LQAgentPlatform/utils_test/Model_Test/test_thinking_mode.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
测试统一模型调用的思考模式配置开关

测试内容：
1. 测试 enable_thinking=False（默认）时，Qwen3.5 模型响应时间是否显著缩短
2. 测试 enable_thinking=True 时，Qwen3.5 模型响应是否包含思考过程
3. 测试非 Qwen3.5 模型不受 enable_thinking 参数影响

运行方式:
    cd D:/wx_work/sichuan_luqiao/LQAgentPlatform
    python utils_test/Model_Test/test_thinking_mode.py
"""

import asyncio
import time
import sys
from pathlib import Path

project_root = Path(__file__).parent.parent.parent

from foundation.ai.agent.generate.model_generate import generate_model_client


# 测试用的简单提示词
TEST_SYSTEM_PROMPT = "你是一个 helpful 的 AI 助手，请简洁回答。"
TEST_USER_PROMPT = "请用一句话回答：1+1等于几？"


async def test_qwen35_with_thinking_disabled():
    """测试 Qwen3.5 模型 - 禁用思考模式（默认）"""
    print("\n" + "=" * 60)
    print("测试 1: Qwen3.5 35B - 禁用思考模式 (enable_thinking=False)")
    print("=" * 60)

    model_name = "qwen3_5_35b_a3b"
    start_time = time.time()

    try:
        response = await generate_model_client.get_model_generate_invoke(
            trace_id=f"test_thinking_disabled_{int(time.time())}",
            system_prompt=TEST_SYSTEM_PROMPT,
            user_prompt=TEST_USER_PROMPT,
            model_name=model_name,
            enable_thinking=False,  # 显式禁用思考模式
            timeout=120
        )

        elapsed_time = time.time() - start_time

        print(f"✅ 调用成功")
        print(f"   模型: {model_name}")
        print(f"   响应时间: {elapsed_time:.2f}s")
        print(f"   响应内容: {response[:100]}...")

        # 验证响应中不包含思考标记（如 <think> 或 思考过程）
        has_think_tag = "<think>" in response or "</think>" in response
        has_thinking_marker = "思考" in response and "过程" in response

        if has_think_tag or has_thinking_marker:
            print(f"   ⚠️ 警告: 响应可能包含思考过程标记")
        else:
            print(f"   ✅ 响应不包含思考过程标记")

        return elapsed_time, True

    except Exception as e:
        elapsed_time = time.time() - start_time
        print(f"❌ 调用失败: {e}")
        return elapsed_time, False


async def test_qwen35_with_thinking_enabled():
    """测试 Qwen3.5 模型 - 启用思考模式"""
    print("\n" + "=" * 60)
    print("测试 2: Qwen3.5 35B - 启用思考模式 (enable_thinking=True)")
    print("=" * 60)

    model_name = "qwen3_5_35b_a3b"
    start_time = time.time()

    # 使用需要推理的问题来激发思考过程
    reasoning_prompt = "请详细解释为什么 1+1=2？请展示你的思考过程。"

    try:
        response = await generate_model_client.get_model_generate_invoke(
            trace_id=f"test_thinking_enabled_{int(time.time())}",
            system_prompt="你是一个善于思考的AI助手，请详细展示你的推理过程。",
            user_prompt=reasoning_prompt,
            model_name=model_name,
            enable_thinking=True,  # 显式启用思考模式
            timeout=300  # 思考模式可能需要更长时间
        )

        elapsed_time = time.time() - start_time

        print(f"✅ 调用成功")
        print(f"   模型: {model_name}")
        print(f"   响应时间: {elapsed_time:.2f}s")
        print(f"   响应长度: {len(response)} 字符")

        # 显示响应的前500字符和后200字符
        display_len = min(500, len(response))
        print(f"   响应开头: {response[:display_len]}...")
        if len(response) > 700:
            print(f"   响应结尾: ...{response[-200:]}")

