图像分类与 OCR 服务

图像分类与 OCR 服务接口

部署环境: 测试, 生产

测试环境目录：10.182.17.2：/data1/tlw/image_classification （环境目录：/data1/tlw/DeepSeek-OCR）

模型训练目录：10.182.17.2：/data1/tlw/timm

生产环境目录：10.51.0.1：/data/image_classification

测试环境启动命令：

环境：DeepSeek-OCR/env

OCR模型启动：

screen -r HunyuanOCR_1

CUDA_VISIBLE_DEVICES=0 vllm serve /data1/models/HunyuanOCR \
    --host 0.0.0.0 \
    --port 8001 \
    --served-model-name HunyuanOCR \
    --no-enable-prefix-caching \
    --mm-processor-cache-gb 0 \
    --gpu-memory-utilization 0.2 \
    --enable-log-requests \
    --enable-log-outputs \
    2>&1 | tee /data1/tlw/image_classification/logs/vllm_8001.log

Ctrl +a +d

screen -r HunyuanOCR_2

CUDA_VISIBLE_DEVICES=1 vllm serve /data1/models/HunyuanOCR \
    --host 0.0.0.0 \
    --port 8002 \
    --served-model-name HunyuanOCR \
    --no-enable-prefix-caching \
    --mm-processor-cache-gb 0 \
    --gpu-memory-utilization 0.2 \
    --enable-log-requests \
    --enable-log-outputs \
    2>&1 | tee /data1/tlw/image_classification/logs/vllm_8002.log

Ctrl +a +d

负载均衡端口：8000

图像分类和主入口启动：

nohup uvicorn app:app --host 0.0.0.0 --port 8005 --workers 1 > logs/app.log 2>&1 &

生产环境启动命令：

OCR模型启动：

10.51.0.1 （生产4090服务器）
screen -r HunyuanOCR_1

CUDA_VISIBLE_DEVICES=2 vllm serve /data/models/HunyuanOCR-1.5 \
    --served-model-name HunyuanOCR \
    -tp 1 \
    --limit-mm-per-prompt '{"image":4,"video":0}' \
    --trust-remote-code \
    --host 0.0.0.0 \
    --port 8001 \
    --gpu-memory-utilization 0.45 \
    --max-model-len 32768 \
    --max-num-batched-tokens 32768 \
    --api-key sk-lakala-hunyuanocr \
    --no-enable-prefix-caching \
    --mm-processor-cache-gb 0 \
    2>&1 | tee /data/HunyuanOCR-1.5/vllm_8001.log

Ctrl +a +d

screen -r HunyuanOCR_2

 CUDA_VISIBLE_DEVICES=3 vllm serve /data/models/HunyuanOCR-1.5 \
    --served-model-name HunyuanOCR \
    -tp 1 \
    --limit-mm-per-prompt '{"image":4,"video":0}' \
    --trust-remote-code \
    --host 0.0.0.0 \
    --port 8002 \
    --gpu-memory-utilization 0.45 \
    --max-model-len 32768 \
    --max-num-batched-tokens 32768 \
    --api-key sk-lakala-hunyuanocr \
    --no-enable-prefix-caching \
    --mm-processor-cache-gb 0 \
    2>&1 | tee /data/HunyuanOCR-1.5/vllm_8002.log

Ctrl +a +d

10.51.0.11 （生产5090D服务器）
/data/tlw/image_classification$ source .venv/bin/activate
screen -r HunyuanOCR_1

CUDA_VISIBLE_DEVICES=2 vllm serve /data/tlw/models/HunyuanOCR-1.5 \
    --served-model-name HunyuanOCR \
    -tp 1 \
    --limit-mm-per-prompt '{"image":4,"video":0}' \
    --trust-remote-code \
    --host 0.0.0.0 \
    --port 8001 \
    --gpu-memory-utilization 0.3 \
    --max-model-len 32768 \
    --max-num-batched-tokens 32768 \
    --api-key sk-lakala-hunyuanocr \
    --no-enable-prefix-caching \
    --mm-processor-cache-gb 0 \
    2>&1 | tee /data/tlw/HunyuanOCR-1.5/vllm_8001.log

Ctrl +a +d

screen -r HunyuanOCR_2

CUDA_VISIBLE_DEVICES=3 vllm serve /data/tlw/models/HunyuanOCR-1.5 \
    --served-model-name HunyuanOCR \
    -tp 1 \
    --limit-mm-per-prompt '{"image":4,"video":0}' \
    --trust-remote-code \
    --host 0.0.0.0 \
    --port 8002 \
    --gpu-memory-utilization 0.3 \
    --max-model-len 32768 \
    --max-num-batched-tokens 32768 \
    --api-key sk-lakala-hunyuanocr \
    --no-enable-prefix-caching \
    --mm-processor-cache-gb 0 \
    2>&1 | tee /data/tlw/HunyuanOCR-1.5/vllm_8002.log

Ctrl +a +d

测试：
curl -X POST http://10.51.0.11:8003/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer sk-lakala-hunyuanocr" \
  -d '{
    "model": "HunyuanOCR",
    "messages": [
      {"role": "system", "content": ""},
      {"role": "user", "content": [
        {"type": "image_url", "image_url": {"url": "http://lkl-zf-private-tms-test.oss-cn-shanghai.aliyuncs.com/partner/C0815159791575658496/20260114/企业授权书申请1.png"}},
        {"type": "text", "text": "提取图中的文字，并以纯文本格式返回"}
      ]}
    ],
    "temperature": 0.0,
    "top_p": 1.0,
    "top_k": -1,
    "repetition_penalty": 1.08,
    "skip_special_tokens": true,
    "max_tokens": 4096,
    "stream": false,
    "seed": 1234
  }'

图像分类和主入口启动：

nohup uvicorn app:app --host 0.0.0.0 --port 8005 --workers 1 > logs/app.log 2>&1 &

负载均衡端口：8003，外部访问：10.61.88.8:31038

接口：

图片分类：POST /predict
OCR： POST /HunyuanOCR

application/json

Bodyapplication/json

示例

图像分类OCR

{
    "image_category_list": [
        {
            "image_seq": "唯一标识或业务序号",
            "image_category": "business_license"
        },
        {
            "image_seq": "另一张",
            "image_category": "其他"
        }
    ]
}

图像分类与 OCR 服务