(Trained with Unsloth)

config.json

@@ -0,0 +1,134 @@
+{
+  "architectures": [
+    "DeepseekOCRForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "modeling_deepseekocr.DeepseekOCRConfig",
+    "AutoModel": "modeling_deepseekocr.DeepseekOCRForCausalLM"
+  },
+  "aux_loss_alpha": 0.001,
+  "bos_token_id": 0,
+  "candidate_resolutions": [
+    [
+      1024,
+      1024
+    ]
+  ],
+  "torch_dtype": "float16",
+  "eos_token_id": 1,
+  "first_k_dense_replace": 1,
+  "global_view_pos": "head",
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 1280,
+  "initializer_range": 0.02,
+  "intermediate_size": 6848,
+  "kv_lora_rank": null,
+  "language_config": {
+    "architectures": [
+      "DeepseekV2ForCausalLM"
+    ],
+    "auto_map": {
+      "AutoConfig": "configuration_deepseekv2.DeepseekV2Config",
+      "AutoModel": "modeling_deepseek.DeepseekV2Model",
+      "AutoModelForCausalLM": "modeling_deepseek.DeepseekV2ForCausalLM"
+    },
+    "bos_token_id": 0,
+    "eos_token_id": 1,
+    "first_k_dense_replace": 1,
+    "hidden_size": 1280,
+    "intermediate_size": 6848,
+    "kv_lora_rank": null,
+    "lm_head": true,
+    "max_position_embeddings": 8192,
+    "moe_intermediate_size": 896,
+    "n_group": 1,
+    "n_routed_experts": 64,
+    "n_shared_experts": 2,
+    "num_attention_heads": 10,
+    "num_experts_per_tok": 6,
+    "num_hidden_layers": 12,
+    "num_key_value_heads": 10,
+    "q_lora_rank": null,
+    "qk_nope_head_dim": 0,
+    "qk_rope_head_dim": 0,
+    "rm_head": false,
+    "topk_group": 1,
+    "topk_method": "greedy",
+    "torch_dtype": "bfloat16",
+    "use_mla": false,
+    "v_head_dim": 0,
+    "vocab_size": 129280
+  },
+  "lm_head": true,
+  "max_position_embeddings": 8192,
+  "mlp_bias": false,
+  "model_type": "DeepseekOCR",
+  "moe_intermediate_size": 896,
+  "n_group": 1,
+  "n_routed_experts": 64,
+  "n_shared_experts": 2,
+  "norm_topk_prob": false,
+  "num_attention_heads": 10,
+  "num_experts_per_tok": 6,
+  "num_hidden_layers": 12,
+  "num_key_value_heads": 10,
+  "pad_token_id": 2,
+  "projector_config": {
+    "input_dim": 2048,
+    "model_type": "mlp_projector",
+    "n_embed": 1280,
+    "projector_type": "linear"
+  },
+  "q_lora_rank": null,
+  "qk_nope_head_dim": 0,
+  "qk_rope_head_dim": 0,
+  "rm_head": false,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 10000.0,
+  "routed_scaling_factor": 1.0,
+  "seq_aux": true,
+  "tie_word_embeddings": false,
+  "tile_tag": "2D",
+  "topk_group": 1,
+  "topk_method": "greedy",
+  "transformers_version": "4.56.2",
+  "unsloth_version": "2025.11.1",
+  "use_cache": true,
+  "use_mla": false,
+  "v_head_dim": 0,
+  "vision_config": {
+    "image_size": 1024,
+    "mlp_ratio": 3.7362,
+    "model_name": "deeplip_b_l",
+    "model_type": "vision",
+    "width": {
+      "clip-l-14-224": {
+        "heads": 16,
+        "image_size": 224,
+        "layers": 24,
+        "patch_size": 14,
+        "width": 1024
+      },
+      "sam_vit_b": {
+        "downsample_channels": [
+          512,
+          1024
+        ],
+        "global_attn_indexes": [
+          2,
+          5,
+          8,
+          11
+        ],
+        "heads": 12,
+        "layers": 12,
+        "width": 768
+      }
+    }
+  },
+  "vocab_size": 129280
+}

conversation.py

@@ -0,0 +1,280 @@
+"""
+From https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
+"""
+
+import dataclasses
+from enum import IntEnum, auto
+from typing import Any, Dict, List
+
+
+class SeparatorStyle(IntEnum):
+    """Separator styles."""
+
+    DeepSeek = auto()
+    DeepSeekV2 = auto()
+    PLAIN = auto()
+    ALIGNMENT = auto()
+
+
+@dataclasses.dataclass
+class Conversation:
+    """A class that manages prompt templates and keeps all conversation history."""
+
+    # The name of this template
+    name: str
+    # The template of the system prompt
+    system_template: str = "{system_message}"
+    # The system message
+    system_message: str = ""
+    # The names of two roles
+    roles: List[str] = (("USER", "ASSISTANT"),)
+    # All messages. Each item is (role, message).
+    messages: List[List[str]] = ()
+    # The number of few shot examples
+    offset: int = 0
+    # The separator style and configurations
+    sep_style: SeparatorStyle = SeparatorStyle.DeepSeek
+    sep: str = "\n"
+    sep2: str = None
+    # Stop criteria (the default one is EOS token)
+    stop_str: str = None
+    # Stops generation if meeting any token in this list
+    stop_token_ids: List[int] = None
+
+    def get_prompt(self) -> str:
+        """Get the prompt for generation."""
+        system_prompt = self.system_template.format(system_message=self.system_message)
+        if self.sep_style == SeparatorStyle.DeepSeek:
+            seps = [self.sep, self.sep2]
+            if system_prompt == "" or system_prompt is None:
+                ret = ""
+            else:
+                ret = system_prompt + seps[0]
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    ret += role + ": " + message + seps[i % 2]
+                else:
+                    ret += role + ":"
+            return ret
+        elif self.sep_style == SeparatorStyle.DeepSeekV2:
+            seps = [self.sep, self.sep2]
+            if system_prompt == "" or system_prompt is None:
+                ret = ""
+            else:
+                ret = system_prompt + seps[0]
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    if role == "User":
+                        ret += "<｜sft▁begin｜>\n" + message + self.sep #<｜sft▁begin｜>User Input<｜sft▁end｜>\nResponse<｜end▁of▁sentence｜>
+                    else:
+                        ret += message + self.sep2
+                else:
+                    ret = ret
+            return ret
+
+        elif self.sep_style == SeparatorStyle.PLAIN:
+            seps = [self.sep, self.sep2]
+            ret = ""
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    if i % 2 == 0:
+                        ret += message + seps[i % 2]
+                    else:
+                        ret += message + seps[i % 2]
+                else:
+                    ret += ""
+            return ret
+        elif self.sep_style == SeparatorStyle.ALIGNMENT:
+            seps = [self.sep, self.sep2]
+            ret = ""
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    if i % 2 == 0:
+                        ret += '<image>\n' + seps[i % 2]
+                    else:
+                        ret += message + seps[i % 2]
+                else:
+                    ret += ""
+            return ret
+        else:
+            raise ValueError(f"Invalid style: {self.sep_style}")
+
+    def set_system_message(self, system_message: str):
+        """Set the system message."""
+        self.system_message = system_message
+
+    def append_message(self, role: str, message: str):
+        """Append a new message."""
+        self.messages.append([role, message])
+
+    def update_last_message(self, message: str):
+        """Update the last output.
+
+        The last message is typically set to be None when constructing the prompt,
+        so we need to update it in-place after getting the response from a model.
+        """
+        self.messages[-1][1] = message
+
+    def reset_message(self):
+        """Reset a new message."""
+        self.messages = []
+
+    def to_gradio_chatbot(self):
+        """Convert the conversation to gradio chatbot format."""
+        ret = []
+        for i, (role, msg) in enumerate(self.messages[self.offset :]):
+            if i % 2 == 0:
+                ret.append([msg, None])
+            else:
+                ret[-1][-1] = msg
+        return ret
+
+    def to_openai_api_messages(self):
+        """Convert the conversation to OpenAI chat completion format."""
+        system_prompt = self.system_template.format(system_message=self.system_message)
+        ret = [{"role": "system", "content": system_prompt}]
+
+        for i, (_, msg) in enumerate(self.messages[self.offset :]):
+            if i % 2 == 0:
+                ret.append({"role": "user", "content": msg})
+            else:
+                if msg is not None:
+                    ret.append({"role": "assistant", "content": msg})
+        return ret
+
+    def copy(self):
+        return Conversation(
+            name=self.name,
+            system_template=self.system_template,
+            system_message=self.system_message,
+            roles=self.roles,
+            messages=[[x, y] for x, y in self.messages],
+            offset=self.offset,
+            sep_style=self.sep_style,
+            sep=self.sep,
+            sep2=self.sep2,
+            stop_str=self.stop_str,
+            stop_token_ids=self.stop_token_ids,
+        )
+
+    def dict(self):
+        return {
+            "template_name": self.name,
+            "system_message": self.system_message,
+            "roles": self.roles,
+            "messages": self.messages,
+            "offset": self.offset,
+        }
+
+
+# A global registry for all conversation templates
+conv_templates: Dict[str, Conversation] = {}
+
+
+def register_conv_template(template: Conversation, override: bool = False):
+    """Register a new conversation template."""
+    if not override:
+        assert template.name not in conv_templates, f"{template.name} has been registered."
+
+    conv_templates[template.name] = template
+
+
+def get_conv_template(name: str) -> Conversation:
+    """Get a conversation template."""
+    return conv_templates[name].copy()
+
+
+register_conv_template(
+    Conversation(
+        name="deepseek",
+        system_template="{system_message}",
+        # system_message="You are a helpful assistant. Please answer truthfully and write out your "
+        # "thinking step by step to be sure you get the right answer.",
+        system_message="",
+        roles=("<|User|>", "<|Assistant|>"),
+        messages=(),
+        offset=0,
+        sep_style=SeparatorStyle.DeepSeek,
+        sep="\n\n",
+        sep2="<｜end▁of▁sentence｜>",
+        stop_token_ids=[100001],
+        stop_str=["User:", "<｜end▁of▁sentence｜>"]
+    )
+)
+register_conv_template(
+    Conversation(
+        name="deepseekv2",
+        system_template="{system_message}",
+        # system_message="You are a helpful assistant. Please answer truthfully and write out your "
+        # "thinking step by step to be sure you get the right answer.",
+        system_message="",
+        roles=("<｜User｜>", "<｜Assistant｜>"),
+        messages=(),
+        offset=0,
+        sep_style=SeparatorStyle.DeepSeek,
+        sep="",
+        sep2="<｜end▁of▁sentence｜>",
+        stop_token_ids=[100001],
+        stop_str=["User:", "<｜end▁of▁sentence｜>"]
+    )
+)
+
+
+register_conv_template(
+    Conversation(
+        name="plain",
+        system_template="",
+        system_message="",
+        roles=("", ""),
+        messages=(),
+        offset=0,
+        sep_style=SeparatorStyle.PLAIN,
+        sep="",
+        sep2="",
+        stop_token_ids=[100001],
+        stop_str=['</s>'],
+    )
+)
+
+
+register_conv_template(
+    Conversation(
+        name="alignment",
+        system_template="",
+        system_message="",
+        roles=("", ""),
+        messages=(),
+        offset=0,
+        sep_style=SeparatorStyle.ALIGNMENT,
+        sep="",
+        sep2="",
+        stop_token_ids=[100001],
+        stop_str=['</s>'],
+    )
+)
+
+
+if __name__ == "__main__":
+    print("deepseek template:")
+    conv = get_conv_template("deepseek")
+    conv.append_message(conv.roles[0], "Hello!")
+    conv.append_message(conv.roles[1], "Hi! This is Tony.")
+    conv.append_message(conv.roles[0], "Who are you?")
+    conv.append_message(conv.roles[1], "I am a helpful assistant.")
+    conv.append_message(conv.roles[0], "How are you?")
+    conv.append_message(conv.roles[1], None)
+    print(conv.get_prompt())
+
+    print("deepseekv2 template:")
+    conv = get_conv_template("deepseekv2")
+    conv.append_message(conv.roles[0], "Hello!")
+    conv.append_message(conv.roles[1], "Hi! This is Tony.")
+    conv.append_message(conv.roles[0], "Who are you?")
+    conv.append_message(conv.roles[1], "I am a helpful assistant.")
+    conv.append_message(conv.roles[0], "How are you?")
+    conv.append_message(conv.roles[1], None)
+    print(conv.get_prompt())

deepencoder.py

@@ -0,0 +1,1058 @@
+import torch.nn as nn
+import torch
+import torch.nn.functional as F
+import copy
+
+from contextlib import nullcontext
+import math
+from typing import Optional, Tuple
+# from megatron.model import LayerNorm
+
+from einops import rearrange
+from easydict import EasyDict as adict
+
+
+from typing import Optional, Tuple, Type
+from functools import partial
+
+
+
+class MlpProjector(nn.Module):
+
+    def __init__(self, cfg):
+
+        super().__init__()
+
+        self.cfg = cfg
+
+        if cfg.projector_type == "identity":
+            modules = nn.Identity()
+
+        elif cfg.projector_type == "linear":
+            modules = nn.Linear(cfg.input_dim, cfg.n_embed)
+
+        elif cfg.projector_type == "mlp_gelu":
+            mlp_depth = cfg.get("depth", 1)
+            modules = [nn.Linear(cfg.input_dim, cfg.n_embed)]
+            for _ in range(1, mlp_depth):
+                modules.append(nn.GELU())
+                modules.append(nn.Linear(cfg.n_embed, cfg.n_embed))
+            modules = nn.Sequential(*modules)
+        
+        elif cfg.projector_type == "normlayer_downsample_mlp_gelu":
+            mlp_depth = cfg.get("depth", 1)
+            mlp_ratio = cfg.get("mlp_ratio", 1)
+            modules = [
+                nn.LayerNorm(cfg.input_dim * cfg.downsample_ratio * cfg.downsample_ratio),
+                nn.Linear(cfg.input_dim * cfg.downsample_ratio * cfg.downsample_ratio, cfg.n_embed * mlp_ratio)
+            ]
+            for _ in range(1, mlp_depth - 1):
+                modules.append(nn.GELU())
+                modules.append(nn.Linear(cfg.n_embed * mlp_ratio, cfg.n_embed * mlp_ratio))
+            modules.append(nn.GELU())
+            modules.append(nn.Linear(cfg.n_embed * mlp_ratio, cfg.n_embed))
+            modules = nn.Sequential(*modules)
+        
+        elif cfg.projector_type == "downsample_mlp_gelu":
+            mlp_depth = cfg.get("depth", 1)
+            mlp_ratio = cfg.get("mlp_ratio", 1)
+            modules = [nn.Linear(cfg.input_dim * cfg.downsample_ratio * cfg.downsample_ratio, cfg.n_embed * mlp_ratio)]
+            for _ in range(1, mlp_depth - 1):
+                modules.append(nn.GELU())
+                modules.append(nn.Linear(cfg.n_embed * mlp_ratio, cfg.n_embed * mlp_ratio))
+            modules.append(nn.GELU())
+            modules.append(nn.Linear(cfg.n_embed * mlp_ratio, cfg.n_embed))
+            modules = nn.Sequential(*modules)
+
+        elif cfg.projector_type == "low_high_hybrid_split_mlp_gelu":
+            mlp_depth = cfg.get("depth", 1)
+            self.high_up_proj = nn.Linear(cfg.input_dim, cfg.n_embed // 2)
+            self.low_up_proj = nn.Linear(cfg.input_dim, cfg.n_embed // 2)
+
+            modules = []
+            for _ in range(1, mlp_depth):
+                modules.append(nn.GELU())
+                modules.append(nn.Linear(cfg.n_embed, cfg.n_embed))
+            modules = nn.Sequential(*modules)
+
+        elif cfg.projector_type == "hybrid_split_feature_mlp_gelu":
+            mlp_depth = cfg.get("depth", 1)
+            channel_div = cfg.get("channel_div", 0.5)
+            self.high_up_proj = nn.Linear(cfg.input_dim[0], int(cfg.n_embed * channel_div))
+            self.low_up_proj = nn.Linear(cfg.input_dim[1], cfg.n_embed - int(cfg.n_embed * channel_div))
+
+            modules = []
+            for _ in range(1, mlp_depth):
+                modules.append(nn.GELU())
+                modules.append(nn.Linear(cfg.n_embed, cfg.n_embed))
+            modules = nn.Sequential(*modules)
+
+        elif cfg.projector_type == "low_high_split_mlp_gelu":
+            mlp_depth = cfg.get("depth", 1)
+            modules = []
+            for _ in range(1, mlp_depth):
+                modules.append(nn.GELU())
+                modules.append(nn.Linear(cfg.n_embed // 2, cfg.n_embed // 2))
+            modules = nn.Sequential(*modules)
+            self.high_layers = nn.Sequential(*modules)
+            self.low_layers = copy.deepcopy(modules)
+
+        else:
+            raise ValueError(f"Unknown projector type: {cfg.projector_type}")
+
+        if cfg.get("token_pooling", False):
+            self.token_pooling_layer = nn.Linear(cfg.input_dim * 4, cfg.input_dim)
+
+        if cfg.get("conv_fusion_high_low_features", False):
+            self.fusion_layer = nn.Linear(cfg.input_dim, cfg.input_dim)
+        self.layers = modules
+
+    def forward(self, x):
+        if self.cfg.get("token_pooling", False):
+            batch_size, wxh, channels = x.shape
+            w = h = int(wxh**0.5)
+            x = x.view(batch_size, w, h, channels)
+            x = x.permute(0, 3, 1, 2)
+            # import ipdb; ipdb.set_trace()
+            patches = x.unfold(2, 2, 2).unfold(3, 2, 2)
+            batch_size, channels, h_patches, w_patches, _, _ = patches.size()
+            # 在通道维度上拼接
+            patches = patches.contiguous().view(batch_size, channels, h_patches * w_patches, -1)
+
+            # 通过线性层
+            patches = patches.permute(0, 2, 1, 3).contiguous()
+            patches = patches.view(batch_size, h_patches * w_patches, channels * 4)
+
+            x = self.token_pooling_layer(patches)
+        
+        if self.cfg.get("conv_fusion_high_low_features", False):
+            x = self.fusion_layer(x[:, 0]) + x[:, 1]
+
+        if self.cfg.projector_type == 'low_high_hybrid_split_mlp_gelu':
+            high_x, low_x = x[0], x[1]
+            high_x = self.high_up_proj(high_x)
+            low_x = self.low_up_proj(low_x)
+            x = torch.concat([high_x, low_x], dim=-1)
+        
+        if self.cfg.projector_type == 'hybrid_split_feature_mlp_gelu':
+            high_x = x[...,:self.cfg.input_dim[0]]
+            low_x = x[...,self.cfg.input_dim[0]:]
+            high_x = self.high_up_proj(high_x)
+            low_x = self.low_up_proj(low_x)
+            x = torch.concat([high_x, low_x], dim=-1)
+        
+        if self.cfg.projector_type == 'low_high_split_mlp_gelu':
+            high_x, low_x = x[0], x[1]
+            high_x = self.high_layers(high_x)
+            low_x = self.low_layers(low_x)
+            x = torch.concat([high_x, low_x], dim=-1)
+            return x
+        
+        if self.cfg.projector_type == 'downsample_mlp_gelu' or self.cfg.projector_type == 'normlayer_downsample_mlp_gelu':
+            bs, hw, input_dim = x.shape
+            h = w = int((hw) ** 0.5)
+
+            """compute padding"""
+            if h % self.cfg.downsample_ratio:
+                pad = self.cfg.downsample_ratio - h % self.cfg.downsample_ratio
+            else:
+                pad = 0
+            x = x.reshape(bs, h, w, input_dim)
+            if pad > 0:
+                x = F.pad(x, (0, 0, 0, pad, 0, pad), "constant", 0)
+
+            """4 to 1 concat"""
+            x = x.permute(0, 3, 1, 2)  # B, C, H, W
+            x = F.unfold(x, kernel_size=self.cfg.downsample_ratio, stride=self.cfg.downsample_ratio, padding=0) # B, C*4, HW // 4
+            x = x.permute(0, 2, 1)
+            
+        return self.layers(x)
+
+    @staticmethod
+    def get_flops_per_sample(cfg):
+        if cfg.projector_type == "linear":
+            fwd = 2 * cfg.input_dim * cfg.n_embed
+
+        elif "mlp_gelu" in cfg.projector_type :
+            mlp_depth = cfg.get("depth", 1)
+            downsample_ratio = cfg.get("downsample_ratio", 1)
+            input_dim = sum(cfg.input_dim) if isinstance(cfg.input_dim, list) else cfg.input_dim
+            input_dim = input_dim * downsample_ratio * downsample_ratio
+            fwd = 2 * input_dim * cfg.n_embed + (mlp_depth - 1) * 2 * cfg.n_embed * cfg.n_embed
+        else:
+            fwd = 0
+
+        return fwd * 3
+    
+
+#===================clip============================================================
+
+class LayerNormfp32(torch.nn.LayerNorm):
+    """Subclass torch's LayerNorm to handle fp16."""
