismaticForConditionalGeneration `forward()` call with provided arguments:\n"
+                f"=> `input_ids` = {input_ids is not None}\n"
+                f"=> `attention_mask` = {attention_mask is not None}\n"
+                f"=> `pixel_values` = {pixel_values is not None}\n"
+                f"=> `labels` = {labels is not None}\n"
+                f"=> `input_embeds` = {inputs_embeds is not None}\n"
+                f"=> `past_key_values` = {past_key_values is not None}\n"
+                f"=> `use_cache` = {use_cache}"
+            )
+
+        # Unpack `language_model_output` and return PrismaticCausalLMOutputWithPast (or tuple if not `return_dict`)
+        if not return_dict:
+            if output_projector_features and (projected_patch_embeddings is not None):
+                return *language_model_output, projected_patch_embeddings
+
+            return language_model_output
+
+        return PrismaticCausalLMOutputWithPast(
+            loss=language_model_output.loss,
+            logits=language_model_output.logits,
+            past_key_values=language_model_output.past_key_values,
+            hidden_states=language_model_output.hidden_states,
+            attentions=language_model_output.attentions,
+            projector_features=projected_patch_embeddings,
+        )
+
+    # === GenerationMixin Methods ===
+    def prepare_inputs_for_generation(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs: str,
+    ) -> Dict[str, torch.Tensor]:
+        """Borrowed from `LlamaForCausalLM` and simplified for batch size = 1; mirrors original PrismaticVLM logic."""
+        if ((input_ids is not None) and (input_ids.shape[0] > 1)) or (
+            (inputs_embeds is not None) and (inputs_embeds.shape[0] > 1)
+        ):
+            raise ValueError("Generation with batch size > 1 is not currently supported!")
+
+        # Handle `past_key_values` (cache) =>> assume `input_ids` just has unprocessed tokens
+        if past_key_values is not None:
+            input_ids = input_ids[:, -1:]
+
+        # If `input_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"input_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        # Make sure `pixel_values` are preserved in `model_inputs`
+        model_inputs.update(
+            {
+                "attention_mask": attention_mask,
+                "pixel_values": pixel_values,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+            }
+        )
+
+        return model_inputs
+
+    # Defer to Language Model (all handle this differently, with different return types)
+    def _reorder_cache(self, *args, **kwargs) -> Any:
+        return self.language_model._reorder_cache(*args, **kwargs)
+
+
+class OpenVLAForActionPrediction(PrismaticForConditionalGeneration):
+    config_class: PretrainedConfig = OpenVLAConfig
+
+    def __init__(self, config: OpenVLAConfig) -> None:
+        super().__init__(config)
+        self.norm_stats = config.norm_stats
+
+        # Compute action bins
+        self.bins = np.linspace(-1, 1, config.n_action_bins)
+        self.bin_centers = (self.bins[:-1] + self.bins[1:]) / 2.0
+
+        # Compute vocab size for de-tokenization -- revert added "multiple of"
+        self.vocab_size = self.config.text_config.vocab_size - self.config.pad_to_multiple_of
+
+    def _prepare_input_for_action_prediction(self, input_ids, attention_mask):
+        """Prepares input for action prediction by adding necessary tokens"""
+        # Add (ACTION_DIM * NUM_ACTIONS_CHUNK) placeholder tokens to input_ids to simulate action tokens
+        placeholder_action_token_ids = (
+            torch.ones((input_ids.shape[0], ACTION_DIM * NUM_ACTIONS_CHUNK)).to(input_ids.device).to(input_ids.dtype)
+        )
+        input_ids = torch.cat([input_ids, placeholder_action_token_ids], dim=-1)
+
+        # Add stop token to sequence (needed in non-causal bi-directional self-attention, as it appears at train time)
+        stop_token_id = torch.ones((input_ids.shape[0], 1)).to(input_ids.device).to(input_ids.dtype) * STOP_INDEX
+        input_ids = torch.cat([input_ids, stop_token_id], dim=-1)
+
+        # Extend the attention mask to fit the new shape of input
+        # Note: Only batch size == 1 supported right now
+        mask_extension = (
+            torch.ones((attention_mask.shape[0], input_ids.shape[-1] - attention_mask.shape[-1]))
+            .to(attention_mask.device)
+            .to(attention_mask.dtype)
+        )
+        attention_mask = torch.cat([attention_mask, mask_extension], dim=-1)
