No module named ‘torch_npu’
这个报错出现在 ComfyUI 启动后的后端探测阶段：ComfyUI 会尝试检测 Ascend 后端，而该后端依赖 torch-npu；如果环境里没有安装 torch-npu，探测就会回退并输出 No module named 'torch_npu' 。如果日志后面仍显示 CUDA 后端可用、程序正
分类： AI 工具排错
[bug]: Per-item If over a collected sibling-iterator zip is rejected as UnsupportedIfFramesError
![[bug]: Per-item If over a collected sibling-iterator zip is rejected as UnsupportedIfFramesError](https://www.chat-gpts.plus/wp-content/uploads/2026/10/9820-ced14590-768x403.jpg)
该报错出现在 InvokeAI 图执行中： If 的条件由 Iterate （left）逐项产生，而 true 分支是一个对 left 与另一个 Iterate（right）配对 zip 后 Collect 的结果。图本身是合法的，但帧判定把 right 误判为“未关闭且与条件无关的分支迭代器”。优
[bug]: Long node chains into an If branch raise RecursionError
![[bug]: Long node chains into an If branch raise RecursionError](https://www.chat-gpts.plus/wp-content/uploads/2026/10/9821-3e34209b-768x403.jpg)
当 Iterate 与 If 分支输入之间存在超长普通节点链（约 480 个以上节点）时， GraphExecutionState.next() 会因递归深度超限抛出 RecursionError ；优先检查 InvokeAI 是否为包含 PR #9843 修复的版本。
[Bug]: Batch retrieve returns 403 for team-scoped models since v1.103.0 (authorize_model_for_key checks internal model_name_{team_id}_{uuid}
![[Bug]: Batch retrieve returns 403 for team-scoped models since v1.103.0 (authorize_model_for_key checks internal model_name_{team_id}_{uuid}](https://www.chat-gpts.plus/wp-content/uploads/2026/10/46958-8bc36f96-768x403.jpg)
该 403 通常发生在 LiteLLM Proxy 使用团队作用域模型（team-scoped model）创建 Batch 后，再调用 GET /v1/batches/{batch_id} 查询或取消时；根因是批量路由用内部模型名 model_name_{team_id}_{uuid} 做 tea
[Bug]: Request Failed Error Code: 500 Message: litellm.APIConnectionError: APIConnectionError: Hosted_vllmException –
![[Bug]: Request Failed Error Code: 500 Message: litellm.APIConnectionError: APIConnectionError: Hosted_vllmException -](https://www.chat-gpts.plus/wp-content/uploads/2026/10/43518-60da76fd-768x403.jpg)
该报错通常出现在 LiteLLM Proxy 以 hosted_vllm 接入 OpenAI 兼容后端、开启流式输出并处于高负载时。流式过程中上游返回 500（或连接异常）后，被异常映射逻辑包装成了误导性的 APIConnectionError: Hosted_vllmException ，优先排查
Misc. bug: generation started to crawl after update
该问题通常出现在升级到含 Vulkan 多设备相关改动的 llama.cpp 版本后（Issue 中指向 b11594 及之后的构建），在多块 Vulkan 设备（尤其双 RADV/AMD 显卡，或 NVIDIA + AMD 混合环境）上跑模型时生成速度骤降。优先排查是否处于多设备 Vulkan 环
Misc. bug: Vulkan ~12% prefill regression on RDNA4 (RX 9070 XT) since #29182 (MoE-aware mul_mat_id tile selection)
这是 llama.cpp Vulkan 后端在 RDNA4（RX 9070 XT）以及 Strix Halo iGPU 上出现的 prefill 性能回退，报错 / 报障核心为 “Misc. bug: Vulkan ~12% prefill regression on RDNA4 (RX 9070
Misc. bug: stack buffer overflow in clip_encode MiniCPM-V position buckets via malicious mmproj (bucket_coords_h[1024] written for i < pos_h
![Misc. bug: stack buffer overflow in clip_encode MiniCPM-V position buckets via malicious mmproj (bucket_coords_h[1024] written for i < pos_h](https://www.chat-gpts.plus/wp-content/uploads/2026/10/27166-45771059-768x403.jpg)
该报错（stack buffer overflow in clip_encode MiniCPM-V position buckets via malicious mmproj (bucket_coords_h[1024] written for i < pos_h)）出现在 llama.cpp/l
Agent app: parallel tool calls leave permanently spinning, empty tool entries in the thought panel
该问题通常出现在自托管 Dify 的 Agent 应用里，当同一轮对话中模型并发调用多个工具（parallel tool calls）时，思考面板（thought panel）会出现一直转圈、结果为空或重复的条目。优先排查前端聊天 hook 的 thought 合并逻辑，而不是工具本身或后端执行环节
[Bug]: cuda_archs_loose_intersection drops plain “10.0” when TORCH_CUDA_ARCH_LIST also has “10.0a”, producing sm_100a-only kernels that fail
![[Bug]: cuda_archs_loose_intersection drops plain "10.0" when TORCH_CUDA_ARCH_LIST also has "10.0a", producing sm_100a-only kernels that fail](https://www.chat-gpts.plus/wp-content/uploads/2026/10/56548-3c41a23b-768x403.jpg)
报错为 no kernel image is available for execution on the device ，通常发生在本机编译 vLLM 时： TORCH_CUDA_ARCH_LIST 同时包含 10.0 和 10.0a ， cuda_archs_loose_intersection