[FA4] Keep stream last in the SM100 MLA kernel so disk-cached objects load (#42926)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Po-Han Huang (NVIDIA)
2026-10-07 15:58:55 +08:00
committed by GitHub
co-authored by Claude Opus 5.5
parent e23093077e
commit ba77d30256
2 changed files with 3 additions and 2 deletions
@@ -396,9 +396,8 @@ class FlashAttentionMLAForwardSm100:
mPageTable: Optional[cute.Tensor] = None,
window_size_left: Int32 | int | None = None,
window_size_right: Int32 | int | None = None,
stream: cuda.CUstream = None,
*,
mValue: Optional[cute.Tensor] = None,
stream: cuda.CUstream = None,
):
# fmt: on
self.store_P = mP is not None
@@ -1589,6 +1589,7 @@ def _flash_attn_fwd(
page_table_tensor,
window_size_left,
window_size_right,
None, # mValue
current_stream,
options="--enable-tvm-ffi",
)
@@ -1687,6 +1688,7 @@ def _flash_attn_fwd(
page_table,
window_size_left,
window_size_right,
None, # mValue
)
else:
call_args = [