Chinese Brief
中文案例导读
easyicetrusty / BerriAI LiteLLM + LibreChat deployment 公开的真实任务执行案例,来源为 公开代码库,复核于 2026-06-27T10:23:35Z。 Model Atlas 将它标记为 A 类证据,因为它同时具备具体使用者、具体任务、公开原始证据和可访问产物。 Model Atlas 不把 benchmark、教程、发布说明或集合页包装成真实案例。
CASE EVIDENCE / A RECORD
easyicetrusty / BerriAI LiteLLM + LibreChat deployment 公开的真实任务执行案例,来源为 公开代码库,复核于 2026-06-27T10:23:35Z。
原始记录:LibreChat/LiteLLM 用户用 Qwen3 Max Thinking 复现并修补 DashScope 流式内容策略回退问题
Chinese Brief
easyicetrusty / BerriAI LiteLLM + LibreChat deployment 公开的真实任务执行案例,来源为 公开代码库,复核于 2026-06-27T10:23:35Z。 Model Atlas 将它标记为 A 类证据,因为它同时具备具体使用者、具体任务、公开原始证据和可访问产物。 Model Atlas 不把 benchmark、教程、发布说明或集合页包装成真实案例。
任务
用户 easyicetrusty 在公开 LiteLLM issue 中说明,其私有 dorx LibreChat 部署通过 LiteLLM proxy、DashScope/Alibaba Cloud OpenAI-compatible 接口调用 `qwen/qwen3-max-thinking` 等 Qwen thinking 模型;任务是在流式 chat-completion 中处理 provider 内容审核 400,并通过 content_policy_fallbacks 路由到 tier-2 模型,而不是让会话 500。
issue 给出可复现路径、环境和补丁:`litellm[proxy]` main-latest、LibreChat streaming consumer、DashScope provider、模型 `qwen/qwen3-max-thinking`;作者报告当前生产环境中已将补丁以 bind-mounted vendored file 方式运行,Streaming DashScope content-policy 400s on Qwen thinking models can now fall through to a tier-2 model instead of returning a misleading `Fallbacks=None` hard error。
Qwen3 Max Thinking 在该案例中是实际部署的 DashScope/Qwen thinking 后端模型;它的内容策略 400 与流式返回行为触发了 LiteLLM fallback 代码路径缺陷,使用户能够围绕真实 Qwen3 Max Thinking 流式会话定位、验证并提出对 ContentPolicyViolationError 的 MidStreamFallbackError 处理修复。
证据来自公开 GitHub issue 的结构化正文,明确列出 Environment 中的 model=`qwen/qwen3-max-thinking`、DashScope provider、LibreChat streaming consumer 和生产 workaround;未公开私有 dorx LibreChat 实例或具体被拦截 prompt,因此 artifact 使用同一公开 issue。不是 benchmark、教程或发布文。