From ba77d30256bfb1d1e0d53cd877c87f0b2e279ef1 Mon Sep 17 00:00:00 2001 From: "Po-Han Huang (NVIDIA)" <53919306+nvpohanh@users.noreply.github.com> Date: Wed, 7 Oct 2026 15:58:55 +0800 Subject: [PATCH] [FA4] Keep stream last in the SM100 MLA kernel so disk-cached objects load (#42926) Co-authored-by: Claude Opus 5.5 --- .../ops/attention/flash_attn/cute/flash_fwd_mla_sm100.py | 3 +-- .../sglang/kernels/ops/attention/flash_attn/cute/interface.py | 2 ++ 2 files changed, 3 insertions(+), 2 deletions(-) diff --git a/python/sglang/kernels/ops/attention/flash_attn/cute/flash_fwd_mla_sm100.py b/python/sglang/kernels/ops/attention/flash_attn/cute/flash_fwd_mla_sm100.py index 3aa7e8e5dd..93c86cef3e 100644 --- a/python/sglang/kernels/ops/attention/flash_attn/cute/flash_fwd_mla_sm100.py +++ b/python/sglang/kernels/ops/attention/flash_attn/cute/flash_fwd_mla_sm100.py @@ -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 diff --git a/python/sglang/kernels/ops/attention/flash_attn/cute/interface.py b/python/sglang/kernels/ops/attention/flash_attn/cute/interface.py index b62b4f3eb5..7f225ae784 100644 --- a/python/sglang/kernels/ops/attention/flash_attn/cute/interface.py +++ b/python/sglang/kernels/ops/attention/flash_attn/cute/interface.py @@ -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 = [