+
+        return input_ids, attention_mask
+
+    def _prepare_labels_for_action_prediction(self, labels, input_ids):
+        """Creates labels tensor for action prediction if not provided"""
+        # Extend labels tensor with fake action labels
+        ARBITRARY_ACTION_TOKEN_IDX = ACTION_TOKEN_BEGIN_IDX + 1
+        labels_extension = (
+            torch.ones((labels.shape[0], input_ids.shape[-1] - labels.shape[-1])).to(labels.device).to(labels.dtype)
+            * ARBITRARY_ACTION_TOKEN_IDX
+        )
+        labels = torch.cat([labels, labels_extension], dim=-1)
+
+        # Replace last label token with stop token
+        labels[:, -1] = STOP_INDEX
+
+        return labels
+
+    def _unnormalize_actions(self, normalized_actions, unnorm_key=None):
+        """Unnormalize actions using dataset statistics"""
+        action_norm_stats = self.get_action_stats(unnorm_key)
+
+        if ACTION_PROPRIO_NORMALIZATION_TYPE == NormalizationType.BOUNDS:
+            mask = action_norm_stats.get("mask", np.ones_like(action_norm_stats["min"], dtype=bool))
+            action_high, action_low = np.array(action_norm_stats["max"]), np.array(action_norm_stats["min"])
+        elif ACTION_PROPRIO_NORMALIZATION_TYPE == NormalizationType.BOUNDS_Q99:
+            mask = action_norm_stats.get("mask", np.ones_like(action_norm_stats["q01"], dtype=bool))
+            action_high, action_low = np.array(action_norm_stats["q99"]), np.array(action_norm_stats["q01"])
+        else:
+            raise ValueError("Unsupported action/proprio normalization type detected!")
+
+        actions = np.where(
+            mask,
+            0.5 * (normalized_actions + 1) * (action_high - action_low + 1e-8) + action_low,
+            normalized_actions,
+        )
+
+        return actions
+
+    def _run_diffusion_prediction(
+        self,
+        input_embeddings,
+        all_actions_mask,
+        noise,
+        action_head,
+        projected_patch_embeddings,
+        labels,
+        attention_mask,
+        NUM_PATCHES,
+        NUM_PROMPT_TOKENS,
+        noisy_action_projector,
+    ):
+        """Run diffusion-based action prediction"""
+        # Clone embedding for reuse in each timestep
+        orig_projected_patch_embeddings = projected_patch_embeddings.clone()
+        curr_noisy_actions = noise
+
+        # Reverse diffusion: Iteratively denoise to generate action prediction
+        for t in action_head.noise_scheduler.timesteps:
+            # Get diffusion model's noise prediction (conditioned on VLA latent embedding, current noisy action
+            # embedding, and diffusion timestep embedding)
+            timesteps = torch.Tensor([t]).to(labels.device)
+            diffusion_timestep_embeddings = (
+                action_head.time_encoder(timesteps).to(curr_noisy_actions.dtype).to(curr_noisy_actions.device)
+            )  # (B, llm_dim)
+            diffusion_timestep_embeddings = diffusion_timestep_embeddings.unsqueeze(1)  # (B, 1, llm_dim)
+
+            # [Diffusion] Replace the embeddings of the action tokens with noisy actions
+            # (Later on, the positional embeddings will be added to them)
+
+            # For simplicity, append diffusion timestep embedding to the end of projected vision tokens
+            projected_patch_embeddings = torch.cat(
+                (orig_projected_patch_embeddings, diffusion_timestep_embeddings), dim=1
+            )
+
+            # Reshape and project noisy actions into language embedding space
+            B = curr_noisy_actions.shape[0]
+            orig_curr_noisy_actions_shape = curr_noisy_actions.shape
+            curr_noisy_actions = curr_noisy_actions.reshape(B, -1).unsqueeze(-1)
+            noisy_action_features = noisy_action_projector(curr_noisy_actions)
+            curr_noisy_actions = curr_noisy_actions.reshape(orig_curr_noisy_actions_shape)
+
+            # Replace action token embeddings with noisy action embeddings
+            input_embeddings = self._replace_input_embeddings(
+                input_embeddings.clone(), all_actions_mask, noisy_action_features
+            )
+
+            # Build multimodal embeddings and attention mask
+            multimodal_embeddings, multimodal_attention_mask = self._build_multimodal_attention(
+                input_embeddings, projected_patch_embeddings, attention_mask
+            )
+
+            # Forward pass through language model
+            language_model_output = self.language_model(
+                input_ids=None,
