Disable pytorch sccache for PRs from forks#5738
Closed
ScottTodd wants to merge 1 commit into
Closed
Conversation
ScottTodd
commented
Jun 10, 2026
Comment on lines
163
to
+164
| - name: Configure AWS Credentials for sccache | ||
| if: ${{ inputs.cache_type == 'sccache' && github.repository_owner == 'ROCm' }} | ||
| if: ${{ inputs.cache_type == 'sccache' && github.repository_owner == 'ROCm' && (github.event_name != 'pull_request' || github.event.pull_request.head.repo.owner.login == 'ROCm') }} |
Member
Author
There was a problem hiding this comment.
I guess we can limit these pull_request changes to just the _ci.yml workflows.
This isn't working though, other steps fail: https://github.com/ROCm/TheRock/actions/runs/27240729872/job/80461488687?pr=5738#step:14:73
sccache: Starting the server...
sccache: error: Server startup failed: cache storage failed to read: PermissionDenied (permanent) at read => S3Error { code: "AccessDenied", message: "User: arn:aws:iam::692859939525:user/therock-external-upload is not authorized to perform: s3:GetObject on resource: \"arn:aws:s3:::therock-pytorch-sccache-ci/linux/multi-arch-release/.sccache_check\" because no identity-based policy allows the s3:GetObject action", resource: "", request_id: "RQ3NQ6XYPNWWZQAA" }
Could compute earlier in the workflow whether or not to enable sccache based on inputs.cache_type and other values. Really this all needs to go through scripts and not have so many available options handled via yml code (different cache types that may or may not work in different conditions).
4 tasks
Member
Author
|
Closing in favor of #5816 |
subodh-dubey-amd
added a commit
that referenced
this pull request
Jun 14, 2026
## Summary Fixes #5737 (supersedes the Draft #5738). PyTorch CI jobs fail on PRs from forks at the sccache steps: ``` User: arn:aws:iam::692859939525:user/therock-external-upload is not authorized to perform: sts:TagSession on resource: arn:aws:iam::324352301041:role/therock-ci ... AccessDenied ... s3:GetObject on therock-pytorch-sccache-ci/.../.sccache_check ``` ### Root cause sccache uses an S3 bucket reached via an OIDC-assumed IAM role (`therock-ci`). Fork PRs don't get OIDC tokens, so the role can't be assumed. The gate `github.repository_owner == 'ROCm'` is **always true** on fork PRs (it's the base repo), and it only guarded the credentials step — `Verify sccache` and the build still passed `--use-sccache`, so sccache started and failed on S3 access. ### Fix Compute the effective cache type **once**, in [build_tools/github_actions/compute_pytorch_cache_type.py](https://github.com/ROCm/TheRock/blob/main/build_tools/github_actions/compute_pytorch_cache_type.py) (reusing `_is_current_run_pr_from_fork()` from `s3_buckets.py`), downgrading `sccache` -> `none` on fork PRs / non-ROCm repos. Every sccache step in the two `*_pytorch_wheels_ci.yml` workflows now keys off that single `steps.cache.outputs.cache_type` instead of repeating fork checks. `ccache`/`none` pass through unchanged; in-org runs keep sccache and its hard-fail semantics. This follows the review direction on #5738 ("compute earlier… this all needs to go through scripts, not yml code") and limits changes to the `_ci.yml` workflows (the release / `multi_arch_build_*` workflows are never fork-triggered). ## Files - `build_tools/github_actions/compute_pytorch_cache_type.py` (new) - `build_tools/github_actions/tests/compute_pytorch_cache_type_test.py` (new, 8 cases) - `build_portable_linux_pytorch_wheels_ci.yml`, `build_windows_pytorch_wheels_ci.yml` — add "Determine cache type" step; gate all sccache steps on its output. ## Test plan - [x] Unit tests (fork / in-org / non-PR / non-ROCm / none / ccache). - [x] Real-script test against fork-shaped event payloads: fork+sccache->none, in-org+sccache->sccache, fork+ccache->ccache. - [ ] In-org CI run on this PR: sccache still active + cache hits (no regression). - [ ] Fork PR (#5729) after merge: sccache skipped, build succeeds with cache_type=none.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Motivation
Tentative fix for #5737, to unblock #5729.
Workflow runs from fork PRs outside of the ROCm organization are not authorized to assume the
therock-ciAWS IAM role, so PyTorch builds are failing on fork PRs with errors likeTechnical Details
I guess
github.repository_owner == 'ROCm'is true for PRs from forks. I thinkgithub.event.pull_request.head.repo.owner.login == 'ROCm'should be a more precise check.We should add base credentials to these runners that can access some cache like how artifacts can upload to
therock-ci-artifacts-external(see https://github.com/ROCm/TheRock/blob/main/docs/development/s3_buckets.md#ci-buckets).Test Plan
Watch to see if workflow runs on this PR successfully skip sccache setup and build pytorch.
Submission Checklist