apps/worker/worker_app/tasks/viral_video.py

@@ -451,7 +451,7 @@ def _analyze_single_image(
 
     # 第二次：pro 降级重试
     if pro_fallback_model and pro_fallback_model != vision_model:
-        pro_raw = _call(pro_fallback_model, 25)  # #2173: pro 25s 上限（lite 已失败过，快速兜底）
+        pro_raw = _call(pro_fallback_model, 60)  # #2180: pro VLM 实测也需25-38s，原25s太短，提到60s
         pro_result = _normalize(pro_raw, "pro_fallback")
         if _is_vision_result_usable(pro_result):
             pro_result["_fallback_used"] = True
@@ -485,7 +485,7 @@ def _step_image_analysis(job: ViralVideoJob) -> dict:
         if _s.doubao_vision_use_lite:
             vision_model = _s.doubao_vision_lite_model
             pro_model = _s.doubao_vision_model
-            vision_timeout = 15  # #2173: lite 15s 快速失败转 pro（实测 lite 持续超时时白等137s是P0）
+            vision_timeout = 45  # #2180: 方舟 VLM 实测服务端处理24-38s，原15s必超时3次重试全挂，提到45s
         else:
             vision_model = _s.doubao_vision_model
             pro_model = None  # 已经是 pro，不再降级
@@ -598,7 +598,7 @@ def _step_intent_parsing(job: ViralVideoJob, image_analysis: dict) -> dict:
     for _m, _lbl in [(_fast, "fast"), (_pro, "pro-fallback")]:
         try:
             logger.info("[爆款视频] 意图解析 model=%s label=%s", _m, _lbl)
-            result = call_llm(prompt, temperature=0.4, max_tokens=800, model=_m, timeout=25)
+            result = call_llm(prompt, temperature=0.4, max_tokens=800, model=_m, timeout=45)
             return (
                 result
                 if isinstance(result, dict)
@@ -1033,16 +1033,16 @@ def _step_script_generation(job: ViralVideoJob, intent: dict, image_analysis: di
 
     try:
         # 第一次：快模型 25s
-        normalized = _try_gen(_fast, 0.8, 2500, "fast-first", tmo=25)
+        normalized = _try_gen(_fast, 0.8, 2500, "fast-first", tmo=45)
         if normalized is not None:
             return normalized
         # 第二次：快模型降温度+加大 max_tokens，25s
-        normalized = _try_gen(_fast, 0.6, 3200, "fast-retry", tmo=25)
+        normalized = _try_gen(_fast, 0.6, 3200, "fast-retry", tmo=45)
         if normalized is not None:
             return normalized
         # 第三次：用主力模型兜底，给 40s
         if _pro and _pro != _fast:
-            normalized = _try_gen(_pro, 0.7, 3500, "pro-fallback", tmo=40)
+            normalized = _try_gen(_pro, 0.7, 3500, "pro-fallback", tmo=60)
             if normalized is not None:
                 return normalized
         logger.warning("[爆款视频] 编导脚本三次都未生成合格结果，使用兜底脚本")
@@ -1079,7 +1079,7 @@ def _step_review(job: ViralVideoJob, copy_result: dict) -> dict:
     for _m, _lbl in [(_fast, "fast"), (_pro, "pro-fallback")]:
         try:
             logger.info("[爆款视频] 合规审核 model=%s label=%s", _m, _lbl)
-            result = call_llm(prompt, temperature=0.1, max_tokens=500, model=_m, timeout=20)
+            result = call_llm(prompt, temperature=0.1, max_tokens=500, model=_m, timeout=30)
             return result if isinstance(result, dict) else {"passed": True, "score": 80, "details": {}}
         except Exception as e:
             logger.warning("[爆款视频] 合规审核失败 label=%s err=%s", _lbl, e)

packages/config/base.py

@@ -95,8 +95,8 @@ class SharedSettings(BaseSettings):
         "doubao-seed-2-1-lite-260915"  # 快速模型（Seed 2.1 Lite，高 RPM，编导/审核/VLM lite；原 1-5-pro-32k 已 Retiring）
     )
     doubao_base_url: str = "https://ark.cn-beijing.volces.com/api/v3"
-    doubao_timeout: int = 30
-    doubao_max_retries: int = 2
+    doubao_timeout: int = 45  # #2180: 方舟LLM高峰期响应6-8s，原30s太紧提到45s
+    doubao_max_retries: int = 1  # #2180: timeout调大后一次调用就够，1次重试防偶发抖动；避免6次重试叠加到351s
     doubao_vision_model: str = (
         "doubao-seed-2-1-pro-260915"  # 高精度视觉（Seed 2.1 Pro 原生多模态；原 vision-pro-250328 已下线）
     )