Ethanfel 4c6c38f05a Fix attention dispatcher: use 4D tensors for SDPA, add math backend ...

SDPA with 3D xformers-BMK tensors cannot use Flash Attention and falls
back to efficient_attention/math kernels that miscompute on Ada Lovelace
GPUs (e.g. RTX 6000 Pro), producing brownish line artifacts.  Unsqueeze
to 4D (1, B*H, N, D) so Flash Attention is eligible.  Also add a naive
"math" backend (chunked bmm) as a guaranteed-correct diagnostic baseline.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

2026-02-15 01:05:51 +01:00

17 KiB

Raw Blame History

View Raw