[reland][ROCm] remove caffe2 from hipify #151845
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/151845
Note: Links to docs will display an error until the docs builds have been completed. ✅ You can merge normally! (10 Unrelated Failures)As of commit 9cd9145 with merge base 4f425a0 ( FLAKY - The following jobs failed but were likely due to flakiness present on trunk:
BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
UNSTABLE - The following jobs are marked as unstable, possibly due to flakiness on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
| @@ -1 +1 @@ | |||
| __version__ = '1.0.0' | |||
| __version__ = '2.0.0' | |||
There was a problem hiding this comment.
What's the impact of this?
There was a problem hiding this comment.
The version has never been updated even as we added new cuda-to-hip mappings. The change is needed now because we're making BC-breaking changes to torch's hipify. We have a related project hipify_torch that C++ projects can use --- it was an attempt years ago to make hipify into a standalone tool that could be a torch thirdparty submodule but we decided it would have negatively impacted developer experience for torch by making it a submodule. So we have both the hipify that's inside torch and the hipify_torch c++ cmake helper project. The hipify_torch project might query the torch.utils.hipify.version so it can conditionally perform the corresponding hipify steps, v1 or v2.
Same as the last PR, we used a script to rewrite the cuda to hip mappings file. New script is https://gist.github.com/jeffdaily/d0497b33135fb5e476e150e7aaea62c6 . It's the biggest diff of this PR, but it's automated. The important part is that we dropped the caffe2-specific mappings. |
These changes correspond to pytorch/pytorch#151845 (and syncs changes from pytorch's hipify as part of this PR). torch's hipify version 2 significantly simplifies the the cuda to hip mappings, including a BC-breaking change of no longer using caffe2 mappings and getting rid of the Masquerading files. In order to ease the transition, hipify_torch will contain both versions of the mappings. An attempt is made to check which version of torch is installed and use the right mappings, otherwise the v1 mappings are used unless the caller specifies --v2.
|
@xw285cornell has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
Reland of #137157.
Needs pytorch/FBGEMM#4028 to be merged first to keep FBGEMM functional.
cc @H-Huang 8000 @awgu @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @dllehr-amd @jataylo @hongxiayang @naromero77amd @albanD @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov