Skip to content

[GDN] Fused all preprocessing into one kernel #244

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
apinge wants to merge 5 commits into
base: Qwen3.5_v0.5.9
Choose a base branch
from
Open

[GDN] Fused all preprocessing into one kernel #244

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
24d4e92
add impl
apinge Apr 9, 2026
2c39a7b
correct comments
apinge Apr 9, 2026
51829b2
correct if statement
apinge Apr 9, 2026
1cc4445
correct if statement
apinge Apr 9, 2026
3f7b5f1
rectify if statement
apinge Apr 9, 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
11 changes: 10 additions & 1 deletion python/sglang/srt/layers/attention/fla/chunk.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,6 +16,7 @@
from sglang.srt.layers.attention.fla.cumsum import chunk_local_cumsum
from sglang.srt.layers.attention.fla.fused_cumsum_kkt import fused_cumsum_kkt
from sglang.srt.layers.attention.fla.fused_merge_recompute import fused_merge_recompute
from sglang.srt.layers.attention.fla.fused_preprocessing import fused_preprocessing_fwd
from sglang.srt.layers.attention.fla.index import prepare_chunk_indices
from sglang.srt.layers.attention.fla.l2norm import fused_l2norm_qk, l2norm_fwd
from sglang.srt.layers.attention.fla.solve_tril import (
Expand All @@ -32,6 +33,9 @@

_is_hip = is_hip()

# When True, preprocessing uses the fused Triton kernel
USE_GDN_PREFILL_FUSION = True


def chunk_gated_delta_rule_fwd(
q: torch.Tensor,
Expand All @@ -47,7 +51,12 @@ def chunk_gated_delta_rule_fwd(
B, T = q.shape[0], q.shape[1]
Hv = g.shape[2]

if _is_hip and T >= 64:
if USE_GDN_PREFILL_FUSION and T == 64:
g, w, u = fused_preprocessing_fwd(
k=k, v=v, beta=beta, g=g, cu_seqlens=cu_seqlens
)
A = None
elif _is_hip and T >= 64:
g, A = fused_cumsum_kkt(g, k, beta, chunk_size=64, cu_seqlens=cu_seqlens)
chunk_indices_16 = (
prepare_chunk_indices(cu_seqlens, 16) if cu_seqlens is not None else None
Expand Down
Loading