Skip to content

Local window attention deterministic mode efficiency improvement #1972

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
wants to merge 2 commits into
base: main
Choose a base branch
from
+7 −11
Draft

Local window attention deterministic mode efficiency improvement #1972

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
18faf12
Enable LPT scheduler for sliding window attention
GD06 Oct 29, 2025
ae4dadf
Early exit for the local window attn in the deterministic mode
GD06 Oct 29, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion hopper/flash_bwd_launch_template.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -94,7 +94,7 @@ void run_flash_bwd(Flash_bwd_params &params, cudaStream_t stream) {
flash::CollectiveEpilogueBwdGQA<TileShape_MNK, ElementAccum, ArchTag, CollectiveMainloop::NumMmaThreads, Varlen, Deterministic>
>;
using Scheduler = std::conditional_t<
Is_causal,
Is_causal || Is_local,
flash::SingleTileBwdLPTScheduler<Varlen, kBlockN, Is_causal && Deterministic /*SPT*/>,
flash::SingleTileScheduler<Varlen, false /*Split*/, false /*PackGQA*/, kBlockN>
>;
Expand Down
16 changes: 6 additions & 10 deletions hopper/mainloop_bwd_sm90_tma_gmma_ws.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -607,8 +607,7 @@ struct CollectiveMainloopBwdSm90 {
seqlen_info, n_block, bidb, params.window_size_left,
params.window_size_right, 0 /*sink_token_length*/);
// It's possible to have m_block_max <= m_block_min. Exit early
// Though if local and deterministic, still need to increment dq semaphore
if constexpr ((Is_causal || Is_local || Varlen) && !(Is_local && Deterministic)) {
if constexpr (Is_causal || Is_local || Varlen) {
if (m_block_max <= m_block_min) { return; }
}

Expand Down Expand Up @@ -637,7 +636,11 @@ struct CollectiveMainloopBwdSm90 {
int n_block_max_for_m_block = std::min(n_block_global_max, cute::ceil_div((m_block + 1) * kBlockM + seqlen_info.seqlen_k - seqlen_info.seqlen_q, kBlockN));
Barrier::wait_eq(lock_ptr, threadIdx.x % cutlass::NumThreadsPerWarp, m_block * num_batch * num_head, n_block_max_for_m_block - 1 - n_block);
} else {
Barrier::wait_eq(lock_ptr, threadIdx.x % cutlass::NumThreadsPerWarp, m_block * num_batch * num_head, n_block);
int n_block_min = 0;
if constexpr(Is_local) {
n_block_min = std::max(n_block_min, (m_block * kBlockM + seqlen_info.seqlen_k - seqlen_info.seqlen_q - params.window_size_left) / kBlockN);
}
Barrier::wait_eq(lock_ptr, threadIdx.x % cutlass::NumThreadsPerWarp, m_block * num_batch * num_head, n_block - n_block_min);
}
}
#pragma unroll
Expand All @@ -657,13 +660,6 @@ struct CollectiveMainloopBwdSm90 {
Barrier::arrive_inc(lock_ptr, threadIdx.x % cutlass::NumThreadsPerWarp, m_block * num_batch * num_head);
}
}
if constexpr (Is_local && Deterministic) {
int const m_block_global_max = cute::ceil_div(seqlen_info.seqlen_q, kBlockM);
#pragma unroll 2
for (; m_block < m_block_global_max; ++m_block) {
Barrier::arrive_inc(lock_ptr, threadIdx.x % cutlass::NumThreadsPerWarp, m_block * num_batch * num_head);
}
}
}

CUTLASS_DEVICE void
Expand Down