added DeepSeek OCR API by liushengtong

DeepSeek-OCR-vllm/.env

+# DeepSeek OCR 服务配置
+# 请根据你的实际情况修改以下配置
+
+# 模型路径（必需）
+MODEL_PATH=/home/lst/deepseek_ocr
+
+# GPU 配置
+GPU_ID=0
+
+# 服务端口
+PORT=8002
+
+# CPU 工作线程数
+CPU_WORKERS=2
+
+# Conda 环境配置
+# 方式1: 使用 conda 环境名称（需要 conda 命令）
+#CONDA_ENV_NAME=dso
+
+# 方式2: 直接使用 miniconda 环境路径（推荐用于 miniconda）
+PYTHON_PATH=/usr/bin/python3
+
+# PYTHON 版本（仅在使用 conda 环境名称时需要）
+PYTHON_VERSION=3.10
+
+# 服务配置
+HOST=0.0.0.0

DeepSeek-OCR-vllm/.env.example

+# DeepSeek OCR 服务配置示例
+# 复制此文件为 .env 并修改以下配置
+
+# ============================================================================
+# 核心配置（必须修改）
+# ============================================================================
+
+# 模型路径 - 请修改为你的实际模型路径
+MODEL_PATH=/path/to/your/DeepSeek-OCR
+
+# ============================================================================
+# GPU 和服务配置（可选）
+# ============================================================================
+
+# GPU ID（默认使用 GPU 3）
+GPU_ID=3
+
+# 服务端口（默认 8708）
+PORT=8708
+
+# 服务监听地址（默认 0.0.0.0，允许外部访问）
+HOST=0.0.0.0
+
+# CPU 工作线程数（默认 2，用于图像预处理）
+CPU_WORKERS=2
+
+# ============================================================================
+# Python 环境配置（二选一）
+# ============================================================================
+
+# 方式1: 使用 conda 环境名称（需要 conda 命令）
+# CONDA_ENV_NAME=deepseek-ocr-vllm
+
+# 方式2: 直接使用 miniconda 环境路径（推荐用于 miniconda）
+# PYTHON_PATH=/home/data/nongwa/miniconda3/envs/your_env/bin/python
+
+# Python 版本（仅在使用 conda 环境名称时需要）
+PYTHON_VERSION=3.10
\ No newline at end of file

DeepSeek-OCR-vllm/config.py

+# TODO: change modes
+# Tiny: base_size = 512, image_size = 512, crop_mode = False
+# Small: base_size = 640, image_size = 640, crop_mode = False
+# Base: base_size = 1024, image_size = 1024, crop_mode = False
+# Large: base_size = 1280, image_size = 1280, crop_mode = False
+# Gundam: base_size = 1024, image_size = 640, crop_mode = True
+
+BASE_SIZE = 1024
+IMAGE_SIZE = 640
+CROP_MODE = True
+MIN_CROPS= 2
+MAX_CROPS= 6 # max:9; If your GPU memory is small, it is recommended to set it to 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 = '/home/lst/deepseek_ocr' # 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 = './use.pdf' 
+OUTPUT_PATH = './output'
+
+PROMPT = '<image>\n<|grounding|>Convert the document to markdown.'
+# PROMPT = '<image>\nFree OCR.'
+# 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.
+# '先天下之忧而忧'
+# .......