+
+    def forward(self, x: torch.Tensor):
+        orig_type = x.dtype
+        ret = super().forward(x.type(torch.float32))
+        return ret.type(orig_type)
+
+
+def get_abs_pos(abs_pos, tgt_size):
+    # abs_pos: L, C
+    # tgt_size: M
+    # return: M, C
+
+    # print(tgt_size)
+    # print(abs_pos.shape)
+    # exit()
+    dim = abs_pos.size(-1)
+    # print(dim)
+    abs_pos_new = abs_pos.squeeze(0)
+    cls_token, old_pos_embed = abs_pos_new[:1], abs_pos_new[1:]
+
+
+
+    src_size = int(math.sqrt(abs_pos_new.shape[0] - 1))
+    tgt_size = int(math.sqrt(tgt_size))
+    dtype = abs_pos.dtype
+
+    if src_size != tgt_size:
+        old_pos_embed = old_pos_embed.view(1, src_size, src_size, dim).permute(0, 3, 1,
+                                                                                    2).contiguous()
+        old_pos_embed = old_pos_embed.to(torch.float32)
+        new_pos_embed = F.interpolate(
+            old_pos_embed,
+            size=(tgt_size, tgt_size),
+            mode='bicubic',
+            antialias=True,
+            align_corners=False,
+        ).to(dtype)
+        new_pos_embed = new_pos_embed.permute(0, 2, 3, 1)
+        new_pos_embed = new_pos_embed.view(tgt_size * tgt_size, dim)
+        vision_pos_embed = torch.cat([cls_token, new_pos_embed], dim=0)
+        vision_pos_embed = vision_pos_embed.view(1, tgt_size * tgt_size + 1, dim)
+        return vision_pos_embed
+    else:
+        return abs_pos
+
+@torch.jit.script
+def quick_gelu(x):
+    return x * torch.sigmoid(1.702 * x)
+
+
+
+class CLIPVisionEmbeddings(nn.Module):
+    def __init__(self, hidden_size=1024, image_size=224, patch_size=14, num_channels=3):
+        super().__init__()
+        self.embed_dim = hidden_size
+        self.image_size = image_size
+        self.patch_size = patch_size
+
+        self.class_embedding = torch.nn.Parameter(torch.randn(self.embed_dim))
+
+        self.patch_embedding = torch.nn.Conv2d(
+            in_channels=num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+            bias=False,
+        )
+
+        self.num_patches = (self.image_size // self.patch_size) ** 2
+        self.num_positions = self.num_patches + 1
+        self.position_embedding = torch.nn.Embedding(self.num_positions, self.embed_dim)
+        self.register_buffer(
+            "position_ids", torch.arange(self.num_positions).expand((1, -1))
+        )
+
+    def forward(self, pixel_values, patch_embeds):
+        batch_size = pixel_values.shape[0]
+        # patch_embeds = self.patch_embedding(
+        #     pixel_values
+        # )  # shape = [*, width, grid, grid]
+
+
+        if patch_embeds is not None:
+            patch_embeds = patch_embeds
+            # print(patch_embeds.shape)
+        else:
+            patch_embeds = self.patch_embedding(pixel_values)  
+            # print(111111)
+        # shape = [*, width, grid, grid]
+        # patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
+
+        patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
+
+
+        class_embeds = self.class_embedding.expand(batch_size, 1, -1)
+        embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
+
+        # x = torch.cat([cls_token, x], dim=1)
+        embeddings = embeddings + get_abs_pos(self.position_embedding(self.position_ids), embeddings.size(1))
+        # embeddings = embeddings + self.position_embedding(self.position_ids)
+        return embeddings
+
+
+class NoTPFeedForward(nn.Module):
+    def __init__(
+            self,
+            cfg,
+            dim: int,
+            hidden_dim: int,
+    ):
+        super().__init__()
+
+        self.fc1 = torch.nn.Linear(dim, hidden_dim, bias=True)
+        self.fc2 = torch.nn.Linear(hidden_dim, dim, bias=True)
+
+    def forward(self, x):
+        output = self.fc2(quick_gelu(self.fc1(x)))
+        return output
+
+
+
+
+class NoTPAttention(torch.nn.Module):
+    def __init__(self, cfg):
+        super().__init__()
+        self.num_heads = cfg.num_attention_heads
+        self.n_local_heads = cfg.num_attention_heads
+        self.head_dim = cfg.hidden_size // cfg.num_attention_heads
+        self.max_seq_len = cfg.seq_length
+        self.use_flash_attention = cfg.use_flash_attn
+
+        self.qkv_proj = torch.nn.Linear(cfg.hidden_size, cfg.hidden_size * 3, bias=True)
+        self.out_proj = torch.nn.Linear(cfg.hidden_size, cfg.hidden_size, bias=True)
+
+        # self.core_attention = CoreAttention(cfg, AttnType.self_attn)
+
+        self.attn_drop = cfg.attention_dropout
+
+    def forward(
+            self,
+            x: torch.Tensor,
+    ):
+        bsz, seqlen, _ = x.shape
+        xqkv = self.qkv_proj(x)
+        xqkv = xqkv.view(bsz, seqlen, 3, self.num_heads, self.head_dim)
+
+        if self.use_flash_attention:
+
+            xq, xk, xv = torch.split(xqkv, 1, dim=2)
+            xq = xq.squeeze(2)
+            xk = xk.squeeze(2)
+            xv = xv.squeeze(2)
+            # xq, xk, xv = xqkv[:, :, 0, ...], xqkv[:, :, 1, ...], xqkv[:, :, 2, ...]
+
+            # （B, num_head, S, head_size)
+            xq = xq.permute(0, 2, 1, 3)
+            xk = xk.permute(0, 2, 1, 3)
+            xv = xv.permute(0, 2, 1, 3)
+            # with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
+            output = torch.nn.functional.scaled_dot_product_attention(xq, xk, xv, attn_mask=None)
+            output = output.permute(0, 2, 1, 3).reshape(bsz, seqlen, -1)
+                # output = output.permute(0, 2, 1, 3).contiguous().view(bsz, seqlen, -1)
+        else:
+            # print(22222)
+            xq, xk, xv = torch.split(xqkv, 1, dim=2)
+            xq = xq.squeeze(2)
+            xk = xk.squeeze(2)
+            xv = xv.squeeze(2)
+            # xq, xk, xv = xqkv[:, :, 0, ...], xqkv[:, :, 1, ...], xqkv[:, :, 2, ...]
+
+            # （B, num_head, S, head_size)
+            xq = xq.permute(0, 2, 1, 3)
+            xk = xk.permute(0, 2, 1, 3)
+            xv = xv.permute(0, 2, 1, 3)
+            # with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
+            output = torch.nn.functional.scaled_dot_product_attention(xq, xk, xv, attn_mask=None)
+            output = output.permute(0, 2, 1, 3).reshape(bsz, seqlen, -1)
+            # output = output.permute(0, 2, 1, 3).contiguous().view(bsz, seqlen, -1)
+        output = self.out_proj(output)
+        return output
+
+class NoTPTransformerBlock(nn.Module):
+    def __init__(self, cfg, layer_id: int, multiple_of=256):
+        super().__init__()
+
+        self.n_heads = cfg.num_attention_heads
+        self.dim = cfg.hidden_size
+        self.head_dim = cfg.hidden_size // cfg.num_attention_heads
+        self.self_attn = NoTPAttention(cfg)
+        self.mlp = NoTPFeedForward(
+            cfg, dim=cfg.hidden_size, hidden_dim=cfg.ffn_hidden_size
+        )
+        self.layer_id = layer_id
+        self.layer_norm1 = torch.nn.LayerNorm(
+            cfg.hidden_size, eps=cfg.layernorm_epsilon
+        )
+        self.layer_norm2 = torch.nn.LayerNorm(
+            cfg.hidden_size, eps=cfg.layernorm_epsilon
+        )
+
+    def forward(self, x: torch.Tensor):
+        residual = self.self_attn.forward(self.layer_norm1(x))
+        h = x + residual
+        out = h + self.mlp.forward(self.layer_norm2(h))
+        return out
+
+
+class NoTPTransformer(nn.Module):
+    def __init__(self, cfg):
+        super().__init__()
+
+        self.cfg = cfg
+        # self.recompute_list = self.cfg.get("recompute_list", [])
+        self.num_layers = cfg.num_layers  # _get_num_layers(cfg)
+
+        self.layers = torch.nn.ModuleList()
+        for layer_id in range(self.num_layers):
+            self.layers.append(
+                NoTPTransformerBlock(
+                    cfg,
+                    layer_id + 1,
+                )
+            )
+
+    def forward(
+            self,
+            hidden_states,
+    ):
+
+        for lid, layer in enumerate(self.layers):
+            # if lid in self.recompute_list:
+            #     def custom(layer_id):
+            #         def custom_forward(*args, **kwargs):
+            #             x_ = self.layers[layer_id](*args, **kwargs)
+            #             return x_
+
+            #         return custom_forward
+
+            #     assert hidden_states.requires_grad == True, logger.warning(
+            #         "When using recalculation, the input must have grad fn"
+            #     )
+            #     hidden_states = tensor_parallel.checkpoint(
+            #         custom(lid),
+            #         False,
+            #         hidden_states.contiguous()
+            #     )
+            # else:
+            hidden_states = layer(hidden_states)
+
+        return hidden_states
+
+
+# from megatron.core.tensor_parallel.layers import non_tensor_paralleled, local_dp_reduce, local_dp_scatter
+
+class VitModel(nn.Module):
+    def __init__(
+            self,
+            cfg,
+            freeze_embed=False,
+            freeze_pre_norm=False
+    ) -> None:
+        super().__init__()
+
+        self.embeddings = CLIPVisionEmbeddings(hidden_size=cfg.hidden_size, image_size=cfg.image_size, patch_size=cfg.patch_size)
+
+        if freeze_embed:
+            for name, param in self.embeddings.named_parameters():
+                param.requires_grad = False
+
+        self.transformer = NoTPTransformer(cfg=cfg)
+
+        if cfg.get("fp32norm", False):
+            logger.info("Load fp32 layernorm for ViT.")
