Function load_models

pasa/pasa_server.py:155–210

Initialize vLLM-backed agents and validate endpoints.

()

def load_models() -> None:
    """Initialize vLLM-backed agents and validate endpoints."""
    global _CRAWLER_AGENT, _SELECTOR_AGENT, _MODEL_LOAD_ERROR

    try:
        if PASA_INFERENCE_BACKEND != "vllm":
            raise RuntimeError(
                f"Unsupported PASA_INFERENCE_BACKEND={PASA_INFERENCE_BACKEND!r}. "
                "This server expects vLLM backend."
            )

        if not DEFAULT_CRAWLER_PATH:
            raise RuntimeError('PASA_CRAWLER_PATH is required.')
        if not DEFAULT_SELECTOR_PATH:
            raise RuntimeError('PASA_SELECTOR_PATH is required.')
        if not os.path.exists(DEFAULT_CRAWLER_PATH):
            raise FileNotFoundError(f"CRAWLER model path not found: {DEFAULT_CRAWLER_PATH}")
        if not os.path.exists(DEFAULT_SELECTOR_PATH):
            raise FileNotFoundError(f"SELECTOR model path not found: {DEFAULT_SELECTOR_PATH}")
        if not os.path.exists(DEFAULT_PROMPTS_PATH):
            raise FileNotFoundError(f"PROMPTS path not found: {DEFAULT_PROMPTS_PATH}")

        logger.info("=" * 70)
        logger.info("🚀 Starting PASA vLLM backend initialization")
        logger.info(f"GPU (expected): {os.getenv('PASA_GPU_ID', '1')}")
        logger.info(f"CRAWLER vLLM:  {PASA_VLLM_CRAWLER_URL} (model={PASA_VLLM_CRAWLER_MODEL_NAME})")
        logger.info(f"SELECTOR vLLM: {PASA_VLLM_SELECTOR_URL} (model={PASA_VLLM_SELECTOR_MODEL_NAME})")

        _CRAWLER_AGENT = VLLMAgent(
            base_url=PASA_VLLM_CRAWLER_URL,
            model_name=PASA_VLLM_CRAWLER_MODEL_NAME,
            tokenizer_path=DEFAULT_CRAWLER_PATH,
            use_chat_template=True,
            request_timeout=PASA_VLLM_REQUEST_TIMEOUT,
            logprobs=PASA_VLLM_LOGPROBS,
        )
        _SELECTOR_AGENT = VLLMAgent(
            base_url=PASA_VLLM_SELECTOR_URL,
            model_name=PASA_VLLM_SELECTOR_MODEL_NAME,
            tokenizer_path=None,
            use_chat_template=False,
            request_timeout=PASA_VLLM_REQUEST_TIMEOUT,
            logprobs=PASA_VLLM_LOGPROBS,
        )

        _CRAWLER_AGENT.ensure_ready()
        _SELECTOR_AGENT.ensure_ready()

        logger.info("✅ PASA vLLM backend is ready")
        logger.info("=" * 70)

    except Exception as e:
        _MODEL_LOAD_ERROR = str(e)
        logger.error(f"❌ Failed to initialize PASA vLLM backend: {e}")
        logger.error(traceback.format_exc())
        raise