release-update-docker-image-pin

작성자: pytorch

(또는 재고정) PyTorch/pytorch의 nightly/release 바이너리 빌드 워크플로우에서 사용되는 Linux manywheel 빌더 도커 이미지를 고정된 .ci/docker 빌드로…

npx skills add https://github.com/pytorch/test-infra --skill release-update-docker-image-pin

Release: Update Docker Image Pin

Pins the Linux manywheel builder docker images that the binary build workflows in pytorch/pytorch use, so a release branch builds against a fixed, reproducible toolchain instead of the floating tags that main tracks (e.g. pytorch/manylinux2_28-builder:cuda12.6). The pin lives in the workflow generator, so it survives regenerate.sh and the lint "generated files are up to date" check.

This skill operates on a pytorch/pytorch checkout (usually a release/X.Y branch), even though the skill itself lives in test-infra.

Inputs

InputRequiredExampleNotes
Release versionyes2.13The release/X.Y being pinned. Drives RELEASE_VERSION_TAG when regenerating and is used to sanity-check the checkout.
pytorch/pytorch pathonly if ambiguous~/pytorchA checkout with the matching release/X.Y branch checked out.

The release version is the primary argument. If it was not supplied when the skill was invoked, ask for it before making any changes -- do not guess. Confirm it matches the checkout with cut -d'.' -f1-2 version.txt.

When to use this skill

Use when the user asks to:

  • Pin / re-pin the manywheel (builder) docker images for a release
  • Update or bump DOCKER_IMAGE_PIN
  • Freeze the docker image the nightly/release binaries build with
  • Refresh the docker image pin after the release branch's .ci/docker changed

Background: how the pin maps to a published image

When the builder images are built (.github/workflows/build-manywheel-images.yml and build-manywheel-images-s390x.yml via the binary-docker-build action), each image is pushed with several tags, including:

docker.io/pytorch/<image>:<prefix>-${CI_FOLDER_SHA}

where CI_FOLDER_SHA="$(git rev-parse HEAD:.ci/docker)" -- the git tree hash of the .ci/docker directory. This is exactly the value test-infra/.github/actions/calculate-docker-image computes at runtime. Pinning simply freezes that hash as a literal so the release stops tracking floating tags. Example published image:

pytorch/manylinux2_28_aarch64-builder:cpu-aarch64-<CI_FOLDER_SHA>

Only linux manywheel builds run inside these containers, so only the linux images are pinned. Windows and macOS keep the plain tag prefix.

Target

Repo: pytorch/pytorch (the release branch you are pinning).

PathRole
.github/scripts/generate_binary_build_matrix.pyHolds DOCKER_IMAGE_PIN; the only file you edit by hand
.github/workflows/generated-linux-binary-manywheel-nightly.ymlRegenerated output (x86)
.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.ymlRegenerated output (aarch64)
.github/workflows/generated-linux-s390x-binary-manywheel-nightly.ymlRegenerated output (s390x)

Instructions

Step 1: Resolve inputs

Resolve the Release version and pytorch/pytorch path from the Inputs table above. If the release version was not provided, ask for it now. Confirm the checkout is on the matching branch and the version agrees:

cut -d'.' -f1-2 version.txt   # must equal the release version, e.g. 2.13
git rev-parse --abbrev-ref HEAD   # should be release/<version>

Step 2: Compute the pin hash

From the root of the pytorch/pytorch checkout, on the release branch:

git rev-parse HEAD:.ci/docker

This 40-char hash is the new DOCKER_IMAGE_PIN. It is the same value the builder images are published under. Do NOT invent or hand-edit it.

If the user instead points at a specific published image tag (e.g. from a known good nightly on https://hud.pytorch.org), use the suffix from that tag and verify it equals the git rev-parse output above; they should match on a fresh release cut.

Step 3 (optional): Verify the image exists

If the user wants confirmation, check that the tag exists on Docker Hub (a human can open the URL; automated egress may be restricted):

https://hub.docker.com/r/pytorch/manylinux2_28_aarch64-builder/tags?name=cpu-aarch64-<HASH>

Step 4: Update or add DOCKER_IMAGE_PIN

Open .github/scripts/generate_binary_build_matrix.py.

