ocr_platform/ocr_tools/daemons/glmocr_vllm_daemon.sh

#!/bin/bash
# filepath: ocr_platform/ocr_tools/daemons/glmocr_vllm_daemon.sh
# 对应: GLM-OCR 自部署 vLLM 服务，SDK 自托管模式 (config: maas.enabled=false, ocr_api → 本服务)

# 保证transformers与vllm的兼容性, 注意先后次序，vllm安装后的transformers需重新安装
# uv pip install -U vllm --torch-backend=auto --extra-index-url https://wheels.vllm.ai/nightly
# uv pip install -U transformers
# GLM-OCR vLLM 服务守护进程脚本

LOGDIR="/home/ubuntu/zhch/logs"
mkdir -p $LOGDIR
PIDFILE="$LOGDIR/glmocr_vllm.pid"
LOGFILE="$LOGDIR/glmocr_vllm.log"

# 配置参数
CONDA_ENV="mineru_2_7_1"
PORT="20036"
HOST="0.0.0.0"
# 本地模型目录（与 config-zhch.yaml 自托管模式 ocr_api.api_port 一致）
MODEL_PATH="/home/ubuntu/models/modelscope_cache/models/ZhipuAI/GLM-OCR"
# 也可使用 HuggingFace 模型 id: zai-org/GLM-OCR
SERVED_MODEL_NAME="glm-ocr"
ALLOWED_LOCAL_MEDIA_PATH="/"

# GPU 配置
GPU_MEMORY_UTILIZATION="0.7"
CUDA_VISIBLE_DEVICES="7"
# 可选：开启 MTP 推测解码以提升推理性能
ENABLE_MTP="0"
MTP_NUM_SPECULATIVE_TOKENS="1"

# 环境变量（按需取消注释）
# export HF_HOME="/home/ubuntu/models/hf_home"
# export HF_ENDPOINT="https://hf-mirror.com"
# export MODELSCOPE_CACHE="/home/ubuntu/models/modelscope_cache"

# 正确初始化和激活 conda 环境
if [ -f "/home/ubuntu/anaconda3/etc/profile.d/conda.sh" ]; then
    source /home/ubuntu/anaconda3/etc/profile.d/conda.sh
    conda activate $CONDA_ENV
elif [ -f "/opt/conda/etc/profile.d/conda.sh" ]; then
    source /opt/conda/etc/profile.d/conda.sh
    conda activate $CONDA_ENV
else
    echo "Warning: Using direct conda path activation"
    export PATH="/home/ubuntu/anaconda3/envs/$CONDA_ENV/bin:$PATH"
fi

start() {
    if [ -f $PIDFILE ] && kill -0 $(cat $PIDFILE) 2>/dev/null; then
        echo "GLM-OCR vLLM is already running"
        return 1
    fi

    echo "Starting GLM-OCR vLLM daemon..."
    echo "Host: $HOST, Port: $PORT"
    echo "Model path: $MODEL_PATH"
    echo "Served model name: $SERVED_MODEL_NAME"
    echo "GPU memory utilization: $GPU_MEMORY_UTILIZATION"
    echo "CUDA devices: $CUDA_VISIBLE_DEVICES"

    # 检查模型：本地路径需存在目录，HuggingFace id (含 / 且非绝对路径) 不检查
    if [[ "$MODEL_PATH" == /* ]]; then
        if [ ! -d "$MODEL_PATH" ]; then
            echo "❌ Model path not found: $MODEL_PATH"
            echo "Use a local path or HuggingFace id (e.g. zai-org/GLM-OCR). Edit MODEL_PATH in this script."
            return 1
        fi
    fi

    # 检查 Python / vLLM
    if ! command -v python >/dev/null 2>&1; then
        echo "❌ Python not found. Check conda environment activation."
        return 1
    fi

    if ! python -c "import vllm" 2>/dev/null; then
        echo "❌ vllm not found. Install: uv pip install -U vllm --torch-backend=auto --extra-index-url https://wheels.vllm.ai/nightly"
        return 1
    fi

    echo "🔧 Using Python: $(which python)"
    echo "🔧 vLLM: $(python -c 'import vllm; print(vllm.__file__)' 2>/dev/null || true)"

    echo "📊 GPU 状态检查:"
    if command -v nvidia-smi >/dev/null 2>&1; then
        nvidia-smi --query-gpu=index,name,memory.used,memory.total --format=csv,noheader,nounits | \
        awk -F',' '{printf "  GPU %s: %s - 内存: %sMB/%sMB\n", $1, $2, $3, $4}'
    else
        echo "⚠️  nvidia-smi not available"
    fi

    # 构建 vllm serve 参数
    if [ "$ENABLE_MTP" = "1" ]; then
        CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES TRANSFORMERS_USE_FAST=false nohup vllm serve "$MODEL_PATH" \
            --host $HOST \
            --port $PORT \
            --allowed-local-media-path $ALLOWED_LOCAL_MEDIA_PATH \
            --served-model-name $SERVED_MODEL_NAME \
            --gpu-memory-utilization $GPU_MEMORY_UTILIZATION \
            --speculative-config "{\"method\": \"mtp\", \"num_speculative_tokens\": $MTP_NUM_SPECULATIVE_TOKENS}" \
            > $LOGFILE 2>&1 &
    else
        CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES TRANSFORMERS_USE_FAST=false nohup vllm serve "$MODEL_PATH" \
            --host $HOST \
            --port $PORT \
            --allowed-local-media-path $ALLOWED_LOCAL_MEDIA_PATH \
            --served-model-name $SERVED_MODEL_NAME \
            --gpu-memory-utilization $GPU_MEMORY_UTILIZATION \
            > $LOGFILE 2>&1 &
    fi

    echo $! > $PIDFILE
    echo "✅ GLM-OCR vLLM started with PID: $(cat $PIDFILE)"
    echo "📋 Log file: $LOGFILE"
    echo "🌐 Service URL: http://$HOST:$PORT"
    echo "📖 OpenAI-compatible API: http://localhost:$PORT/v1 (chat/completions, models)"
    echo ""
    echo "Waiting for service to start..."
    sleep 5
    status
}

stop() {
    if [ ! -f $PIDFILE ]; then
        echo "GLM-OCR vLLM is not running"
        return 1
    fi

    PID=$(cat $PIDFILE)
    echo "Stopping GLM-OCR vLLM (PID: $PID)..."

    kill $PID

    for i in {1..30}; do
        if ! kill -0 $PID 2>/dev/null; then
            break
        fi
        echo "Waiting for process to stop... ($i/30)"
        sleep 1
    done

    if kill -0 $PID 2>/dev/null; then
        echo "Force killing process..."
        kill -9 $PID
    fi

    rm -f $PIDFILE
    echo "✅ GLM-OCR vLLM stopped"
}

status() {
    if [ -f $PIDFILE ] && kill -0 $(cat $PIDFILE) 2>/dev/null; then
        PID=$(cat $PIDFILE)
        echo "✅ GLM-OCR vLLM is running (PID: $PID)"
        echo "🌐 Service URL: http://$HOST:$PORT"
        echo "📋 Log file: $LOGFILE"

        if command -v ss >/dev/null 2>&1; then
            if ss -tuln | grep -q ":$PORT "; then
                echo "🔗 Port $PORT is being listened"
            else
                echo "⚠️  Port $PORT is not being listened (service may be starting up)"
            fi
        elif command -v netstat >/dev/null 2>&1; then
            if netstat -tuln | grep -q ":$PORT "; then
                echo "🔗 Port $PORT is being listened"
            else
                echo "⚠️  Port $PORT is not being listened (service may be starting up)"
            fi
        fi

        if command -v curl >/dev/null 2>&1; then
            if curl -s --connect-timeout 2 http://127.0.0.1:$PORT/v1/models > /dev/null 2>&1; then
                echo "🎯 API 响应正常"
            else
                echo "⚠️  API 无响应 (service may be starting up)"
            fi
        fi

        if command -v nvidia-smi >/dev/null 2>&1; then
            echo "📊 GPU 使用情况:"
            nvidia-smi --query-gpu=index,utilization.gpu,utilization.memory,memory.used,memory.total --format=csv,noheader,nounits | \
            awk -F',' '{printf "  GPU %s: GPU利用率 %s%%, 内存利用率 %s%%, 显存 %sMB/%sMB\n", $1, $2, $3, $4, $5}'
        fi

        if [ -f $LOGFILE ]; then
            echo "📄 Latest logs (last 3 lines):"
            tail -3 $LOGFILE | sed 's/^/  /'
        fi
    else
        echo "❌ GLM-OCR vLLM is not running"
        if [ -f $PIDFILE ]; then
            echo "Removing stale PID file..."
            rm -f $PIDFILE
        fi
    fi
}

logs() {
    if [ -f $LOGFILE ]; then
        echo "📄 GLM-OCR vLLM logs:"
        echo "====================="
        tail -f $LOGFILE
    else
        echo "❌ Log file not found: $LOGFILE"
    fi
}

config() {
    echo "📋 Current configuration:"
    echo "  Conda Environment: $CONDA_ENV"
    echo "  Host: $HOST"
    echo "  Port: $PORT"
    echo "  Model Path: $MODEL_PATH"
    echo "  Served Model Name: $SERVED_MODEL_NAME"
    echo "  Allowed Local Media Path: $ALLOWED_LOCAL_MEDIA_PATH"
    echo "  GPU Memory Utilization: $GPU_MEMORY_UTILIZATION"
    echo "  CUDA Visible Devices: $CUDA_VISIBLE_DEVICES"
    echo "  Enable MTP: $ENABLE_MTP"
    echo "  PID File: $PIDFILE"
    echo "  Log File: $LOGFILE"

    if [ -d "$MODEL_PATH" ]; then
        echo "✅ Model path exists"
        echo "  Model files:"
        ls -la "$MODEL_PATH" | head -10 | sed 's/^/    /'
        if [ $(ls -1 "$MODEL_PATH" 2>/dev/null | wc -l) -gt 10 ]; then
            echo "    ... and more files"
        fi
    else
        echo "❌ Model path not found (use HuggingFace id like zai-org/GLM-OCR by setting MODEL_PATH)"
    fi

