| 87 | system = "You are a precise coding assistant. " * 200 |
| 88 | |
| 89 | def chat(user_msg, max_tokens=8): |
| 90 | payload = { |
| 91 | "model": "luce-dflash", |
| 92 | "messages": [ |
| 93 | {"role": "system", "content": system}, |
| 94 | {"role": "user", "content": user_msg}, |
| 95 | ], |
| 96 | "max_tokens": max_tokens, "stream": False, |
| 97 | } |
| 98 | body = json.dumps(payload).encode() |
| 99 | req = urllib.request.Request( |
| 100 | f"{base_url}/v1/chat/completions", |
| 101 | data=body, headers={"Content-Type": "application/json"}) |
| 102 | t0 = time.time() |
| 103 | resp = urllib.request.urlopen(req, timeout=600) |
| 104 | data = json.loads(resp.read()) |
| 105 | dt = time.time() - t0 |
| 106 | return dt, data["choices"][0]["message"]["content"] |
| 107 | |
| 108 | # Turn 1: cold (cache miss → snapshot taken at end) |
| 109 | print("\n=== Turn 1 (cold) ===", flush=True) |