Add deepseek-ocr-2

DeepSeek-OCR2-hf/run_dpsk_ocr2.py

+from transformers import AutoModel, AutoTokenizer
+import torch
+import os
+import argparse
+
+parse = argparse.ArgumentParser()
+parse.add_argument('--model_name_or_path', type=str, default='deepseek-ai/DeepSeek-OCR-2')
+parse.add_argument('--image_file', type=str, default='doc/docstructbench_dianzishu_zhongwenzaixian-o.O-63686436.pdf_57.jpg')
+parse.add_argument('--output_path', type=str, default='output/image')
+args = parse.parse_args()
+
+if __name__ == '__main__':
+    tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, trust_remote_code=True)
+    model = AutoModel.from_pretrained(args.model_name_or_path, _attn_implementation='flash_attention_2', trust_remote_code=True, use_safetensors=True)
+    model = model.eval().cuda().to(torch.bfloat16)
+
+    # prompt = "<image>\nFree OCR. "
+    prompt = "<image>\n<|grounding|>Convert the document to markdown. "
+
+    res = model.infer(tokenizer, prompt=prompt, image_file=args.image_file, output_path =args.output_path, base_size = 1024, image_size = 768, crop_mode=True, save_results = True)
+    print("process end, result saved to ", args.output_path)
\ No newline at end of file

DeepSeek-OCR2-vllm/config.py

+
+BASE_SIZE = 1024
+IMAGE_SIZE = 768
+CROP_MODE = True
+MIN_CROPS= 2
+MAX_CROPS= 6 # max:6
+MAX_CONCURRENCY = 100 # If you have limited GPU memory, lower the concurrency count.
+NUM_WORKERS = 64 # image pre-process (resize/padding) workers
+PRINT_NUM_VIS_TOKENS = False
+SKIP_REPEAT = True
+MODEL_PATH = 'deepseek-ai/DeepSeek-OCR-2' # change to your model path
+
+# TODO: change INPUT_PATH
+# .pdf: run_dpsk_ocr_pdf.py;
+# .jpg, .png, .jpeg: run_dpsk_ocr_image.py;
+# Omnidocbench images path: run_dpsk_ocr_eval_batch.py
+
+
+INPUT_PATH = 'doc/docstructbench_dianzishu_zhongwenzaixian-o.O-63686436.pdf_57.jpg'
+OUTPUT_PATH = 'output/image/'
+
+PROMPT = '<image>\n<|grounding|>Convert the document to markdown.'
+# PROMPT = '<image>\nFree OCR.'
+# PROMPT = '<image>\nParse the figure.'
+# TODO commonly used prompts
+# document: <image>\n<|grounding|>Convert the document to markdown.
+# other image: <image>\n<|grounding|>OCR this image.
+# without layouts: <image>\nFree OCR.
+# figures in document: <image>\nParse the figure.
+# general: <image>\nDescribe this image in detail.
+# rec: <image>\nLocate <|ref|>xxxx<|/ref|> in the image.
+# .......
