From f9b5b60c68e8ea47727875cee9890addf46ccd1a Mon Sep 17 00:00:00 2001
From: yumoqing <yumoqing@gmail.com>
Date: Sat, 5 Sep 2026 10:24:44 +0800
Subject: [PATCH] =?UTF-8?q?fix(inference):=20=E8=B6=85=E6=97=B6=E6=A0=B9?=
 =?UTF-8?q?=E5=9B=A0=E5=8F=8C=E4=BF=AE=E2=80=94=E2=80=94=E6=8A=A5=E9=94=99?=
 =?UTF-8?q?=E7=94=A8repr(TimeoutError=E7=A9=BAstr)+=E9=87=8D=E8=AF=95print?=
 =?UTF-8?q?=E8=BF=9B=E5=BA=94=E7=94=A8=E6=97=A5=E5=BF=97;chat=5Finference?=
 =?UTF-8?q?=E6=94=AF=E6=8C=81=5Ftimeout=E5=8D=95=E6=AC=A1=E8=B6=85?=
 =?UTF-8?q?=E6=97=B6=E8=A6=86=E7=9B=96(=E4=B8=8A=E9=99=90900=E7=A7=92,?=
 =?UTF-8?q?=E9=95=BF=E6=96=87=E6=A1=A3=E6=8F=90=E5=8F=96=E7=94=A8)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 pipeline_llm/inference.py | 14 +++++++++++---
 1 file changed, 11 insertions(+), 3 deletions(-)

diff --git a/pipeline_llm/inference.py b/pipeline_llm/inference.py
index 2c2fcab..b287139 100644
--- a/pipeline_llm/inference.py
+++ b/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: