Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

added_tokens.json

@@ -0,0 +1,24 @@
+{
+  "</tool_call>": 151658,
+  "<tool_call>": 151657,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

attention_mask.py

@@ -0,0 +1,81 @@
+from typing import Optional
+
+import torch
+
+
+def _make_causal_mask(
+    attention_mask: torch.Tensor, dtype: torch.dtype, device: torch.device
+):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = attention_mask.shape
+    mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
+    mask_cond = torch.arange(mask.size(-1), device=device)
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len)
+
+
+def _make_2dvison_mask(column_mask, dtype: torch.dtype, device: torch.device):
+    """
+    """
+    bsz, seq_length = column_mask.shape
+    cross_mask = torch.zeros((bsz, 1, seq_length, seq_length), dtype=dtype, device=device)
+
+    # 找到连续的 1 的区间
+    start = None
+    for bsz_idx in range(bsz):
+        for i in range(seq_length):
+            if column_mask[bsz_idx, i] == 1:
+                if start is None:
+                    start = i
+            else:
+                if start is not None:
+                    # 填充区间
+                    cross_mask[bsz_idx, 0, start:i, start:i] = 1
+                    start = None
+
+        # 处理最后一个区间
+        if start is not None:
+            cross_mask[bsz_idx, 0, start:seq_length, start:seq_length] = 1
+
+    return cross_mask
+
+
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill_(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
+
+
+def make_mask(attention_mask: torch.Tensor, dtype: torch.dtype=None, device: torch.device=None, mode: str="default", vision_mask: torch.Tensor=None, ):
+    if dtype is None:
+        dtype = attention_mask.dtype
+    if device is None:
+        device = attention_mask.device      
+    expanded_attn_mask = _expand_mask(attention_mask, dtype).to(device)
+    causal_mask = _make_causal_mask(attention_mask, dtype, device).to(device)
+    if mode == "default":
+        return attention_mask
+    else:
+        assert vision_mask is not None, "vision_mask is None"
+        vision_mask = vision_mask.to(device)
+        bsz, seq_length = attention_mask.shape
+        vision_mask_bg = vision_mask[:, None, :, None]
+        vision_mask_2d = _make_2dvison_mask(vision_mask, dtype, device)
+        if mode == "bidirectional":
+            mask = expanded_attn_mask + causal_mask
+            mask = mask.clone().masked_fill_(vision_mask_2d.to(torch.bool), 0)
+            return mask 
+        else:
+            raise NotImplementedError(f"mode {mode} is not implemented")

chat_template.json

@@ -0,0 +1,3 @@
+{
+  "chat_template": "{% set image_count = namespace(value=0) %}{% set video_count = namespace(value=0) %}{% for message in messages %}{% if loop.first and message['role'] != 'system' %}<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n{% endif %}<|im_start|>{{ message['role'] }}\n{% if message['content'] is string %}{{ message['content'] }}<|im_end|>\n{% else %}{% for content in message['content'] %}{% if content['type'] == 'image' or 'image' in content or 'image_url' in content %}{% set image_count.value = image_count.value + 1 %}{% if add_vision_id %}Picture {{ image_count.value }}: {% endif %}<|vision_start|><|image_pad|><|vision_end|>{% elif content['type'] == 'video' or 'video' in content %}{% set video_count.value = video_count.value + 1 %}{% if add_vision_id %}Video {{ video_count.value }}: {% endif %}<|vision_start|><|video_pad|><|vision_end|>{% elif 'text' in content %}{{ content['text'] }}{% endif %}{% endfor %}<|im_end|>\n{% endif %}{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}"
+}

config.json

@@ -0,0 +1,52 @@
+{
+  "architectures": [
+    "VoRAForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_vora.VoRAConfig",
+    "AutoModelForCausalLM": "modeling_vora.VoRAForCausalLM"
+  },
+  "aux_vision": "",
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "hidden_act": "silu",
+  "hidden_size": 3584,
+  "image_size": 448,
+  "initializer_range": 0.02,
+  "intermediate_size": 18944,
+  "llm": "/mnt/bn/wh-data/data/models/Qwen2.5-7B-Instruct",
+  "lora": {
+    "layers": 24,
+    "r": -1,
+    "target_modules": [
+      "self_attn.q_proj",
+      "self_attn.k_proj",
+      "self_attn.v_proj",
+      "self_attn.o_proj",
+      "mlp.up_proj",
+      "mlp.gate_proj",
+      "mlp.down_proj"
+    ]
+  },
+  "max_position_embeddings": 32768,
+  "max_window_layers": 28,
+  "model_type": "vora",
+  "num_attention_heads": 28,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 4,
+  "patch_size": 14,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000.0,
+  "sliding_window": 131072,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float32",
+  "transformers_version": "4.50.3",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vision_attention_mask": "bidirectional",
+  "vision_embedding_intermediate_size": 1536,
+  "vision_embedding_type": "AIMv2",
+  "vocab_size": 152064
+}

configuration_vora.py

@@ -0,0 +1,35 @@
+from typing import Any
+
+from transformers.configuration_utils import PretrainedConfig
+
+__all__ = ["VoRAConfig"]
+
+
+class VoRAConfig(PretrainedConfig):
+    model_type = "vora"
+    _auto_class = "AutoConfig"
+
+    def __init__(
+        self,
+        llm: str = "",
+        aux_vision: str = "",
+        lora: dict = {}, 
+        image_size: int = 448,
+        vision_embedding_type: str = "",
+        vision_embedding_intermediate_size: int = 1536,
+        patch_size: int = 14,
+        vision_attention_mask: str = "bidirectional",
+        rms_norm_eps: float = 1e-5,
+        **kwargs: Any,
+    ):
+        super().__init__(**kwargs)
+        self.llm = llm
+        self.aux_vision = aux_vision
+        self.lora = lora
+        self.image_size = image_size
+        self.vision_embedding_type = vision_embedding_type
+        self.vision_embedding_intermediate_size = vision_embedding_intermediate_size
+        self.patch_size = patch_size
+        self.vision_attention_mask = vision_attention_mask
+        self.rms_norm_eps = rms_norm_eps
+    

