Skip to content

[Piecewise CUDA Graph] Support Kimi-K2 (non-Thinking) #13466

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
b8zhong merged 3 commits into from
Nov 21, 2025
Merged

[Piecewise CUDA Graph] Support Kimi-K2 (non-Thinking) #13466

Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
816ee1f
upd
b8zhong Nov 18, 2025
815df3f
Merge branch 'main' into brayden/kimi-k2-piecewise
b8zhong Nov 19, 2025
3a0c338
Merge branch 'main' into brayden/kimi-k2-piecewise
b8zhong Nov 20, 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
23 changes: 23 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 @@ -74,6 +74,29 @@
if _is_cuda:
from sgl_kernel import kimi_k2_moe_fused_gate, moe_fused_gate

@torch.library.register_fake("sgl_kernel::kimi_k2_moe_fused_gate")
def _kimi_k2_moe_fused_gate(
input_tensor,
bias,
topk,
renormalize,
routed_scaling_factor,
apply_routed_scaling_factor_on_output,
):
num_rows = input_tensor.shape[0]
topk_weights = input_tensor.new_empty(
num_rows,
topk,
dtype=torch.float32,
)
topk_ids = input_tensor.new_empty(
num_rows,
topk,
dtype=torch.int32,
Comment on lines +87 to +95
Copy link

Copilot AI Nov 18, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[nitpick] The tensor creation pattern is inconsistent with the existing moe_fused_gate fake implementation (lines 1072-1077). The existing implementation uses torch.empty((num_rows, topk), ...) with an explicit device=input_tensor.device parameter, while this uses input_tensor.new_empty(num_rows, topk, ...). For consistency and clarity, consider using the same pattern: torch.empty((num_rows, topk), dtype=torch.float32, device=input_tensor.device).

Suggested change
topk_weights = input_tensor.new_empty(
num_rows,
topk,
dtype=torch.float32,
)
topk_ids = input_tensor.new_empty(
num_rows,
topk,
dtype=torch.int32,
topk_weights = torch.empty(
(num_rows, topk),
dtype=torch.float32,
device=input_tensor.device,
)
topk_ids = torch.empty(
(num_rows, topk),
dtype=torch.int32,
device=input_tensor.device,

Copilot uses AI. Check for mistakes.
)
return topk_weights, topk_ids


if _is_cuda or _is_hip:
from sgl_kernel import topk_softmax
if _use_aiter:
Expand Down
Loading