+
+
+from transformers import AutoTokenizer
+
+TOKENIZER = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True)

DeepSeek-OCR-vllm/deepencoder/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-OCR-vllm/deepencoder/clip_sdpa.py

+from contextlib import nullcontext
+import math
+from typing import Optional, Tuple
+# from megatron.model import LayerNorm
+from easydict import EasyDict as adict
+import torch
+from torch.nn import functional as F
+from torch import nn
+from flash_attn import flash_attn_qkvpacked_func, flash_attn_func
+# from optimus import flash_attn_func
+# from megatron.core import tensor_parallel
+# from megatron.core import parallel_state as mpu
+# from megatron.core.utils import make_viewless_tensor, divide
+# from megatron.model.fused_rms_norm import RMSNorm
+# from megatron.model.transformer import (
+#     FlashSelfAttention,
+#     NoopTransformerLayer,
+#     _cfg_to_kwargs,
+# )
+# from megatron.model.enums import AttnMaskType, AttnType
+# from megatron.model.fused_softmax import FusedScaleMaskSoftmax
+# from megatron.model.utils import attention_mask_func
+
+# from megatron.model.module import MegatronModule
+
+# try:
+#     from einops import rearrange
+# except ImportError:
+#     rearrange = None
+
+# from flash_attn import flash_attn_varlen_func as flash_attn_unpadded_func
+
+# try:
+#     # flash attention 2.x
+#     from flash_attn import flash_attn_varlen_func as flash_attn_unpadded_func
+# except ImportError:
+#     try:
+#         # flash attention 1.x
+#         from flash_attn.flash_attn_interface import flash_attn_unpadded_func
+#     except ImportError:
+#         flash_attn_unpadded_func = None
+
+# try:
+#     from flash_attn.flash_attn_interface import flash_attn_unpadded_relative_attention_bias_func
+# except ImportError:
+#     flash_attn_unpadded_relative_attention_bias_func = None
+
+# try:
+#     from flash_attn.flash_attn_interface import mask_flash_attn_unpadded_func
+# except ImportError:
+#     mask_flash_attn_unpadded_func = None
+
+
+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
+
+
+# from optimus.flash_attn_interface import flash_attn_qkvpacked_func
+
+
+# class NoTPAttention(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:
+#             output = flash_attn_qkvpacked_func(xqkv)
+#             output = output.view(bsz, seqlen, -1)
+#         else:
+#             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)
+
+#             output = torch.nn.functional.scaled_dot_product_attention(xq, xk, xv, attn_mask=None)
+#             utput = output.permute(0, 2, 1, 3).view(bsz, seqlen, -1)
+#         output = self.out_proj(output)
+#         return output
+
+
+# from optimus.flash_attn_interface import flash_attn_qkvpacked_func
+
+
+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:
+            output = flash_attn_qkvpacked_func(xqkv)
+            output = output.view(bsz, seqlen, -1)
+            # 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:
+            # output = flash_attn_qkvpacked_func(xqkv)
+            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 = 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,
+    )
+
+
+if __name__ == '__main__':
+
+    
+    from mmgpt.model.vision_encoder.sam_b import build_sam_vit_b
+
+
+
+    vit_model_cfg = adict(
+        num_layers=24,
+        hidden_size=1024,
+        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 = []
+    )
+
+    sam_model = build_sam_vit_b()
+
+
+    vision_model = VitModel(
+        cfg=vit_model_cfg,
+        freeze_embed=False,
+        freeze_pre_norm=False,
+    )
+
+    # model = VitModel(1344)
+    # x = torch.zeros(2, 3, 224, 224)
+    x = torch.zeros(2, 3, 1024, 1024)
+
+    
+    with torch.no_grad():
+        # y = vision_model(x)
+        patch_embed = sam_model(x)
+        print(patch_embed.shape)
+        y = vision_model(x, patch_embed)
+        print(y.shape)
+
+        image_feature = torch.add(y[:, 1:], patch_embed.flatten(2).permute(0, 2, 1))
+
+        print(image_feature.shape)

DeepSeek-OCR-vllm/deepencoder/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, 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(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-OCR-vllm/deepseek_ocr_server.py

+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+DeepSeek OCR API Server (vLLM) - 极简版 + 优化版
+"""
+import os
+import io
+import re
+import argparse
+import asyncio
+from io import BytesIO
+from typing import List
+from concurrent.futures import ThreadPoolExecutor
+
+import torch
+from PIL import Image
+
+try:
+    import fitz
+except Exception:
+    fitz = None
+
+from fastapi import FastAPI, File, UploadFile, HTTPException, Form
+from fastapi.responses import JSONResponse
+from fastapi.middleware.cors import CORSMiddleware
+import uvicorn
+
+from vllm import LLM, SamplingParams
+from vllm.model_executor.models.registry import ModelRegistry
+from deepseek_ocr import DeepseekOCRForCausalLM
+from process.ngram_norepeat import NoRepeatNGramLogitsProcessor
+from process.image_process import DeepseekOCRProcessor
+
+
+app = FastAPI(title="DeepSeek OCR API (vLLM) - Optimized", version="2.0.0")
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"],
+    allow_credentials=True,
+    allow_methods=["*"],
+    allow_headers=["*"],
+)
+
+
+llm = None
+
+cpu_executor = ThreadPoolExecutor(max_workers=8, thread_name_prefix="CPU-Worker")
+gpu_executor = ThreadPoolExecutor(max_workers=1, thread_name_prefix="GPU-Worker")
+
+vllm_lock = asyncio.Lock()
+
+PROMPT_OCR = "<image>\n<|grounding|>Convert the document to markdown."