        # 检查是否包含思考过程特征
        has_think_tag = "<think>" in response or "</think>" in response
        has_reasoning_markers = any(marker in response for marker in [
            "思考", "推理", "首先", "然后", "第一步", "第二步",
            "让我", "我需要", "我们来", "分析一下"
        ])
        is_long_response = len(response) > 800  # 思考模式通常产生更长响应

        print(f"\n   思考模式检测:")
        print(f"   - 包含 <think> 标签: {'是' if has_think_tag else '否'}")
        print(f"   - 包含推理标记词: {'是' if has_reasoning_markers else '否'}")
        print(f"   - 响应较长 (>800字符): {'是' if is_long_response else '否'}")

        if has_think_tag or (has_reasoning_markers and is_long_response):
            print(f"   ✅ 思考模式似乎已生效")
        else:
            print(f"   ℹ️ 思考模式特征不明显，但调用已返回")

        return elapsed_time, True

    except Exception as e:
        elapsed_time = time.time() - start_time
        print(f"❌ 调用失败: {e}")
        return elapsed_time, False


async def test_non_qwen35_model():
    """测试非 Qwen3.5 模型（如 Doubao）不受 enable_thinking 影响"""
    print("\n" + "=" * 60)
    print("测试 3: 非 Qwen3.5 模型 - enable_thinking 参数不应产生影响")
    print("=" * 60)

    model_name = "doubao-1.5-pro-256k"  # 非 Qwen3.5 模型
    start_time = time.time()

    try:
        response = await generate_model_client.get_model_generate_invoke(
            trace_id=f"test_non_qwen35_{int(time.time())}",
            system_prompt=TEST_SYSTEM_PROMPT,
            user_prompt=TEST_USER_PROMPT,
            model_name=model_name,
            enable_thinking=False,  # 对非 Qwen3.5 模型应被忽略
            timeout=60
        )

        elapsed_time = time.time() - start_time

        print(f"✅ 调用成功")
        print(f"   模型: {model_name}")
        print(f"   响应时间: {elapsed_time:.2f}s")
        print(f"   响应内容: {response[:100]}...")
        print(f"   ✅ 非 Qwen3.5 模型正常响应，enable_thinking 参数被正确忽略")

        return elapsed_time, True

    except Exception as e:
        elapsed_time = time.time() - start_time
        print(f"❌ 调用失败: {e}")
        return elapsed_time, False


async def test_multiple_calls_consistency():
    """测试多次调用的一致性（验证 enable_thinking=False 稳定生效）"""
    print("\n" + "=" * 60)
    print("测试 4: Qwen3.5 多次调用一致性测试 (enable_thinking=False)")
    print("=" * 60)

    model_name = "qwen3_5_35b_a3b"
    call_times = []
    success_count = 0
    num_calls = 3

    for i in range(num_calls):
        start_time = time.time()
        try:
            response = await generate_model_client.get_model_generate_invoke(
                trace_id=f"test_consistency_{i}_{int(time.time())}",
                system_prompt=TEST_SYSTEM_PROMPT,
                user_prompt=TEST_USER_PROMPT,
                model_name=model_name,
                enable_thinking=False,
                timeout=120
            )
            elapsed_time = time.time() - start_time
            call_times.append(elapsed_time)
            success_count += 1
            print(f"   调用 {i+1}/{num_calls}: {elapsed_time:.2f}s - 成功")
        except Exception as e:
            elapsed_time = time.time() - start_time
            call_times.append(elapsed_time)
            print(f"   调用 {i+1}/{num_calls}: {elapsed_time:.2f}s - 失败: {e}")

    if call_times:
        avg_time = sum(call_times) / len(call_times)
        min_time = min(call_times)
        max_time = max(call_times)
        print(f"\n   统计结果:")
        print(f"   - 成功次数: {success_count}/{num_calls}")
        print(f"   - 平均响应时间: {avg_time:.2f}s")
        print(f"   - 最快: {min_time:.2f}s, 最慢: {max_time:.2f}s")

        # 验证响应时间合理性（禁用思考模式应在 60s 内完成）
        if avg_time < 60:
            print(f"   ✅ 平均响应时间在合理范围内（<60s）")
        else:
            print(f"   ⚠️ 平均响应时间较长（>=60s），可能思考模式未正确禁用")

    return success_count == num_calls


async def test_qwen35_122b_model():
    """测试 Qwen3.5 122B 大模型"""
    print("\n" + "=" * 60)
    print("测试 5: Qwen3.5 122B - 禁用思考模式")
    print("=" * 60)

    model_name = "qwen3_5_122b_a10b"
    start_time = time.time()

    try:
        response = await generate_model_client.get_model_generate_invoke(
            trace_id=f"test_122b_disabled_{int(time.time())}",
            system_prompt=TEST_SYSTEM_PROMPT,
            user_prompt=TEST_USER_PROMPT,
            model_name=model_name,
            enable_thinking=False,
            timeout=120
        )

        elapsed_time = time.time() - start_time

        print(f"✅ 调用成功")
        print(f"   模型: {model_name}")
        print(f"   响应时间: {elapsed_time:.2f}s")
        print(f"   响应内容: {response[:100]}...")