generation_config.json

@@ -0,0 +1,14 @@
+{
+  "bos_token_id": 151643,
+  "pad_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "repetition_penalty": 1.05,
+  "temperature": 0.7,
+  "top_p": 0.8,
+  "top_k": 20,
+  "transformers_version": "4.37.0"
+}

lora.py

@@ -0,0 +1,91 @@
+import torch
+import types
+import math
+from torch import nn
+import torch.nn.functional as F
+
+
+QWEN2_TARGET_MODULES = [
+    "self_attn.q_proj",
+    "self_attn.k_proj",
+    "self_attn.v_proj",
+    "self_attn.o_proj",
+    "mlp.up_proj",
+    "mlp.gate_proj",
+    "mlp.down_proj",
+]
+
+
+class LoRALayer(nn.Linear):
+    def __init__(
+        self, 
+        in_features: int, 
+        out_features: int, 
+        r: int = 1024,
+        **kwargs
+    ):
+        nn.Linear.__init__(self, in_features, out_features)
+        if r < 0:
+            self.forward = self.naive_forward
+        else:
+            # we elimate lora_alpha here bc we find it unnecessary in VoRA
+            self.lora_A = nn.Linear(in_features, r, bias=False)
+            self.lora_B = nn.Linear(r, out_features, bias=False)
+            nn.init.kaiming_uniform_(self.lora_A.weight, a=math.sqrt(5))
+            nn.init.zeros_(self.lora_B.weight)
+        
+    def forward(self, x: torch.Tensor):
+        intermediate = F.linear(x, self.weight, bias=self.bias)
+        result = intermediate + self.lora_B(self.lora_A(x))
+        return result
+
+    def naive_forward(self, x: torch.Tensor):
+        return F.linear(x, self.weight, bias=self.bias)
+
+def _get_submodules(self, key):
+    parent = self.get_submodule(".".join(key.split(".")[:-1]))
+    target_name = key.split(".")[-1]
+    target = self.get_submodule(key)
+    return parent, target, target_name
+
+def _find_and_replace(self, lora_params):
+    target_modules = lora_params["target_modules"]
+
+    for llm_module_name in target_modules:
+        parent, target, target_name = self._get_submodules(llm_module_name)
+        vora_layer = LoRALayer(
+            target.in_features,
+            target.out_features,
+            **lora_params
+        )
+        self._replace_module(parent, target_name, vora_layer, target)
+
+def _replace_module(self, parent_module, child_name, new_module, old_module):
+    setattr(parent_module, child_name, new_module)
+    new_module.weight = old_module.weight
+    if old_module.bias is not None:
+        new_module.bias = old_module.bias
+    if getattr(old_module, "state", None) is not None:
+        new_module.state = old_module.state
+        new_module.to(old_module.weight.device)
+
+def apply_lora(llm, lora_params={"layers": "all", "r": 1024, "target_modules": QWEN2_TARGET_MODULES}):
+    llm_num_layers = llm.config.num_hidden_layers
+    total_layers = lora_params.get("layers", "all")
+
+    # -------------------- validation check ---------------------
+    if isinstance(total_layers, str):
+        if total_layers.lower() == "all":
+            total_layers = list(range(llm_num_layers))
+    else:
+        assert isinstance(total_layers, int), "total_layers must be an integer or 'all'"
+        total_layers = list(range(total_layers))
+    # -------------------- validation check ---------------------
+
+    # -------------------- replace llm layers ---------------------
+    for i in total_layers:
+        llm_layer = llm.model.layers[i]
+        llm_layer._get_submodules = types.MethodType(_get_submodules, llm_layer)
+        llm_layer._find_and_replace = types.MethodType(_find_and_replace, llm_layer)
+        llm_layer._replace_module = types.MethodType(_replace_module, llm_layer)
+        llm_layer._find_and_replace(lora_params)

model-00001-of-00007.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a2164771a66a6d1bded7a320400a72dd2d6754c6c52a70a9299db0d1f319a59
+size 4977229368

model-00002-of-00007.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3624d28b4695a26768fe361090f3f126d5e10ec5c44a006d33886cf7e1d6ffff
+size 4779540384

model-00003-of-00007.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3f739dc6bba6a644182f3972e28841e12d375a72d34a3bb6e5d9e4b43b4c01fa
+size 4933723424

model-00004-of-00007.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fa7e59c2059f06ba171f40c09f3ea8375c551ecfb490e35d165342795ab1ac7
+size 4933723464

model-00005-of-00007.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6a9ed61468f6c7a9064b3a516ef5dac9b28ad282cee9ddf044ec511ef1dc60b
+size 4999770352

model-00006-of-00007.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:29fe7750560a75fe7348499eec2cf79d275b00b904c363e5ea300e9ec0b94871
+size 3662865360

model-00007-of-00007.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9bc90d888b4d1fb1e1d10b58f7bc64df1dfb9dfd8d9bda05b06166b1b65fd0e9
+size 2211926704

