LQDeployConfig/prod/models/vllm/test_models.sh

#!/bin/bash

# SGLang 多模型 curl 测试脚本
# 移除 set -e，避免遇到错误就终止

# 颜色定义
GREEN='\033[0;32m'
RED='\033[0;31m'
YELLOW='\033[1;33m'
NC='\033[0m'

# 模型配置（按顺序定义）
MODEL_NAMES=("qwen3-8b" "qwen3.5-35b" "qwen3.5-122b" "qwen3-embedding-8b" "qwen3-reranker-8b")

declare -A MODEL_PORTS=(
    ["qwen3-8b"]="25424"
    ["qwen3.5-35b"]="25427"
    ["qwen3.5-122b"]="25423"
    ["qwen3-embedding-8b"]="25425"
    ["qwen3-reranker-8b"]="25426"
)

declare -A MODEL_PATHS=(
    ["qwen3-8b"]="/model/Qwen3-8B"
    ["qwen3.5-35b"]="/model/Qwen3.5-35B-A3B"
    ["qwen3.5-122b"]="/model/Qwen3.5-122B-A10B"
    ["qwen3-embedding-8b"]="/model/Qwen3-Embedding-8B"
    ["qwen3-reranker-8b"]="/model/Qwen3-Reranker-8B"
)

declare -A MODEL_TYPES=(
    ["qwen3-8b"]="chat"
    ["qwen3.5-35b"]="chat"
    ["qwen3.5-122b"]="chat"
    ["qwen3-embedding-8b"]="embedding"
    ["qwen3-reranker-8b"]="rerank"
)

API_KEY="lq123456"
TIMEOUT=30

echo "========================================"
echo "SGLang 多模型健康检查 (curl)"
echo "时间: $(date '+%Y-%m-%d %H:%M:%S')"
echo "========================================"

TOTAL=0
SUCCESS=0

# 测试对话模型
test_chat_model() {
    local name=$1
    local port=$2
    local model_path=$3
    
    echo ""
    echo "----------------------------------------"
    echo "测试模型: $name (对话模型)"
    echo "端口: $port"
    echo "----------------------------------------"
    
    local response
    local body
    local code
    
    response=$(curl -s -w "\n%{http_code}" \
        --max-time $TIMEOUT \
        -H "Content-Type: application/json" \
        -H "Authorization: Bearer $API_KEY" \
        -d "{
            \"model\": \"$model_path\",
            \"messages\": [{\"role\": \"user\", \"content\": \"你好，请用一句话介绍自己\"}],
            \"temperature\": 0.7,
            \"max_tokens\": 50
        }" \
        "http://localhost:$port/v1/chat/completions" 2>/dev/null || echo -e "\n000")
    
    body=$(echo "$response" | head -n -1)
    code=$(echo "$response" | tail -n 1)
    
    if [ "$code" = "200" ]; then
        local content=$(echo "$body" | grep -o '"content":"[^"]*"' | head -1 | cut -d'"' -f4)
        echo -e "${GREEN}✅ 成功${NC} HTTP $code"
        echo "回复: ${content:0:100}..."
        ((SUCCESS++))
    else
        echo -e "${RED}❌ 失败${NC} HTTP $code"
        echo "响应: ${body:0:200}"
    fi
    ((TOTAL++))
}

# 测试嵌入模型
test_embedding_model() {
    local name=$1
    local port=$2
    local model_path=$3
    
    echo ""
    echo "----------------------------------------"
    echo "测试模型: $name (嵌入模型)"
    echo "端口: $port"
    echo "----------------------------------------"
    
    local response
    local body
    local code
    
    response=$(curl -s -w "\n%{http_code}" \
        --max-time $TIMEOUT \
        -H "Content-Type: application/json" \
        -H "Authorization: Bearer $API_KEY" \
        -d "{
            \"model\": \"$model_path\",
            \"input\": [\"你好，这是一个测试句子\", \"Hello world\"]
        }" \
        "http://localhost:$port/v1/embeddings" 2>/dev/null || echo -e "\n000")
    
    body=$(echo "$response" | head -n -1)
    code=$(echo "$response" | tail -n 1)
    
    if [ "$code" = "200" ]; then
        local dims=$(echo "$body" | grep -o '"embedding":\[[^]]*\]' | head -1 | grep -o ',' | wc -l)
        dims=$((dims + 1))
        echo -e "${GREEN}✅ 成功${NC} HTTP $code"
        echo "向量维度: $dims"
        ((SUCCESS++))
    else
        echo -e "${YELLOW}⚠️  Embedding 接口失败，尝试 Rerank 接口...${NC}"
        response=$(curl -s -w "\n%{http_code}" \
            --max-time $TIMEOUT \
            -H "Content-Type: application/json" \
            -H "Authorization: Bearer $API_KEY" \
            -d "{
                \"model\": \"$model_path\",
                \"query\": \"测试查询\",
                \"documents\": [\"文档1\", \"文档2\"]
            }" \
            "http://localhost:$port/v1/rerank" 2>/dev/null || echo -e "\n000")
        
        code=$(echo "$response" | tail -n 1)
        if [ "$code" = "200" ]; then
            echo -e "${GREEN}✅ 成功${NC} (Rerank 接口可用)"
            ((SUCCESS++))
        else
            echo -e "${RED}❌ 失败${NC} HTTP $code"
        fi
    fi
    ((TOTAL++))
}

