Isolate Poetry validation into its own release-only job

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.

Three coordinated changes:

1. Split the Poetry test into a standalone linux-poetry job
   modeled on the existing linux-amazon-2023 job. The new job
   only runs on inputs.channel == 'release' and exports a fixed
   Python 3.11 / stable-CUDA combination, matching what the old
   inline gate selected.

2. Drop the inline `source validate_poetry.sh` block from the
   main matrix job so Poetry ecosystem churn can no longer flake
   binary validation.

3. In validate_poetry.sh, pin Poetry to a released version via
   POETRY_VERSION (default 2.4.1, env-overridable) instead of
   installing from poetry@main, and replace `--quiet` with
   `--no-interaction` so resolution errors are no longer
   silently swallowed.

Authored with Claude Code.

Author: atalman

.github/scripts/validate_poetry.sh

@@ -1,7 +1,11 @@
 
 conda create -y -n ${ENV_NAME}_poetry python=${MATRIX_PYTHON_VERSION} numpy ffmpeg
 conda activate ${ENV_NAME}_poetry
-curl -sSL https://install.python-poetry.org | python3 - --git https://github.com/python-poetry/poetry.git@main
+# Pin Poetry to a released version. Installing from @main has caused
+# recurring CI flakes whenever Poetry's dev branch regresses. Bump
+# POETRY_VERSION deliberately when you want to upgrade.
+POETRY_VERSION="${POETRY_VERSION:-2.4.1}"
+curl -sSL https://install.python-poetry.org | python3 - --version "${POETRY_VERSION}"
 export PATH="/root/.local/bin:$PATH"
 
 poetry --version
@@ -24,11 +28,11 @@ if [[ ! -z ${RELEASE_VERSION} ]]; then
 fi
 
 if [[ ${TORCH_ONLY} == 'true' ]]; then
-    poetry --quiet add torch${RELEASE_SUFFIX}
+    poetry add --no-interaction torch${RELEASE_SUFFIX}
 elif [[ ${INCLUDE_TORCHAUDIO:-} == 'true' ]]; then
-    poetry --quiet add torch${RELEASE_SUFFIX} torchaudio torchvision
+    poetry add --no-interaction torch${RELEASE_SUFFIX} torchaudio torchvision
 else
-    poetry --quiet add torch${RELEASE_SUFFIX} torchvision
+    poetry add --no-interaction torch${RELEASE_SUFFIX} torchvision
 fi
 
 pushd ${PWD}/../.ci/pytorch/

.github/workflows/validate-linux-binaries.yml

@@ -141,19 +141,42 @@ jobs:
         eval "$(conda shell.bash hook)"
         printf '%s\n' ${{ toJson(inputs.release-matrix) }} > release_matrix.json
 
-        CUDA_VERSION_STABLE=$(python3 ../../test-infra/tools/scripts/get_stable_cuda_version.py --channel ${MATRIX_CHANNEL})
-        # Special case PyPi installation package. And Install of PyPi package via poetry
-        if [[ ${MATRIX_PACKAGE_TYPE} == "manywheel" && \
-              ${MATRIX_GPU_ARCH_VERSION} == "${CUDA_VERSION_STABLE}" && \
-              ${MATRIX_CHANNEL} == "release" && \
-              ${USE_ONLY_DL_PYTORCH_ORG} == "false" && \
-              ${USE_WHEEL_VARIANTS} == "false" ]]; then
-          source ../../test-infra/.github/scripts/validate_poetry.sh
-        fi
-
-        # Validate binaries
+        # Validate binaries. Poetry-based validation now lives in the
+        # standalone linux-poetry job below so Poetry's own ecosystem
+        # churn (it was previously installed from poetry@main) cannot
+        # flake the binary matrix.
         source ../../test-infra/.github/scripts/validate_binaries.sh
 
+  linux-poetry:
+    if: ${{ inputs.channel == 'release' }}
+    uses: ./.github/workflows/linux_job_v2.yml
+    name: poetry-test
+    with:
+      runner: "linux.g5.4xlarge.nvidia.gpu"
+      repository: "pytorch/pytorch"
+      ref: main
+      job-name: "poetry-test"
+      docker-image: 'pytorch/almalinux-builder:cpu-main'
+      docker-build-dir: "skip-docker-build"
+      timeout: 60
+      script: |
+        set -ex
+        export ENV_NAME="conda-env-${{ github.run_id }}"
+        export TORCH_ONLY=${{ inputs.torchonly }}
+        export INCLUDE_TORCHAUDIO=${{ inputs.include-torchaudio }}
+        export RELEASE_VERSION=${{ inputs.version }}
+        export TARGET_OS="linux"
+        eval "$(conda shell.bash hook)"
+        printf '%s\n' ${{ toJson(inputs.release-matrix) }} > release_matrix.json
+
+        # Poetry test is release-channel-only and covers a single
+        # stable-CUDA / Python 3.11 path. The full validate-binaries
+        # matrix above already covers every other combination via pip.
+        export MATRIX_PYTHON_VERSION="3.11"
+        export MATRIX_GPU_ARCH_VERSION=$(python3 ../../test-infra/tools/scripts/get_stable_cuda_version.py --channel ${{ inputs.channel }})
+
+        source ../../test-infra/.github/scripts/validate_poetry.sh
+
   linux-amazon-2023:
     uses: ./.github/workflows/linux_job_v2.yml
     name: amazon-linux-2023-test