model.safetensors.index.json

@@ -0,0 +1,447 @@
+{
+  "metadata": {
+    "total_size": 30498727936
+  },
+  "weight_map": {
+    "llm.lm_head.weight": "model-00007-of-00007.safetensors",
+    "llm.model.embed_tokens.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.input_layernorm.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.mlp.down_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.mlp.down_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.mlp.gate_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.mlp.gate_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.mlp.up_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.mlp.up_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.post_attention_layernorm.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.self_attn.k_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.self_attn.k_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.self_attn.o_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.self_attn.o_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.self_attn.q_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.self_attn.q_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.self_attn.v_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.self_attn.v_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.input_layernorm.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.mlp.down_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.mlp.down_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.mlp.gate_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.mlp.gate_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.mlp.up_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.mlp.up_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.post_attention_layernorm.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.self_attn.k_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.self_attn.k_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.self_attn.o_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.self_attn.o_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.self_attn.q_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.self_attn.q_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.self_attn.v_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.self_attn.v_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.10.input_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.mlp.down_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.mlp.down_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.mlp.gate_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.mlp.gate_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.mlp.up_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.mlp.up_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.post_attention_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.self_attn.k_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.self_attn.k_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.self_attn.o_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.self_attn.o_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.self_attn.q_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.self_attn.q_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.self_attn.v_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.self_attn.v_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.input_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.mlp.down_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.mlp.down_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.mlp.gate_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.mlp.gate_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.mlp.up_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.mlp.up_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.post_attention_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.self_attn.k_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.self_attn.k_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.self_attn.o_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.self_attn.o_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.self_attn.q_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.self_attn.q_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.self_attn.v_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.self_attn.v_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.input_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.mlp.down_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.mlp.down_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.mlp.gate_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.mlp.gate_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.mlp.up_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.mlp.up_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.post_attention_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.self_attn.k_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.self_attn.k_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.self_attn.o_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.self_attn.o_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.self_attn.q_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.self_attn.q_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.self_attn.v_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.self_attn.v_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.13.input_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.13.mlp.down_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.13.mlp.down_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.13.mlp.gate_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.13.mlp.gate_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.13.mlp.up_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.13.mlp.up_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.13.post_attention_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.13.self_attn.k_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.13.self_attn.k_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.13.self_attn.o_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.13.self_attn.o_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.13.self_attn.q_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.13.self_attn.q_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.13.self_attn.v_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.13.self_attn.v_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.14.input_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.14.mlp.down_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.14.mlp.down_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.14.mlp.gate_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.14.mlp.gate_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.14.mlp.up_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.14.mlp.up_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.14.post_attention_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.14.self_attn.k_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.14.self_attn.k_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.14.self_attn.o_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.14.self_attn.o_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.14.self_attn.q_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.14.self_attn.q_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.14.self_attn.v_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.14.self_attn.v_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.input_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.mlp.down_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.mlp.down_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.mlp.gate_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.mlp.gate_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.mlp.up_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.mlp.up_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.post_attention_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.self_attn.k_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.self_attn.k_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.self_attn.o_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.self_attn.o_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.self_attn.q_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.self_attn.q_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.self_attn.v_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.self_attn.v_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.input_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.mlp.down_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.mlp.down_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.mlp.gate_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.mlp.gate_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.mlp.up_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.mlp.up_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.post_attention_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.self_attn.k_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.self_attn.k_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.self_attn.o_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.self_attn.o_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.self_attn.q_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.self_attn.q_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.self_attn.v_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.self_attn.v_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.input_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.mlp.down_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.mlp.down_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.mlp.gate_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.mlp.gate_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.mlp.up_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.mlp.up_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.post_attention_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.self_attn.k_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.self_attn.k_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.self_attn.o_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.self_attn.o_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.self_attn.q_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.self_attn.q_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.self_attn.v_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.self_attn.v_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.input_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.18.mlp.down_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.18.mlp.down_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.18.mlp.gate_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.mlp.gate_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.mlp.up_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.mlp.up_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.post_attention_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.18.self_attn.k_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.self_attn.k_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.self_attn.o_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.self_attn.o_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.self_attn.q_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.self_attn.q_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.self_attn.v_proj.bias": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.self_attn.v_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.19.input_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.19.mlp.down_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.19.mlp.down_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.19.mlp.gate_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.19.mlp.gate_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.19.mlp.up_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.19.mlp.up_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.19.post_attention_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.19.self_attn.k_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.19.self_attn.k_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.19.self_attn.o_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.19.self_attn.o_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.19.self_attn.q_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.19.self_attn.q_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.19.self_attn.v_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.19.self_attn.v_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.2.input_layernorm.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.mlp.down_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.mlp.down_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.mlp.gate_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.mlp.gate_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.mlp.up_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.mlp.up_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.post_attention_layernorm.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.self_attn.k_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.self_attn.k_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.self_attn.o_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.self_attn.o_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.self_attn.q_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.self_attn.q_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.self_attn.v_proj.bias": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.self_attn.v_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.20.input_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.mlp.down_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.mlp.down_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.mlp.gate_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.mlp.gate_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.mlp.up_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.mlp.up_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.post_attention_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.self_attn.k_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.self_attn.k_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.self_attn.o_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.self_attn.o_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.self_attn.q_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.self_attn.q_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.self_attn.v_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.self_attn.v_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.input_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.mlp.down_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.mlp.down_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.mlp.gate_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.mlp.gate_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.mlp.up_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.mlp.up_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.post_attention_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.self_attn.k_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.self_attn.k_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.self_attn.o_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.self_attn.o_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.self_attn.q_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.self_attn.q_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.self_attn.v_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.self_attn.v_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.input_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.mlp.down_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.mlp.down_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.mlp.gate_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.mlp.gate_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.mlp.up_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.mlp.up_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.post_attention_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.self_attn.k_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.self_attn.k_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.self_attn.o_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.self_attn.o_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.self_attn.q_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.self_attn.q_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.self_attn.v_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.self_attn.v_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.input_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.mlp.down_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.mlp.down_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.mlp.gate_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.mlp.gate_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.mlp.up_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.mlp.up_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.post_attention_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.self_attn.k_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.self_attn.k_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.self_attn.o_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.self_attn.o_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.self_attn.q_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.self_attn.q_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.self_attn.v_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.self_attn.v_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.24.input_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.24.mlp.down_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.24.mlp.gate_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.24.mlp.up_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.24.post_attention_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.24.self_attn.k_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.24.self_attn.k_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.24.self_attn.o_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.24.self_attn.q_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.24.self_attn.q_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.24.self_attn.v_proj.bias": "model-00005-of-00007.safetensors",
+    "llm.model.layers.24.self_attn.v_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.25.input_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.25.mlp.down_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.25.mlp.gate_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.25.mlp.up_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.25.post_attention_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.25.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
+    "llm.model.layers.25.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.25.self_attn.o_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.25.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
+    "llm.model.layers.25.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.25.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
+    "llm.model.layers.25.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.input_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.mlp.down_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.mlp.gate_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.mlp.up_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.post_attention_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.self_attn.o_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.input_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.mlp.down_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.mlp.gate_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.mlp.up_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.post_attention_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.self_attn.o_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.3.input_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.mlp.down_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.mlp.down_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.mlp.gate_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.mlp.gate_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.mlp.up_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.mlp.up_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.post_attention_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.self_attn.k_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.self_attn.k_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.self_attn.o_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.self_attn.o_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.self_attn.q_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.self_attn.q_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.self_attn.v_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.self_attn.v_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.input_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.mlp.down_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.mlp.down_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.mlp.gate_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.mlp.gate_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.mlp.up_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.mlp.up_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.post_attention_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.self_attn.k_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.self_attn.k_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.self_attn.o_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.self_attn.o_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.self_attn.q_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.self_attn.q_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.self_attn.v_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.self_attn.v_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.input_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.mlp.down_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.mlp.down_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.mlp.gate_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.mlp.gate_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.mlp.up_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.mlp.up_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.post_attention_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.self_attn.k_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.self_attn.k_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.self_attn.o_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.self_attn.o_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.self_attn.q_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.self_attn.q_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.self_attn.v_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.self_attn.v_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.input_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.mlp.down_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.mlp.down_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.mlp.gate_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.mlp.gate_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.mlp.up_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.mlp.up_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.post_attention_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.self_attn.k_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.self_attn.k_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.self_attn.o_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.self_attn.o_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.self_attn.q_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.self_attn.q_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.self_attn.v_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.self_attn.v_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.input_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.mlp.down_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.mlp.down_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.mlp.gate_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.mlp.gate_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.mlp.up_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.mlp.up_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.post_attention_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.self_attn.k_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.self_attn.k_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.self_attn.o_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.self_attn.o_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.self_attn.q_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.self_attn.q_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.self_attn.v_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.self_attn.v_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.8.input_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.8.mlp.down_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.8.mlp.down_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.8.mlp.gate_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.8.mlp.gate_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.8.mlp.up_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.8.mlp.up_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.8.post_attention_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.8.self_attn.k_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.8.self_attn.k_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.8.self_attn.o_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.8.self_attn.o_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.8.self_attn.q_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.8.self_attn.q_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.8.self_attn.v_proj.bias": "model-00002-of-00007.safetensors",
+    "llm.model.layers.8.self_attn.v_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.9.input_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.mlp.down_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.mlp.down_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.mlp.gate_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.mlp.gate_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.mlp.up_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.mlp.up_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.post_attention_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.self_attn.k_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.self_attn.k_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.self_attn.o_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.self_attn.o_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.self_attn.q_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.self_attn.q_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.self_attn.v_proj.bias": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.self_attn.v_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.norm.weight": "model-00006-of-00007.safetensors",
+    "vision_embedding.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vision_embedding.patchifier.norm.weight": "model-00007-of-00007.safetensors",
+    "vision_embedding.patchifier.proj.bias": "model-00007-of-00007.safetensors",
+    "vision_embedding.patchifier.proj.weight": "model-00007-of-00007.safetensors",
+    "vision_embedding.pos_embed": "model-00007-of-00007.safetensors"
+  }
+}