# 测试重排序模型
test_rerank_model() {
    local name=$1
    local port=$2
    local model_path=$3
    
    echo ""
    echo "----------------------------------------"
    echo "测试模型: $name (重排序模型)"
    echo "端口: $port"
    echo "----------------------------------------"
    
    local response
    local body
    local code
    
    response=$(curl -s -w "\n%{http_code}" \
        --max-time $TIMEOUT \
        -H "Content-Type: application/json" \
        -H "Authorization: Bearer $API_KEY" \
        -d "{
            \"model\": \"$model_path\",
            \"query\": \"什么是机器学习\",
            \"documents\": [\"机器学习是AI的分支\", \"Python是编程语言\", \"深度学习使用神经网络\"],
            \"top_n\": 2
        }" \
        "http://localhost:$port/v1/rerank" 2>/dev/null || echo -e "\n000")
    
    body=$(echo "$response" | head -n -1)
    code=$(echo "$response" | tail -n 1)
    
    if [ "$code" = "200" ]; then
        local top_doc=$(echo "$body" | grep -o '"text":"[^"]*"' | head -1 | cut -d'"' -f4)
        local score=$(echo "$body" | grep -o '"relevance_score":[0-9.]*' | head -1 | cut -d':' -f2)
        echo -e "${GREEN}✅ 成功${NC} HTTP $code"
        echo "Top1: ${top_doc:0:50}... (得分: $score)"
        ((SUCCESS++))
    else
        echo -e "${YELLOW}⚠️  Rerank 接口失败，尝试 Chat 接口...${NC}"
        response=$(curl -s -w "\n%{http_code}" \
            --max-time $TIMEOUT \
            -H "Content-Type: application/json" \
            -H "Authorization: Bearer $API_KEY" \
            -d "{
                \"model\": \"$model_path\",
                \"messages\": [{\"role\": \"user\", \"content\": \"你好\"}],
                \"max_tokens\": 20
            }" \
            "http://localhost:$port/v1/chat/completions" 2>/dev/null || echo -e "\n000")
        
        code=$(echo "$response" | tail -n 1)
        if [ "$code" = "200" ]; then
            echo -e "${GREEN}✅ 成功${NC} (Chat 接口可用)"
            ((SUCCESS++))
        else
            echo -e "${RED}❌ 失败${NC} HTTP $code"
        fi
    fi
    ((TOTAL++))
}

# 快速检查
quick_check() {
    echo ""
    echo "========================================"
    echo "快速检查模式"
    echo "========================================"
    
    for key in "${MODEL_NAMES[@]}"; do
        local port=${MODEL_PORTS[$key]}
        
        local code
        code=$(curl -s -o /dev/null -w "%{http_code}" \
            --max-time 5 \
            -H "Authorization: Bearer $API_KEY" \
            "http://localhost:$port/v1/models" 2>/dev/null || echo "000")
        
        if [ "$code" = "200" ]; then
            echo -e "${GREEN}✅${NC} $key (端口 $port)"
            ((SUCCESS++))
        else
            echo -e "${RED}❌${NC} $key (端口 $port) HTTP $code"
        fi
        ((TOTAL++))
    done
}

# 主函数
main() {
    if [ "$1" = "--quick" ]; then
        quick_check
    elif [ "$1" = "--model" ] && [ -n "$2" ]; then
        local key=$2
        local port=${MODEL_PORTS[$key]}
        local path=${MODEL_PATHS[$key]}
        local mtype=${MODEL_TYPES[$key]}
        
        case $mtype in
            chat) test_chat_model "$key" "$port" "$path" ;;
            embedding) test_embedding_model "$key" "$port" "$path" ;;
            rerank) test_rerank_model "$key" "$port" "$path" ;;
        esac
    else
        # 按顺序测试所有模型
        for key in "${MODEL_NAMES[@]}"; do
            local port=${MODEL_PORTS[$key]}
            local path=${MODEL_PATHS[$key]}
            local mtype=${MODEL_TYPES[$key]}
            
            case $mtype in
                chat) test_chat_model "$key" "$port" "$path" ;;
                embedding) test_embedding_model "$key" "$port" "$path" ;;
                rerank) test_rerank_model "$key" "$port" "$path" ;;
            esac
        done
    fi
    
    echo ""
    echo "========================================"
    echo "测试结果摘要"
    echo "========================================"
    echo "总计: $SUCCESS / $TOTAL 个模型正常"
    
    if [ $SUCCESS -eq $TOTAL ]; then
        echo -e "${GREEN}所有模型运行正常！${NC}"
        exit 0
    else
        echo -e "${RED}部分模型异常，请检查日志${NC}"
        exit 1
    fi
}

main "$@"