Migrating a workflow from EC2 to OSDC

OSDC (= ARC = on-site data center, EKS-hosted self-hosted runners) replaces EC2 runners. Migration touches the workflow file and requires a few inputs to flow into the reusable linux-build.yml / linux-test.yml so the OSDC code paths (build-osdc, test-osdc) activate. EC2 paths (build, test) and OSDC paths gate on inputs.use-arc (!inputs.use-arc vs inputs.use-arc), so flipping that input swaps execution lanes.

This playbook is for workflows that call the reusable linux-build.yml / linux-test.yml (e.g. test-b200.yml, pull.yml, trunk.yml, tsan.yml, operator_microbenchmark.yml). Standalone OSDC workflows (raw ARC label + container: directive) follow a different pattern not covered here.

Decide: dial-up vs. 100% opt-in

The two pieces — runner_prefix and use-arc — must move together. Hardcoding one but driving the other off the determinator is a bug.

Migration steps (dial-up pattern)

Worked reference: PR #181544 / commit f156b7ddfd1 ("Migrate smoke test on B200 to OSDC"). The diff was 10 lines.

1. Make sure get-label-type opts into the ARC experiment

In the get-label-type job that calls _runner-determinator.yml, add:

check_experiments: arc,lf

Without this the determinator won't consider the ARC experiment and use-arc will always be false.

2. Plumb five inputs into the build job

On the call to ./.github/workflows/_linux-build.yml:

with:
  runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"  # likely already present
  ...existing inputs...
  ci-docker-hash: ${{ needs.get-label-type.outputs.ci-docker-hash }}
  use-arc: ${{ needs.get-label-type.outputs.use-arc == 'true' }}
  python-version: "3.10"     # match the docker-image-name's python
  compiler: gcc11            # match the docker-image-name's compiler
  cuda-version: "13.0"       # match the docker-image-name's cuda (or "cpu" for CPU)

ci-docker-hash is the git rev-parse HEAD:.ci/docker hash that the determinator computes. _linux-build.yml appends it to the ghcr.io image tag (ghcr.io/pytorch/<docker-image-name>-<hash>); without it the OSDC build pulls the unhashed tag, which may not exist or may resolve to a stale image. The build job must have needs: get-label-type for this expression to resolve.

The last three (python-version / compiler / cuda-version) feed setup-linux so it can configure the OSDC container env. Read them off the existing docker-image-name (e.g. pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11 → py3.10 / gcc11 / cuda13.0).

3. Add get-label-type to the test job's needs and plumb the use-arc + setup-linux inputs

On the call to ./.github/workflows/_linux-test.yml:

needs:
  - <existing-build-job>
  - get-label-type           # add this
with:
  ...existing inputs...
  use-arc: ${{ needs.get-label-type.outputs.use-arc == 'true' }}
  python-version: "3.10"
  compiler: gcc11
  cuda-version: "13.0"

The test job needs a direct needs: get-label-type dependency to read its outputs (it can't rely on transitive needs through the build job). ci-docker-hash is not an input to _linux-test.yml — the test job picks the right image up via needs.<build>.outputs.docker-image.

4. Drop the OSDC-incompatible inputs

If the EC2 path used inputs that don't apply on OSDC (e.g. aws-role-to-assume: for ECR pulls), remove them. OSDC has its own AWS role wired into setup-linux (arn:aws:iam::308535385114:role/arc) and pulls images from ghcr.io/pytorch, not ECR.

In #181544 this meant deleting:

aws-role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only

5. Verify

• gh workflow run or push a PR with the relevant ciflow/* tag.
• Look for the build-osdc / test-osdc jobs running (they're separate jobs in linux-build.yml / linux-test.yml, gated on inputs.use-arc). When use-arc is false, the original build / test jobs run instead.

Migration steps (100% opt-in pattern)

For a job that should always run on OSDC, hardcode runner_prefix and use-arc instead of pulling them from the determinator. The build job still needs ci-docker-hash from get-label-type, so keep needs: get-label-type:

needs: get-label-type
with:
  runner_prefix: "mt-"
  ci-docker-hash: ${{ needs.get-label-type.outputs.ci-docker-hash }}
  use-arc: true
  python-version: "3.10"
  compiler: gcc11
  cuda-version: "13.0"

Both jobs (build and test) need the use-arc + python-version / compiler / cuda-version inputs; only the build job needs ci-docker-hash. The build job's runner: stays as the EC2 label (e.g. linux.r7i.4xlarge); linux-build.yml's build-osdc job translates that to the right ARC runner via mapec2toarc.py.

Reference: operator_microbenchmark.yml jobs opmicrobenchmark-build-b200 / opmicrobenchmark-test-b200.

What the inputs do (mental model)

linux-build.yml and linux-test.yml each define two jobs:

• build / test — EC2 path. Runs directly on the runner host VM. Gated on !inputs.use-arc.
• build-osdc / test-osdc — OSDC path. Runs inside a container: from ghcr.io/pytorch/${docker-image-name}${ci-docker-hash ? '-' + ci-docker-hash : ''}. Gated on inputs.use-arc.

The OSDC jobs call setup-linux with use-arc: true and pass python-version / compiler / cuda-version so the container env matches the build environment. That's why the migration must pass these three inputs — setup-linux errors out without them on the OSDC path. ci-docker-hash is what pins the ghcr.io image tag to a specific .ci/docker tree state so the build doesn't drift onto an unhashed (or wrong) image.

Common gotchas

• Forgetting check_experiments: arc,lf → use-arc is always false, so OSDC never activates and you'll think the migration silently failed.
• Forgetting needs: get-label-type on the test job → ${{ needs.get-label-type.outputs.use-arc }} evaluates to empty, dial-up doesn't engage.
• Forgetting ci-docker-hash on the build job → OSDC pulls the unhashed ghcr.io tag (<docker-image-name> with no -<hash> suffix), which may not exist or may resolve to a stale image. The failure mode is a manifest unknown pull error in the OSDC container step.
• **Plumbing ci-docker-hash from needs.get-label-type.outputs.* without adding needs: get-label-type** → the expression resolves to empty, so it's a silent no-op equivalent to not setting it at all.
• Mismatched python-version / compiler / cuda-version vs. docker-image-name → container has one toolchain, setup-linux configures another, build fails confusingly.
• Leaving aws-role-to-assume: for ECR → harmless on OSDC (it's only used by EC2 path) but stale and misleading; remove it.
• Mixing patterns (dynamic runner_prefix + hardcoded use-arc: true, or vice versa). Pick one pattern and apply it consistently.