(name, session, prompts, counter, latencies, pbar, queue)
| 54 | |
| 55 | # ============ Worker ============ |
| 56 | async def worker(name, session, prompts, counter, latencies, pbar, queue): |
| 57 | while True: |
| 58 | i = await queue.get() |
| 59 | if i is None: # 毒丸退出 |
| 60 | queue.task_done() |
| 61 | break |
| 62 | |
| 63 | payload = { |
| 64 | "model": "eb", |
| 65 | "messages": [{"role": "user", "content": prompts[i % len(prompts)]}], |
| 66 | "max_prompt_len": 2048, |
| 67 | "max_dec_len": 1024, |
| 68 | "min_dec_len": 32, |
| 69 | "top_p": 1.0, |
| 70 | "temperature": 1.0, |
| 71 | "repetition_penalty": 1.0, |
| 72 | "rollout_quant_type": "weight_only_int8", |
| 73 | "disable_chat_template": True, |
| 74 | } |
| 75 | |
| 76 | success, latency, status, error = await send_request(session, payload) |
| 77 | if success: |
| 78 | counter["success"] += 1 |
| 79 | latencies.append(latency) |
| 80 | else: |
| 81 | # print(f"Request failed ({status}): {error}") |
| 82 | counter["fail"] += 1 |
| 83 | counter[f"error_{error or 'client'}"] += 1 |
| 84 | |
| 85 | pbar.update(1) |
| 86 | queue.task_done() |
| 87 | |
| 88 | |
| 89 | # ============ 主流程 ============ |
no test coverage detected