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

[Benchmarks] Deprecate import intel_extension_for_pytorch in benchmarks including decoupling XeTLA build which require ipex.h #1911

Closed
vlad-penkin opened this issue Aug 17, 2024 · 3 comments
Closed

[Benchmarks] Deprecate import intel_extension_for_pytorch in benchmarks including decoupling XeTLA build which require ipex.h #1911

vlad-penkin opened this issue Aug 17, 2024 · 3 comments
Assignees
@ESI-SYD
Labels
dependencies: pytorch enhancement New feature or request upstream: rebase PR to be up-streamed
Milestone
0.1.1 [Triton / P...

Comments

@vlad-penkin
Copy link
Contributor

No description provided.

@vlad-penkin vlad-penkin added enhancement New feature or request upstream: rebase PR to be up-streamed dependencies labels Aug 17, 2024
@vlad-penkin vlad-penkin changed the title [Benchmarks] Deprecate import intel_extension_for_pytorch in benchmarks including decoupling XeTLA build dependency on ipex.h [Benchmarks] Deprecate import intel_extension_for_pytorch in benchmarks including decoupling XeTLA build which require ipex.h Aug 17, 2024
@vlad-penkin vlad-penkin mentioned this issue Aug 18, 2024
Closed
@ZzEeKkAa ZzEeKkAa mentioned this issue Aug 21, 2024
Closed
@vlad-penkin vlad-penkin linked a pull request Aug 23, 2024 that will close this issue
Build benchmarks with upstream pytorch #1905
Closed
@vlad-penkin vlad-penkin assigned chengjunlu and unassigned ZzEeKkAa Aug 26, 2024
@chengjunlu chengjunlu removed a link to a pull request Aug 30, 2024
Build benchmarks with upstream pytorch #1905
Closed
@ESI-SYD
Copy link
Contributor

ESI-SYD commented Sep 3, 2024 •
edited
Loading

Status:

We have changed the codes and have an initial performance impact evaluation. There are 3 points need to be addressed for now.

  1. Upstream pytorch attention as golden reference for triton not works which is a blocker [FA][Upstream PT] XPU out of memory raised by FA kernel with upstream pytorch #2042.
  2. Ratio of Triton/ XeTLA keep same except for attention caused by XeTLA attention absolute number degraded
  3. Both Triton and XeTLA softmax cases degraded, so Triton/ XeTLA not changed.

details: #1905 (comment)

@vlad-penkin vlad-penkin linked a pull request Sep 3, 2024 that will close this issue
Build benchmarks with upstream pytorch #1905
Closed
@vlad-penkin
Copy link
Contributor Author

@ESI-SYD please create a new ticket for items 2 and 3.

@vlad-penkin vlad-penkin assigned ESI-SYD and unassigned chengjunlu Sep 3, 2024
@ESI-SYD
Copy link
Contributor

ESI-SYD commented Sep 4, 2024

#2106

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

@ESI-SYD ESI-SYD

Labels
dependencies: pytorch enhancement New feature or request upstream: rebase PR to be up-streamed
Projects
None yet
Milestone
0.1.1 [Triton / Pytorch] Switch to Py...
Development

Successfully merging a pull request may close this issue.

Build benchmarks with upstream pytorch intel/intel-xpu-backend-for-triton
4 participants
@ZzEeKkAa @chengjunlu @ESI-SYD @vlad-penkin