Isolate Poetry validation into its own release-only job

[atalman]

Summary

The Poetry-based validation was sourced inline inside the main manywheel matrix job in validate-linux-binaries.yml, gated by a 5-condition check, and Poetry itself was installed from its dev branch (poetry@main). When Poetry's main branch regressed, the whole binary-validation job went red for the stable-CUDA Python 3.11 combination, masking the actual binary signal. See this failed run for the most recent example.

Three coordinated changes:

1. Split Poetry into a standalone linux-poetry job

In .github/workflows/validate-linux-binaries.yml, the inline source validate_poetry.sh block (with its 5-condition gate) is removed from the linux: matrix job. A new linux-poetry: top-level job is added, modeled on the existing linux-amazon-2023: job:

linux-poetry:
  if: ${{ inputs.channel == 'release' }}
  uses: ./.github/workflows/linux_job_v2.yml
  name: poetry-test
  with:
    runner: "linux.g5.4xlarge.nvidia.gpu"
    repository: "pytorch/pytorch"
    ref: main
    job-name: "poetry-test"
    docker-image: 'pytorch/almalinux-builder:cpu-main'
    docker-build-dir: "skip-docker-build"
    timeout: 60

It runs only on inputs.channel == 'release' and exports a fixed Python 3.11 / stable-CUDA combination, matching what the old inline gate selected.

2. Pin Poetry to a released version

In .github/scripts/validate_poetry.sh:

-curl -sSL https://install.python-poetry.org | python3 - --git https://github.com/python-poetry/poetry.git@main
+POETRY_VERSION="${POETRY_VERSION:-2.4.1}"
+curl -sSL https://install.python-poetry.org | python3 - --version "${POETRY_VERSION}"

Installing from poetry@main has caused recurring CI flakes whenever Poetry's dev branch regresses. POETRY_VERSION is env-overridable; bump deliberately to upgrade.

3. Surface Poetry's actual error

poetry --quiet add ... silenced resolution errors (today's exit 1 had no further output). Replaced with poetry add --no-interaction ... so resolution errors print but Poetry still doesn't prompt.

Test plan

• bash -n .github/scripts/validate_poetry.sh passes.
• The new linux-poetry: job mirrors the existing linux-amazon-2023: job structure; both reuse linux_job_v2.yml with the same input shape.
• Verify via dispatch on the release channel that the new linux-poetry job appears and runs; verify on nightly / test channels that it is skipped.
• Verify that the next validate-linux-binaries run is no longer red when Poetry's main is broken.

Effect on vllm#40077-style flakes

Before	After
Poetry-main breakage → red manywheel-py3_11-cuda13_0 job → looks like a binary regression	Poetry-main breakage → red poetry-test job only; binary validation is unaffected
--quiet hides the real error	--no-interaction keeps prompts off but lets errors print
Poetry installed from @main	Poetry pinned to ${POETRY_VERSION:-2.4.1}
Poetry runs on every matrix entry matching 5 conditions	Poetry runs once, on channel == 'release' only

Authored with Claude Code.

[vercel]

@atalman is attempting to deploy a commit to the Meta Open Source Team on Vercel.

A member of the Team first needs to authorize it.

[huydhn]

LGTM!