Add Apex to the Pytorch Manifest file

[erman-gurses]

Motivation

Add Apex to the Pytorch Manifest file
Follow up PR for #2547

See the comment here: #2547 (comment)

Test Plan

Test on CI

Test Result

The test runs: https://github.com/ROCm/TheRock/actions/runs/22070943413

Example Manifest from the test:

{
  "pytorch": {
    "commit": "f466ea7ac710374ce7066d9ead92e57dc42dc76f",
    "repo": "https://github.com/ROCm/pytorch.git",
    "branch": "release/2.7"
  },
  "pytorch_audio": {
    "commit": "95c61b4168fc5133be8dd8c1337d929d066ae6cf",
    "repo": "https://github.com/pytorch/audio"
  },
  "pytorch_vision": {
    "commit": "59a3e1f9f78cfe44cb989877cc6f4ea77c8a75ca",
    "repo": "https://github.com/pytorch/vision"
  },
  "triton": {
    "commit": "95abd4daa02d789ed257178ba35b4ef2d746ca29",
    "repo": "https://github.com/ROCm/triton.git"
  },
  "apex": {
    "commit": "7a57becffa31a4394369152e978fa6f8a0f005c2",
    "repo": "https://github.com/ROCm/apex"
  },
  "therock": {
    "commit": "02024427089134e951fe4063100b93c28e7e01a8",
    "repo": "https://github.com/ROCm/TheRock.git",
    "branch": "users/erman-gurses/add-apex-to-pytorch-manifest"
  }
}

Link: https://therock-dev-artifacts.s3.us-east-2.amazonaws.com/22070943413-linux/manifests/gfx94X-dcgpu/therock-manifest_torch_py3.10_release-2.7.json

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[marbre]

@erman-gurses can you either file the sections in the PR description or remove the empty ones?

[erman-gurses]

@erman-gurses can you either file the sections in the PR description or remove the empty ones?

@marbre, I will fill out fully after I finish my tests then convert it from draft to a full PR.

[HereThereBeDragons]

just looking at the pytorch build failures we are having in nightlies: could we create the manifest anyway independent if the build succeeds or not? so we have an idea what exactly got checked out? even if it is not all components of torch?

[erman-gurses]

just looking at the pytorch build failures we are having in nightlies: could we create the manifest anyway independent if the build succeeds or not? so we have an idea what exactly got checked out? even if it is not all components of torch?

https://github.com/ROCm/TheRock/actions/runs/22070943413/job/63775043033
That is a good point. I generate the manifest file Generate PyTorch Manifest after after Build PyTorch Wheels so for now, we need to finish PyTorch Build successfully. However, this will change in the near future. It should generate the manifest file right after the checkouts. We have ongoing discussion for that.

[HereThereBeDragons]

lgtm

here some thoughts about resilience of this code:
is this a manifest to know what is in the build or mainly for us for debugging?

for debugging i would suggest the following:

• run this after the build
• run this always
• check which repos were checked out and add in the manifest if the build was successful (== wheel exists).

if we just want to have the manifest for the wheels that have successfully built: add repo/commit for failing/excluded wheels in the github output

context:
the apex problem on linux got resolved, so right now there is no problem other that apex is not build on windows. with the issue before, apex would have only been built for torch >= 2.10. this decision would have been done in build_prod_wheels.py and as such the apex repo checkout would always happen - but not the build. (see PR #3458 note the changes also needed in write_torch_versions.py)

[erman-gurses]

lgtm

here some thoughts about resilience of this code: is this a manifest to know what is in the build or mainly for us for debugging?

for debugging i would suggest the following:

• run this after the build
• run this always
• check which repos were checked out and add in the manifest if the build was successful (== wheel exists).

if we just want to have the manifest for the wheels that have successfully built: add repo/commit for failing/excluded wheels in the github output

context: the apex problem on linux got resolved, so right now there is no problem other that apex is not build on windows. with the issue before, apex would have only been built for torch >= 2.10. this decision would have been done in build_prod_wheels.py and as such the apex repo checkout would always happen - but not the build. (see PR #3458 note the changes also needed in write_torch_versions.py)

Thanks, this is helpful context.

At the moment the manifest is meant to capture source provenance (repos/commits used for the build), not strictly the set of produced wheels. That’s why it runs after the build but still reflects checked-out sources even if some components are conditionally skipped.

In this PR I only wired Apex into the Linux PyTorch workflow, since that’s where it’s currently built; Windows still skips Apex today.

I agree that if we want the manifest to reflect only shipped artifacts, we’d need to tie entries to actual wheel outputs. That could also let us report skipped components explicitly via workflow outputs, which would help debugging.

For now I’d prefer to keep this PR focused on consistent provenance capture, but this sounds like a reasonable follow-up improvement.

[erman-gurses]

@amd-sriram - let me know if you have any comment for the PR that I can address. Thank you!