    echo ""
    echo "🔧 Environment:"
    echo "  Python: $(which python 2>/dev/null || echo 'Not found')"
    echo "  vLLM: $(python -c 'import vllm; print(vllm.__file__)' 2>/dev/null || echo 'Not found')"
    echo "  Conda: $(which conda 2>/dev/null || echo 'Not found')"
    echo "  CUDA: $(which nvcc 2>/dev/null || echo 'Not found')"

    if command -v nvidia-smi >/dev/null 2>&1; then
        echo ""
        echo "🔥 GPU Information:"
        nvidia-smi --query-gpu=index,name,driver_version,memory.total --format=csv,noheader,nounits | \
        awk -F',' '{printf "  GPU %s: %s (Driver: %s, Memory: %sMB)\n", $1, $2, $3, $4}'
    fi
}

test_api() {
    echo "🧪 Testing GLM-OCR vLLM API..."

    if [ ! -f $PIDFILE ] || ! kill -0 $(cat $PIDFILE) 2>/dev/null; then
        echo "❌ GLM-OCR vLLM service is not running"
        return 1
    fi

    if ! command -v curl >/dev/null 2>&1; then
        echo "❌ curl command not found"
        return 1
    fi

    echo "📡 Testing /v1/models endpoint..."
    response=$(curl -s --connect-timeout 10 http://127.0.0.1:$PORT/v1/models)
    if [ $? -eq 0 ]; then
        echo "✅ Models endpoint accessible"
        echo "$response" | python -m json.tool 2>/dev/null || echo "$response"
    else
        echo "❌ Models endpoint not accessible"
    fi
}

test_client() {
    echo "🧪 Testing GLM-OCR SDK with vLLM server (self-hosted mode)..."

    if [ ! -f $PIDFILE ] || ! kill -0 $(cat $PIDFILE) 2>/dev/null; then
        echo "❌ GLM-OCR vLLM service is not running. Start it first: $0 start"
        return 1
    fi

    TEST_IMAGE="/home/ubuntu/zhch/data/test/sample.png"
    if [ ! -f "$TEST_IMAGE" ]; then
        echo "⚠️  Test image not found: $TEST_IMAGE"
        echo "Update TEST_IMAGE in this script or create the file."
        return 1
    fi

    echo "📄 Test image: $TEST_IMAGE"
    echo "Run GLM-OCR with config that has maas.enabled=false and ocr_api pointing to 127.0.0.1:$PORT"
    echo "Example: glmocr parse $TEST_IMAGE --config /path/to/config.yaml"
    echo ""
    echo "Or start GLM-OCR Flask server (layout+OCR) that uses this vLLM backend:"
    echo "  glmocr server --config /path/to/config-zhch.yaml  # with maas.enabled=false, ocr_api.api_port=$PORT"
    echo "Then: curl -X POST http://localhost:5002/glmocr/parse -H 'Content-Type: application/json' -d '{\"images\": [\"file://$TEST_IMAGE\"]}'"
}

usage() {
    echo "GLM-OCR vLLM Service Daemon"
    echo "==========================="
    echo "Usage: $0 {start|stop|restart|status|logs|config|test|test-client}"
    echo ""
    echo "Commands:"
    echo "  start       - Start the GLM-OCR vLLM service"
    echo "  stop        - Stop the GLM-OCR vLLM service"
    echo "  restart     - Restart the GLM-OCR vLLM service"
    echo "  status      - Show service status and resource usage"
    echo "  logs        - Show service logs (follow mode)"
    echo "  config      - Show current configuration"
    echo "  test        - Test /v1/models API endpoint"
    echo "  test-client - Show how to test SDK/Flask with this vLLM backend"
    echo ""
    echo "Configuration (edit script to modify):"
    echo "  Host: $HOST"
    echo "  Port: $PORT"
    echo "  Model Path: $MODEL_PATH"
    echo "  Served Model Name: $SERVED_MODEL_NAME"
    echo "  GPU Memory: $GPU_MEMORY_UTILIZATION"
    echo "  CUDA Devices: $CUDA_VISIBLE_DEVICES"
    echo "  Enable MTP: $ENABLE_MTP"
    echo ""
    echo "Examples:"
    echo "  ./glmocr_vllm_daemon.sh start"
    echo "  ./glmocr_vllm_daemon.sh status"
    echo "  ./glmocr_vllm_daemon.sh logs"
    echo "  ./glmocr_vllm_daemon.sh test"
}

case "$1" in
    start)
        start
        ;;
    stop)
        stop
        ;;
    restart)
        stop
        sleep 3
        start
        ;;
    status)
        status
        ;;
    logs)
        logs
        ;;
    config)
        config
        ;;
    test)
        test_api
        ;;
    test-client)
        test_client
        ;;
    *)
        usage
        exit 1
        ;;
esac