+
+
+from transformers import AutoTokenizer
+
+TOKENIZER = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True)
\ No newline at end of file

DeepSeek-OCR2-vllm/deepencoderv2/build_linear.py

+import torch.nn as nn
+import torch
+import torch.nn.functional as F
+import copy
+
+
+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
+

DeepSeek-OCR2-vllm/deepencoderv2/qwen2_d2e.py

+import torch
+import torch.nn as nn
+import transformers
+
+
+
+
+class CustomQwen2Decoder(nn.Module):
+    """
+    Qwen2 visual encoder
+    non-causal attention + causal attention
+    token_type_ids ：0=non-causal, 1=causal
+    """
+    
+    def __init__(
+        self,
+        decoder_layer: int = 24,
+        max_position_embeddings: int = 131072,
+        hidden_dimension: int = 896,
+        num_attention_heads: int = 14,
+        num_key_value_heads: int = 2,
+        intermediate_size: int = 4864,
+        vocab_size: int = 151936,
+        attn_implementation: str = "sdpa",  # ⭐ 
+        rms_norm_eps: float = 1e-06,
+        rope_theta: float = 1000000.0,
+        attention_dropout: float = 0.0,
+        hidden_act: str = "silu",
+        initializer_range: float = 0.02,
+    ):
+        super().__init__()
+        
+        # attn_implementation check
+        if attn_implementation == "flash_attention_2":
+            raise ValueError(
+                "CustomQwen2Decoder do not support flash_attention_2，"
+                "new attention mask needs 'sdpa' or 'eager'"
+            )
+        
+        # load
+        Qwen2Model = getattr(transformers.models.qwen2.modeling_qwen2, 'Qwen2Model')
+        Qwen2Config = getattr(transformers, 'Qwen2Config')
+        
+        # config
+        config = Qwen2Config(
+            hidden_size=hidden_dimension,
+            num_hidden_layers=decoder_layer,
+            num_attention_heads=num_attention_heads,
+            num_key_value_heads=num_key_value_heads,
+            intermediate_size=intermediate_size,
+            max_position_embeddings=max_position_embeddings,
+            vocab_size=vocab_size,
+            rms_norm_eps=rms_norm_eps,
+            rope_theta=rope_theta,
+            attention_dropout=attention_dropout,
+            hidden_act=hidden_act,
+            initializer_range=initializer_range,
+            _attn_implementation=attn_implementation,  # ⭐ 
+        )
+        
+        # 
+        self.model = self._create_custom_model(Qwen2Model, config)
+
+        del self.model.embed_tokens 
+    
+    def _create_custom_model(self, Qwen2Model, config):
+        """ Qwen2Model """
+        
+        class CustomQwen2ModelInner(Qwen2Model):
+
+            
+            def forward(
+                self,
+                input_ids=None,
+                attention_mask=None,
+                position_ids=None,
+                past_key_values=None,
+                inputs_embeds=None,
+                token_type_ids=None,  # ⭐
+                use_cache=None,
+                output_attentions=None,
+                output_hidden_states=None,
+                return_dict=None,
+                cache_position=None,
+            ):
+                # token_type_ids
+                self._current_token_type_ids = token_type_ids
+                
+                outputs = super().forward(
+                    input_ids=input_ids,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_values=past_key_values,
+                    inputs_embeds=inputs_embeds,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                    output_hidden_states=output_hidden_states,
+                    return_dict=return_dict,
+                    cache_position=cache_position,
+                )
+                
+                return outputs
+            
+            def _update_causal_mask(
+                self,
+                attention_mask,
+                input_tensor,
+                cache_position,
+                past_key_values,
+                output_attentions,
+            ):
+                dtype, device = input_tensor.dtype, input_tensor.device
+                min_dtype = torch.finfo(dtype).min
+                batch_size, sequence_length = input_tensor.shape[0], input_tensor.shape[1]
+                
+                token_type_ids = self._current_token_type_ids
+                
+                # attention mask
+                causal_mask = self._create_custom_4d_mask(
+                    sequence_length=sequence_length,
+                    dtype=dtype,
+                    device=device,
+                    batch_size=batch_size,
