Skip to content

[GPU] Fix accuracy issues for mvn and sdpa_micro #30698

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

Jump to bottom
Open
wants to merge 4 commits into
base: master
Choose a base branch
from
Open

[GPU] Fix accuracy issues for mvn and sdpa_micro #30698

wants to merge 4 commits into from
+285 −2

Conversation

ahnyoung-paul
Copy link
Contributor

Details:

  • Add boundary check enabled to MVN fused ops if boundary check is needed
  • Fix attn mask tile loading issue when attn mask shape has single query dimension [batch, num_heads, 1, sequence_length of k]

Tickets:

  • 166691

@ahnyoung-paul ahnyoung-paul requested review from a team as code owners May 23, 2025 13:52
@ahnyoung-paul ahnyoung-paul added category: GPU OpenVINO GPU plugin pr: needs tests PR needs tests updating under_perf_check labels May 23, 2025
yeonbok
yeonbok reviewed May 23, 2025
View reviewed changes
src/plugins/intel_gpu/src/kernel_selector/cl_kernels/sdpa_micro.cl
if (MSK_D2 == 1) {
// Define mask dimensions for single Query dimension
uint mask_m = MSK_D1; // num_heads
uint mask_n = MSK_D3; // sequence_length
Copy link
Contributor

Choose a reason for hiding this comment

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

why num_heads? masks are usually [q, seqlen]..

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@yeonbok yeonbok yeonbok left review comments

At least 1 approving review is required to merge this pull request.

Assignees
No one assigned
Labels
category: GPU OpenVINO GPU plugin pr: needs tests PR needs tests updating under_perf_check
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
2 participants
@ahnyoung-paul @yeonbok