modeling_vora.py

@@ -0,0 +1,219 @@
+import os.path as osp
+
+import torch
+import torch.distributed as dist
+from transformers import (
+    AutoModelForCausalLM,
+    AutoTokenizer,
+    AutoConfig,
+    PreTrainedModel,
+    PretrainedConfig,
+    Qwen2ForCausalLM,
+)
+
+from .attention_mask import make_mask
+from .configuration_vora import VoRAConfig
+from .lora import apply_lora
+from .vision_embedding import build_vision_embedding
+from .vora_generation_utils import (
+    VoraGenerationMixin,
+    custom_prepare_4d_causal_attention_mask_with_cache_position,
+)
+
+
+class VoRAForCausalLM(PreTrainedModel):   
+    config_class = VoRAConfig 
+    _auto_class = 'AutoModelForCausalLM'
+    supports_gradient_checkpointing = True
+
+    def __init__(self, config: PretrainedConfig = VoRAConfig()):
+        super().__init__(config)
+        self.config = config
+        # -------------- Setup LLM ---------------------
+        self.llm = Qwen2ForCausalLM(config)
+        
+        # monkey patch for generation
+        self.llm.__class__ = type(self.llm.__class__.__name__, (self.llm.__class__, VoraGenerationMixin), {})
+        self.llm.model._prepare_4d_causal_attention_mask_with_cache_position = staticmethod(custom_prepare_4d_causal_attention_mask_with_cache_position)
+        dtype = self.llm.dtype
+        
+        # hacking for multi-processor infer
+        self._tp_plan = self.llm._tp_plan
+
+        self.config.update(self.llm.config.to_dict())
+        # ----------------------------------------------
+
+
+        # -------------- Setup LoRA -------------------
+        if config.lora:
+            for _, param in self.llm.named_parameters():
+                param.requires_grad = False
+            apply_lora(self.llm, config.lora)
+            self.llm.to(dtype)
+        # ----------------------------------------------
+
+        # ------------ Setup Vision Embedding ----------
+        self.vision_embedding = build_vision_embedding(config, self.llm.config.hidden_size)
+        # ----------------------------------------------
+
+    def _encode_vision(self, images, n_frames):
+        # TODO: we need a more elegant way here to deal with mixed image and pure text training
+        if images.size(0) > 0:
+            vision_embeds = self.vision_embedding(images)
+        else:
+            # FIXME: hacking for deepspeed training
+            # we feed a dummy image tensor (1, 3, H, W) into vision_encoder when training a pure-text batch
+            images = images.new_zeros((1, *images.shape[1:]))
+            vision_embeds = self.vision_embedding(images)[0:0]
+        vision_embeds = vision_embeds.split(n_frames, dim=0)
+        attention_mask = [torch.ones(feature.size()[:-1], dtype=torch.long).to(feature.device) for feature in vision_embeds]
+        vision_targets = [torch.ones(feature.size(), dtype=torch.long).to(feature.device).fill_(-100) for feature in attention_mask]
+
+        image_shapes = images.shape[-2:]
+
+        return vision_embeds, attention_mask, vision_targets, image_shapes
+
+    def _concat_embedding(self, vision_encode_out, batch, vision_placeholder_index, left_padding=False, pad_token_id=0):
+        """ concat vision and text
+        """
+
+        vision_embeds, vision_atts, vision_targets, _ = vision_encode_out
+
+        input_embeds = []
+        attention_mask = []
+        targets = []
+        vision_mask = []  # only set vision embeds as 1, text as 0, for aux loss
+
+        for cur_batch_idx, cur_input_ids in enumerate(batch["input_ids"]):
+            cur_vision_embeds = vision_embeds[cur_batch_idx]
+            cur_vision_attn = vision_atts[cur_batch_idx]
+            cur_vision_targets = vision_targets[cur_batch_idx]
+            cur_attn_masks = batch["attention_mask"][cur_batch_idx]
+
+            image_token_indices = torch.where(cur_input_ids == vision_placeholder_index)[0]
+            cur_image_num = len(image_token_indices)
+            image_token_indices = list(image_token_indices) + [cur_input_ids.shape[0]]
+
+            cur_input_embeds = []
+            cur_attention_mask = []
+            cur_target = []
+            cur_vision_mask = []
+
+            # convert text before 1st <image> to embedding
+            image_token_index = image_token_indices[0]
+
+            cur_input_embeds.append(
+                self.llm.get_input_embeddings()(cur_input_ids[:image_token_index]),
+            )
+            cur_attention_mask.append(
+                cur_attn_masks[:image_token_index],
+            )
+            cur_vision_mask.append(
+                torch.zeros_like(cur_attn_masks[:image_token_index]).to(cur_attn_masks.device),
+            )
+            if "labels" in batch:
+                cur_target.append(
+                    batch["labels"][cur_batch_idx, :image_token_index],
+                )
+
+            if batch.get("vison_placeholder_mode", 0) == 1:
+                assert cur_image_num <= 1, "multiple video input is not supported"
+                cur_vision_embeds = cur_vision_embeds.unsqueeze(0)
+                cur_vision_attn = cur_vision_attn.unsqueeze(0)
+                cur_vision_targets = cur_vision_targets.unsqueeze(0)
+            assert cur_image_num == len(cur_vision_embeds), \
+                f"Size mismatch! cur_image_num: {cur_image_num}, len(cur_vision_embeds): {len(cur_vision_embeds)} {len(cur_vision_embeds)} \
+                    in {batch['prompt'][cur_batch_idx]} & {batch['gt'][cur_batch_idx]} & {batch['input_ids'][cur_batch_idx]}"
+            # convert each <image> xxx group into embedding
+            text_embedding = self.llm.get_input_embeddings()(cur_input_ids.relu())
+            for i in range(0, cur_image_num):
+                image_token_index = image_token_indices[i]
+                cur_input_embeds.extend([
+                    cur_vision_embeds[i],
+                    text_embedding[image_token_index+1:image_token_indices[i+1]]
+                ])
+                cur_attention_mask.extend([
+                    cur_vision_attn[i],
+                    cur_attn_masks[image_token_index+1:image_token_indices[i+1]]
+                ])
+                cur_vision_mask.extend([
+                    torch.ones_like(cur_vision_attn[i]).to(cur_vision_attn[i].device),
+                    torch.zeros_like(cur_attn_masks[image_token_index+1:image_token_indices[i+1]]).to(cur_vision_attn[i].device),
+                ])
+                if "labels" in batch:
+                    cur_target.extend([
+                        cur_vision_targets[i],
+                        batch["labels"][cur_batch_idx, image_token_index+1:image_token_indices[i+1]],
+                    ])
+
+            input_embeds.append(torch.cat(cur_input_embeds))
+            attention_mask.append(torch.cat(cur_attention_mask))
+            vision_mask.append(torch.cat(cur_vision_mask))
+            if "labels" in batch:
+                targets.append(torch.cat(cur_target))
+
+        # padding
+        n_tokens = [embed.shape[0] for embed in input_embeds]
+
+        max_token = max(n_tokens)
+
+        for i in range(len(input_embeds)):
+            if max_token > n_tokens[i]:
+                pad_token = torch.tensor([pad_token_id] * (max_token - n_tokens[i]))
+                pad_embedding = self.llm.get_input_embeddings()(pad_token.to(batch["attention_mask"][i].device))
+                pad_attention = torch.zeros(pad_embedding.shape[0], dtype=torch.long).to(batch["attention_mask"][i].device)
+                pad_targets = torch.ones(pad_attention.size(), dtype=torch.long).to(batch["attention_mask"][i].device).fill_(-100)
+
+                if left_padding:
+                    input_embeds[i] = torch.cat([pad_embedding, input_embeds[i]])
+                    attention_mask[i] = torch.cat([pad_attention, attention_mask[i]])
+                    vision_mask[i] = torch.cat([pad_attention, vision_mask[i]])
+                    if "labels" in batch:
+                        targets[i] = torch.cat([pad_targets, targets[i]])
+                else:
+                    input_embeds[i] = torch.cat([input_embeds[i], pad_embedding])
+                    attention_mask[i] = torch.cat([attention_mask[i], pad_attention])
+                    vision_mask[i] = torch.cat([vision_mask[i], pad_attention])
+                    if "labels" in batch:
+                        targets[i] = torch.cat([targets[i], pad_targets])
+
+        inputs_embeds = torch.stack(input_embeds, dim=0).type(self.llm.dtype)
+        attention_mask = torch.stack(attention_mask, dim=0)
+        vision_mask = torch.stack(vision_mask, dim=0).to(attention_mask.device)
+
+        if len(targets) > 0:
+            targets = torch.stack(targets, dim=0)
+
+        attention_mask = make_mask(
+            attention_mask, 
+            mode=self.config.vision_attention_mask, 
+            vision_mask=vision_mask, 
+            dtype=inputs_embeds.dtype
+        )
+
+        return inputs_embeds, attention_mask, targets, vision_mask
+        
+    def generate(self, batch, **generate_params):
+
+        with torch.amp.autocast(
+            device_type="cuda",
+            enabled=(self.device != torch.device("cpu"))
+        ):
+            # get vision token
+            vision_placeholder_index = batch.pop("vision_placeholder_index")
+
+            # get vision features
+            images, n_frames = batch["frames"], batch["n_frames"]
+            vision_encode_out = self._encode_vision(images, n_frames)
+
+            inputs_embeds, attention_mask, _, _ = self._concat_embedding(
+                vision_encode_out, batch, vision_placeholder_index, left_padding=False, pad_token_id=generate_params["eos_token_id"])
+
+        outputs = self.llm.generate(
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            output_attentions=True,
+            **generate_params
+        )
+
+        return outputs

