Function inference_single_task

examples/run_leetcode_agent.py:133–245

仅执行单个任务的推理阶段（不包括评测）。 推理使用 semaphore 限制并发数。 时间记录： - inference_start_time: 推理开始时间（获取信号量后） - inference_time: 推理耗时

(
    task: Task,
    save_dir: str,
    semaphore: asyncio.Semaphore
)

    return content_parts

async def inference_single_task(
    task: Task,
    save_dir: str,
    semaphore: asyncio.Semaphore
) -> Task:
    """
    仅执行单个任务的推理阶段（不包括评测）。
    推理使用 semaphore 限制并发数。
    
    时间记录：
    - inference_start_time: 推理开始时间（获取信号量后）
    - inference_time: 推理耗时
    """
    task_id = task.task_id
    
    # 使用信号量限制并发推理数量
    async with semaphore:
        # ✅ 在获取信号量后立即记录推理开始时间
        inference_start_time = time.time()
        task.extra["inference_start_time"] = inference_start_time
        
        try:
            logger.info(f"| 🚀 [Task {task_id}] Starting inference...")
            
            # --- 1. Prepare Prompt ---
            question_text = task.input
            # 检查模型是否支持视觉/是否被禁用视觉
            if TARGET_MODEL in NON_VISION_MODELS:
                logger.info(f"| 🙈 [Task {task_id}] Vision disabled for model {TARGET_MODEL}, using text only.")
                question_content = question_text
            else:
                question_content = parse_markdown_with_images(question_text)
            
            system_prompt_text = task.system_prompt
            
            logger.info(f"| 📋 [Task {task_id}] Input length: {len(question_text)}")
            
            if isinstance(question_content, list):
                image_count = sum(1 for part in question_content if isinstance(part, ContentPartImage))
                logger.info(f"| 🖼️ [Task {task_id}] Found {image_count} image(s) in question")

            messages = [
                SystemMessage(content=system_prompt_text),
                HumanMessage(content=question_content)
            ]
            # --- 2. Model Inference (Structured Output) ---
            logger.info(f"| ⏳ [Task {task_id}] Model inferencing...")
            
            try:
                response = await model_manager(
                    model=TARGET_MODEL,
                    messages=messages,
                    response_format=Response,
                    max_completion_tokens=65536
                )
                
                if response.success:
                    response_model = response.extra.parsed_model