fix(inference): 超时根因双修——报错用repr(TimeoutError空str)+重试print进应用日志;chat_inference支持_timeout单次超时覆盖(上限900秒,长文档提取用)

pipeline_llm/inference.py

@@ -318,12 +318,16 @@ async def _call_upstream_chat(ctx, payload):
                     return data
         except (asyncio.TimeoutError, aiohttp.ClientError) as e:
             last = e
+            # TimeoutError 的 str() 是空串，用 repr 保证消息真实可行动
+            detail = repr(e) or type(e).__name__
+            # print 确保进应用日志（模块 logger 未接应用日志管道，2026-09-04 实测）
+            print("[inference] 上游瞬时错误重试 %d/%d: %s (url=%s)" % (
+                attempt + 1, _MAX_ATTEMPTS, detail, url.split('?')[0]))
             if attempt < _MAX_ATTEMPTS - 1:
-                logger.warning("inference: 瞬时错误重试 %d/%d: %s",
-                               attempt + 1, _MAX_ATTEMPTS, e)
                 await asyncio.sleep(2 * (attempt + 1))
                 continue
-            raise ValueError('上游调用失败（重试 %d 次仍失败）: %s' % (_MAX_ATTEMPTS, e))
+            raise ValueError('上游调用失败（重试 %d 次仍失败，%s）: %s' % (
+                _MAX_ATTEMPTS, type(e).__name__, detail))
     raise last if last else ValueError('上游调用失败')
 
 
@@ -346,12 +350,16 @@ async def chat_inference(org_id, user_id, payload, model_name='', task_ref=''):
 
     # 用途标记（辅助任务走辅助模型链）：取出后从 payload 剥离，不透传给上游
     purpose = (payload.pop('_purpose', '') or '')
+    # 单次调用超时覆盖（长文本提取等慢任务）：剥离不透传，上限 900 秒
+    req_timeout = int(_fnum(payload.pop('_timeout', 0)) or 0)
 
     est = max(_texts_len(payload.get('messages')) // EST_TOKENS_PER_CHAR, 200)
     db, dbname = _get_db()
     async with db.sqlorContext(dbname) as sor:
         ctx = await _resolve_call(
             sor, org_id, user_id or '', model_name, 't2t', est, task_ref, purpose)
+    if req_timeout > 0:
+        ctx['timeout'] = min(req_timeout, 900)
     try:
         data = await _call_upstream_chat(ctx, payload)
     except Exception as e: