Repository navigation
Conversation
Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags:
|
@scotts has updated the pull request. You must reimport the pull request before landing. |
|
@scotts has updated the pull request. You must reimport the pull request before landing. |
|
|
||
| jobs: | ||
| pr-test: | ||
| runs-on: linux.rocm.gpu.gfx942 |
There was a problem hiding this comment.
Should be linux.rocm.gpu.gfx942.1 if you want a single-GPU runner
There was a problem hiding this comment.
@jithunnair-amd, will that enable the job to run? It hasn't been successfully scheduled yet.
|
@scotts has updated the pull request. You must reimport the pull request before landing. |
|
@scotts has updated the pull request. You must reimport the pull request before landing. |
|
@scotts has updated the pull request. You must reimport the pull request before landing. |
|
@jithunnair-amd, my workflow is now running, but I haven't been able to successfully pull a ROCm image. Can you advise me on what image to pull and through what mechanisms? |
|
|
||
| # This is the docker image that pytorch/pytorch uses for ROCM CI. | ||
| env: | ||
| DOCKER_IMAGE: "ghcr.io/pytorch/ci-image:pytorch-linux-noble-rocm-n-py3" |
There was a problem hiding this comment.
@scotts If we want to use the same image that's used by PyTorch CI for ROCm, there's a sha suffix that is also part of the tag e.g. pytorch-linux-noble-rocm-n-py3-f2bf68fc1a48c33b01facc4b2e0227aa3da827e0. However, this is ever-changing, and is not maintainable. Alternatives:
- Work with @atalman to update the publishing workflow, so that every time it pushes out a new image for
pytorch-linux-noble-rocm-n-py3-*tag, it also publishes it aspytorch-linux-noble-rocm-n-py3-latesttag or something. - Use a different image such as
pytorch/manylinux2_28-builder:rocm7.2, which will still be a maintenance burden to update the ROCm versions, but with less frequency than the SHA updates in the CI image. That part can again be addressed by automatically publishing the same image with a tagpytorch/manylinux2_28-builder:rocm-latest. Also, this image usually updates ROCm version before the CI image, so you might have potential days of mismatch in the ROCm version between pytorch CI and kineto CI, if that matters.
There was a problem hiding this comment.
@jithunnair-amd, thanks for taking a look!
Since my short-term goal is to just get some tests up and running to check regressions, I'm happy doing the easiest thing for now. I tried pulling pytorch/manylinux2_28-builder:rocm7.2, but I'm getting these failures: https://github.com/pytorch/kineto/actions/runs/21452690799/job/61785302020?pr=1231
It looks like I don't have permissions. Is this because this is running in pytorch/kineto and not pytorch/pytorch? There's also some aws: command not found errors, but these same patterns work fine for the CPU and CUDA workflows. Any help is greatly appreciated!
There was a problem hiding this comment.
It's because it's assuming the image is on ghcr.io, whereas it's actually on dockerhub...
The "aws: command not found" errors can be ignored, but you'll probably need an AWS login via OIDC step to access any S3 buckets as part of the workflow e.g. https://github.com/pytorch/pytorch/blob/2b610902b46d0cbbb64923544f37feec0139cdfa/.github/actions/setup-rocm/action.yml#L115
There was a problem hiding this comment.
You'll also need to add a permissions section to your workflow with id-token: write for the OIDC step to work successfully e.g.
https://github.com/pytorch/pytorch/blob/2b610902b46d0cbbb64923544f37feec0139cdfa/.github/workflows/_rocm-test.yml#L73
There was a problem hiding this comment.
I see you're already calling setup-rocm in your workflow, so I think it'll try to do the AWS OIDC login step as well. But you'll probably still need the permissions clause.
|
@scotts has updated the pull request. You must reimport the pull request before landing. |
|
@scotts has updated the pull request. You must reimport the pull request before landing. |
|
@scotts has updated the pull request. You must reimport the pull request before landing. |
|
@scotts has updated the pull request. You must reimport the pull request before landing. |
| on: | ||
| push: | ||
| tags: | ||
| - ciflow/rocm/* |
There was a problem hiding this comment.
@scotts A few more steps to get this system working:
- Create a
ciflow/rocmlabel on kineto repo (to be used on PRs) - Add pytorch-probot.yaml to kineto repo like https://github.com/meta-pytorch/monarch/blob/main/.github/pytorch-probot.yml
There was a problem hiding this comment.
I can't add labels on Kineto PRs. Please try adding the ciflow/rocm label and we can see if it works as expected
|
@scotts has updated the pull request. You must reimport the pull request before landing. |
| - ciflow/rocm/* | ||
| branches: | ||
| - main | ||
| pull_request: |
There was a problem hiding this comment.
Lets remove pull_request and try to land this
|
@scotts has updated the pull request. You must reimport the pull request before landing. |
Summary: Depends on pytorch/kineto#1222. Pull Request resolved: pytorch/kineto#1231 Reviewed By: atalman Differential Revision: D91166410 Pulled By: scotts fbshipit-source-id: e237b9c22815d9725bf4aa548085f73098939658 Signed-off-by: Antoni Viros i Martin <aviros@ibm.com>
Depends on #1222.