+                attention_mask=multimodal_attention_mask,
+                position_ids=None,
+                past_key_values=None,
+                inputs_embeds=multimodal_embeddings,
+                labels=None,
+                use_cache=None,
+                output_attentions=False,
+                output_hidden_states=True,
+                return_dict=True,
+            )
+
+            # Extract hidden states for action portion of response
+            last_hidden_states = language_model_output.hidden_states[-1]  # (B, seq_len, D)
+            actions_hidden_states = last_hidden_states[
+                :,
+                NUM_PATCHES + NUM_PROMPT_TOKENS : NUM_PATCHES + NUM_PROMPT_TOKENS + ACTION_DIM * NUM_ACTIONS_CHUNK,
+                :,
+            ]  # (B, act_chunk_len, D)
+
+            # Predict noise and update noisy actions: x_t -> x_{t-1}
+            noise_pred = action_head.predict_noise(actions_hidden_states)
+            curr_noisy_actions = action_head.noise_scheduler.step(noise_pred, t, curr_noisy_actions).prev_sample
+
+        curr_noisy_actions = curr_noisy_actions.reshape(NUM_ACTIONS_CHUNK, ACTION_DIM)
+
+        # Return final actions
+        return curr_noisy_actions.float().cpu().detach().numpy(), actions_hidden_states
+
+    def _regression_or_discrete_prediction(
+        self,
+        input_embeddings,
+        all_actions_mask,
+        projected_patch_embeddings,
+        attention_mask,
+        labels,
+        NUM_PATCHES,
+        NUM_PROMPT_TOKENS,
+        action_head=None,
+    ):
+        """Run L1 regression-based continuous action prediction or discrete action tokens prediction."""
+        # Zero out action token embeddings
+        all_actions_mask = all_actions_mask.unsqueeze(-1)  # (B, seq_len, 1)
+        input_embeddings = input_embeddings * ~all_actions_mask
+
+        # Build multimodal embeddings and attention mask
+        multimodal_embeddings, multimodal_attention_mask = self._build_multimodal_attention(
+            input_embeddings, projected_patch_embeddings, attention_mask
+        )
+
+        # Forward pass through language model
+        language_model_output = self.language_model(
+            input_ids=None,
+            attention_mask=multimodal_attention_mask,
+            position_ids=None,
+            past_key_values=None,
+            inputs_embeds=multimodal_embeddings,
+            labels=None,
+            use_cache=None,
+            output_attentions=False,
+            output_hidden_states=True,
+            return_dict=True,
+        )
+
+        # Extract hidden states for action tokens
+        last_hidden_states = language_model_output.hidden_states[-1]  # (B, seq_len, D)
+        actions_hidden_states = last_hidden_states[
+            :,
+            NUM_PATCHES + NUM_PROMPT_TOKENS : NUM_PATCHES + NUM_PROMPT_TOKENS + ACTION_DIM * NUM_ACTIONS_CHUNK,
+            :,
+        ]  # (B, act_chunk_len, D)
+
+        # Handle different prediction methods
+        if action_head is not None:
+            # L1 regression prediction
+            normalized_actions = action_head.predict_action(actions_hidden_states)
+            normalized_actions = normalized_actions.reshape(NUM_ACTIONS_CHUNK, ACTION_DIM)
+            normalized_actions = normalized_actions.float().cpu().detach().numpy()
+        else:
+            # Discrete token-based prediction
+            predicted_action_token_ids = (
+                language_model_output.logits[
+                    :,
+                    NUM_PATCHES + NUM_PROMPT_TOKENS : NUM_PATCHES + NUM_PROMPT_TOKENS + ACTION_DIM * NUM_ACTIONS_CHUNK,
+                ]
+                .argmax(dim=2)
+                .cpu()
+                .numpy()
+            )
+            discretized_actions = self.vocab_size - predicted_action_token_ids
+            discretized_actions = np.clip(discretized_actions - 1, a_min=0, a_max=self.bin_centers.shape[0] - 1)
+            normalized_actions = self.bin_centers[discretized_actions]
+            normalized_actions = normalized_actions.reshape(NUM_ACTIONS_CHUNK, ACTION_DIM)
+
+        return normalized_actions, actions_hidden_states
+
+    def predict_action(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        unnorm_key: Optional[str] = None,
+        proprio=None,
+        proprio_projector=None,
+        action_head=None,
+        noisy_action_projector=None,
+        use_film: bool = False,
+        **kwargs: str,
+    ) -> np.ndarray:
+        """Predict actions from input sequence, with options for different prediction methods.
+
+        Args:
+            input_ids: Input token ids
+            unnorm_key: Key for unnormalization statistics
+            proprio: Proprioceptive features
+            proprio_projector: Projector for proprioceptive features
+            action_head: Optional head for L1 regression or diffusion-based prediction
+            noisy_action_projector: Projector for noisy actions in diffusion-based prediction
+            use_film: Whether to use FiLM conditioning
+            **kwargs: Additional arguments including pixel_values and attention_mask
+
+        Returns:
+            Tuple of (unnormalized_actions, action_hidden_states)
+        """
+        # If the special empty token ('') does not already appear after the colon (':') token in the prompt
+        # (after "OUT:" or "ASSISTANT:"), insert it to match the inputs seen at training time
+        if not torch.all(input_ids[:, -1] == 29871):
+            input_ids = torch.cat(
+                (input_ids, torch.unsqueeze(torch.Tensor([29871]).long(), dim=0).to(input_ids.device)), dim=1
+            )
+
+        pixel_values = kwargs["pixel_values"]
+        attention_mask = kwargs["attention_mask"]
+
+        # Create fake labels tensor (needed for action mask)
+        labels = input_ids.clone()
+        labels[:] = IGNORE_INDEX
+
+        # Get number of tokens in prompt (excluding the start token)
+        NUM_PROMPT_TOKENS = input_ids.shape[-1] - 1  # Subtract action tokens and stop token
+
+        # Prepare inputs by adding necessary tokens
+        input_ids, attention_mask = self._prepare_input_for_action_prediction(input_ids, attention_mask)
+
+        # Update labels tensor for action mask computation later
+        labels = self._prepare_labels_for_action_prediction(labels, input_ids)
+
+        # Get input embeddings and action masks
+        input_embeddings = self.get_input_embeddings()(input_ids)
+        all_actions_mask = self._process_action_masks(labels)
+
+        # Extract language embeddings
+        language_embeddings = input_embeddings[~all_actions_mask].reshape(
+            input_embeddings.shape[0], -1, input_embeddings.shape[2]
+        )
+
+        # Process vision features
+        projected_patch_embeddings = self._process_vision_features(pixel_values, language_embeddings, use_film)
+
+        # Add proprioceptive features if provided
+        use_proprio = proprio_projector is not None and proprio is not None
+        if use_proprio:
+            proprio = torch.Tensor(proprio).to(projected_patch_embeddings.device, dtype=projected_patch_embeddings.dtype)
+            projected_patch_embeddings = self._process_proprio_features(
+                projected_patch_embeddings, proprio, proprio_projector
+            )
+
+        # Use diffusion if provided, otherwise use regression or discrete prediction
+        use_diffusion = noisy_action_projector is not None and hasattr(action_head, "noise_scheduler")
+
+        # Calculate number of patches (including proprio token and/or diffusion timestep embedding if present)
+        NUM_PATCHES = self.vision_backbone.get_num_patches() * self.vision_backbone.get_num_images_in_input()
+        if use_proprio:
+            NUM_PATCHES += 1
+        if use_diffusion:
+            NUM_PATCHES += 1
+
+        if use_diffusion:
+            # Sample random noise with shape equal to output action, used as the starting state for reverse diffusion
+            noise = torch.randn(
+                size=(1, NUM_ACTIONS_CHUNK, ACTION_DIM), device=input_embeddings.device, dtype=input_embeddings.dtype
+            )
+
+            # Run diffusion-based prediction
+            normalized_actions, actions_hidden_states = self._run_diffusion_prediction(
+                input_embeddings,
+                all_actions_mask,
+                noise,
+                action_head,
+                projected_patch_embeddings,
+                labels,
+                attention_mask,
+                NUM_PATCHES,
+                NUM_PROMPT_TOKENS,
+                noisy_action_projector,
+            )
+        else:
+            # Run regression or discrete token-based prediction
+            normalized_actions, actions_hidden_states = self._regression_or_discrete_prediction(
+                input_embeddings,
+                all_actions_mask,
+                projected_patch_embeddings,
+                attention_mask,
+                labels,
+                NUM_PATCHES,
+                NUM_PROMPT_TOKENS,
+                action_head,
+            )
+
+        # Unnormalize predicted actions
+        actions = self._unnormalize_actions(normalized_actions, unnorm_key)
+
+        return actions, actions_hidden_states
+
+    @staticmethod
+    def _check_unnorm_key(norm_stats: Dict[str, Dict[str, Any]], unnorm_key: Optional[str]) -> str:
+        """Validate and resolve the unnormalization key for action statistics"""
+        if unnorm_key is None:
+            assert len(norm_stats) == 1, (
+                f"Your model was trained on more than one dataset, "
+                f"please pass a `unnorm_key` from the following options to choose the statistics "
+                f"used for un-normalizing actions: {norm_stats.keys()}"
+            )
+            unnorm_key = next(iter(norm_stats.keys()))
+
+        assert unnorm_key in norm_stats, (
+            f"The `unnorm_key` you chose is not in the set of available dataset statistics, "
+            f"please choose from: {norm_stats.keys()}"
+        )
+        return unnorm_key
+
+    def get_action_dim(self, unnorm_key: Optional[str] = None) -> int:
+        """Get the dimensionality of the policy's action space."""