+            self.pre_layrnorm = LayerNormfp32(
+                cfg.hidden_size,
+                eps=cfg.get("pre_layernorm_epsilon", 1e-5),
+            )
+        else:
+            self.pre_layrnorm = torch.nn.LayerNorm(
+                cfg.hidden_size,
+                eps=cfg.get("pre_layernorm_epsilon", 1e-5),
+            )
+
+        # self.pre_layrnorm = RMSNorm(
+        #     cfg.hidden_size,
+        #     eps=cfg.get("pre_layernorm_epsilon", 1e-5),
+        #     sequence_parallel=False,
+        #     use_fp32=True,
+        #     use_optimus=True,
+        # )
+
+        if freeze_pre_norm:
+            for name, param in self.pre_layrnorm.named_parameters():
+                param.requires_grad = False
+
+        for p in self.parameters():
+            p.micro_dp = True
+
+    def set_input_tensor(self, input_tensor):
+        if not isinstance(input_tensor, list):
+            input_tensor = [input_tensor]
+        self.transformer.set_input_tensor(input_tensor[0])
+
+    def __str__(self) -> str:
+        return "open_clip"
+
+    def forward(
+            self,
+            x,
+            patch_embeds
+    ):
+        x = self.embeddings(x, patch_embeds)
+        hidden_states = self.pre_layrnorm(x)
+
+        # hidden_states, dis = local_dp_scatter(hidden_states)
+        output = self.transformer(hidden_states)
+
+        # output = local_dp_reduce(output, dis)
+
+        return output
+
+
+vit_model_cfg = adict(
+    num_layers=24,
+    hidden_size=1024,
+    num_heads = 16,
+    num_attention_heads=16,
+    ffn_hidden_size=4096,
+    seq_length=256,
+    max_position_embeddings=256,
+    use_flash_attn=False,
+    understand_projector_stride=2,
+    hidden_dropout = 0.0,
+    attention_dropout = 0.0,
+    no_persist_layer_norm = False,
+    layernorm_epsilon = 1e-5,
+    pre_layernorm_epsilon = 1e-5,
+    image_size = 224,
+    patch_size = 14,
+    recompute_list = []
+)
+
+def build_clip_l():
+    return VitModel(
+        cfg=vit_model_cfg,
+        freeze_embed=False,
+        freeze_pre_norm=False,
+    )
+
+
+
+
+
+#=========================Sam-Vary=================================
+
+
+def get_abs_pos_sam(abs_pos, tgt_size):
+
+    dtype = abs_pos.dtype
+
+    src_size = abs_pos.size(1)
+
+    if src_size != tgt_size:
+        old_pos_embed = abs_pos.permute(0, 3, 1, 2)
+        old_pos_embed = old_pos_embed.to(torch.float32)
+        new_pos_embed = F.interpolate(
+            old_pos_embed,
+            size=(tgt_size, tgt_size),
+            mode='bicubic',
+            antialias=True,
+            align_corners=False,
+        ).to(dtype)
+        new_pos_embed = new_pos_embed.permute(0, 2, 3, 1)
+        return new_pos_embed
+    else:
+        return abs_pos
+
+
+
+
+class MLPBlock(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        mlp_dim: int,
+        act: Type[nn.Module] = nn.GELU,
+    ) -> None:
+        super().__init__()
+        self.lin1 = nn.Linear(embedding_dim, mlp_dim)
+        self.lin2 = nn.Linear(mlp_dim, embedding_dim)
+        self.act = act()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.lin2(self.act(self.lin1(x)))
+
+
+# From https://github.com/facebookresearch/detectron2/blob/main/detectron2/layers/batch_norm.py # noqa
+# Itself from https://github.com/facebookresearch/ConvNeXt/blob/d1fa8f6fef0a165b27399986cc2bdacc92777e40/models/convnext.py#L119  # noqa
+class LayerNorm2d(nn.Module):
+    def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(num_channels))
+        self.bias = nn.Parameter(torch.zeros(num_channels))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        u = x.mean(1, keepdim=True)
+        s = (x - u).pow(2).mean(1, keepdim=True)
+        x = (x - u) / torch.sqrt(s + self.eps)
+        x = self.weight[:, None, None] * x + self.bias[:, None, None]
+        return x
+
+
+# This class and its supporting functions below lightly adapted from the ViTDet backbone available at: https://github.com/facebookresearch/detectron2/blob/main/detectron2/modeling/backbone/vit.py # noqa
+class ImageEncoderViT(nn.Module):
+    def __init__(
+        self,
+        img_size: int = 1024,
+        patch_size: int = 16,
+        in_chans: int = 3,
+        embed_dim: int = 768,
+        depth: int = 12,
+        num_heads: int = 12,
+        mlp_ratio: float = 4.0,
+        out_chans: int = 256,
+        qkv_bias: bool = True,
+        norm_layer: Type[nn.Module] = nn.LayerNorm,
+        act_layer: Type[nn.Module] = nn.GELU,
+        use_abs_pos: bool = True,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        window_size: int = 0,
+        global_attn_indexes: Tuple[int, ...] = (),
+    ) -> None:
+        """
+        Args:
+            img_size (int): Input image size.
+            patch_size (int): Patch size.
+            in_chans (int): Number of input image channels.
+            embed_dim (int): Patch embedding dimension.
+            depth (int): Depth of ViT.
+            num_heads (int): Number of attention heads in each ViT block.
+            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+            qkv_bias (bool): If True, add a learnable bias to query, key, value.
+            norm_layer (nn.Module): Normalization layer.
+            act_layer (nn.Module): Activation layer.
+            use_abs_pos (bool): If True, use absolute positional embeddings.
+            use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            window_size (int): Window size for window attention blocks.
+            global_attn_indexes (list): Indexes for blocks using global attention.
+        """
+        super().__init__()
+        self.img_size = img_size
+
+        self.patch_embed = PatchEmbed(
+            kernel_size=(patch_size, patch_size),
+            stride=(patch_size, patch_size),
+            in_chans=in_chans,
+            embed_dim=embed_dim,
+        )
+
+        self.pos_embed: Optional[nn.Parameter] = None
+        if use_abs_pos:
+            # Initialize absolute positional embedding with pretrain image size.
+            self.pos_embed = nn.Parameter(
+                torch.zeros(1, img_size // patch_size, img_size // patch_size, embed_dim)
+            )
+
+        self.blocks = nn.ModuleList()
+        for i in range(depth):
+            block = Block(
+                dim=embed_dim,
+                num_heads=num_heads,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                norm_layer=norm_layer,
+                act_layer=act_layer,
+                use_rel_pos=use_rel_pos,
+                rel_pos_zero_init=rel_pos_zero_init,
+                window_size=window_size if i not in global_attn_indexes else 0,
+                input_size=(img_size // patch_size, img_size // patch_size),
+            )
+            self.blocks.append(block)
+
+        self.neck = nn.Sequential(
+            nn.Conv2d(
+                embed_dim,
+                out_chans,
+                kernel_size=1,
+                bias=False,
+            ),
+            LayerNorm2d(out_chans),
+            nn.Conv2d(
+                out_chans,
+                out_chans,
+                kernel_size=3,
+                padding=1,
+                bias=False,
+            ),
+            LayerNorm2d(out_chans),
+        )
+
+        self.net_2 = nn.Conv2d(256, 512, kernel_size=3, stride=2, padding=1, bias=False)
+        self.net_3 = nn.Conv2d(512, 1024, kernel_size=3, stride=2, padding=1, bias=False)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.patch_embed(x)
+        if self.pos_embed is not None:
+            # x = x + self.pos_embed
+            x = x + get_abs_pos_sam(self.pos_embed, x.size(1))
+
+        for blk in self.blocks:
+            x = blk(x)
+
+        x = self.neck(x.permute(0, 3, 1, 2))
+        x2 = self.net_2(x)
+        x3 = self.net_3(x2.clone())
+
+        return x3
+
+
+class Block(nn.Module):
+    """Transformer blocks with support of window attention and residual propagation blocks"""
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        norm_layer: Type[nn.Module] = nn.LayerNorm,
+        act_layer: Type[nn.Module] = nn.GELU,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        window_size: int = 0,
+        input_size: Optional[Tuple[int, int]] = None,
+    ) -> None:
+        """
+        Args:
+            dim (int): Number of input channels.
+            num_heads (int): Number of attention heads in each ViT block.
+            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+            qkv_bias (bool): If True, add a learnable bias to query, key, value.
+            norm_layer (nn.Module): Normalization layer.
+            act_layer (nn.Module): Activation layer.
+            use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            window_size (int): Window size for window attention blocks. If it equals 0, then
+                use global attention.
+            input_size (tuple(int, int) or None): Input resolution for calculating the relative
+                positional parameter size.
+        """
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            use_rel_pos=use_rel_pos,
+            rel_pos_zero_init=rel_pos_zero_init,
+            input_size=input_size if window_size == 0 else (window_size, window_size),
+        )
+
+        self.norm2 = norm_layer(dim)
+        self.mlp = MLPBlock(embedding_dim=dim, mlp_dim=int(dim * mlp_ratio), act=act_layer)
+
+        self.window_size = window_size
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        shortcut = x
+        x = self.norm1(x)
+        # Window partition
+        if self.window_size > 0:
+            H, W = x.shape[1], x.shape[2]
+            x, pad_hw = window_partition(x, self.window_size)
+
+        x = self.attn(x)
+        # Reverse window partition
+        if self.window_size > 0:
+            x = window_unpartition(x, self.window_size, pad_hw, (H, W))
+
+        x = shortcut + x
+        x = x + self.mlp(self.norm2(x))
+
+        return x
+
+
+class Attention(nn.Module):
+    """Multi-head Attention block with relative position embeddings."""
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int = 8,
+        qkv_bias: bool = True,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        input_size: Optional[Tuple[int, int]] = None,
+    ) -> None:
+        """
+        Args:
+            dim (int): Number of input channels.
+            num_heads (int): Number of attention heads.
+            qkv_bias (bool):  If True, add a learnable bias to query, key, value.
+            rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            input_size (tuple(int, int) or None): Input resolution for calculating the relative
+                positional parameter size.
+        """
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = head_dim**-0.5
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.proj = nn.Linear(dim, dim)
+
+        self.use_rel_pos = use_rel_pos
+        if self.use_rel_pos:
+            assert (
+                input_size is not None
+            ), "Input size must be provided if using relative positional encoding."
+            # initialize relative positional embeddings
+            self.rel_pos_h = nn.Parameter(torch.zeros(2 * input_size[0] - 1, head_dim))
+            self.rel_pos_w = nn.Parameter(torch.zeros(2 * input_size[1] - 1, head_dim))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        B, H, W, _ = x.shape
+        # qkv with shape (3, B, nHead, H * W, C)
+        qkv = self.qkv(x).reshape(B, H * W, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        # q, k, v with shape (B * nHead, H * W, C)
+        q, k, v = qkv.reshape(3, B * self.num_heads, H * W, -1).unbind(0)
+
+        rel_h, rel_w = None, None
+        if self.use_rel_pos:
+            rel_h, rel_w = add_decomposed_rel_pos(q, self.rel_pos_h, self.rel_pos_w, (H, W), (H, W))
+
+        q = q.view(B, self.num_heads, H * W, -1)
+        k = k.view(B, self.num_heads, H * W, -1)
+        v = v.view(B, self.num_heads, H * W, -1)
+
+        if self.use_rel_pos:
+            rel_h = rel_h.view(B, self.num_heads, rel_h.size(1), rel_h.size(2), rel_h.size(3))
+            rel_w = rel_w.view(B, self.num_heads, rel_w.size(1), rel_w.size(2), rel_w.size(3))
+            attn_bias = (rel_h + rel_w).view(B, self.num_heads, rel_h.size(2), rel_h.size(3) * rel_w.size(4))
+            x = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=attn_bias)
+            # x = _attention_rel_h_rel_w(q, k, v, rel_h, rel_w)
+        else:
+            x = torch.nn.functional.scaled_dot_product_attention(q, k, v)
+
+        x = x.view(B, self.num_heads, H, W, -1).permute(0, 2, 3, 1, 4).reshape(B, H, W, -1)
+
+        x = self.proj(x)
+
+        return x
+
+
+def window_partition(x: torch.Tensor, window_size: int) -> Tuple[torch.Tensor, Tuple[int, int]]:
+    """
+    Partition into non-overlapping windows with padding if needed.
