Skip to content
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

[FA2 performance] support HEAD_SIZE=128 #1917

Closed
Dewei-Wang-sh opened this issue Aug 19, 2024 · 5 comments · Fixed by #2180
Closed

[FA2 performance] support HEAD_SIZE=128 #1917

Dewei-Wang-sh opened this issue Aug 19, 2024 · 5 comments · Fixed by #2180
Assignees
@Dewei-Wang-sh
Labels
codegen: attention enhancement New feature or request
Milestone
4.0 [Performance]...

Comments

@Dewei-Wang-sh
Copy link
Contributor

need to change several passes to make it end2end work.
@Dewei-Wang-sh Dewei-Wang-sh self-assigned this Aug 19, 2024
@Dewei-Wang-sh Dewei-Wang-sh added this to the 4.0 [Performance] Core milestone Aug 19, 2024
@vlad-penkin vlad-penkin added the enhancement New feature or request label Aug 19, 2024
@Dewei-Wang-sh
Copy link
Contributor Author

WIP, modify code in convert-to-tritongpu-warp to accommodate dim128.
still need changes in match-target-size and convert-to-llvm to make it work.

@Dewei-Wang-sh
Copy link
Contributor Author

  1. this relies on [FIX]: fix FlashAttension in functionality #2061 to be merged
  2. convert-to-tritongpu-warp now in draft.
  3. still need to modify other passes to make it end2en work.

@Dewei-Wang-sh
Copy link
Contributor Author

finish changes in other passes.
result do not match. need to check why.
depend on #2098 to make it under review and merge.

@Dewei-Wang-sh Dewei-Wang-sh linked a pull request Sep 11, 2024 that will close this issue
Fix broadcast #2180
Merged
@Dewei-Wang-sh Dewei-Wang-sh reopened this Sep 23, 2024
@Dewei-Wang-sh
Copy link
Contributor Author

first pr merged.
depend on #2158 to get correct result.

@Dewei-Wang-sh
Copy link
Contributor Author

false alert, used wrong reference data.

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

@Dewei-Wang-sh Dewei-Wang-sh

Labels
codegen: attention enhancement New feature or request
Projects
None yet
Milestone
4.0 [Performance] Core
Development

Successfully merging a pull request may close this issue.

Fix broadcast Dewei-Wang-sh/intel-xpu-backend-for-triton
2 participants
@vlad-penkin @Dewei-Wang-sh