        if elapsed_time < 100:
            print(f"   ✅ 响应时间合理（<100s），思考模式可能已禁用")
        else:
            print(f"   ⚠️ 响应时间较长（>=100s）")

        return elapsed_time, True

    except Exception as e:
        elapsed_time = time.time() - start_time
        print(f"❌ 调用失败: {e}")
        return elapsed_time, False


async def run_all_tests():
    """运行所有测试"""
    print("\n" + "=" * 70)
    print(" 统一模型调用 - 思考模式配置开关测试")
    print("=" * 70)
    print("\n测试说明:")
    print("- 本测试验证 generate_model_client.get_model_generate_invoke()")
    print("- 的 enable_thinking 参数是否正确控制 Qwen3.5 模型的思考模式")
    print("- 预期: enable_thinking=False 时响应时间显著缩短（<60s）")

    results = {}

    # 测试 1: Qwen3.5 35B 禁用思考模式
    time1, success1 = await test_qwen35_with_thinking_disabled()
    results["qwen35_35b_disabled"] = {"time": time1, "success": success1}

    # 测试 2: Qwen3.5 35B 启用思考模式（对比测试）
    print("\n   [对比测试] 启用思考模式 - 预计耗时 60-300s，请耐心等待...")
    time2, success2 = await test_qwen35_with_thinking_enabled()
    results["qwen35_35b_enabled"] = {"time": time2, "success": success2}

    # 测试 3: 非 Qwen3.5 模型
    time3, success3 = await test_non_qwen35_model()
    results["non_qwen35"] = {"time": time3, "success": success3}

    # 测试 4: 多次调用一致性
    success4 = await test_multiple_calls_consistency()
    results["consistency"] = {"success": success4}

    # 测试 5: Qwen3.5 122B 大模型
    time5, success5 = await test_qwen35_122b_model()
    results["qwen35_122b_disabled"] = {"time": time5, "success": success5}

    # 汇总结果
    print("\n" + "=" * 70)
    print(" 测试结果汇总")
    print("=" * 70)

    for test_name, result in results.items():
        status = "✅ 通过" if result.get("success") else "❌ 失败"
        time_info = f" ({result['time']:.2f}s)" if "time" in result else ""
        print(f"   {test_name}: {status}{time_info}")

    # 思考模式对比分析
    if "qwen35_35b_disabled" in results and "qwen35_35b_enabled" in results:
        print("\n" + "-" * 70)
        print(" 思考模式性能对比")
        print("-" * 70)
        disabled_time = results["qwen35_35b_disabled"]["time"]
        enabled_time = results["qwen35_35b_enabled"]["time"]
        speedup = enabled_time / disabled_time if disabled_time > 0 else 0
        print(f"   禁用思考模式: {disabled_time:.2f}s")
        print(f"   启用思考模式: {enabled_time:.2f}s")
        print(f"   性能差异: {speedup:.1f}倍")
        if speedup > 3:
            print(f"   ✅ 思考模式开关效果显著，禁用后提速 {speedup:.1f} 倍")
        else:
            print(f"   ℹ️ 性能差异不明显")

    all_passed = all(r.get("success") for r in results.values())
    print("\n" + "=" * 70)
    if all_passed:
        print(" 🎉 所有测试通过！思考模式配置开关工作正常")
    else:
        print(" ⚠️ 部分测试失败，请检查配置")
    print("=" * 70 + "\n")

    return all_passed


if __name__ == "__main__":
    try:
        success = asyncio.run(run_all_tests())
        sys.exit(0 if success else 1)
    except KeyboardInterrupt:
        print("\n\n测试被用户中断")
        sys.exit(1)
    except Exception as e:
        print(f"\n\n测试运行出错: {e}")
        import traceback
        traceback.print_exc()
        sys.exit(1)