+    Args:
+        x (tensor): input tokens with [B, H, W, C].
+        window_size (int): window size.
+
+    Returns:
+        windows: windows after partition with [B * num_windows, window_size, window_size, C].
+        (Hp, Wp): padded height and width before partition
+    """
+    B, H, W, C = x.shape
+
+    pad_h = (window_size - H % window_size) % window_size
+    pad_w = (window_size - W % window_size) % window_size
+    if pad_h > 0 or pad_w > 0:
+        x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
+    Hp, Wp = H + pad_h, W + pad_w
+
+    x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows, (Hp, Wp)
+
+
+def window_unpartition(
+    windows: torch.Tensor, window_size: int, pad_hw: Tuple[int, int], hw: Tuple[int, int]
+) -> torch.Tensor:
+    """
+    Window unpartition into original sequences and removing padding.
+    Args:
+        windows (tensor): input tokens with [B * num_windows, window_size, window_size, C].
+        window_size (int): window size.
+        pad_hw (Tuple): padded height and width (Hp, Wp).
+        hw (Tuple): original height and width (H, W) before padding.
+
+    Returns:
+        x: unpartitioned sequences with [B, H, W, C].
+    """
+    Hp, Wp = pad_hw
+    H, W = hw
+    B = windows.shape[0] // (Hp * Wp // window_size // window_size)
+    x = windows.view(B, Hp // window_size, Wp // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, Hp, Wp, -1)
+
+    if Hp > H or Wp > W:
+        x = x[:, :H, :W, :].contiguous()
+    return x
+
+
+def get_rel_pos(q_size: int, k_size: int, rel_pos: torch.Tensor) -> torch.Tensor:
+    """
+    Get relative positional embeddings according to the relative positions of
+        query and key sizes.
+    Args:
+        q_size (int): size of query q.
+        k_size (int): size of key k.
+        rel_pos (Tensor): relative position embeddings (L, C).
+
+    Returns:
+        Extracted positional embeddings according to relative positions.
+    """
+    max_rel_dist = int(2 * max(q_size, k_size) - 1)
+    # Interpolate rel pos if needed.
+    if rel_pos.shape[0] != max_rel_dist:
+        # Interpolate rel pos.
+        dtype = rel_pos.dtype
+        rel_pos = rel_pos.to(torch.float32)
+        rel_pos_resized = F.interpolate(
+            rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),
+            size=max_rel_dist,
+            mode="linear",
+        ).to(dtype)
+        rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)
+    else:
+        rel_pos_resized = rel_pos
+
+    # Scale the coords with short length if shapes for q and k are different.
+    q_coords = torch.arange(q_size, device=rel_pos.device)[:, None] * max(k_size / q_size, 1.0)
+    k_coords = torch.arange(k_size, device=rel_pos.device)[None, :] * max(q_size / k_size, 1.0)
+    relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)
+
+    return rel_pos_resized[relative_coords.long()]
+
+
+def add_decomposed_rel_pos(
+    q: torch.Tensor,
+    rel_pos_h: torch.Tensor,
+    rel_pos_w: torch.Tensor,
+    q_size: Tuple[int, int],
+    k_size: Tuple[int, int],
+) -> torch.Tensor:
+    """
+    Calculate decomposed Relative Positional Embeddings from :paper:`mvitv2`.
+    https://github.com/facebookresearch/mvit/blob/19786631e330df9f3622e5402b4a419a263a2c80/mvit/models/attention.py   # noqa B950
+    Args:
+        q (Tensor): query q in the attention layer with shape (B, q_h * q_w, C).
+        rel_pos_h (Tensor): relative position embeddings (Lh, C) for height axis.
+        rel_pos_w (Tensor): relative position embeddings (Lw, C) for width axis.
+        q_size (Tuple): spatial sequence size of query q with (q_h, q_w).
+        k_size (Tuple): spatial sequence size of key k with (k_h, k_w).
+
+    Returns:
+        attn (Tensor): attention map with added relative positional embeddings.
+    """
+    q_h, q_w = q_size
+    k_h, k_w = k_size
+    Rh = get_rel_pos(q_h, k_h, rel_pos_h)
+    Rw = get_rel_pos(q_w, k_w, rel_pos_w)
+
+    B, _, dim = q.shape
+    r_q = q.reshape(B, q_h, q_w, dim)
+    rel_h = torch.einsum("bhwc,hkc->bhwk", r_q, Rh)
+    rel_w = torch.einsum("bhwc,wkc->bhwk", r_q, Rw)
+    rel_h = rel_h.unsqueeze(-1)
+    rel_w = rel_w.unsqueeze(-2)
+    rel_h = rel_h.reshape(B, q_h * q_w, k_h, 1)
+    rel_w = rel_w.reshape(B, q_h * q_w, 1, k_w)
+
+    return rel_h, rel_w
+
+
+class PatchEmbed(nn.Module):
+    """
+    Image to Patch Embedding.
+    """
+
+    def __init__(
+        self,
+        kernel_size: Tuple[int, int] = (16, 16),
+        stride: Tuple[int, int] = (16, 16),
+        padding: Tuple[int, int] = (0, 0),
+        in_chans: int = 3,
+        embed_dim: int = 768,
+    ) -> None:
+        """
+        Args:
+            kernel_size (Tuple): kernel size of the projection layer.
+            stride (Tuple): stride of the projection layer.
+            padding (Tuple): padding size of the projection layer.
+            in_chans (int): Number of input image channels.
+            embed_dim (int): Patch embedding dimension.
+        """
+        super().__init__()
+
+        self.proj = nn.Conv2d(
+            in_chans, embed_dim, kernel_size=kernel_size, stride=stride, padding=padding
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.proj(x)
+        # B C H W -> B H W C
+        x = x.permute(0, 2, 3, 1)
+        return x
+
+
+def build_sam_vit_b(checkpoint=None):
+    return _build_sam(
+        encoder_embed_dim=768,
+        encoder_depth=12,
+        encoder_num_heads=12,
+        encoder_global_attn_indexes=[2, 5, 8, 11],
+        checkpoint=checkpoint,
+    )
+
+def build_sam_fast_vit_b(checkpoint=None, compile_mode='max-autotune', dtype=torch.bfloat16):
+    image_encoder = build_sam_vit_b(checkpoint).eval().to(dtype)
+    # sam = _apply_eval_dtype_sam(sam, dtype)
+    image_encoder = torch.compile(image_encoder, mode=compile_mode)
+    return image_encoder
+
+
+def _build_sam(
+    encoder_embed_dim,
+    encoder_depth,
+    encoder_num_heads,
+    encoder_global_attn_indexes,
+    checkpoint=None,
+):
+    prompt_embed_dim = 256
+    image_size = 1024
+    vit_patch_size = 16
+    image_embedding_size = image_size // vit_patch_size
+    image_encoder=ImageEncoderViT(
+            depth=encoder_depth,
+            embed_dim=encoder_embed_dim,
+            img_size=image_size,
+            mlp_ratio=4,
+            norm_layer=partial(torch.nn.LayerNorm, eps=1e-6),
+            num_heads=encoder_num_heads,
+            patch_size=vit_patch_size,
+            qkv_bias=True,
+            use_rel_pos=True,
+            global_attn_indexes=encoder_global_attn_indexes,
+            window_size=14,
+            out_chans=prompt_embed_dim,
+        )
+    image_encoder.eval()
+    if checkpoint is not None:
+        # with open(checkpoint, "rb") as f:
+        state_dict = torch.load(checkpoint)
+        # print(state_dict.keys())
+        # for key in state_dict:
+        # image_encoder.load_state_dict({k[14:]: v for k, v in state_dict.items() if 'image_encoder' in k}, strict=False)
+        # ocr-anyting
+        # image_encoder.load_state_dict(state_dict, strict=True)
+        # tob
+        image_encoder.load_state_dict({k[30:]: v for k, v in state_dict.items() if 'vision_tower_high' in k}, strict=True)
+        print(checkpoint)
+    return image_encoder

modeling_deepseekocr.py

@@ -0,0 +1,1043 @@
+# MIT License modified from prithivMLmods/DeepSeek-OCR-Latest-BF16.I64
+import os
+import math
+import re
+from tqdm import tqdm
+from abc import ABC
+from typing import List, Optional, Tuple, Union
+
+from addict import Dict
+from PIL import Image, ImageOps, ImageDraw, ImageFont
+import numpy as np
+
+import torch
+import torch.nn as nn
+from torch.nn import CrossEntropyLoss
+from torchvision import transforms
+
+from transformers.cache_utils import Cache
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers import DeepseekV2Model, DeepseekV2ForCausalLM
+from transformers import DeepseekV2Config
+from transformers.models.deepseek_v2.modeling_deepseek_v2 import (
+    DeepseekV2Attention, DeepseekV2MLP, DeepseekV2MoE, DeepseekV2RMSNorm, DeepseekV2DecoderLayer)
+from transformers.models.llama.modeling_llama import LlamaAttention, LlamaRotaryEmbedding
+from transformers import TextStreamer
+from .deepencoder import build_sam_vit_b, build_clip_l, MlpProjector
+from .conversation import get_conv_template
+
+torch_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
+
+def load_image(image_path):
+
+    try:
+        image = Image.open(image_path)
+        
+        corrected_image = ImageOps.exif_transpose(image)
+        
+        return corrected_image
+        
+    except Exception as e:
+        print(f"error: {e}")
+        try:
+            return Image.open(image_path)
+        except:
+            return None
+
+
+def re_match(text):
+    pattern = r'(<\|ref\|>(.*?)<\|/ref\|><\|det\|>(.*?)<\|/det\|>)'
+    matches = re.findall(pattern, text, re.DOTALL)
+
+    # pattern1 = r'<\|ref\|>.*?<\|/ref\|>\n'
+    # new_text1 = re.sub(pattern1, '', text, flags=re.DOTALL)
+
+    mathes_image = []
+    mathes_other = []
+    for a_match in matches:
+        if '<|ref|>image<|/ref|>' in a_match[0]:
+            mathes_image.append(a_match[0])
+        else:
+            mathes_other.append(a_match[0])
+    return matches, mathes_image, mathes_other
+
+
+def extract_coordinates_and_label(ref_text, image_width, image_height):
+
+    try:
+        label_type = ref_text[1]
+        cor_list = eval(ref_text[2])
+    except Exception as e:
+        print(e)
+        return None
+
+    return (label_type, cor_list)
+
+
+def draw_bounding_boxes(image, refs, ouput_path):
+
+    image_width, image_height = image.size
+    
+    img_draw = image.copy()
+    draw = ImageDraw.Draw(img_draw)
+
+    overlay = Image.new('RGBA', img_draw.size, (0, 0, 0, 0))
+    draw2 = ImageDraw.Draw(overlay)
+    
+    # try:
+    # except IOError:
+    #     try:
+    #         font = ImageFont.truetype("DejaVuSans.ttf", 20) 
+    #     except IOError:
+    font = ImageFont.load_default()
+
+    img_idx = 0
+    
+    for i, ref in enumerate(refs):
+        try:
+            result = extract_coordinates_and_label(ref, image_width, image_height)
+            if result:
+                label_type, points_list = result
+                
+                color = (np.random.randint(0, 200), np.random.randint(0, 200), np.random.randint(0, 255))
+
+                color_a = color + (20, )
+                for points in points_list:
+                    x1, y1, x2, y2 = points
+
+                    x1 = int(x1 / 999 * image_width)
+                    y1 = int(y1 / 999 * image_height)
+
+                    x2 = int(x2 / 999 * image_width)
+                    y2 = int(y2 / 999 * image_height)
+
+                    if label_type == 'image':
+                        try:
+                            cropped = image.crop((x1, y1, x2, y2))