+PROMPT_DESC = "<image>\nDescribe this image in detail."
+
+# -----------------------
+# Monkey Patch
+# -----------------------
+_original_tokenize = DeepseekOCRProcessor.tokenize_with_images
+
+def _patched_tokenize(self, images, bos=True, eos=True, cropping=True, prompt=None):
+    if prompt is not None:
+        import config
+        old = config.PROMPT
+        config.PROMPT = prompt
+        try:
+            return _original_tokenize(self, images, bos, eos, cropping)
+        finally:
+            config.PROMPT = old
+    return _original_tokenize(self, images, bos, eos, cropping)
+
+DeepseekOCRProcessor.tokenize_with_images = _patched_tokenize
+
+
+def pdf_to_images_sync(pdf_bytes: bytes, dpi: int = 144) -> List[Image.Image]:
+    """PDF 转图片 """
+    if fitz is None:
+        raise RuntimeError("Please install PyMuPDF")
+    
+    images = []
+    doc = fitz.open(stream=pdf_bytes, filetype="pdf")
+    matrix = fitz.Matrix(dpi / 72.0, dpi / 72.0)
+    
+    for page in doc:
+        pix = page.get_pixmap(matrix=matrix, alpha=False)
+        img = Image.open(io.BytesIO(pix.tobytes("png")))
+        
+        if img.mode != "RGB":
+            if img.mode in ("RGBA", "LA"):
+                bg = Image.new("RGB", img.size, (255, 255, 255))
+                bg.paste(img, mask=img.split()[-1])
+                img = bg
+            else:
+                img = img.convert("RGB")
+        
+        images.append(img)
+    
+    doc.close()
+    return images
+
+
+def image_open_sync(image_bytes: bytes) -> Image.Image:
+    """打开图片 (同步版本)"""
+    return Image.open(BytesIO(image_bytes)).convert("RGB")
+
+
+def clear_vllm_cache_sync():
+    """清理 vLLM 缓存 (同步版本)"""
+    if llm is None:
+        return
+    try:
+        if hasattr(llm.llm_engine, 'input_preprocessor'):
+            prep = llm.llm_engine.input_preprocessor
+            if hasattr(prep, '_mm_processor_cache'):
+                prep._mm_processor_cache.clear()
+    except:
+        pass
+
+
+def tokenize_image_sync(image: Image.Image, prompt: str):
+    """
+    图像 tokenize (同步版本, CPU 密集)
+    WARNING: 这是最大的优化点!
