Yestino - The Signal
← Back to events
ActiveAIRelease

trunk/68d20d4ee3956ceb5fecbc1676112aa32d2ad7d9: Disable USE_FLASH_ATTENTION when no CUDA arch supports it (#194502)

What happened

USE_FLASH_ATTENTION stays ON for any CUDA build regardless of target architecture, even though the kernels need sm80+ and can_use_flash_attention() rejects anything below that at runtime. A build pinned to older hardware compiles the whole kernel set (the largest source tree under third_party) into a binary that can never use it. Disable the option when none of the resolved CUDA architectures reach sm80. Reuses NVCC…

Summary assembled by rule from the sources below

Why it's spreading

Timeline

  1. First appeared on PyTorch ReleasesPyTorch Releases

Sources

Release