+                            cropped.save(f"{ouput_path}/images/{img_idx}.jpg")
+                        except Exception as e:
+                            print(e)
+                            pass
+                        img_idx += 1
+                        
+                    try:
+                        if label_type == 'title':
+                            draw.rectangle([x1, y1, x2, y2], outline=color, width=4)
+                            draw2.rectangle([x1, y1, x2, y2], fill=color_a, outline=(0, 0, 0, 0), width=1)
+                        else:
+                            draw.rectangle([x1, y1, x2, y2], outline=color, width=2)
+                            draw2.rectangle([x1, y1, x2, y2], fill=color_a, outline=(0, 0, 0, 0), width=1)
+                        text_x = x1
+                        text_y = max(0, y1 - 15)
+                            
+                        
+                        text_bbox = draw.textbbox((0, 0), label_type, font=font)
+                        text_width = text_bbox[2] - text_bbox[0]
+                        text_height = text_bbox[3] - text_bbox[1]
+                        draw.rectangle([text_x, text_y, text_x + text_width, text_y + text_height], 
+                                    fill=(255, 255, 255, 30))
+                        
+                        draw.text((text_x, text_y), label_type, font=font, fill=color)
+                    except:
+                        pass
+        except:
+            continue
+    img_draw.paste(overlay, (0, 0), overlay)
+    return img_draw
+
+
+def process_image_with_refs(image, ref_texts, output_path):
+
+    result_image = draw_bounding_boxes(image, ref_texts, output_path)
+    
+    return result_image
+
+
+
+
+
+def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
+    best_ratio_diff = float('inf')
+    best_ratio = (1, 1)
+    area = width * height
+    for ratio in target_ratios:
+        target_aspect_ratio = ratio[0] / ratio[1]
+        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
+        if ratio_diff < best_ratio_diff:
+            best_ratio_diff = ratio_diff
+            best_ratio = ratio
+        elif ratio_diff == best_ratio_diff:
+            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
+                best_ratio = ratio
+    # print(f'width: {width}, height: {height}, best_ratio: {best_ratio}')
+    return best_ratio
+
+
+def dynamic_preprocess(image, min_num=2, max_num=9, image_size=640, use_thumbnail=False):
+    orig_width, orig_height = image.size
+    aspect_ratio = orig_width / orig_height
+
+    # calculate the existing image aspect ratio
+    target_ratios = set(
+        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
+        i * j <= max_num and i * j >= min_num)
+    # print(target_ratios)
+    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
+
+    # find the closest aspect ratio to the target
+    target_aspect_ratio = find_closest_aspect_ratio(
+        aspect_ratio, target_ratios, orig_width, orig_height, image_size)
+
+    # print(target_aspect_ratio)
+    # calculate the target width and height
+    target_width = image_size * target_aspect_ratio[0]
+    target_height = image_size * target_aspect_ratio[1]
+    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
+
+    # resize the image
+    resized_img = image.resize((target_width, target_height))
+    processed_images = []
+    for i in range(blocks):
+        box = (
+            (i % (target_width // image_size)) * image_size,
+            (i // (target_width // image_size)) * image_size,
+            ((i % (target_width // image_size)) + 1) * image_size,
+            ((i // (target_width // image_size)) + 1) * image_size
+        )
+        # split the image
+        split_img = resized_img.crop(box)
+        processed_images.append(split_img)
+    assert len(processed_images) == blocks
+    if use_thumbnail and len(processed_images) != 1:
+        thumbnail_img = image.resize((image_size, image_size))
+        processed_images.append(thumbnail_img)
+    return processed_images, target_aspect_ratio
+
+
+
+def normalize_transform(mean, std):
+    if mean is None and std is None:
+        transform = None
+    elif mean is None and std is not None:
+        mean = [0.] * len(std)
+        transform = transforms.Normalize(mean=mean, std=std)
+    elif mean is not None and std is None:
+        std = [1.] * len(mean)
+        transform = transforms.Normalize(mean=mean, std=std)
+    else:
+        transform = transforms.Normalize(mean=mean, std=std)
+
+    return transform
+
+
+
+def format_messages(
+        conversations: List[Dict[str, str]],
+        sft_format: str = "deepseek",
+        system_prompt: str = "",
+):
+    """
+    Applies the SFT template to conversation.
+
+    Args:
+        conversations (List[Dict]): A List of messages.
+        sft_format (str, optional): The format of the SFT template to use. Defaults to "deepseek".
+        system_prompt (str, optional): The system prompt to use in the SFT template. Defaults to "".
+
+    Returns:
+        sft_prompt (str): The formatted text.
+    """
+
+    conv = get_conv_template(sft_format)
+    conv.set_system_message(system_prompt)
+    for message in conversations:
+        conv.append_message(message["role"], message["content"].strip())
+    sft_prompt = conv.get_prompt().strip()
+
+    return sft_prompt
+
+
+def text_encode(tokenizer, text: str, bos: bool = True, eos: bool = False):
+    t = tokenizer.encode(text, add_special_tokens=False)
+    bos_id = 0
+    eos_id = 1
+    if bos:
+        t = [bos_id] + t
+    if eos:
+        t = t + [eos_id]
+
+    return t
+
+def load_pil_images(conversations: List[Dict[str, str]]) -> List[Image.Image]:
+    """
+
+    Args:
+        conversations (List[Dict[str, str]]): the conversations with a list of messages. An example is :
+            [
+                {
+                    "role": "User",
+                    "content": "<image_placeholder>\nExtract all information from this image and convert them into markdown format.",
+                    "images": ["./examples/table_datasets.png"]
+                },
+                {"role": "Assistant", "content": ""},
+            ]
+
+    Returns:
+        pil_images (List[PIL.Image.Image]): the list of PIL images.
+
+    """
+
+    pil_images = []
+
+    for message in conversations:
+        if "images" not in message:
+            continue
+
+        for image_path in message["images"]:
+            # print('----------------')
+            # print(image_path)
+            # print('----------------')
+            # exit()
+            
+            # pil_img = Image.open(image_path)
+            pil_img = load_image(image_path)
+            pil_img = pil_img.convert("RGB")
+            pil_images.append(pil_img)
+
+    return pil_images
+
+
+class BaseTransform(ABC):
+
+    def set_rng(self, *args, **kwargs):
+        pass
+
+    def __call__(self, *args, **kwargs) -> torch.Tensor:
+        pass
+
+    @property
+    def default_shape(self):
+        raise NotImplementedError
+
+
+class BasicImageTransform(BaseTransform):
+    def __init__(
+        self, 
+        mean: Optional[Tuple[float, float, float]] = (0.5, 0.5, 0.5),
+        std: Optional[Tuple[float, float, float]] = (0.5, 0.5, 0.5),
+        normalize: bool = True
+    ):
+        self.mean = mean
+        self.std = std
+    
+        transform_pipelines = [
+            transforms.ToTensor()
+        ]
+
+        normalize = normalize_transform(mean, std) if normalize else nn.Identity()
+        if normalize is not None:
+            transform_pipelines.append(normalize)
+
+        self.transform = transforms.Compose(transform_pipelines)
+    
+    def __call__(self, x):
+        x = self.transform(x)
+        return x
+
+class NoEOSTextStreamer(TextStreamer):
+    def on_finalized_text(self, text: str, stream_end: bool = False):
+
+        eos_text = self.tokenizer.decode([self.tokenizer.eos_token_id], skip_special_tokens=False)
+        text = text.replace(eos_text, "\n")
+        print(text, flush=True, end="")
+
+
+def decoder_layer_init(self, config: DeepseekV2Config, layer_idx: int):
+    nn.Module.__init__(self)
+    self.hidden_size = config.hidden_size
+
+    if config.use_mla:
+        self.self_attn = DeepseekV2Attention(config=config, layer_idx=layer_idx)
+    else:
+        config.head_dim = config.hidden_size // config.num_attention_heads
+        self.self_attn = LlamaAttention(config, layer_idx)
+    self.mlp = DeepseekV2MoE(config) if layer_idx >= config.first_k_dense_replace else DeepseekV2MLP(config)
+
+    self.input_layernorm = DeepseekV2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+    self.post_attention_layernorm = DeepseekV2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+
+DeepseekV2DecoderLayer.__init__ = decoder_layer_init
+
+class DeepseekOCRConfig(DeepseekV2Config):
+    model_type = "DeepseekOCR"
+
+class DeepseekOCRModel(DeepseekV2Model):
+    config_class = DeepseekOCRConfig
+
+    def __init__(self, config: DeepseekV2Config):
+        super(DeepseekOCRModel, self).__init__(config)
+
+        self.sam_model = build_sam_vit_b()
+        self.vision_model = build_clip_l()
+        # self.conv_2 = nn.Conv2d(in_channels=1024, out_channels=2048, kernel_size=2, stride=2)
+        n_embed = 1280
+        self.projector =  MlpProjector(Dict(projector_type="linear", input_dim=2048, n_embed=n_embed))
+        embed_std = 1 / torch.sqrt(torch.tensor(n_embed, dtype=torch.float32))
+        self.image_newline = nn.Parameter(torch.randn(n_embed) * embed_std)
+        self.view_seperator = nn.Parameter(torch.randn(n_embed) * embed_std)
+
+        self.rotary_emb = LlamaRotaryEmbedding(config=config)
+    
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        images: Optional[torch.FloatTensor] = None,
+        images_seq_mask: Optional[torch.FloatTensor] = None,
+        images_spatial_crop: Optional[torch.FloatTensor] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+
+
+
+        if inputs_embeds is None:
+            # inputs_embeds = self.embed_tokens(input_ids)
+            inputs_embeds = self.get_input_embeddings()(input_ids)
+
+        inputs_embeds = inputs_embeds.clone()
+
+        sam_model = getattr(self, 'sam_model', None)
+        # sam_model = self.sam_model
+        vision_model = getattr(self, 'vision_model', None)
+
+
+
+        if sam_model is not None and (input_ids.shape[1] != 1 or self.training) and torch.sum(images[0][1]).item() != 0:
+
+            idx = 0
+            
+            # sam_model = torch.jit.script(sam_model)
+            
+            # start_time = time.time()
+            for image, crop_shape in zip(images, images_spatial_crop):
+                images_in_this_batch = []
+
+                patches = image[0]
+                image_ori = image[1]
+
+                with torch.no_grad():
+                # with torch.inference_mode(): 
