Function _pd_process_generate

lightllm/server/httpserver/pd_loop.py:163–176

(
    manager: HttpServerManager, prompt, sampling_params, multimodal_params, forwarding_queue: AsyncQueue
)

# 触发推理的task
async def _pd_process_generate(
    manager: HttpServerManager, prompt, sampling_params, multimodal_params, forwarding_queue: AsyncQueue
):
    try:
        async for sub_req_id, request_output, metadata, finish_status in manager.generate(
            prompt, sampling_params, multimodal_params, None
        ):
            # p d 模式下，将 token 数据放入到转发队列中, 请求id 小于0的请求是health探测请求，不用转发。
            is_health_check_req = sub_req_id < 0
            if not is_health_check_req:
                await forwarding_queue.put((sub_req_id, request_output, metadata, finish_status))

    except BaseException as e:
        logger.error(str(e))


# 转发token的task