Skip to content

[hipBLASLt] Fix numeric regression of MXFP4#4808

Closed
amd-chunxlin wants to merge 3 commits into
gfx950_mx_rebasefrom
users/chunxlin/workaround_fix
Closed

[hipBLASLt] Fix numeric regression of MXFP4#4808
amd-chunxlin wants to merge 3 commits into
gfx950_mx_rebasefrom
users/chunxlin/workaround_fix

Conversation

@amd-chunxlin
Copy link
Copy Markdown
Contributor

@amd-chunxlin amd-chunxlin commented Feb 23, 2026

Motivation

This PR fixes the numeric regression caused by #4599

Technical Details

Revert attaching scaleA and scaleB data pointers when MX block size is non-zero.

Test Plan

Manually build tensileLite and ran the test with mx32f4_tn.yaml and it passed.

Test Result

Submission Checklist

@amd-chunxlin amd-chunxlin changed the title [hipBLASLt] Workaround for solving the numeric regression [hipBLASLt] Fix numeric regression of MXFP4 Feb 23, 2026
@amd-chunxlin
Copy link
Copy Markdown
Contributor Author

Fixed in another PR.

@amd-chunxlin amd-chunxlin deleted the users/chunxlin/workaround_fix branch February 27, 2026 16:40
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants