-
Notifications
You must be signed in to change notification settings - Fork 228
Fixing MAGMA with HIP issues #695
Fixing MAGMA with HIP issues #695
Conversation
@jithunnair-amd @pruthvistony @jeffdaily @sunway513 - please review |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Can confirm that the ./manywheel/deploy.sh
is able to successfully build the pytorch/manylinux-rocm:4.1
image with these changes.
Summary: Fixes #55552 * Root-caused issue to MAGMA kernels * Issue is fixed on master of MAGMA MAGMA issue: https://bitbucket.org/icl/magma/issues/43/zgetrf_batched-shfl-kernel-failure-seen-on * Changing PyTorch to use particular commit sha from master of MAGMA project * ~~Reactivating skipped ROCm tests~~ : We will reactivate tests in a different PR Corresponding PyTorch builder PR: pytorch/builder#695 Pull Request resolved: #56448 Reviewed By: seemethere Differential Revision: D27974563 Pulled By: janeyx99 fbshipit-source-id: 25e6f95a20a06d27a5199a623dd7c5db7ca8d6ea
@seemethere We have the exact same changes in PR pytorch/pytorch#56448 merged upstream. This PR is to make sure the wheels also have the correct version of Magma built. The only difference from the upstream PR change is the Please merge this PR if it looks okay. |
Summary: Fixes pytorch#55552 * Root-caused issue to MAGMA kernels * Issue is fixed on master of MAGMA MAGMA issue: https://bitbucket.org/icl/magma/issues/43/zgetrf_batched-shfl-kernel-failure-seen-on * Changing PyTorch to use particular commit sha from master of MAGMA project * ~~Reactivating skipped ROCm tests~~ : We will reactivate tests in a different PR Corresponding PyTorch builder PR: pytorch/builder#695 Pull Request resolved: pytorch#56448 Reviewed By: seemethere Differential Revision: D27974563 Pulled By: janeyx99 fbshipit-source-id: 25e6f95a20a06d27a5199a623dd7c5db7ca8d6ea
Fixes pytorch/pytorch#55552
Main PyTorch PR : pytorch/pytorch#56448
MAGMA issue: https://bitbucket.org/icl/magma/issues/43/zgetrf_batched-shfl-kernel-failure-seen-on