preprocessor_config.json

@@ -0,0 +1,23 @@
+{
+  "do_center_crop": false,
+  "do_convert_rgb": true,
+  "do_normalize": true,
+  "do_rescale": true,
+  "do_resize": false,
+  "image_mean": [
+    0.48145466,
+    0.4578275,
+    0.40821073
+  ],
+  "image_processor_type": "CLIPImageProcessor",
+  "image_std": [
+    0.26862954,
+    0.26130258,
+    0.27577711
+  ],
+  "resample": 3,
+  "rescale_factor": 0.00392156862745098,
+  "size": {
+    "shortest_edge": 224
+  }
+}

processing_vora.py

@@ -0,0 +1,139 @@
+import torch
+import math
+
+from typing import List, Union
+from PIL import Image
+
+from transformers.feature_extraction_utils import BatchFeature
+from transformers.image_utils import ImageInput
+from transformers.processing_utils import ProcessingKwargs, ProcessorMixin, Unpack, _validate_images_text_input_order
+from transformers.tokenization_utils_base import PreTokenizedInput, TextInput
+
+from .modeling_vora import VoRAForCausalLM
+
+
+def smart_resize(
+    height: int, width: int, factor: int = 14, min_pixels: int = 14 * 14, max_pixels: int = 14 * 14 * 160 * 160 
+):
+    """Rescales the image so that the following conditions are met:
+
+    1. Both dimensions (height and width) are divisible by 'factor'.
+
+    2. The total number of pixels is within the range ['min_pixels', 'max_pixels'].
+
+    3. The aspect ratio of the image is maintained as closely as possible.
+
+    """
+    if height < factor or width < factor:
+        raise ValueError(f"height:{height} or width:{width} must be larger than factor:{factor}")
+    elif max(height, width) / min(height, width) > 200:
+        raise ValueError(
+            f"absolute aspect ratio must be smaller than 200, got {max(height, width) / min(height, width)}"
+        )
+    h_bar = round(height / factor) * factor
+    w_bar = round(width / factor) * factor
+    if h_bar * w_bar > max_pixels:
+        beta = math.sqrt((height * width) / max_pixels)
+        h_bar = math.floor(height / beta / factor) * factor
+        w_bar = math.floor(width / beta / factor) * factor
+    elif h_bar * w_bar < min_pixels:
+        beta = math.sqrt(min_pixels / (height * width))
+        h_bar = math.ceil(height * beta / factor) * factor
+        w_bar = math.ceil(width * beta / factor) * factor
+    return h_bar, w_bar
+
+
+class VoRAProcessorKwargs(ProcessingKwargs, total=False):
+    _defaults = {
+        "text_kwargs": {
+            "padding": False,
+        },
+        "images_kwargs": {},
+    }
+
+
+class VoRAProcesser(ProcessorMixin):
+    attributes = ["image_processor", "tokenizer"]
+    valid_kwargs = [
+        "chat_template",
+        "image_token",
+    ]
+    image_processor_class = "AutoImageProcessor"
+    tokenizer_class = "AutoTokenizer"
+
+    def __init__(
+        self,
+        image_processor=None,
+        tokenizer=None,
+        chat_template=None,
+        image_token="<image>",  # set the default and let users change if they have peculiar special tokens in rare cases
+        image_token_index = -200,
+        **kwargs,
+    ):
+        self.image_token = image_token
+        self.image_token_index = image_token_index
+        super().__init__(image_processor, tokenizer, chat_template=chat_template)
+
+    def __call__(
+        self,
+        images: ImageInput = None,
+        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
+        **kwargs: Unpack[VoRAProcessorKwargs],
+    ):
+        if images is None and text is None:
+            raise ValueError("You have to specify at least one of `images` or `text`.")
+
+        images, text = _validate_images_text_input_order(images, text)
+        output_kwargs = self._merge_kwargs(
+            VoRAProcessorKwargs,
+            tokenizer_init_kwargs=self.tokenizer.init_kwargs,
+            **kwargs,
+        )
+
+        if images is not None:
+            images = [[self.anyres_resize(image[0])] for image in images]
+            image_inputs = self.image_processor(images, **output_kwargs["images_kwargs"])
+        else:
+            image_inputs = {}
+        
+        if isinstance(text, str):
+            text = [text]
+        elif not isinstance(text, list) and not isinstance(text[0], str):
+            raise ValueError("Invalid input text. Please provide a string, or a list of strings")
+        
+        input_ids = [self.tokenizer_vision_placeholder(t) for t in text]
+        attention_mask = [
+            [1] * len(input_ids[i]) for i in range(len(input_ids))
+        ]
+        text_inputs = dict(
+            input_ids=torch.as_tensor(input_ids, dtype=torch.int64),
+            attention_mask=torch.as_tensor(attention_mask, dtype=torch.int64),
+        )
+        image_inputs['frames'] = image_inputs.pop('pixel_values')
+        image_inputs['n_frames'] = [len(_images) for _images in images]
+        image_inputs['vision_placeholder_index'] = self.image_token_index
+        return BatchFeature(data={**text_inputs, **image_inputs})
+
+    def anyres_resize(self, pil_img: Image.Image):
+        h, w = pil_img.size
+        h, w = smart_resize(h, w)
+        image = pil_img.resize((w, h))
+        return image
+
+    def tokenizer_vision_placeholder(self, prompt, add_bos=False):
+        def join_lists(*lists, sep):
+            result = []
+            for i, lst in enumerate(lists):
+                if i > 0 and sep:
+                    result.extend([sep])
+                result.extend(lst)
+            return result
+
+        prompt_chunks = [self.tokenizer.encode(
+            chunk) for chunk in prompt.split(self.image_token)]
+        input_ids = join_lists(*prompt_chunks, sep=self.image_token_index)
+        if add_bos:
+            input_ids = [self.tokenizer.bos_token_id] + input_ids
+
+        return input_ids
+

processor_config.json

@@ -0,0 +1,6 @@
+{
+  "image_token": "<image>",
+  "image_token_index": -200,
+  "processor_class": "VoRAProcessing",
+  "auto_map": {"AutoProcessor": "processing_vora.VoRAProcesser"}
+}

special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c5ae00e602b8860cbd784ba82a8aa14e8feecec692e7076590d014d7b7fdafa
+size 11421896

tokenizer_config.json

@@ -0,0 +1,209 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "chat_template": "{%- if tools %}\n    {{- '<|im_start|>system\\n' }}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- messages[0]['content'] }}\n    {%- else %}\n        {{- 'You are Qwen, created by Alibaba Cloud. You are a helpful assistant.' }}\n    {%- endif %}\n    {{- \"\\n\\n# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\n</tools>\\n\\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\\n<tool_call>\\n{\\\"name\\\": <function-name>, \\\"arguments\\\": <args-json-object>}\\n</tool_call><|im_end|>\\n\" }}\n{%- else %}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- '<|im_start|>system\\n' + messages[0]['content'] + '<|im_end|>\\n' }}\n    {%- else %}\n        {{- '<|im_start|>system\\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- for message in messages %}\n    {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) or (message.role == \"assistant\" and not message.tool_calls) %}\n        {{- '<|im_start|>' + message.role + '\\n' + message.content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {{- '<|im_start|>' + message.role }}\n        {%- if message.content %}\n            {{- '\\n' + message.content }}\n        {%- endif %}\n        {%- for tool_call in message.tool_calls %}\n            {%- if tool_call.function is defined %}\n                {%- set tool_call = tool_call.function %}\n            {%- endif %}\n            {{- '\\n<tool_call>\\n{\"name\": \"' }}\n            {{- tool_call.name }}\n            {{- '\", \"arguments\": ' }}\n            {{- tool_call.arguments | tojson }}\n            {{- '}\\n</tool_call>' }}\n        {%- endfor %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- message.content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n{%- endif %}\n",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "processor_class": "VoRAProcessing",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

