Skip to content

Fix fused DeepSeek top-k capability fallback #35347

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
nv-tusharma wants to merge 5 commits into
base: main
Choose a base branch
from
Open

Fix fused DeepSeek top-k capability fallback #35347

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
34a7502
fix: fall back when fused top-k lacks device support
trsharm25 Aug 18, 2026
5330def
style: fix fused top-k test imports
trsharm25 Aug 18, 2026
c9de8ab
test: remove fused top-k capability unit tests
trsharm25 Aug 19, 2026
87673c7
Merge branch 'main' into codex/fallback-unsupported-fused-topk
nv-tusharma Aug 20, 2026
ef1d7d3
Merge branch 'main' into codex/fallback-unsupported-fused-topk
nv-tusharma 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
14 changes: 14 additions & 0 deletions python/sglang/srt/layers/moe/topk.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -125,6 +125,7 @@ def routing(
_is_npu = is_npu()
_use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
_is_musa = is_musa()
_fused_topk_deepseek = None

# Epsilon added to the top-k weight sum before renormalization, matching the
# DeepSeek reference gate (modeling_deepseek.py: `topk_weight.sum(...) + 1e-20`)
Expand All @@ -145,6 +146,18 @@ def routing(
_skip_hip_pad_mask = get_bool_env_var("SGLANG_MORI_NO_PAD_MASK", "False")


def _flashinfer_fused_topk_supports_device(device: torch.device) -> bool:
"""Return whether FlashInfer's DeepSeek top-k kernel supports ``device``."""
support_check = getattr(
_fused_topk_deepseek, "is_compute_capability_supported", None
)
if support_check is None:
return True

major, minor = torch.cuda.get_device_capability(device)
return support_check(major * 10 + minor)


if _is_cuda:
try:
from flashinfer.fused_moe import fused_topk_deepseek as _fused_topk_deepseek
Expand Down Expand Up @@ -1532,6 +1545,7 @@ def biased_grouped_topk_gpu(
if (
_is_cuda
and fused_topk_deepseek is not None
and _flashinfer_fused_topk_supports_device(gating_output.device)
and is_power_of_two(num_experts)
# flashinfer constraints (applied to routed experts only)
and topk_routed <= 8
Expand Down
Loading