From fbc1df36e016db7f5d5bb41b1c438e9849bc6789 Mon Sep 17 00:00:00 2001
From: xiaoxia <xiaoxia@xiaoxiajianji.com>
Date: Mon, 5 Oct 2026 15:14:42 +0800
Subject: [PATCH] =?UTF-8?q?fix(viral=5Fvideo):=20#2194=20VLM=E5=8D=95?=
 =?UTF-8?q?=E5=9B=BE=E6=9C=80=E5=9D=8F240s=E2=86=9270s=EF=BC=8C=E7=A6=81?=
 =?UTF-8?q?=E7=94=A8httpx=E5=BA=95=E5=B1=82=E9=87=8D=E8=AF=95+=E6=94=B6?=
 =?UTF-8?q?=E7=B4=A7timeout?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

P0根因：用户实测图片分析卡4分钟+。代码分析发现：
- _analyze_single_image 底层 DoubaoClient.max_retries=1（共2次请求）
- lite timeout=30s × 2次重试 = 60s
- lite失败后 pro_fallback timeout=90s × 2次重试 = 180s
- 单图最坏 60+180 = 240s，3图2并发最坏 ~240s+，恰好命中用户4分钟描述

修复：
1. _call() 临时将 client.max_retries=0 单次调用，不重复请求
2. lite timeout 30s → 25s（超时立即降级pro）
3. pro fallback timeout 90s → 45s（pro稳定响应<30s）
4. 单图最坏 25+45=70s，3图2并发最坏约70s（含排队）
5. 每张图VLM调用结束打印elapsed日志，方便后续排查
---
 apps/worker/worker_app/tasks/viral_video.py | 60 +++++++++++++++------
 1 file changed, 45 insertions(+), 15 deletions(-)

diff --git a/apps/worker/worker_app/tasks/viral_video.py b/apps/worker/worker_app/tasks/viral_video.py
index da598d833..d1adf6c58 100644
--- a/apps/worker/worker_app/tasks/viral_video.py
+++ b/apps/worker/worker_app/tasks/viral_video.py
@@ -422,18 +422,46 @@ def _analyze_single_image(
     )
 
     def _call(model: str, tmo: int):
+        # #2194: 单次调用临时关闭 httpx 层重试，超时/失败由外层 pro_fallback 统一兜底，
+        # 避免底层 max_retries=1 导致 lite timeout × 2 + pro timeout × 2 最坏 240s+
+        _t0 = time.time()
         try:
-            return call_vision(
-                image_url=img_url,
-                prompt=user,
-                model=model,
-                max_tokens=1200,  # #2188: 结构化 XML 输出 600-900 字足够，2048 翻倍耗时
-                temperature=0.3,
-                timeout=tmo,
-                system_prompt=system,
-            )
+            from packages.shared.ai_client import get_doubao_client as _gdc
+            import json as _json
+            _client = _gdc()
+            _messages = [
+                {"role": "system", "content": system},
+                {"role": "user", "content": user},
+            ]
+            _orig_retries = _client.max_retries
+            _client.max_retries = 0
+            try:
+                raw = _client.vision_completion(
+                    messages=_messages,
+                    images=[img_url],
+                    temperature=0.3,
+                    max_tokens=1200,  # #2188: 结构化 XML 输出 600-900 字足够
+                    timeout=tmo,
+                    model=model,
+                )
+            finally:
+                _client.max_retries = _orig_retries
+            _elapsed = time.time() - _t0
+            logger.info("[爆款视频] 图片 #%d VLM(%s) 完成 elapsed=%.1fs timeout=%d", idx, model, _elapsed, tmo)
+            if raw is None:
+                return None
+            stripped = raw.strip()
+            if stripped.startswith("```"):
+                stripped = stripped.strip("`")
+                if stripped.startswith("json"):
+                    stripped = stripped[4:].lstrip()
+            try:
+                return _json.loads(stripped)
+            except (_json.JSONDecodeError, TypeError):
+                return raw
         except Exception as e:
-            logger.warning("[爆款视频] 图片 #%d call_vision(%s) 异常 err=%s", idx, model, e)
+            _elapsed = time.time() - _t0
+            logger.warning("[爆款视频] 图片 #%d call_vision(%s) 异常 elapsed=%.1fs err=%s", idx, model, _elapsed, e)
             return None
 
     def _xml_to_product(nodes: list, raw_text: str) -> dict:
@@ -731,7 +759,7 @@ def _analyze_single_image(
         return first_result
 
     if pro_fallback_model and pro_fallback_model != vision_model:
-        pro_raw = _call(pro_fallback_model, 90)  # #2188: pro fallback 给 90s 余量
+        pro_raw = _call(pro_fallback_model, 45)  # #2194: pro 单次 45s 封顶，禁用重试，避免最坏 240s
         pro_result = _normalize(pro_raw, "pro_fallback")
         if _is_vision_result_usable(pro_result):
             pro_result["_fallback_used"] = True
@@ -742,10 +770,12 @@ def _analyze_single_image(
 
 def _step_image_analysis(job: ViralVideoJob) -> dict:
     """步骤 1: 图片 VLM 分析 — 识别产品特征（v1.6 优化：并行 + lite 模型提速）。
-    #2188: (1) 所有图片 URL 先归一化（storage_key→公网URL+空值报400）
+    #2188/#2194: (1) 所有图片 URL 先归一化（storage_key→公网URL+空值报400）
            (2) 爆款视频强制 lite-first，不依赖 .env USE_LITE 开关
            (3) max_tokens=1200，max_workers=min(2,n) 防方舟限流
-           (4) lite timeout=30s，pro fallback timeout=90s
+           (4) lite 单次25s封顶、pro单次45s封顶，底层httpx重试关闭，
+               单图最坏 25+45=70s，3图2并发最坏约70s（含排队），不超4min
+           (5) 每张图 VLM 调用结束打印 elapsed 耗时日志便于排查
     """
     try:
         from packages.shared.ai_service import call_vision  # noqa: F401
@@ -776,7 +806,7 @@ def _step_image_analysis(job: ViralVideoJob) -> dict:
         lite_model = "doubao-seed-2-1-lite-260915"
         pro_model = "doubao-seed-2-1-pro-260915"
     vision_model = lite_model  # 永远 lite 主跑
-    vision_timeout = 30  # lite 目标 20-30s
+    vision_timeout = 25  # #2194: lite 单次 25s 封顶，超时立即降级 pro，不让用户等 4min+
 
     results: list[dict] = [None] * len(normalized_urls)  # type: ignore
     max_workers = min(2, max(1, len(normalized_urls)))  # #2188: 并发≤2 防方舟限流
@@ -786,7 +816,7 @@ def _step_image_analysis(job: ViralVideoJob) -> dict:
         vision_model,
         pro_model,
         vision_timeout,
-        90,
+        45,
         max_workers,
     )
     with ThreadPoolExecutor(max_workers=max_workers) as pool:
-- 
2.54.0