+                    
+                    if torch.sum(patches).item() != 0:
+                        # P, C, H, W = patches.shape
+                        crop_flag = 1
+                        local_features_1 = sam_model(patches)
+
+                        local_features_2 = vision_model(patches, local_features_1)  
+                        # vit_time = time.time()
+                        local_features = torch.cat((local_features_2[:, 1:], local_features_1.flatten(2).permute(0, 2, 1)), dim=-1) 
+                        local_features = self.projector(local_features)
+
+
+                        global_features_1 = sam_model(image_ori)
+                        global_features_2 = vision_model(image_ori, global_features_1) 
+                        global_features = torch.cat((global_features_2[:, 1:], global_features_1.flatten(2).permute(0, 2, 1)), dim=-1) 
+                        global_features = self.projector(global_features)
+
+                        print('=====================')
+                        print('BASE: ', global_features.shape)
+                        print('PATCHES: ', local_features.shape)
+                        print('=====================')
+
+                        _, hw, n_dim = global_features.shape
+                        h = w = int(hw ** 0.5)
+
+                        _2, hw2, n_dim2 = local_features.shape
+                        h2 = w2 = int(hw2 ** 0.5)
+
+                        width_crop_num, height_crop_num = crop_shape[0], crop_shape[1]
+
+                        global_features = global_features.view(h, w, n_dim)
+
+                        global_features = torch.cat(
+                            [global_features, self.image_newline[None, None, :].expand(h, 1, n_dim)], dim=1
+                        )
+
+                        global_features = global_features.view(-1, n_dim)
+
+
+                        local_features = local_features.view(height_crop_num, width_crop_num, h2, w2, n_dim2).permute(0, 2, 1, 3, 4).reshape(height_crop_num*h2, width_crop_num*w2, n_dim2)
+                        local_features = torch.cat(
+                            [local_features, self.image_newline[None, None, :].expand(height_crop_num * h2, 1, n_dim2)], dim=1
+                        )
+                        local_features = local_features.view(-1, n_dim2)
+
+                        global_local_features = torch.cat([local_features, global_features, self.view_seperator[None, :]], dim=0)
+
+                        # end_time = time.time()
+
+                        # print('sam: ', sam_time - start_time)
+                        # print('vit: ', vit_time - sam_time)
+                        # print('all: ', end_time - start_time)
+
+                        # exit()
+                   
+                    else:
+                        global_features_1 = sam_model(image_ori)
+                        global_features_2 = vision_model(image_ori, global_features_1) 
+                        global_features = torch.cat((global_features_2[:, 1:], global_features_1.flatten(2).permute(0, 2, 1)), dim=-1) 
+                        global_features = self.projector(global_features)
+                        _, hw, n_dim = global_features.shape
+                        h = w = int(hw ** 0.5)
+
+
+                        global_features = global_features.view(h, w, n_dim)
+
+                        global_features = torch.cat(
+                            [global_features, self.image_newline[None, None, :].expand(h, 1, n_dim)], dim=1
+                        )
+
+                        global_features = global_features.view(-1, n_dim)
+
+                        global_local_features = torch.cat([global_features, self.view_seperator[None, :]], dim=0)
+
+                    images_in_this_batch.append(global_local_features)
+                
+
+                if images_in_this_batch:
+                    images_in_this_batch = torch.cat(images_in_this_batch, dim=0)
+                    images_in_this_batch = images_in_this_batch.to(
+                        device=inputs_embeds.device, dtype=inputs_embeds.dtype
+                    )
+                    mask = images_seq_mask[idx].unsqueeze(-1).to(inputs_embeds.device)   # bool [T, 1]
+                    updated_row = inputs_embeds[idx].masked_scatter(mask, images_in_this_batch)
+                    inputs_embeds[idx] = updated_row
+
+                idx += 1
+            
+        return super(DeepseekOCRModel, self).forward(
+            input_ids=None, attention_mask=attention_mask, past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds, use_cache=use_cache, position_ids = position_ids,
+            output_attentions=output_attentions, output_hidden_states=output_hidden_states,
+            return_dict=return_dict
+        )
+    
+
+class DeepseekOCRForCausalLM(DeepseekV2ForCausalLM):
+
+    config_class = DeepseekOCRConfig
+    # supports_gradient_checkpointing = True
+
+    def __init__(self, config):
+        super(DeepseekV2ForCausalLM, self).__init__(config)
+        self.model = DeepseekOCRModel(config)
+
+        self.vocab_size = config.vocab_size
+
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_model(self):
+        return self.model
+
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        images: Optional[torch.FloatTensor] = None,
+        images_seq_mask: Optional[torch.FloatTensor] = None,
+        images_spatial_crop: Optional[torch.FloatTensor] = None,
+        return_dict: Optional[bool] = None,
+        
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        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
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+
+
+        outputs  = self.model(
+            input_ids=input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            images=images,
+            images_seq_mask = images_seq_mask,
+            images_spatial_crop = images_spatial_crop,
+            return_dict=return_dict
+            
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+        logits = logits.float()
+
+        # logits
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+    def prepare_inputs_for_generation(
+        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+        # Omit tokens covered by past_key_values
+        past_length = 0
+        if past_key_values is not None:
+            if isinstance(past_key_values, Cache):
+                cache_length = past_key_values.get_seq_length()
+                past_length = past_key_values.get_seq_length()
+                max_cache_length = None
+            else:
+                cache_length = past_length = past_key_values[0][0].shape[2]
+                max_cache_length = None
+
+            # Keep only the unprocessed tokens:
+            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
+            # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
+            # input)
+            if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
+                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
+            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
+            # input_ids based on the past_length.
+            elif past_length < input_ids.shape[1]:
+                input_ids = input_ids[:, past_length:]
+            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
+
+            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
+            if (
+                max_cache_length is not None
+                and attention_mask is not None
+                and cache_length + input_ids.shape[1] > max_cache_length
+            ):
+                attention_mask = attention_mask[:, -max_cache_length:]
+
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+        # if self.generation_config.cache_implementation == "static":
+        #     # generation with static cache
+        #     cache_position = kwargs.get("cache_position", None)
+        #     if cache_position is None:
+        #         past_length = 0
+        #     else:
+        #         past_length = cache_position[-1] + 1
+        #     input_ids = input_ids[:, past_length:]
+        #     position_ids = position_ids[:, past_length:]
+
+        # TODO @gante we should only keep a `cache_position` in generate, and do +=1.
+        # same goes for position ids. Could also help with continued generation.
+        cache_position = torch.arange(past_length, past_length + position_ids.shape[-1], device=position_ids.device)
+
+        # if `inputs_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 = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+                "images": kwargs.get("images", None),
+                "images_seq_mask": kwargs.get("images_seq_mask", None),
+                "images_spatial_crop": kwargs.get("images_spatial_crop", None),
+            }
+        )
+        return model_inputs
+    
+
+    def disable_torch_init(self):
+        """
+        Disable the redundant torch default initialization to accelerate model creation.
+        """
+        import torch
+        setattr(torch.nn.Linear, "reset_parameters", lambda self: None)
+        setattr(torch.nn.LayerNorm, "reset_parameters", lambda self: None)
+
+
+
+    def infer(self, tokenizer, prompt='', image_file='', output_path = '', base_size=1024, image_size=640, crop_mode=True, test_compress=False, save_results=False, eval_mode=False):
+        self.disable_torch_init()
+
+        os.makedirs(output_path, exist_ok=True)
+        os.makedirs(f'{output_path}/images', exist_ok=True)
+
+        if prompt and image_file:
+            conversation = [
+                {
+                    "role": "<|User|>",
+                    # "content": "<image>\n<|grounding|>Given the layout of the image. ",
+                    "content": f'{prompt}',
+                    # "content": "君不见黄河之水天上来的下一句是什么？",
+                    # "content": "<image>\nFree OCR. ",
+                    # "content": "<image>\nParse the figure. ",
+                    # "content": "<image>\nExtract the text in the image. ",
+                    "images": [f'{image_file}'],
+                },
+                {"role": "<|Assistant|>", "content": ""},
+            ]
+        
+        elif prompt:
+            conversation = [
+                {
+                    "role": "<|User|>",
+                    # "content": "<image>\n<|grounding|>Given the layout of the image. ",
+                    "content": f'{prompt}',
+                    # "content": "君不见黄河之水天上来的下一句是什么？",
+                    # "content": "<image>\nFree OCR. ",
+                    # "content": "<image>\nParse the figure. ",
+                    # "content": "<image>\nExtract the text in the image. ",
+                    # "images": [f'{image_file}'],
+                },
+                {"role": "<|Assistant|>", "content": ""},
+            ]
+        else:
+            assert False, f'prompt is none!'
