1. 安装过程

1.1. Cuda-11.8环境

• GPU：3090（24G） 、cuda驱动13.0
• torch==2.6.0,cuda==11.8

export PATH=/usr/local/cuda-11.8/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-11.8/lib64:$LD_LIBRARY_PATH

1.2. 拉取DeepSeek-OCR

git clone https://github.com/deepseek-ai/DeepSeek-OCR.git

1.3. torch-2.6.0、python-3.12.9 环境

conda create -n deepseek-ocr python=3.12.9 -y
conda activate deepseek-ocr
pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu118

1.4. vllm-0.8.5

download the vllm-0.8.5 whl

pip install vllm-0.8.5+cu118-cp38-abi3-manylinux1_x86_64.whl
pip install -r requirements.txt # 报错没关系

1.5. flash_attn-2.7.3

pip install flash-attn==2.7.3 --no-build-isolation

若报错，采用下面方法下载离线包至本地

wget https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.3/flash_attn-2.7.3+cu11torch2.6cxx11abiFALSE-cp312-cp312-linux_x86_64.whl
pip install flash_attn-2.7.3+cu11torch2.6cxx11abiFALSE-cp312-cp312-linux_x86_64.whl

1.6. vllm配置config.py

cd DeepSeek-OCR-master/DeepSeek-OCR-vllm
mkdir images #存放测试图片
mkdir output
vim config.py

把一张测试图片放入images内，我的是一张户型图

MODEL_PATH = '/model/Deepseek/DeepSeek-OCR/' # 填你自己下载的DeepSeek-OCR模型路径
INPUT_PATH = 'images/test.jpg' 
OUTPUT_PATH = 'output'
PROMPT = '<image>\nDescribe this image in detail.'

1.7. 运行

python run_dpsk_ocr_image.py

PROMPT = '<image>\nDescribe this image in detail.' # 对户型图的描述

PROMPT = '<image>\nLocate <|ref|>小孩房<|/ref|> in the image.' #标注小孩房

识别得不是很准。还有待研究