+                    token_type_ids=token_type_ids,
+                )
+                
+                #  padding mask
+                if attention_mask is not None and attention_mask.dim() == 2:
+                    padding_mask = attention_mask[:, None, None, :].to(dtype=dtype)
+                    padding_mask = (1.0 - padding_mask) * min_dtype
+                    causal_mask = causal_mask + padding_mask
+                
+                return causal_mask
+            
+            def _create_custom_4d_mask(
+                self,
+                sequence_length,
+                dtype,
+                device,
+                batch_size,
+                token_type_ids,
+            ):
+                min_dtype = torch.finfo(dtype).min
+                
+                masks = []
+                for b in range(batch_size):
+                    mask = torch.full(
+                        (sequence_length, sequence_length),
+                        fill_value=min_dtype,
+                        dtype=dtype,
+                        device=device
+                    )
+                    
+                    type_ids = token_type_ids[b]
+                    
+                    image_positions = (type_ids == 0).nonzero(as_tuple=True)[0]
+                    text_positions = (type_ids == 1).nonzero(as_tuple=True)[0]
+                    
+                    # non-casual
+                    if len(image_positions) > 0:
+                        mask[image_positions[:, None], image_positions] = 0.0
+                    
+                    # causal
+                    for i, text_pos in enumerate(text_positions):
+                        if len(image_positions) > 0:
+                            mask[text_pos, image_positions] = 0.0
+                        mask[text_pos, text_positions[:i+1]] = 0.0
+                    
+                    masks.append(mask)
+                
+                mask = torch.stack(masks, dim=0).unsqueeze(1)
+                return mask
+        
+        return CustomQwen2ModelInner(config)
+    
+    def forward(
+        self,
+        inputs_embeds,
+        token_type_ids,
+        attention_mask=None,
+        **kwargs
+    ):
+        """
+        Args:
+            inputs_embeds: [batch_size, seq_len, hidden_dim]
+            token_type_ids: [batch_size, seq_len], 0=non-causal, 1=causal
+            attention_mask: [batch_size, seq_len], optional
+        """
+        return self.model(
+            inputs_embeds=inputs_embeds,
+            token_type_ids=token_type_ids,
+            attention_mask=attention_mask,
+            **kwargs
+        )
+
+
+
+
+
+# batch_size = 2
+# inputs_embeds = torch.randn(batch_size, 512, 896).cuda()
+
+# inputs_embeds = torch.randn(batch_size, 512, 896).cuda()
+# token_type_ids = torch.cat([
+#     torch.zeros(batch_size, 256, dtype=torch.long),
+#     torch.ones(batch_size, 256, dtype=torch.long),
+# ], dim=1).cuda()
+
+# # start = time.time()
+# with torch.no_grad():
+#     outputs_sdpa = decoder_sdpa(inputs_embeds, token_type_ids)
+#     print(outputs_sdpa[0].shape)
+# print(f"SDPA time: {time.time() - start:.4f}s")
+
+
+
+class Qwen2Decoder2Encoder(nn.Module):
+    """
+    Decoder based on Multilingual BART
+    Set the initial weights and configuration with a pretrained multilingual BART model,
+    and modify the detailed configurations as a Nougat decoder
+    """
+
+    def __init__(
+        self,
+        decoder_layer: int,
+        hidden_dimension: int,
+        num_attention_heads: int,
+        num_key_value_heads: int,
+        intermediate_size: int,
+        max_query: int,
+    ):
+        super().__init__()
+
+        self.model = CustomQwen2Decoder(
+            decoder_layer=decoder_layer,
+            hidden_dimension=hidden_dimension,
+            num_attention_heads=num_attention_heads,
+            num_key_value_heads=num_key_value_heads,
+            intermediate_size=intermediate_size,
+            attn_implementation="sdpa", 
+        )
+
+
+
+
+        self.query_768 = nn.Embedding(144, hidden_dimension)
+        self.query_1024 = nn.Embedding(256, hidden_dimension)
+
+
+        # self.query_refixation = nn.Embedding(int(math.sqrt(max_query)), hidden_dimension)
+
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x  = x.flatten(2).transpose(1, 2)
+
+        bs, n_query, _ = x.shape
+
+        if n_query == 144:
+            param_img = self.query_768.weight
+        elif n_query == 256:
+            param_img = self.query_1024.weight
+
+        batch_query_imgs = param_img.unsqueeze(0).expand(
