Eval bug: CUDA: MUL_MAT_ID via MMVQ fails with “invalid argument” on Turing when device code is compiled for an older arch (e.g. 61-virtual)

在 llama.cpp 的 CUDA 后端上运行 MoE 模型时,如果二进制或构建过程按较旧的 compute capability 生成 PTX(例如用 61-virtual ),再在 Turing/Volta 等较新的显卡上通过 JIT 运行,那么遇到 5–8 个 token 的 MoE 小批次


![Muse tool parser JSON-decodes Optional[str] parameters, so "10001" becomes 10001](https://www.chat-gpts.plus/wp-content/uploads/2026/11/64546-39a741f7-768x403.jpg)




![`prompt_cache_options.ttl` is typed `Literal["30m"]` but the API accepts and returns `"12h"`; `cache_write_tokens_details` is untyped](https://www.chat-gpts.plus/wp-content/uploads/2026/11/4156-a7aeeaf6-768x403.jpg)