vision_embedding.py

@@ -0,0 +1,134 @@
+import torch
+import torch.nn as nn
+
+from .configuration_vora import VoRAConfig
+
+
+def _get_1d_sincos_pos_embed_from_grid(
+    embed_dim: int, pos: torch.Tensor, device: torch.device
+) -> torch.Tensor:
+    omega = torch.arange(embed_dim // 2).float().to(device)
+    omega /= embed_dim / 2.0
+    omega = 1.0 / 10000**omega  # (D / 2,)
+    pos = pos.reshape(-1)  # (M,)
+    out = pos[:, None] * omega[None, :]  # (M, D / 2), outer product
+    emb_sin, emb_cos = torch.sin(out).to(device), torch.cos(out).to(device)  # (M, D / 2)
+    emb = torch.cat([emb_sin, emb_cos], dim=1)  # (M, D)
+    return emb
+
+
+def get_sincos_pos_embed(h: int, w: int, embed_dim: int, device: torch.device) -> torch.Tensor:
+    assert embed_dim % 2 == 0, embed_dim
+    grid_h = torch.arange(h).float().to(device)
+    grid_w = torch.arange(w).float().to(device)
+    grid = torch.meshgrid(grid_w, grid_h, indexing="xy")
+    grid = torch.stack(grid, dim=0).to(device)
+    grid = grid.reshape([2, 1, h, w])
+    emb_h = _get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0], device)
+    emb_w = _get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1], device)
+    pos_embed = torch.cat([emb_h, emb_w], dim=1)  # (H * W, D)
+    return pos_embed
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(dim))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        output = self._norm(x.float()).type_as(x)
+        return output * self.weight
+
+    def extra_repr(self) -> str:
+        return f"{tuple(self.weight.shape)}, eps={self.eps}"
+
+    def _norm(self, x: torch.Tensor) -> torch.Tensor:
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+
+class VisionEmbedding(nn.Module):
+    def __init__(self,
+        config: VoRAConfig = None,
+        hidden_size: int = 4096,
+    ):
+        super().__init__()
+        self.patch_size = config.patch_size
+        self.proj = nn.Conv2d(
+            3,
+            hidden_size,
+            kernel_size=(self.patch_size, self.patch_size),
+            stride=(self.patch_size, self.patch_size),
+            bias=True,
+        )
+        self.norm = RMSNorm(hidden_size, eps=1e-05)
+        self.embed_dim = hidden_size
+
+    def forward(self, pixel_values: torch.Tensor):
+        _, _, H, W = pixel_values.shape
+        tokens = self.norm(self.proj(pixel_values).flatten(2).transpose(1, 2))
+        pos_embed = get_sincos_pos_embed(
+            H // self.patch_size, W // self.patch_size, embed_dim=self.embed_dim, device=tokens.device
+        )
+        tokens = tokens + pos_embed.to(tokens.device)
+        return tokens
+
+
+class AIMv2PatchEmbed(nn.Module):
+    def __init__(self, config: VoRAConfig):
+        super().__init__()
+        self.proj = nn.Conv2d(
+            3,
+            config.vision_embedding_intermediate_size,
+            kernel_size=(config.patch_size, config.patch_size),
+            stride=(config.patch_size, config.patch_size),
+        )
+        self.norm = RMSNorm(config.vision_embedding_intermediate_size, eps=config.rms_norm_eps)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.proj(x).flatten(2).transpose(1, 2)
+        x = self.norm(x)
+        return x
+
+
+class AIMv2ViTPreprocessor(nn.Module):
+    def __init__(self,         
+        config: VoRAConfig = None,
+        hidden_size: int = 4096,
+    ):
+        super().__init__()
+        num_patches = (config.image_size // config.patch_size) ** 2
+        self.config = config
+
+        self.patchifier = AIMv2PatchEmbed(config)
+        self.pos_embed = nn.Parameter(torch.zeros((1, num_patches, config.vision_embedding_intermediate_size)))
+        self.out_proj = nn.Linear(config.vision_embedding_intermediate_size, hidden_size, bias=False)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        B, C, H, W = x.shape
+        h_token = H // self.config.patch_size
+        w_token = W // self.config.patch_size
+        tokens = self.patchifier(x)
+        _, N, _ = tokens.shape
+        pos_embed = self.pos_embed.to(tokens.device)
+
+        if N <= pos_embed.size(1):
+            # 如果 N 小于或等于 num_patches，直接相加
+            tokens = tokens + pos_embed[:, :N]
+        else:
+            # 如果 N 大于 num_patches，使用双线性插值
+            # 将 pos_embed 调整为 (1, num_patches, hidden_size) 的形状
+            pos_embed = pos_embed.view(1, int(pos_embed.size(1)**0.5), int(pos_embed.size(1)**0.5), -1).permute(0, 3, 1, 2)
+            # 使用双线性插值调整大小
+            pos_embed = F.interpolate(pos_embed, size=(h_token, w_token), mode='bilinear', align_corners=False).permute(0, 2, 3, 1)
+            # 重塑为 (1, N, hidden_size) 形状
+            pos_embed = pos_embed.view(1, N, pos_embed.size(-1))
+            tokens = tokens + pos_embed
+
+        return self.out_proj(tokens)
+
+
+def build_vision_embedding(config: VoRAConfig, hidden_size):
+    if config.vision_embedding_type == "AIMv2":
+        return AIMv2ViTPreprocessor(config, hidden_size)
+    return VisionEmbedding(config, hidden_size)

vora_generation_utils.py

@@ -0,0 +1,101 @@
+from typing import Any, Dict, Optional
+
+import torch
+from transformers import GenerationMixin
+from transformers.cache_utils import Cache
+from transformers.utils import ModelOutput
+
+
+class VoraGenerationMixin(GenerationMixin):
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids: torch.LongTensor,
+        past_key_values: Optional[Cache] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ):
+        if attention_mask is not None and attention_mask.ndim == 4:
+            attention_mask_2d = (attention_mask[:, 0, :, :] == 0).any(dim=1).long().to(attention_mask.device)
+            model_input = super().prepare_inputs_for_generation(
+                input_ids,
+                past_key_values=past_key_values,
+                attention_mask=attention_mask_2d,
+                inputs_embeds=inputs_embeds,
+                cache_position=cache_position,
+                **kwargs,
+            )
+            model_input['attention_mask'] = attention_mask
+            return model_input
+        else:
+            return super().prepare_inputs_for_generation(
+                input_ids,
+                past_key_values=past_key_values,
+                attention_mask=attention_mask,
+                inputs_embeds=inputs_embeds,
+                cache_position=cache_position,
+                **kwargs,
+            )
+
+    def _update_model_kwargs_for_generation(
+        self,
+        outputs: ModelOutput,
+        model_kwargs: Dict[str, Any],
+        is_encoder_decoder: bool = False,
+        num_new_tokens: int = 1,
+    ) -> Dict[str, Any]:
+        if "attention_mask" in model_kwargs and model_kwargs["attention_mask"].ndim == 4:
+            attention_mask = model_kwargs.pop("attention_mask")
+            model_kwargs = super()._update_model_kwargs_for_generation(
+                outputs, model_kwargs, is_encoder_decoder=is_encoder_decoder, num_new_tokens=num_new_tokens
+            )
+            bs, _, seq_len, tgt_len = attention_mask.shape
+            dtype = attention_mask.dtype
+            min_dtype = torch.finfo(dtype).min
+            new_col = attention_mask.new_zeros((bs, 1, seq_len, 1)).fill_(min_dtype)
+            new_row = attention_mask.new_zeros((bs, 1, 1, tgt_len + 1))
+            model_kwargs["attention_mask"] = torch.cat([
+                torch.cat([attention_mask, new_col], dim=-1),
+                new_row
+            ], dim=2)
+            return model_kwargs
+        else:
+            return super()._update_model_kwargs_for_generation(
+                outputs, model_kwargs, is_encoder_decoder=is_encoder_decoder, num_new_tokens=num_new_tokens
+            )
+
+
+def custom_prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    cache_position: torch.Tensor,
+    batch_size: int,
+    **kwargs,
+):
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask[:, :, -sequence_length:, -target_length:]
+    else:
+        min_dtype = torch.finfo(dtype).min
+        causal_mask = torch.full(
+            (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
+        )
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask