Skip to content

fix: align ck_moe_stage1 split-K tmp_out buffer with CK kernel #2508

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

Closed
karverma-amd wants to merge 4 commits into from
+10 −8
Closed

fix: align ck_moe_stage1 split-K tmp_out buffer with CK kernel #2508

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
f7ddc47
fix: align ck_moe_stage1 split-K tmp_out buffer with CK kernel
karverma-amd Mar 27, 2026
3f50525
Merge branch 'main' into main
karverma-amd Mar 30, 2026
857c9d2
Update fused_moe.py
karverma-amd Mar 30, 2026
5063b5b
Merge branch 'main' into main
karverma-amd Mar 31, 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
18 changes: 10 additions & 8 deletions aiter/fused_moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1707,13 +1707,14 @@ def ck_moe_stage1(
):
token_num = hidden_states.shape[0]
is_splitk = quant_type is aiter.QuantType.per_1x128 and splitk > 1
tmp_out = (
Copy link
Copy Markdown

@ChuanLi1101 ChuanLi1101 Mar 31, 2026

Choose a reason for hiding this comment

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

cktile_moe_stage1 has the identical undersized buffer pattern for split_k > 1. It calls aiter.moe_cktile2stages_gemm1 which is a different kernel, but if that kernel has the same hipMemsetAsync behavior, it would have the same overflow. The TODO at line 1723 (# TODO: support fp32 splitk) suggests this path may not be actively used yet, but it's worth either applying the same fix or adding a guard/comment.

torch.zeros(
(token_num, topk, w1.shape[1]), dtype=dtypes.fp32, device=out.device
if is_splitk:
# CK splitK kernel hipMemsetAsync zeros sorted_size * w1.shape[1] floats
sorted_size = min(token_num * topk * block_m, sorted_token_ids.shape[0])
tmp_out = torch.zeros(
Copy link
Copy Markdown
Collaborator

@valarLip valarLip Mar 31, 2026

Choose a reason for hiding this comment

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

not good to me, looks like we do zeros twice, 1 here, 1 ck, we need remove one

Copy link
Copy Markdown
Contributor

@rbrugaro-amd rbrugaro-amd Mar 31, 2026 •
edited
Loading

Choose a reason for hiding this comment

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

This ROCm/rocm-libraries#5225 already addressed the proper allocation from the C++ side. I have swap to torch.empty to avoid the double zeroing #2551 (created new PR because i cannot edit current branch)
@ChuanLi1101 i see you also created another PR but I think the double zeroing still there?

(sorted_size, w1.shape[1]), dtype=dtypes.fp32, device=out.device
)
if is_splitk
else out
)
else:
tmp_out = out
aiter.ck_moe_stage1_fwd(
hidden_states,
w1,
Expand All @@ -1735,10 +1736,11 @@ def ck_moe_stage1(
out.dtype,
)
if is_splitk:
valid_out = tmp_out[: token_num * topk, :].contiguous()
if activation == ActivationType.Silu:
aiter.silu_and_mul(out, tmp_out.view(dtypes.fp32))
aiter.silu_and_mul(out, valid_out.view(dtypes.fp32))
else:
aiter.gelu_and_mul(out, tmp_out.view(dtypes.fp32))
aiter.gelu_and_mul(out, valid_out.view(dtypes.fp32))
return out


Expand Down
Loading