+            bs, -1, -1
+        )  # (batch_size, num_queries, hidden_size)
+
+
+
+        x_combined = torch.cat([x, batch_query_imgs], dim=1)
+
+        token_type_ids = torch.cat([
+            torch.zeros(bs, n_query, dtype=torch.long),
+            torch.ones(bs, n_query, dtype=torch.long),
+        ], dim=1)
+
+
+        y = self.model(x_combined, token_type_ids)[0]
+
+
+        y = y[:, n_query:, :] # causal flow query
+
+
+        return y
+
+
+def build_qwen2_decoder_as_encoder(
+    decoder_layer=24,
+    hidden_dimension=896,
+    num_attention_heads=14,
+    num_key_value_heads=2,
+    intermediate_size=4864,
+    max_query = 400,
+    checkpoint=None,
+):
+
+    decoder_as_encoder = Qwen2Decoder2Encoder(
+            decoder_layer=decoder_layer,
+            hidden_dimension = hidden_dimension,
+            num_attention_heads = num_attention_heads,
+            num_key_value_heads = num_key_value_heads,
+            intermediate_size = intermediate_size,
+            max_query = max_query
+        )
+
+
+
+    
+    if checkpoint is not None:
+        # with open(checkpoint, "rb") as f:
+        state_dict = torch.load(checkpoint)
+
+        decoder_as_encoder.load_state_dict(state_dict, strict=True)
+        # tob
+        print(checkpoint)
+    return decoder_as_encoder
+
+
+if __name__ == '__main__':
+
+    x = torch.zeros(2, 896, 16, 16).cuda()
+
+    net = build_qwen2_decoder_as_encoder(checkpoint = '').cuda()
+
+
+    y = net(x)
+
+    # y = y.flatten(2).permute(0, 2, 1)
+
+    print('-------shape---------')
+    print(y.shape)
+    print('-------------------')
\ No newline at end of file

DeepSeek-OCR2-vllm/deepencoderv2/sam_vary_sdpa.py

+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from typing import Optional, Tuple, Type
+from functools import partial
+from flash_attn import flash_attn_qkvpacked_func
+# from .common import LayerNorm2d, MLPBlock
+
+# from mmgpt.model.vision_encoder.flash_4 import _attention_rel_h_rel_w
+
+
+def get_abs_pos(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, 896, 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(self.pos_embed, x.size(1))
+
+        for blk in self.blocks:
+            x = blk(x)
+
+        neck_output  = self.neck(x.permute(0, 3, 1, 2))
+        conv2_output  = self.net_2(neck_output)
+        # print(f"conv2_output shape: {conv2_output.shape}")
+        conv3_output  = self.net_3(conv2_output)
+
+        return conv3_output 
+
+
+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)
+            # qkv = torch.stack([q, k, v], dim=1).transpose(1, 3).reshape(B, H * W, 3, self.num_heads, -1)
+            # x = flash_attn_qkvpacked_func(qkv, dropout_p=0.0, causal=False).transpose(1, 2)
+
+        
+
+        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(
+    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,
+        )
+    
+    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
\ No newline at end of file

DeepSeek-OCR2-vllm/deepseek_ocr2.py

+# SPDX-License-Identifier: Apache-2.0
+
+# adapted from https://github.com/deepseek-ai/DeepSeek-VL2/blob/faf18023f24b962b32d9f0a2d89e402a8d383a78/deepseek_vl2/models/modeling_deepseek_vl_v2.py
+"""Inference-only Deepseek-VL2 model compatible with HuggingFace weights."""
+import math
+from collections.abc import Iterable, Mapping, Sequence
+from typing import List, Literal, Optional, Set, Tuple, TypedDict, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange, repeat
+from transformers import BatchFeature
+
+from vllm.config import VllmConfig
+from vllm.model_executor import SamplingMetadata
+from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.model_loader.utils import set_default_torch_dtype
+from vllm.multimodal import MULTIMODAL_REGISTRY
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    MultiModalKwargs, NestedTensors)
+from vllm.multimodal.parse import (ImageEmbeddingItems, ImageProcessorItems,
+                                   ImageSize, MultiModalDataItems)
+from vllm.multimodal.processing import (BaseMultiModalProcessor,
+                                        BaseProcessingInfo, PromptReplacement,
+                                        PromptUpdate)
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
+from vllm.sequence import IntermediateTensors
+from vllm.transformers_utils.configs.deepseek_vl2 import (DeepseekVLV2Config,
+                                                          MlpProjectorConfig,
+                                                          VisionEncoderConfig)
+from process.image_process import (
