Skip to content

Conversation

@leonling-ll
Copy link
Contributor

@leonling-ll leonling-ll commented Nov 15, 2024

The cause of this error is that we should initialize the results tensor when using atomic_add, otherwise it would read dirty memory from previous benchmarking cases.

@leonling-ll leonling-ll self-assigned this Nov 15, 2024
@leonling-ll leonling-ll linked an issue Nov 15, 2024 that may be closed by this pull request
[512, 32768, 8192],
[1024, 28672, 8192],
[3072, 4096, 3072],
[4096, 4096, 4096],
Copy link
Contributor

Choose a reason for hiding this comment

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

Adding a new combination breaks the CI and seems a bit out of topic for this pull request. Maybe we should move this change to a separate pull request?

Copy link
Contributor

Choose a reason for hiding this comment

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

Agree. Split that change in a separate PR please @LiyangLingIntel

Copy link
Contributor

Choose a reason for hiding this comment

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

Other than that LGTM.

Copy link
Contributor

Choose a reason for hiding this comment

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

I disagree, as this PR supposes to fix the 4k functional error.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

I've added 4k shape to XeTLA splitk list, benchmark CI works now.

[512, 32768, 8192],
[1024, 28672, 8192],
[3072, 4096, 3072],
[4096, 4096, 4096],
Copy link
Contributor

Choose a reason for hiding this comment

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

Other than that LGTM.

Copy link
Contributor

@whitneywhtsang whitneywhtsang left a comment

Choose a reason for hiding this comment

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

please add gemm_splitk_shape_4096_4096_4096 to XeTLA to fix the CI failure.

@etiotto etiotto merged commit 932d0be into main Nov 18, 2024
5 checks passed
@etiotto etiotto deleted the liyang/fix_splitk_gemm_assertion branch November 18, 2024 14:22
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Assertion error on gemm_splitk_benchmark.py

6 participants