viable/strict/1789757464: [Windows] Fix FlashAttention CUDA build with MSVC C++20 (#197382)
- PyTorch 近 90 天出现 1156 次
- PyTorch 近 90 天第 1142 次发版
- 上一次：同一天稍早 · viable/strict/1789753839: [MPS] Guard MPP attention include on header availability (#196104)
发生了什么
Summary Follow-up to #186343 , which enables the native FlashAttention SDPA backend for Windows CUDA builds. MSVC implicitly enables /permissive- when compiling as C++20 with CTK13.4. This triggers a CUTLASS/CuTe template parsing failure ( C3545 ) while compiling the FlashAttention CUDA kernels. Pass /permissive only to CUDA sources in the flash_attention target, leaving conformance settings for the rest of PyTorch…
摘要按规则整理自下方来源原文