+    DeepseekOCR2Processor, count_tiles)
+from vllm.transformers_utils.tokenizer import cached_tokenizer_from_config
+# from vllm.utils import is_list_of
+
+from vllm.model_executor.models.interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
+from vllm.model_executor.models.utils import (AutoWeightsLoader, WeightsMapper, flatten_bn,
+                    init_vllm_registered_model, maybe_prefix,
+                    merge_multimodal_embeddings)
+
+from deepencoderv2.sam_vary_sdpa import build_sam_vit_b
+# from deepencoder.clip_sdpa import build_clip_l
+from deepencoderv2.qwen2_d2e import build_qwen2_decoder_as_encoder
+from deepencoderv2.build_linear import MlpProjector
+from addict import Dict
+# import time
+from config import IMAGE_SIZE, BASE_SIZE, CROP_MODE, PRINT_NUM_VIS_TOKENS, PROMPT
+# The image token id may be various
+_IMAGE_TOKEN = "<image>"
+
+
+class DeepseekOCR2ProcessingInfo(BaseProcessingInfo):
+
+    def get_hf_config(self):
+        return self.ctx.get_hf_config(DeepseekVLV2Config)
+
+    def get_hf_processor(self, **kwargs: object):
+        return self.ctx.get_hf_processor(DeepseekOCR2Processor, **kwargs)
+
+    def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
+        return {"image": None}
+
+    def get_num_image_tokens(self,
+                             *,
+                             image_width: int,
+                             image_height: int,
+                             cropping: bool = True) -> int:
+        hf_processor = self.get_hf_processor()
+
+
+        # image_size = hf_processor.image_size
+        # patch_size = hf_processor.patch_size
+        # downsample_ratio = hf_processor.downsample_ratio
+
+        image_size = IMAGE_SIZE
+        base_size = BASE_SIZE
+        patch_size = 16
+        downsample_ratio = 4
+
+        if CROP_MODE:
+            if image_width <= 768 and image_height <= 768:
+                crop_ratio = [1, 1]
+            else:
+                # images_crop_raw, crop_ratio = hf_processor.dynamic_preprocess(image)
+
+                # find the closest aspect ratio to the target
+                crop_ratio = count_tiles(image_width, image_height, image_size=IMAGE_SIZE)
+
+                # print('===========')
+                # print('crop_ratio ', crop_ratio)
+                # print('============')
+                
+            num_width_tiles, num_height_tiles = crop_ratio
+        else:
+            num_width_tiles = num_height_tiles = 1
+
+        h = w = math.ceil((base_size // patch_size) / downsample_ratio)
+
+        h2 = w2 = math.ceil((image_size // patch_size) / downsample_ratio)
+
+        global_views_tokens = h * (w)
+        if num_width_tiles >1 or num_height_tiles>1:
+            local_views_tokens = (num_height_tiles * h2) * (num_width_tiles * w2)
+        else:
+            local_views_tokens = 0
+
+
+        return global_views_tokens + local_views_tokens + 1
+
+    def get_image_size_with_most_features(self) -> ImageSize:
+
+        if IMAGE_SIZE == 1024 and BASE_SIZE == 1280:
+            return ImageSize(width=1024*2, height=1024*2)
+        return ImageSize(width=768*2, height=768*2)
+
+
+class DeepseekOCR2DummyInputsBuilder(
+        BaseDummyInputsBuilder[DeepseekOCR2ProcessingInfo]):
+
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        num_images = mm_counts.get("image", 0)
+
+        processor = self.info.get_hf_processor()
+        image_token = processor.image_token
+
+        return image_token * num_images
+
+    def get_dummy_mm_data(
+        self,
+        seq_len: int,
+        mm_counts: Mapping[str, int],
+    ) -> MultiModalDataDict:
+        num_images = mm_counts.get("image", 0)
+
+        max_image_size = self.info.get_image_size_with_most_features()
+
+        if '<image>' in PROMPT:
+            return {
+                "image":
+                DeepseekOCR2Processor().tokenize_with_images(images = self._get_dummy_images(width=max_image_size.width,
+                                    height=max_image_size.height,
+                                    num_images=num_images), bos=True, eos=True, cropping=CROP_MODE)
+            }
+        else:
+            return {
+                "image": []
+            }
+
+
+
+
+class DeepseekOCR2MultiModalProcessor(
+        BaseMultiModalProcessor[DeepseekOCR2ProcessingInfo]):
+    
+
+    def _call_hf_processor(
+        self,
+        prompt: str,
+        mm_data: Mapping[str, object],
+        mm_kwargs: Mapping[str, object],
+    ) -> BatchFeature:
+        
+        
+        # print(mm_data)
+        if mm_data:
