refactor(vision): 唯一后端DashScope(qwen3.8-flash+qwen3.7-plus),删除ARK/doubao冗余代码 #2207
Reference in New Issue
Block a user
Delete Branch "refactor/vision-v2-dashscope-only"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
根因:上版同时传 thinking={type:disabled} + reasoning_effort=low,方舟返回400 Invalid combination of reasoning_effort and thinking type。降级重试分支 把两个参数都pop掉,模型回到默认thinking开启→响应9-12s超时,fast全败。 修复: 1. 只传 thinking={"type":"disabled"},去掉互斥的 reasoning_effort 2. 400降级只pop thinking,保留精简payload重试(不pop多个) 3. 先单独验证 lite 单次HTTP 200且reasoning_tokens=0再跑E2E,省一轮部署- 新增_provider.py统一后端切换:VISION_V2_PROVIDER=ark|dashscope(默认dashscope对比测试) - vlm_fast_json.py:双后端httpx直连,ark用thinking={type:disabled}、dashscope用enable_thinking=false,400降级逻辑按provider处理 - vlm_fallback.py:dashscope用qwen3.7-plus+精简JSON prompt(<20s),ark保留原有ai_client+XML/JSON双解析 - fast_path.py:超时默认值从provider读取(dashscope pro 20s、ark pro 45s) - 通过环境变量DASHSCOPE_API_KEY配置key(敏感信息不入代码) - 本地curl验证:qwen3.8-flash关thinking单图2.6s、3图并发7.3s、8图并发10.3s(7/8可用)🚀 预览环境已部署
🗑️ 预览环境已清理
PR #2207 已关闭或合并,对应的预览环境已被清理。