-
Notifications
You must be signed in to change notification settings - Fork 21.5k
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.
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
[Reopen] Upgrade submodule oneDNN to v3.4.2 #126137
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/126137
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New FailureAs of commit 1995054 with merge base 3c4058c ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
Hi @snadampal Please help check if the aarch64 failure is related to oneDNN upgrade. If yes, please help investigate. If no, we will start validation on x86 CPU. |
I think the problem here is that when we are building wheel (is this script used: https://github.com/pytorch/builder/blob/ead2e21b24038ed9117060bb2c00bc79c4e91d85/aarch64_linux/aarch64_wheel_ci_build.py#L81 to build @snadampal is this right? Should CI build script be updated to use ACL 23.11+ to build with oneDNN 3.4+? |
@atalman has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
@Xia-Weiwen , the CI failure is because oneDNN3.4 requires ACL version update, here is the PR:#125549 Hi @atalman , could you please suggest how we can handle this dependency. one option is to merge aarch64-linux CI PR (#125549) first, so that the CI will be ready to test oneDNN 3.4 on aarch64-linux. |
Yes, @milpuz01 , it was the ACL version, i added the details above. |
hi @snadampal and @Xia-Weiwen merging both PRs sounds like a good solution. @Xia-Weiwen could you please add this commit: 091f84e on top of this PR ? |
This is required for oneDNN 3.4.x version
Hi @atalman @snadampal @milpuz01 I have added the commit 091f84e on top of this PR. |
@atalman has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
@pytorchmergebot merge -f "All required workflows are green" |
Merge startedYour change will be merged immediately since you used the force (-f) flag, bypassing any CI checks (ETA: 1-5 minutes). Please use Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
Update oneDNN v3.4.2 acceptance test reports
All tests geomean ratio is good, but detected 1 model AMP training performance drop with oneDNN v3.4.2 as below. The issue has been reported to oneDNN team. @atalman any comments on it?
|
Thanks for Chuanqi's testing result, @vpirogov, could you also help comment the potential impact of this outlier? |
Reopen of pytorch#122472 ## Improvements This upgrade fixes the following issues: - pytorch#120982 This upgrade brings the following new features: - Introduced memory descriptor serialization API. This API is needed to support freezing on CPU in AOTInductor (pytorch#114450) ## Validation results on CPU Original results with oneDNN v3.4.1 are here: pytorch#122472 (comment) Need to rerun validation and update results. Co-authored-by: Sunita Nadampalli <nadampal@amazon.com> Pull Request resolved: pytorch#126137 Approved by: https://github.com/jgong5, https://github.com/snadampal, https://github.com/atalman
Reopen of #122472
Improvements
This upgrade fixes the following issues:
This upgrade brings the following new features:
Validation results on CPU
Original results with oneDNN v3.4.1 are here: #122472 (comment)
Need to rerun validation and update results.
Results of reruns
All tests geomean ratio is good, but detected 1 model AMP training performance drop with oneDNN v3.4.2 as below. The issue has been reported to oneDNN team.
Dynamo benchmarks
Inductor benchmarks
cc @gujinghui @PenghuiCheng @XiaobingSuper @jianyuh @jgong5 @mingfeima @sanchitintel @ashokei @jingxu10 @min-jean-cho @yanbing-j @Guobing-Chen @snadampal