[ROCm] Add attention sink support to AITer flash attention backend (#43817)

Signed-off-by: Xiaoran Chen <xiaoran@fb.com>
Co-authored-by: Xiaoran Chen <xiaoran@fb.com>
This commit is contained in:
Xiaoran
2026-05-30 18:13:18 +08:00
committed by GitHub
co-authored by Xiaoran Chen
parent 124fac10cb
commit 3becc5db40
3 changed files with 25 additions and 5 deletions
+2
View File
@@ -2394,6 +2394,7 @@ class rocm_aiter_ops:
alibi_slopes: torch.Tensor | None = None,
return_lse: bool = False,
out: torch.Tensor | None = None,
sink_ptr: torch.Tensor | None = None,
):
"""
Flash attention with variable length sequences.
@@ -2422,6 +2423,7 @@ class rocm_aiter_ops:
alibi_slopes=alibi_slopes,
return_lse=return_lse,
out=out,
sink_ptr=sink_ptr,
)
@staticmethod