+        
+        prompt = format_messages(conversations=conversation, sft_format='plain', system_prompt='')
+
+        patch_size = 16
+        downsample_ratio = 4
+        images = load_pil_images(conversation)
+
+        valid_img_tokens = 0
+        ratio = 1
+
+        image_draw = images[0].copy()
+
+        w,h = image_draw.size
+        # print(w, h)
+        ratio = 1 - ((max(w, h) - min(w, h)) / (max(w, h)))
+    
+
+        image_transform=BasicImageTransform(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), normalize=True)
+        images_seq_mask = []
+
+        image_token = '<image>'
+        image_token_id = 128815
+        text_splits = prompt.split(image_token)
+
+        images_list, images_crop_list, images_seq_mask = [], [], []
+        tokenized_str = []
+        images_spatial_crop = []
+        for text_sep, image in zip(text_splits, images):
+
+            tokenized_sep = text_encode(tokenizer, text_sep, bos=False, eos=False)
+            tokenized_str += tokenized_sep
+            images_seq_mask += [False] * len(tokenized_sep)
+
+            if crop_mode:
+
+                if image.size[0] <= 640 and image.size[1] <= 640:
+                    crop_ratio = [1, 1]
+
+                else:
+                    if crop_mode:
+                        # best_width, best_height = select_best_resolution(image.size, self.candidate_resolutions)
+                        images_crop_raw, crop_ratio = dynamic_preprocess(image)
+                    else:
+                        # best_width, best_height = self.image_size, self.image_size
+                        crop_ratio = [1, 1]
+                
+                """process the global view"""
+                # image = image.resize((base_size, base_size))
+                global_view = ImageOps.pad(image, (base_size, base_size),
+                                        color=tuple(int(x * 255) for x in image_transform.mean))
+                
+                if base_size == 1024:
+                    valid_img_tokens += int(256 * ratio)
+                elif base_size == 1280:
+                    valid_img_tokens += int(400 * ratio)
+                # elif base_size == 640:
+                #     valid_img_tokens += int(100 * ratio)
+                
+
+
+
+                
+                images_list.append(image_transform(global_view).to(torch_dtype))
+
+                # global_view_tensor = image_transform(global_view).to(torch_dtype)
+
+                width_crop_num, height_crop_num = crop_ratio
+
+                images_spatial_crop.append([width_crop_num, height_crop_num])
+                
+                
+                if width_crop_num > 1 or height_crop_num > 1:
+                    """process the local views"""
+                    
+                    for i in range(len(images_crop_raw)):
+                        images_crop_list.append(image_transform(images_crop_raw[i]).to(torch_dtype))
+                
+                if image_size == 640:
+                    valid_img_tokens += len(images_crop_list) * 100
+
+                num_queries = math.ceil((image_size // patch_size) / downsample_ratio)
+                num_queries_base = math.ceil((base_size // patch_size) / downsample_ratio)
+
+
+
+                """add image tokens"""
+
+                
+
+                tokenized_image = ([image_token_id] * num_queries_base + [image_token_id]) * num_queries_base
+                tokenized_image += [image_token_id]
+                if width_crop_num > 1 or height_crop_num > 1:
+                    tokenized_image += ([image_token_id] * (num_queries * width_crop_num) + [image_token_id]) * (
+                                num_queries * height_crop_num)
+                tokenized_str += tokenized_image
+                images_seq_mask += [True] * len(tokenized_image)
+                # num_image_tokens.append(len(tokenized_image))
+
+            else:
+                # best_width, best_height = self.image_size, self.image_size
+                # print(image.size, (best_width, best_height)) # check the select_best_resolutions func
+
+                """process the global view"""
+                if image_size <= 640:
+                    print('directly resize')
+                    image = image.resize((image_size, image_size))
+                # else:
+                global_view = ImageOps.pad(image, (image_size, image_size),
+                                        color=tuple(int(x * 255) for x in image_transform.mean))
+                images_list.append(image_transform(global_view).to(torch_dtype))
+
+                if base_size == 1024:
+                    valid_img_tokens += int(256 * ratio)
+                elif base_size == 1280:
+                    valid_img_tokens += int(400 * ratio)
+                elif base_size == 640:
+                    valid_img_tokens += int(100 * 1)
+                elif base_size == 512:
+                    valid_img_tokens += int(64 * 1)
+
+                width_crop_num, height_crop_num = 1, 1
+
+                images_spatial_crop.append([width_crop_num, height_crop_num])
+
+
+                """add image tokens"""
+                num_queries = math.ceil((image_size // patch_size) / downsample_ratio)
+
+                tokenized_image = ([image_token_id] * num_queries + [image_token_id]) * num_queries
+                tokenized_image += [image_token_id]
+                # tokenized_image += ([self.image_token_id] * (num_queries * width_crop_num) + [self.image_token_id]) * (
+                #             num_queries * height_crop_num)
+                tokenized_str += tokenized_image
+                images_seq_mask += [True] * len(tokenized_image)
+                # num_image_tokens.append(len(tokenized_image))
+        
+
+        """process the last text split"""
+        tokenized_sep = text_encode(tokenizer, text_splits[-1], bos=False, eos=False)
+        tokenized_str += tokenized_sep
+        images_seq_mask += [False] * len(tokenized_sep)
+
+        """add the bos tokens"""
+        bos_id = 0
+        tokenized_str = [bos_id] + tokenized_str 
+        images_seq_mask = [False] + images_seq_mask
+
+
+
+        input_ids = torch.LongTensor(tokenized_str)
+
+        images_seq_mask = torch.tensor(images_seq_mask, dtype=torch.bool)
+
+
+        if len(images_list) == 0:
+            images_ori = torch.zeros((1, 3, image_size, image_size))
+            images_spatial_crop = torch.zeros((1, 2), dtype=torch.long)
+            images_crop = torch.zeros((1, 3, base_size, base_size))
+
+        else:
+            images_ori = torch.stack(images_list, dim=0)
+            images_spatial_crop = torch.tensor(images_spatial_crop, dtype=torch.long)
+            if images_crop_list:
+                images_crop = torch.stack(images_crop_list, dim=0)
+            else:
+                images_crop = torch.zeros((1, 3, base_size, base_size))
+
+
+
+        if not eval_mode:
+            streamer = NoEOSTextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=False)
+            with torch.autocast("cuda", dtype=torch_dtype):
+                with torch.no_grad():
+                    output_ids = self.generate(
+                        input_ids.unsqueeze(0).cuda(),
+                        images=[(images_crop.cuda(), images_ori.cuda())],
+                        images_seq_mask = images_seq_mask.unsqueeze(0).cuda(),
+                        images_spatial_crop = images_spatial_crop,
+                        # do_sample=False,
+                        # num_beams = 1,
+                        temperature=0.0,
+                        eos_token_id=tokenizer.eos_token_id,
+                        streamer=streamer,
+                        max_new_tokens=8192,
+                        no_repeat_ngram_size = 20,
+                        use_cache = True
+                        )
+
+        else:
+            with torch.autocast("cuda", dtype=torch_dtype):
+                with torch.no_grad():
+                    output_ids = self.generate(
+                        input_ids.unsqueeze(0).cuda(),
+                        images=[(images_crop.cuda(), images_ori.cuda())],
+                        images_seq_mask = images_seq_mask.unsqueeze(0).cuda(),
+                        images_spatial_crop = images_spatial_crop,
+                        # do_sample=False,
+                        # num_beams = 1,
+                        temperature=0.0,
+                        eos_token_id=tokenizer.eos_token_id,
+                        max_new_tokens=8192,
+                        no_repeat_ngram_size = 35,
+                        use_cache = True
+                        )
+                
+
+        if '<image>' in conversation[0]['content'] and eval_mode:
+                outputs = tokenizer.decode(output_ids[0, input_ids.unsqueeze(0).cuda().shape[1]:])
+                stop_str = '<｜end▁of▁sentence｜>'
+                if outputs.endswith(stop_str):
+                    outputs = outputs[:-len(stop_str)]
+                # re_match
+                outputs = outputs.strip()
+
+                return outputs
+        
+        if '<image>' in conversation[0]['content'] and test_compress:
+            outputs = tokenizer.decode(output_ids[0, input_ids.unsqueeze(0).cuda().shape[1]:])
+            pure_texts_outputs_token_length = len(text_encode(tokenizer, outputs, bos=False, eos=False))
+            print('='*50)
+            print('image size: ', (w, h))
+            print('valid image tokens: ', int(valid_img_tokens))
+            print('output texts tokens (valid): ', pure_texts_outputs_token_length)
+            print('compression ratio: ', round(pure_texts_outputs_token_length/valid_img_tokens, 2))
+            print('='*50)
+
+
+        if '<image>' in conversation[0]['content'] and save_results:
+            outputs = tokenizer.decode(output_ids[0, input_ids.unsqueeze(0).cuda().shape[1]:])
+            stop_str = '<｜end▁of▁sentence｜>'
+
+            print('='*15 + 'save results:' + '='*15)
+            
+            # # # # conv.messages[-1][-1] = outputs
+            if outputs.endswith(stop_str):
+                outputs = outputs[:-len(stop_str)]
+            outputs = outputs.strip()
+
+            matches_ref, matches_images, mathes_other = re_match(outputs)
+            # print(matches_ref)
+            result = process_image_with_refs(image_draw, matches_ref, output_path)
+
+
+            for idx, a_match_image in enumerate(tqdm(matches_images, desc="image")):
+                outputs = outputs.replace(a_match_image, '![](images/' + str(idx) + '.jpg)\n')
+            
+            for idx, a_match_other in enumerate(tqdm(mathes_other, desc="other")):
+                outputs = outputs.replace(a_match_other, '').replace('\\coloneqq', ':=').replace('\\eqqcolon', '=:')
+
+
+            # if 'structural formula' in conversation[0]['content']:
+            #     outputs = '<smiles>' + outputs + '</smiles>'
+            with open(f'{output_path}/result.mmd', 'w', encoding = 'utf-8') as afile:
+                afile.write(outputs)
+
+            if 'line_type' in outputs:
+                import matplotlib.pyplot as plt
+                lines = eval(outputs)['Line']['line']
+
+                line_type = eval(outputs)['Line']['line_type']
+                # print(lines)
+
+                endpoints = eval(outputs)['Line']['line_endpoint']
+
+                fig, ax = plt.subplots(figsize=(3,3), dpi=200)
+                ax.set_xlim(-15, 15)
+                ax.set_ylim(-15, 15)
+
+                for idx, line in enumerate(lines):
+                    try:
+                        p0 = eval(line.split(' -- ')[0])
+                        p1 = eval(line.split(' -- ')[-1])
+
+                        if line_type[idx] == '--':
+                            ax.plot([p0[0], p1[0]], [p0[1], p1[1]], linewidth=0.8, color='k')
+                        else:
+                            ax.plot([p0[0], p1[0]], [p0[1], p1[1]], linewidth = 0.8, color = 'k')
+
+                        ax.scatter(p0[0], p0[1], s=5, color = 'k')
+                        ax.scatter(p1[0], p1[1], s=5, color = 'k')
+                    except:
+                        pass
+
+                for endpoint in endpoints:
+
+                    label = endpoint.split(': ')[0]
+                    (x, y) = eval(endpoint.split(': ')[1])
+                    ax.annotate(label, (x, y), xytext=(1, 1), textcoords='offset points', 
+                                fontsize=5, fontweight='light')
+                
+
+                plt.savefig(f'{output_path}/geo.jpg')
+                plt.close()
+
+            result.save(f"{output_path}/result_with_boxes.jpg")

special_tokens_map.json

@@ -0,0 +1,27 @@
+{
+  "additional_special_tokens": [
+    "<|User|>",
+    "<|Assistant|>"
+  ],
+  "bos_token": {
+    "content": "<｜begin▁of▁sentence｜>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<｜end▁of▁sentence｜>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<｜▁pad▁｜>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}