+        unnorm_key = self._check_unnorm_key(self.norm_stats, unnorm_key)
+        return len(self.norm_stats[unnorm_key]["action"]["min"])
+
+    def get_action_stats(self, unnorm_key: Optional[str] = None) -> Dict[str, Any]:
+        """Get all the logged statistics for the given dataset."""
+        unnorm_key = self._check_unnorm_key(self.norm_stats, unnorm_key)
+        return self.norm_stats[unnorm_key]["action"]

preprocessor_config.json

@@ -0,0 +1,114 @@
+{
+  "auto_map": {
+    "AutoImageProcessor": "processing_prismatic.PrismaticImageProcessor",
+    "AutoProcessor": "processing_prismatic.PrismaticProcessor"
+  },
+  "image_processor_type": "PrismaticImageProcessor",
+  "image_resize_strategy": "resize-naive",
+  "input_sizes": [
+    [
+      3,
+      224,
+      224
+    ],
+    [
+      3,
+      224,
+      224
+    ]
+  ],
+  "interpolations": [
+    "bicubic",
+    "bicubic"
+  ],
+  "means": [
+    [
+      0.485,
+      0.456,
+      0.406
+    ],
+    [
+      0.5,
+      0.5,
+      0.5
+    ]
+  ],
+  "processor_class": "PrismaticProcessor",
+  "stds": [
+    [
+      0.229,
+      0.224,
+      0.225
+    ],
+    [
+      0.5,
+      0.5,
+      0.5
+    ]
+  ],
+  "tvf_crop_params": [
+    {
+      "output_size": [
+        224,
+        224
+      ]
+    },
+    {
+      "output_size": [
+        224,
+        224
+      ]
+    }
+  ],
+  "tvf_do_letterbox": false,
+  "tvf_letterbox_fill": null,
+  "tvf_normalize_params": [
+    {
+      "inplace": false,
+      "mean": [
+        0.484375,
+        0.455078125,
+        0.40625
+      ],
+      "std": [
+        0.228515625,
+        0.2236328125,
+        0.224609375
+      ]
+    },
+    {
+      "inplace": false,
+      "mean": [
+        0.5,
+        0.5,
+        0.5
+      ],
+      "std": [
+        0.5,
+        0.5,
+        0.5
+      ]
+    }
+  ],
+  "tvf_resize_params": [
+    {
+      "antialias": true,
+      "interpolation": 3,
+      "max_size": null,
+      "size": [
+        224,
+        224
+      ]
+    },
+    {
+      "antialias": true,
+      "interpolation": 3,
+      "max_size": null,
+      "size": [
+        224,
+        224
+      ]
+    }
+  ],
+  "use_fused_vision_backbone": true
+}

processing_prismatic.py

@@ -0,0 +1,257 @@
+"""
+processing_prismatic.py
+
+HuggingFace-style preprocessor definitions for Prismatic VLMs, inheriting from `ProcessorMixin`. Default configuration
+specifies `siglip-224px+7b`.
+"""
+
+from typing import Any, ClassVar, List, Optional, Tuple, Union
+
+import timm.data
+import torch
+import torchvision.transforms.functional as TVF
+from PIL import Image
+from torchvision.transforms import CenterCrop, Compose, Normalize, Resize, ToTensor
+from transformers import PreTrainedTokenizerBase
+from transformers.image_processing_utils import BatchFeature, ImageProcessingMixin
+from transformers.processing_utils import ProcessorMixin
+from transformers.tokenization_utils import PaddingStrategy, PreTokenizedInput, TextInput, TruncationStrategy
+from transformers.utils import TensorType
+
+
+# === Image Processing ===
+def letterbox_pad_transform(image: Image.Image, padding_fill_value: Tuple[int, int, int]) -> Image.Image:
+    """Given a PIL.Image, pad to square by adding a symmetric border around the height/width."""
+    (w, h), max_wh = image.size, max(image.size)
+    horizontal_pad, vertical_pad = int((max_wh - w) / 2), int((max_wh - h) / 2)
+    padding = (horizontal_pad, vertical_pad, horizontal_pad, vertical_pad)
+
+    return TVF.pad(image, padding, fill=padding_fill_value, padding_mode="constant")
+
+
+class PrismaticImageProcessor(ImageProcessingMixin):
+    model_input_names: ClassVar[List[str]] = ["pixel_values"]
+
+    def __init__(
+        self,
+        use_fused_vision_backbone: bool = False,
+        image_resize_strategy: str = "letterbox",
+        input_sizes: Optional[List[Tuple[int, int, int]]] = None,
+        interpolations: Optional[List[str]] = None,
+        means: Optional[List[Tuple[float, float, float]]] = None,
+        stds: Optional[List[Tuple[float, float, float]]] = None,
+        **kwargs: str,
+    ) -> None:
+        """
+        Initialize a PrismaticImageProcessor as a wrapper around a torchvision transform; this transform will be
+        created by TIMM, and edited to follow our custom `image_resize_strategy` logic.
+
+        @param use_fused_vision_backbone: Boolean indicating single or fused (dual) vision backbone
+        @param image_resize_strategy: Prismatic image resize strategy in < resize-naive | resize-crop | letterbox >
+        @param input_size: [TIMM :: `data_cfg`] Input image size as tuple (channels, width, height)
+        @param interpolation: [TIMM :: `data_cfg`] Interpolation as string (default: "bicubic")
+        @param mean: [TIMM :: `data_cfg`] Normalization mean as float tuple (or two-tuple if `fused_backbone`)
+        @param std: [TIMM :: `data_cfg`] Normalization std as float tuple (or two-tuple if `fused_backbone`)
+        """
+        self.use_fused_vision_backbone = use_fused_vision_backbone
+        self.image_resize_strategy = image_resize_strategy
+
+        # Handle `None` default values
+        input_sizes = [(3, 224, 224)] if input_sizes is None else input_sizes
+        means = [(0.5, 0.5, 0.5)] if means is None else means
+        stds = [(0.5, 0.5, 0.5)] if stds is None else stds
+
+        # TIMM `data_cfg` Parameters
+        self.input_sizes, self.interpolations, self.means, self.stds = input_sizes, interpolations, means, stds
+
+        # Grab torchvision transforms via TIMM =>> need to parse for specific "functional" transform values!
+        self.tvf_resize_params, self.tvf_crop_params, self.tvf_normalize_params = [], [], []
+        self.tvf_do_letterbox, self.tvf_letterbox_fill = False, None
+
+        for idx in range(len(input_sizes)):
+            transform = timm.data.create_transform(
+                input_size=self.input_sizes[idx],
+                interpolation=self.interpolations[idx],
+                mean=self.means[idx],
+                std=self.stds[idx],
+                crop_pct=1.0,  # Set to 1.0 to ignore cropping (initial Resize sets `input_size`)
+                crop_mode="center",  # Default crop mode -- no-op when `crop_pct == 1.0`
+                is_training=False,  # No image augmentations when loading the transform!
+            )
+
+            # [Validation] Ensure appropriate transform structure, expected sizes
+            if not (
+                isinstance(transform, Compose)
+                and (len(transform.transforms) == 4)
+                and isinstance(transform.transforms[0], Resize)
+                and isinstance(transform.transforms[1], CenterCrop)
+                and isinstance(transform.transforms[2], ToTensor)
+                and isinstance(transform.transforms[3], Normalize)
+                and (transform.transforms[0].size == self.input_sizes[idx][-1])
+                and (transform.transforms[1].size == self.input_sizes[idx][-2:])
+            ):
+                raise ValueError(f"Unexpected TIMM image transformation structure/sizes: `{transform}`")
+
+            # HF Image Processors *must* be JSON-serializable; as such, cannot have torchvision. as an attribute.
+            #   => Instead, we're going to parse the transform and call "torchvision.transforms.functional" (`tvf`)
+            resize_t, crop_t, norm_t = transform.transforms[0], transform.transforms[1], transform.transforms[3]
+            self.tvf_resize_params.append(
+                {
+                    "size": resize_t.size,
+                    "interpolation": TVF.pil_modes_mapping[resize_t.interpolation],
+                    "max_size": None,
+                    "antialias": True,
+                }
+            )
+            self.tvf_crop_params.append({"output_size": crop_t.size})
+            self.tvf_normalize_params.append(
+                {
+                    "mean": norm_t.mean.float().numpy().tolist(),
+                    "std": norm_t.std.float().numpy().tolist(),
+                    "inplace": False,
+                }
+            )
+            self.tvf_do_letterbox, self.tvf_letterbox_fill = False, None
+
+            # Handle Prismatic `image_resize_strategy`
+            if self.image_resize_strategy == "resize-naive":
+                self.tvf_resize_params[idx]["size"] = (resize_t.size, resize_t.size)
+            elif self.image_resize_strategy == "letterbox":
+                self.tvf_do_letterbox, self.tvf_letterbox_fill = True, tuple([int(x * 255) for x in self.means[idx]])
+            elif self.image_resize_strategy == "resize-crop":
+                pass
+            else:
+                raise ValueError(f"Image resize strategy `{self.image_resize_strategy}` is not supported!")
+
+        # Dispatch **kwargs to super()
+        super().__init__(**kwargs)
+
+    def apply_transform(self, img: Image.Image) -> torch.Tensor:
+        """Apply `functional` variant of TIMM's Transform = Compose([Resize -> CenterCrop -> ToTensor -> Normalize])"""
+        if self.tvf_do_letterbox:
+            img = letterbox_pad_transform(img, self.tvf_letterbox_fill)
+
+        # [Contract] Fused Backbones expect "channel-stacked" inputs; we'll unpack on the model side!
+        imgs_t = []
+        for idx in range(len(self.input_sizes)):
+            img_idx = TVF.resize(img, **self.tvf_resize_params[idx])
+            img_idx = TVF.center_crop(img_idx, **self.tvf_crop_params[idx])
+            img_idx_t = TVF.to_tensor(img_idx)
+            img_idx_t = TVF.normalize(img_idx_t, **self.tvf_normalize_params[idx])
+            imgs_t.append(img_idx_t)
+
+        # [Contract] `imgs_t` is a list of Tensors of shape [3, input_size, input_size]; stack along dim = 0
+        img_t = torch.vstack(imgs_t)
+
+        return img_t
+
+    def preprocess(
+        self,
+        images: Union[Image.Image, List[Image.Image]],
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        **_: str,
+    ) -> BatchFeature:
+        """
+        Preprocess an image (or batch of images); note that unlike the `transformers :: BaseImageProcessor` we
+        explicitly only handle PIL.Image.Image instances for simplicity.
+
+        @param images: A (batch of) PIL.Image.Image instance(s) to preprocess.
