โœ… Temporary fixes for FA3, which is currently broken; fallback to SDPA.

#10
by prithivMLmods - opened

FA3 is currently broken on Blackwell (sm_100) GPUs; this module detects that
at import time and falls back to PyTorch scaled-dot-product attention (SDPA)
automatically. The public class name / call signature are unchanged.

prithivMLmods changed pull request title from Temporary fixes for FA3, which is currently broken; fallback to SDPA. to โœ… Temporary fixes for FA3, which is currently broken; fallback to SDPA.
prithivMLmods changed pull request status to merged

Sign up or log in to comment