Skip to content

[ROCm] Bound the sliding-window prefix loop and the unified-kernel KV loop (bit-identical; no e2e win claimed) #34468

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

Open
alexnails wants to merge 3 commits into
base: main
Choose a base branch
from
Open

[ROCm] Bound the sliding-window prefix loop and the unified-kernel KV loop (bit-identical; no e2e win claimed) #34468

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
9ed808e
[ROCm] Bound the sliding-window extend-attention prefix loop
alexnails Aug 11, 2026
4f2cd17
[ROCm] Bound the sliding-window KV loop in _fwd_kernel_unified
alexnails Aug 11, 2026
2677ced
Merge branch 'main' into perf/gfx950-swa-prefix-trim
alexnails Aug 24, 2026
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
29 changes: 27 additions & 2 deletions python/sglang/kernels/ops/attention/extend_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -458,7 +458,19 @@ def _fwd_kernel(
e_max = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")

prefix_end = 0 if SKIP_PREFIX else cur_seq_len_prefix
for start_n in range(0, prefix_end, BLOCK_N):
# The extend stage's floor, with q at its absolute position
# (cur_seq_len_prefix + cur_block_m * BLOCK_M) because kv is indexed within the
# prefix; tight for any BLOCK_M/BLOCK_N, and SKIP_TILE already made the skipped
# tiles no-ops. Past the first window's m-blocks no query reaches the prefix.
prefix_start = 0
if SLIDING_WINDOW_SIZE > 0:
prefix_start = (
tl.maximum(
cur_seq_len_prefix + cur_block_m * BLOCK_M - SLIDING_WINDOW_SIZE, 0
)
// BLOCK_N
) * BLOCK_N
for start_n in range(prefix_start, prefix_end, BLOCK_N):
start_n = tl.multiple_of(start_n, BLOCK_N)
mask_n = (start_n + offs_n) < cur_seq_len_prefix

Expand Down Expand Up @@ -1088,7 +1100,20 @@ def _fwd_kernel_unified(
e_max = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")

# Unified loop: process all KV tokens (prefix + extend)
for start_n in range(0, cur_seq_kv_len, BLOCK_N):
#
# The two-stage kernel's floor. cur_window_start appears on both sides of the
# mask below and cancels, so A = cur_seq_prefix_len + cur_block_m * BLOCK_M -
# SLIDING_WINDOW_SIZE. This loop spans prefix and extend KV alike, so the one
# floor drops the fully-masked tiles of both.
unified_start = 0
if SLIDING_WINDOW_SIZE > 0:
unified_start = (
tl.maximum(
cur_seq_prefix_len + cur_block_m * BLOCK_M - SLIDING_WINDOW_SIZE, 0
)
// BLOCK_N
) * BLOCK_N
for start_n in range(unified_start, cur_seq_kv_len, BLOCK_N):
start_n = tl.multiple_of(start_n, BLOCK_N)
mask_n = (start_n + offs_n) < cur_seq_kv_len

Expand Down
Loading