+        @param return_tensors: BatchFeature default Tensor format (e.g., "pt" for torch); if None, returns np.ndarray
+
+        @return: Instance of `transformers :: BatchFeature` with a single key "pixel_values"
+        """
+        if not isinstance(images, list):
+            images = [images]
+
+        # Apply `self.img_transform` to each image (will return list of torch.Tensors); stack into "batched" Tensor
+        pixel_values = torch.stack([self.apply_transform(img.convert("RGB")) for img in images])
+
+        # Return BatchFeature =>> note that for compatibility, constructor expects Dict[str, np.ndarray], so we convert
+        return BatchFeature(data={"pixel_values": pixel_values.float().numpy()}, tensor_type=return_tensors)
+
+    def __call__(self, images: Union[Image.Image, List[Image.Image]], **kwargs) -> BatchFeature:
+        return self.preprocess(images, **kwargs)
+
+
+# === PrismaticProcessor =>> Wraps both ImageProcessor and Tokenizer ===
+#   =>> https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava/processing_llava.py
+class PrismaticProcessor(ProcessorMixin):
+    attributes: ClassVar[List[str]] = ["image_processor", "tokenizer"]
+    image_processor_class: str = "AutoImageProcessor"
+    tokenizer_class: str = "AutoTokenizer"
+
+    def __init__(
+        self,
+        image_processor: Optional[ImageProcessingMixin] = None,
+        tokenizer: Optional[PreTrainedTokenizerBase] = None,
+    ) -> None:
+        super().__init__(image_processor, tokenizer)
+
+    def __call__(
+        self,
+        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]],
+        images: Union[Image.Image, List[Image.Image]],
+        padding: Union[bool, str, PaddingStrategy] = False,
+        truncation: Optional[Union[bool, str, TruncationStrategy]] = None,
+        max_length: Optional[int] = None,
+        return_tensors: Optional[Union[str, TensorType]] = TensorType.PYTORCH,
+    ) -> BatchFeature:
+        """
+        Preprocess a given (batch) of text/images for a Prismatic VLM; forwards text to the underlying LLM's tokenizer,
+        forwards images to PrismaticImageProcessor.
+
+        @param text: The (batch) of text to encode; must be a string or list of strings.
+        @param images: A (batch of) PIL.Image.Image instance(s) to preprocess.
+        @param padding: Sequence padding strategy (if multiple specified) in < True = "longest" | "max_length" | False >
+        @param truncation: Truncation strategy for the output sequences; requires `max_length` to be specified
+        @param max_length: Maximum length (in tokens) to truncate
+        @param return_tensors: Type of return tensors (usually "pt" or TensorType.PYTORCH)
+
+        @return: BatchFeature with keys for `input_ids`, `attention_mask` and `pixel_values`.
+        """
+        pixel_values = self.image_processor(images, return_tensors=return_tensors)["pixel_values"]
+        text_inputs = self.tokenizer(
+            text, return_tensors=return_tensors, padding=padding, truncation=truncation, max_length=max_length
+        )
+
+        # [Validate] Need same number of images and text inputs!
+        if pixel_values.shape[0] != text_inputs.input_ids.shape[0]:
+            raise ValueError("Batch is malformed; expected same number of images and text inputs!")
+
+        return BatchFeature(data={**text_inputs, "pixel_values": pixel_values})
+
+    # === Tokenizer Dispatch Utilities =>> check `PreTrainedTokenizerBase` for documentation ===
+    def batch_decode(
+        self,
+        sequences: Union[List[int], List[List[int]], torch.Tensor, Any],  # `Any` = np.ndarray | tf.Tensor
+        skip_special_tokens: bool = False,
+        clean_up_tokenization_spaces: Optional[bool] = None,
+        **kwargs: str,
+    ) -> List[str]:
+        return self.tokenizer.batch_decode(
+            sequences=sequences,
+            skip_special_tokens=skip_special_tokens,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            **kwargs,
+        )
+
+    def decode(
+        self,
+        token_ids: Union[int, List[int], torch.Tensor, Any],  # `Any` = np.ndarray | tf.Tensor
+        skip_special_tokens: bool = False,
+        clean_up_tokenization_spaces: Optional[bool] = None,
+        **kwargs: str,
+    ) -> str:
+        return self.tokenizer.decode(
+            token_ids=token_ids,
+            skip_special_tokens=skip_special_tokens,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            **kwargs,
+        )
+
+    @property
+    def model_input_names(self) -> List[str]:
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+
+        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))

processor_config.json

@@ -0,0 +1,6 @@
+{
+  "auto_map": {
+    "AutoProcessor": "processing_prismatic.PrismaticProcessor"
+  },
+  "processor_class": "PrismaticProcessor"
+}

special_tokens_map.json

@@ -0,0 +1,30 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<PAD>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347
+size 499723

tokenizer_config.json

@@ -0,0 +1,53 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32000": {
+      "content": "<PAD>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "auto_map": {
+    "AutoProcessor": "processing_prismatic.PrismaticProcessor"
+  },
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "legacy": false,
+  "model_max_length": 2048,
+  "pad_token": "<PAD>",
+  "padding_side": "right",
+  "processor_class": "PrismaticProcessor",
+  "sp_model_kwargs": {},
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false
+}