+    """
+    processor = DeepseekOCRProcessor()
+    return processor.tokenize_with_images(images=[image], prompt=prompt)
+
+
+def vllm_generate_sync(tokenized, prompt: str) -> str:
+    """
+    vLLM 推理 (同步版本, GPU 密集)
+    注意: tokenized 已经在 CPU 线程池完成
+    """
+    batch_inputs = [{
+        "prompt": prompt,
+        "multi_modal_data": {"image": tokenized}
+    }]
+    
+    if prompt == PROMPT_OCR:
+        logits_proc = [NoRepeatNGramLogitsProcessor(20, 50, {128821, 128822})]
+        params = SamplingParams(
+            temperature=0.0,
+            max_tokens=8192,
+            skip_special_tokens=False,
+            logits_processors=logits_proc,
+            repetition_penalty=1.05,
+        )
+    else:
+        params = SamplingParams(
+            temperature=0.0,
+            max_tokens=8192,
+            skip_special_tokens=False,
+        )
+    
+    outputs = llm.generate(batch_inputs, params)
+    return outputs[0].outputs[0].text
+
+
+def clean_markdown_sync(text: str) -> str:
+    """清理 Markdown (同步版本)"""
+    text = re.sub(r'<\|ref\|>.*?<\|/ref\|>', '', text)
+    text = re.sub(r'<\|det\|>.*?<\|/det\|>', '', text)
+    text = re.sub(r'<\|.*?\|>', '', text)
+    text = re.sub(r'\[\[.*?\]\]', '', text)
+    text = re.sub(r'={50,}.*?={50,}', '', text, flags=re.DOTALL)
+    text = re.sub(r'\n{3,}', '\n\n', text)
+    return text.strip()
+
+
+
+async def pdf_to_images_async(pdf_bytes: bytes, dpi: int = 144) -> List[Image.Image]:
+    """PDF 转图片 (异步)"""
+    loop = asyncio.get_event_loop()
+    return await loop.run_in_executor(cpu_executor, pdf_to_images_sync, pdf_bytes, dpi)
+
+
+async def image_open_async(image_bytes: bytes) -> Image.Image:
+    """打开图片 (异步)"""
+    loop = asyncio.get_event_loop()
+    return await loop.run_in_executor(cpu_executor, image_open_sync, image_bytes)
+
+
+async def tokenize_image_async(image: Image.Image, prompt: str):
+    """
+    图像 tokenize (异步)
+    NOTE: 关键优化: 在 CPU 线程池执行
+    """
+    loop = asyncio.get_event_loop()
+    return await loop.run_in_executor(cpu_executor, tokenize_image_sync, image, prompt)
+
+
+async def vllm_generate_async(image: Image.Image, prompt: str) -> str:
+    """
+    完整的 vLLM 推理流程 (异步)
+    优化: 分离 tokenize (CPU) 和 generate (GPU)
+    """
+    # 步骤1: tokenize (CPU 密集, 在 CPU 线程池执行)
+    tokenized = await tokenize_image_async(image, prompt)
+    
+    # 步骤2: GPU 推理 (GPU 密集, 在 GPU 线程池执行, 有锁保护)
+    async with vllm_lock:
+        # 清理缓存 (在 GPU 线程池执行)
+        loop = asyncio.get_event_loop()
+        await loop.run_in_executor(gpu_executor, clear_vllm_cache_sync)
+        
+        # GPU 推理
+        result = await loop.run_in_executor(
+            gpu_executor,
+            vllm_generate_sync,
+            tokenized,
+            prompt
+        )
+        
+        return result
+
+
+async def clean_markdown_async(text: str) -> str:
+    """清理 Markdown (异步)"""
+    loop = asyncio.get_event_loop()
+    return await loop.run_in_executor(cpu_executor, clean_markdown_sync, text)
+
+
+async def generate_image_description_async(image: Image.Image) -> str:
+    """生成图片描述 (异步)"""
+    try:
+        # GPU 推理
+        result = await vllm_generate_async(image, PROMPT_DESC)
+        
+        # CPU 后处理
+        loop = asyncio.get_event_loop()
+        
+        def process_desc(text):
+            desc = re.sub(r'<\|ref\|>.*?<\|/ref\|>', '', text)
+            desc = re.sub(r'<\|det\|>.*?<\|/det\|>', '', desc)
+            desc = re.sub(r'<\|.*?\|>', '', desc)
+            desc = re.sub(r'\[\[.*?\]\]', '', desc)
+            desc = re.sub(r'\s+', ' ', desc).strip()
+            
+            if len(desc) > 200:
+                cutoff = desc[:200].rfind('.')
+                if cutoff > 100:
+                    desc = desc[:cutoff + 1]
+                else:
+                    desc = desc[:200].rsplit(' ', 1)[0] + '...'