+            processed_outputs = self.info.ctx.call_hf_processor(
+                self.info.get_hf_processor(**mm_kwargs),
+                dict(prompt=prompt, **mm_data),
+                mm_kwargs,
+            )
+
+        else:
+            tokenizer = self.info.get_tokenizer()
+            processed_outputs = tokenizer(prompt,
+                                          add_special_tokens=True,
+                                          return_tensors="pt")
+
+        return processed_outputs
+
+    def _get_mm_fields_config(
+        self,
+        hf_inputs: BatchFeature,
+        hf_processor_mm_kwargs: Mapping[str, object],
+    ) -> Mapping[str, MultiModalFieldConfig]:
+        return dict(
+            pixel_values=MultiModalFieldConfig.batched("image"),
+            images_spatial_crop=MultiModalFieldConfig.batched("image"),
+            # image_embeds=MultiModalFieldConfig.batched("image2"),
+            images_crop=MultiModalFieldConfig.batched("image"),
+        )
+
+    def _get_prompt_updates(
+        self,
+        mm_items: MultiModalDataItems,
+        hf_processor_mm_kwargs: Mapping[str, object],
+        out_mm_kwargs: MultiModalKwargs,
+    ) -> Sequence[PromptUpdate]:
+        hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
+
+        image_token_id = hf_processor.image_token_id
+        assert isinstance(image_token_id, int)
+
+        def get_replacement_deepseek_vl2(item_idx: int):
+            images = mm_items.get_items(
+                "image", (ImageEmbeddingItems, ImageProcessorItems))
+
+
+
+            if isinstance(images, ImageEmbeddingItems):
+                num_image_tokens = images.get_feature_size(item_idx)
+            else:
+
+                
+                width = images[0][-1][0][0]
+                height = images[0][-1][0][1]
+
+                num_image_tokens = self.info.get_num_image_tokens(
+                    image_width=width,
+                    image_height=height,
+                    # flag = True,
+                    cropping=CROP_MODE,
+                )
+            return [image_token_id] * num_image_tokens
+
+        return [
+            PromptReplacement(
+                modality="image",
+                target=[image_token_id],
+                replacement=get_replacement_deepseek_vl2,
+            )
+        ]
+
+    def _cached_apply_hf_processor(
+        self,
+        prompt: Union[str, list[int]],
+        mm_data_items: MultiModalDataItems,
+        hf_processor_mm_kwargs: Mapping[str, object],
+    ) -> tuple[list[int], MultiModalKwargs, bool]:
+        # The processor logic is different for len(images) <= 2 vs > 2
+        # Since the processing cache assumes that the processor output is
+        # invariant of how many images are passed per prompt, we only
+        # perform caching for the most common case
+        if mm_data_items.get_count("image", strict=False) > 2:
+            # This code path corresponds to the cache being disabled
+            return self._apply_hf_processor_main(
+                prompt=prompt,
+                mm_items=mm_data_items,
+                hf_processor_mm_kwargs=hf_processor_mm_kwargs,
+                enable_hf_prompt_update=True,
+            )
+
+        return super()._cached_apply_hf_processor(
+            prompt=prompt,
+            mm_data_items=mm_data_items,
+            hf_processor_mm_kwargs=hf_processor_mm_kwargs,
+        )
+
+
+@MULTIMODAL_REGISTRY.register_processor(
+    DeepseekOCR2MultiModalProcessor,
+    info=DeepseekOCR2ProcessingInfo,
+    dummy_inputs=DeepseekOCR2DummyInputsBuilder)
+class DeepseekOCR2ForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
+
+    hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={
+        "language.": "language_model.",
+    })
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+
+        config: DeepseekVLV2Config = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
+        multimodal_config = vllm_config.model_config.multimodal_config
+
+        # config.model_type ='deepseek_vl_v2'
+
+        self.config = config
+        self.multimodal_config = multimodal_config
+
+
+        self.vision_config = config.vision_config
+        self.projector_config = config.projector_config
+        self.text_config = config.text_config
+
+        model_config = vllm_config.model_config
+        tokenizer = cached_tokenizer_from_config(model_config)
+        self.image_token_id = tokenizer.vocab[_IMAGE_TOKEN]
+
+        self.sam_model = build_sam_vit_b()
+        self.qwen2_model = build_qwen2_decoder_as_encoder()
+
+        n_embed = 1280
