Skip to content

[opt kimi k2 2/n] apply kimi k2 thinking moe_fused_gate #13332

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

Merged
ispobock merged 3 commits into from
Nov 16, 2025
Merged

[opt kimi k2 2/n] apply kimi k2 thinking moe_fused_gate #13332

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
8e5699a
apply kimi k2 thinking moe_fused_gate
BBuf Nov 15, 2025
9afb429
Add CUDA check for Kimi K2 optimized path
BBuf Nov 16, 2025
c705784
Merge branch 'main' into apply_kimi_k2_moe_fused_gate
BBuf Nov 16, 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
15 changes: 6 additions & 9 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 @@ -72,7 +72,7 @@
_use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip

if _is_cuda:
from sgl_kernel import moe_fused_gate
from sgl_kernel import kimi_k2_moe_fused_gate, moe_fused_gate

if _is_cuda or _is_hip:
from sgl_kernel import topk_softmax
Expand Down Expand Up @@ -817,16 +817,13 @@ def biased_grouped_topk_gpu(
else:
# Use optimized path for Kimi K2 (384 experts with num_expert_group=1)
num_experts = gating_output.shape[1]
if num_experts == 384 and num_expert_group == 1:
return kimi_k2_biased_topk_impl(
hidden_states,
gating_output,
if _is_cuda and num_experts == 384 and num_expert_group == 1:
return kimi_k2_moe_fused_gate(
gating_output.to(dtype=torch.float32),
correction_bias,
topk,
renormalize,
topk=topk,
renormalize=renormalize,
routed_scaling_factor=routed_scaling_factor,
num_token_non_padded=num_token_non_padded,
expert_location_dispatch_info=expert_location_dispatch_info,
apply_routed_scaling_factor_on_output=apply_routed_scaling_factor_on_output,
)
else:
Expand Down
Loading