+            
+            return desc
+        
+        desc = await loop.run_in_executor(cpu_executor, process_desc, result)
+        return desc
+    
+    except Exception as e:
+        print(f"WARNING: 图片描述失败: {e}")
+        return ""
+
+
+# -----------------------
+# 模型初始化
+# -----------------------
+def initialize_model(model_path: str, gpu_id: int):
+    global llm
+    
+    ModelRegistry.register_model("DeepseekOCRForCausalLM", DeepseekOCRForCausalLM)
+    
+    os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu_id)
+    
+    os.environ['VLLM_USE_V1'] = '0'
+    
+    print(f"[INFO] 加载模型: {model_path}")
+    
+    llm = LLM(
+        model=model_path,
+        hf_overrides={"architectures": ["DeepseekOCRForCausalLM"]},
+        block_size=64,
+        enforce_eager=False,
+        trust_remote_code=True,
+        max_model_len=8192,
+        tensor_parallel_size=1,
+        gpu_memory_utilization=0.9,
+        max_num_seqs=20,
+        disable_mm_preprocessor_cache=True,
+    )
+    
+    print("[SUCCESS] 模型加载完成")
+    #print(f"[INFO] 线程池配置:")
+    #print(f"   - CPU 线程池: {cpu_executor._max_workers} 线程")
+    #print(f"   - GPU 线程池: {gpu_executor._max_workers} 线程")
+
+
+# -----------------------
+# API 路由
+# -----------------------
+@app.get("/")
+async def root():
+    return {
+        "service": "DeepSeek OCR (vLLM) - Optimized",
+        "version": "2.0.0",
+        "status": "running"
+    }
+
+
+@app.get("/health")
+async def health():
+    return {
+        "status": "healthy",
+        "model_ready": llm is not None,
+        "cpu_workers": cpu_executor._max_workers,
+        "gpu_workers": gpu_executor._max_workers,
+    }
+
+
+async def vllm_generate_batch_async(
+    images: List[Image.Image], 
+    prompt: str,
+    show_progress: bool = True
+) -> List[str]:
+    """
+    批量 vLLM 推理 - 真正的批处理优化
+    
+    Args:
+        images: 图片列表
+        prompt: 提示词
+        show_progress: 是否显示进度
+    
+    Returns:
+        生成的文本列表
+    """
+    total = len(images)
+    
+    # 步骤1: 并发 tokenize
+    # 标准化图片 -> Vision Encoder (ViT) -> 图像特征向量 (例如：[196, 1024] - 196个位置，每个1024维)
+    if show_progress:
+        print(f"   [1/3] Tokenize {total} 页...")
+    
+    tokenize_tasks = [tokenize_image_async(img, prompt) for img in images]
+    all_tokenized = await asyncio.gather(*tokenize_tasks)
+    
+    if show_progress:
+        print(f"   [1/3] Tokenize 完成")
+    
+    # 步骤2: 构造批量输入
+    batch_inputs = [
+        {
+            "prompt": prompt,
+            "multi_modal_data": {"image": tok}
+        }
+        for tok in all_tokenized
+    ]
+    
+    # 步骤3: 批量 GPU 推理
+    async with vllm_lock:
+        if show_progress:
+            print(f"   [2/3] GPU 批量推理 {total} 页...")
+        
+        loop = asyncio.get_event_loop()
+        
+        # 清理缓存
+        await loop.run_in_executor(gpu_executor, clear_vllm_cache_sync)
+        
+        # 批量推理
+        def batch_generate():
+            # 根据 prompt 类型选择参数
+            if prompt == PROMPT_OCR:
+                logits_proc = [NoRepeatNGramLogitsProcessor(20, 50, {128821, 128822})]
+                params = SamplingParams(
+                    temperature=0.0,
+                    max_tokens=8192,
+                    skip_special_tokens=False,
+                    logits_processors=logits_proc,
+                    repetition_penalty=1.05,
+                )
+            else:
+                params = SamplingParams(
+                    temperature=0.0,
+                    max_tokens=8192,
+                    skip_special_tokens=False,
+                )
+            
+            # NOTE: 关键: 批量调用
+            outputs = llm.generate(batch_inputs, params)
+            return [out.outputs[0].text for out in outputs]
+        
+        results = await loop.run_in_executor(gpu_executor, batch_generate)
+        
+        if show_progress:
+            print(f"   [2/3] GPU 推理完成")
+        
+        return results
+
+
+@app.post("/ocr")
+async def ocr(
+    file: UploadFile = File(...),
+    enable_description: bool = Form(False),
+):
+    """OCmR 接口 (批量处理)"""
+    if llm is None:
+        raise HTTPException(503, "模型未加载")
+    
+    import time
+    start_time = time.time()
+    
+    try:
+        # 1. 读取文件
+        contents = await file.read()
+        t1 = time.time()
+        
+        # 2. 解析文件
+        if file.filename.lower().endswith('.pdf'):
+            # 如果是PDF，则转换为图片列表
+            images = await pdf_to_images_async(contents)
+        else:
+            # 如果是图片，则直接打开
+            images = [await image_open_async(contents)]
+        
+        t2 = time.time()
+  
+        # 3. 批量 OCR
+        raw_results = await vllm_generate_batch_async(images, PROMPT_OCR)
+        t3 = time.time()
+        print(f"   OCR 耗时: {t3 - t2:.2f}s")
+        
+        # 4. 后处理
+        print(f"   [3/3] 后处理...")
+        
+        async def postprocess(idx: int, raw: str, img: Image.Image) -> str:
+            # 图片描述
+            if enable_description:
+                img_pattern = r'<\|ref\|>image<\|/ref\|><\|det\|>\[\[.*?\]\]<\|/det\|>'
+                matches = list(re.finditer(img_pattern, raw))
+                
+                for match in matches:
+                    desc = await generate_image_description_async(img)
+                    replacement = f"[图片: {desc}]" if desc else "[图片]"
+                    raw = raw.replace(match.group(0), replacement)
+            
+            # 清理 Markdown
+            cleaned = await clean_markdown_async(raw)
+            return cleaned if cleaned else ""
+        
+        tasks = [postprocess(i, raw, img) for i, (raw, img) in enumerate(zip(raw_results, images))]
+        md_parts = await asyncio.gather(*tasks)
+        
+        t4 = time.time()
+        print(f"   [3/3] 后处理完成 ({t4 - t3:.2f}s)")
+        
+        # 5. 合并结果
+        final_md = "\n\n".join([md for md in md_parts if md])
+        
+        total_time = time.time() - start_time
+        print(f"{'='*60}")
+        print(f"[SUCCESS] 全部完成")
+        print(f"   总耗时: {total_time:.2f}s")
+        print(f"   平均: {total_time / len(images):.2f}s/页")
+        print(f"{'='*60}\n")
+        
+        return JSONResponse({
+            "markdown": final_md,
+            "page_count": len(images),
+            "processing_time": round(total_time, 2),
+        })
+    
+    except Exception as e:
+        import traceback
+        print(f"[ERROR] 处理失败: {e}")
+        print(traceback.format_exc())
+        raise HTTPException(500, f"处理失败: {e}")
+
+
+# -----------------------
+# 优雅关闭
+# -----------------------
+@app.on_event("shutdown")
+async def shutdown_event():
+    print("[INFO] 关闭线程池...")
+    cpu_executor.shutdown(wait=True)
+    gpu_executor.shutdown(wait=True)
+    print("[SUCCESS] 线程池已关闭")
+
+
+# -----------------------
+# 启动
+# -----------------------
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", required=True, help="模型路径")
+    parser.add_argument("--gpu-id", type=int, default=0, help="GPU ID")
+    parser.add_argument("--port", type=int, default=8002, help="端口")
+    parser.add_argument("--host", default="0.0.0.0", help="监听地址")
+    parser.add_argument("--cpu-workers", type=int, default=2, help="CPU 线程池大小")
+    
+    args = parser.parse_args()
+    
+    # 更新线程池大小
+    global cpu_executor
+    cpu_executor = ThreadPoolExecutor(
+        max_workers=args.cpu_workers,
+        thread_name_prefix="CPU-Worker"
+    )
+    
+    initialize_model(args.model_path, args.gpu_id)
+    
+    print(f"\n[INFO] 服务启动: http://{args.host}:{args.port}")
+    print(f"[INFO] 接口文档: http://{args.host}:{args.port}/docs\n")
+    
+    uvicorn.run(app, host=args.host, port=args.port, workers=1)
+
+
+if __name__ == "__main__":
+    main()