Commit 6bb896f
committed
Adds flash sparse attention utils
Introduces lazy import plumbing for flash sparse attention kernels to streamline future integrations.
Prepares padding-aware helpers and kwarg validation so padding-free flows and PEFT casting stay compatible with the kernels.1 parent 0402b39 commit 6bb896f
File tree
1 file changed
+596
-0
lines changed- flash_sparse_attn/integrations
1 file changed
+596
-0
lines changed
0 commit comments