+        self.projector =  MlpProjector(Dict(projector_type="linear", input_dim=896, n_embed=n_embed))
+        self.tile_tag = config.tile_tag
+        self.global_view_pos = config.global_view_pos
+    
+        # self.sam_model = torch.compile(self.sam_model, mode="reduce-overhead")
+        # self.vision_model = torch.compile(self.vision_model, mode="reduce-overhead")
+        # self.projector = torch.compile(self.projector, mode="max-autotune")
+
+
+
+
+        # special token for image token sequence format
+        embed_std = 1 / torch.sqrt(torch.tensor(n_embed, dtype=torch.float32))
+        if self.tile_tag == "2D":
+            # <|view_separator|>, <|\n|>
+            # self.image_newline = nn.Parameter(torch.randn(n_embed) * embed_std)
+            self.view_seperator = nn.Parameter(torch.randn(n_embed) * embed_std)
+        else:
+            raise ValueError(
+                f"Only 2D tile_tag is supported currently, got: {self.tile_tag}"
+            )
+
+        if self.text_config.topk_method == "noaux_tc":
+            architectures = ["DeepseekV3ForCausalLM"]
+            # architectures = ["DeepseekForCausalLM"]
+        elif not self.text_config.use_mla:
+            architectures = ["DeepseekForCausalLM"]
+        else:
+            architectures = ["DeepseekV2ForCausalLM"]
+
+        self.language_model = init_vllm_registered_model(
+            vllm_config=vllm_config,
+            hf_config=self.text_config,
+            prefix=maybe_prefix(prefix, "language"),
+            architectures=architectures,
+        )
+
+        self.make_empty_intermediate_tensors = (
+            self.language_model.make_empty_intermediate_tensors)
+        
+
+
+        self.sam_model.to(dtype=torch.bfloat16)
+
+
+
+    def _parse_and_validate_image_input(
+            self, **kwargs: object):
+        
+        pixel_values = kwargs.pop("pixel_values", None)
+        images_spatial_crop = kwargs.pop("images_spatial_crop", None)
+        images_crop = kwargs.pop("images_crop", None)
+
+
+        if pixel_values is None or torch.sum(pixel_values).item() == 0:
+            return None
+
+        if pixel_values is not None:
+            if not isinstance(pixel_values, (torch.Tensor, list)):
+                raise ValueError("Incorrect type of pixel values. "
+                                 f"Got type: {type(pixel_values)}")
+
+            if not isinstance(images_spatial_crop, (torch.Tensor, list)):
+                raise ValueError("Incorrect type of image sizes. "
+                                 f"Got type: {type(images_spatial_crop)}")
+            
+            if not isinstance(images_crop, (torch.Tensor, list)):
+                raise ValueError("Incorrect type of image crop. "
+                                 f"Got type: {type(images_crop)}")
+
+            return [pixel_values, images_crop, images_spatial_crop]
+
+
+        raise AssertionError("This line should be unreachable.")
+    
+
+
+    def _pixel_values_to_embedding(
+        self,
+        pixel_values: torch.Tensor,
+        images_crop: torch.Tensor,
+        images_spatial_crop: torch.Tensor,
+    ) -> NestedTensors:
+
+        # Pixel_values (global view): [n_image, batch_size, 3, height, width]
+        # images_spatial_crop: [n_image, batch_size, [num_tiles_w, num_tiles_h]]
+        # images_crop (local view): [n_image, batch_size, num_pathes, 3, h, w]
+        # split the pixel and image_crop, all batch_size = 1
+
+        images_in_this_batch = []
+
+
+        with torch.no_grad():
+            for jdx in range(images_spatial_crop.size(0)):
+                # with torch.set_grad_enabled(False):
+                patches = images_crop[jdx][0].to(torch.bfloat16) # batch_size = 1
+                # patches = images_crop[jdx][0]
+                image_ori = pixel_values[jdx]
+                crop_shape = images_spatial_crop[jdx][0]
+
+                if torch.sum(patches).item() != 0:  # if all values = 0, no crop
+                    # P, C, H, W = patches.shape
+                    # crop_flag = 1
+                    local_features_1 = self.sam_model(patches)
+                    #TODO del patches 
+                    # torch.compiler.cudagraph_mark_step_begin()
+                    local_features_2 = self.qwen2_model(local_features_1)  
+
+
+                    # 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_2)
+
+
+                    global_features_1 = self.sam_model(image_ori)
+                    global_features_2 = self.qwen2_model(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_2)
+
+                    if PRINT_NUM_VIS_TOKENS:
+                        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)
+                
+                else:
+                    global_features_1 = self.sam_model(image_ori)
+                    global_features_2 = self.qwen2_model(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_2)
+
+                    if PRINT_NUM_VIS_TOKENS:
+                        print('=====================')
+                        print('BASE: ', global_features.shape)
+                        print('NO PATCHES')
+                        print('=====================')
+
+                    _, 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)
+
+        return images_in_this_batch
+
+    def _process_image_input(
+            self, image_input) -> torch.Tensor:
+        
+
+        # image_input: [pixel_values, images_crop, images_spatial_crop]
+    
+        pixel_values = image_input[0].to(torch.bfloat16)
+
+        # images_crop = image_input[1].to(torch.bfloat16)
+        images_crop = image_input[1]
+        # images_crop = image_input[1]
+        images_spatial_crop = image_input[2].to(dtype=torch.long)
+
+        # local_start = time.time()
+        vision_features = self._pixel_values_to_embedding(
+            pixel_values=pixel_values, images_crop = images_crop,  images_spatial_crop=images_spatial_crop)
+
+
+        return vision_features
+
+    def get_language_model(self) -> torch.nn.Module:
+        return self.language_model
+
+    def get_multimodal_embeddings(
+            self, **kwargs: object) -> Optional[MultiModalEmbeddings]:
+        image_input = self._parse_and_validate_image_input(**kwargs)
+        if image_input is None:
+            return None
+        vision_embeddings = self._process_image_input(image_input)
+        return vision_embeddings
+    
+
+
+    def get_input_embeddings(
+        self,
+        input_ids: torch.Tensor,
+        multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
+    ) -> torch.Tensor:
+        
+
+
+        inputs_embeds = self.language_model.get_input_embeddings(input_ids)
+        # input_ids.to(torch.bfloat16)
+        # self.image_token_id.to(torch.bfloat16)
+
+        if multimodal_embeddings is not None:
+            # multimodal_embeddings = multimodal_embeddings.to(torch.bfloat16)
+            # multimodal_embeddings = [emb.to(torch.bfloat16) for emb in multimodal_embeddings]
+            inputs_embeds = merge_multimodal_embeddings(
+                input_ids, inputs_embeds, multimodal_embeddings,
+                self.image_token_id)
+
+            
+        return inputs_embeds
+
+    def forward(self,
+                input_ids: torch.Tensor,
+                positions: torch.Tensor,
+                intermediate_tensors: Optional[IntermediateTensors] = None,
+                inputs_embeds: Optional[torch.Tensor] = None,
+                **kwargs: object):
+
+        if intermediate_tensors is not None:
+            inputs_embeds = None
+
+        # NOTE: In v1, inputs_embeds is always generated at model runner, this
+        # condition is for v0 compatibility
+        elif inputs_embeds is None:
+            vision_embeddings = self.get_multimodal_embeddings(**kwargs)
+            inputs_embeds = self.get_input_embeddings(input_ids,
+                                                      vision_embeddings)
+            input_ids = None
+
+        hidden_states = self.language_model(input_ids,
+                                            positions,
+                                            intermediate_tensors,
+                                            inputs_embeds=inputs_embeds)
+
+        return hidden_states
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[torch.Tensor]:
+        return self.language_model.compute_logits(hidden_states,
+                                                  sampling_metadata)
+
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]) -> Set[str]:
+        processed_weights = []
+        
+        for name, tensor in weights:
+            if 'sam_model' in name or 'qwen2_model' in name or 'projector' in name or 'view_seperator' in name:
+                new_name = name.replace('model.', '', 1)
+            else:
+                new_name = 'language.' + name
+            
+            # tensor = tensor.to(torch.bfloat16)
+
+            processed_weights.append((new_name, tensor))
+        
+        loader = AutoWeightsLoader(self)
+        autoloaded_weights = loader.load_weights(processed_weights, mapper=self.hf_to_vllm_mapper)
+
+
+
+
+
+        return autoloaded_weights