If the pin block already exists (re-pinning), just replace the hash:

DOCKER_IMAGE_PIN = "<NEW_40_CHAR_HASH>"

If pinning for the first time on this release branch, add the block right after the WHEEL_CONTAINER_IMAGES dict, and route the wheel tag prefix through the helper. Add:

# RELEASE-ONLY: pin the manywheel builder images to a fixed build so the release
# uses a reproducible toolchain instead of main's floating tags. The suffix is
# the .ci/docker tree hash (`git rev-parse HEAD:.ci/docker`), i.e. the same tag
# .github/actions/binary-docker-build publishes. Only linux manywheel builds run
# inside these containers, so only those images are pinned.
DOCKER_IMAGE_PIN = "<NEW_40_CHAR_HASH>"
MANYWHEEL_OSES = ("linux", "linux-aarch64", "linux-s390x")


def wheel_container_image_tag_prefix(arch_version: str, os: str) -> str:
    tag_prefix = WHEEL_CONTAINER_IMAGES[arch_version].split(":")[1]
    if os in MANYWHEEL_OSES:
        return f"{tag_prefix}-{DOCKER_IMAGE_PIN}"
    return tag_prefix

Then replace BOTH occurrences of the inline tag-prefix computation in generate_wheels_matrix with a call to the helper:

                        "container_image_tag_prefix": wheel_container_image_tag_prefix(
                            arch_version, os
                        ),

(The original reads WHEEL_CONTAINER_IMAGES[arch_version].split(":")[1].)

Step 5: Regenerate the workflows

Run the generator in release mode from the repo root:

RELEASE_VERSION_TAG=<RELEASE_VERSION> python3 .github/scripts/generate_ci_workflows.py

(Equivalent: RELEASE_VERSION_TAG=<ver> ./.github/regenerate.sh.)

Step 6: Verify

# Only the three linux manywheel files should change:
git status --short .github/workflows/generated-*.yml

# Every linux builder image should carry the new hash; none floating:
grep -rhE "image: pytorch/manylinux" .github/workflows/generated-linux-*.yml | grep -v "<HASH>" || echo "none floating (good)"

# Windows/macOS must be unchanged (plain cpu / cuda12.6 prefixes):
grep -nE "docker_image_tag_prefix:" .github/workflows/generated-macos-arm64-binary-wheel-nightly.yml | grep -v '\${{'

Then confirm idempotency -- regenerating a second time must produce byte-identical output (this is what the lint up-to-date check enforces):

md5sum .github/workflows/generated-linux-*.yml > /tmp/a
RELEASE_VERSION_TAG=<ver> python3 .github/scripts/generate_ci_workflows.py
md5sum .github/workflows/generated-linux-*.yml > /tmp/b
diff /tmp/a /tmp/b && echo "idempotent"

Step 7: Lint and commit

Run lintrunner -a on the changed files, then commit. The commit message should record the hash and that it equals git rev-parse HEAD:.ci/docker. This is a release-only change; do not port it to main.

Common pitfalls

  • Wrong hash: always derive it with git rev-parse HEAD:.ci/docker on the release branch. A hand-typed or stale hash points at a nonexistent tag and every linux binary build will fail to pull its container.
  • Pinning windows/macOS: those builds do not run in these containers. Keep the pin scoped to MANYWHEEL_OSES; a pinned windows/macOS prefix is wrong.
  • Editing the generated YAML directly: the lint check regenerates and asserts no diff, so hand edits get reverted. Always change the generator.
  • Forgetting RELEASE_VERSION_TAG: without it the regenerated reusable workflow refs/version may differ from the committed release files.

Example usage

Pin the manywheel docker images for release/2.13
Re-pin the builder images, .ci/docker changed on the release branch

pytorch의 다른 스킬

zephyr
pytorch
임베디드 보드용 Zephyr RTOS 모듈로 ExecuTorch를 빌드하고 구성합니다. ET로 Zephyr 워크스페이스를 설정하거나 보드 지원(오버레이 등)을 추가할 때 사용합니다.
aoti-debug
pytorch
AOTInductor(AOTI) 오류 및 충돌을 디버깅합니다. AOTI 세그폴트, 장치 불일치 오류, 상수 로딩 실패 또는 런타임 오류가 발생할 때 사용하세요.
skill-writer
pytorch
Claude Code를 위한 잘 구조화된 Agent Skill 생성 가이드로, 모범 사례와 검증을 포함합니다. Skill의 전체 수명 주기(범위 설정, 파일 구조, YAML 프론트매터 검증, 콘텐츠 구성, 테스트 절차)를 다룹니다. 엄격한 명명 규칙(소문자, 하이픈, 최대 64자)과 설명 요구 사항(특정 트리거, 파일 유형, "무엇" 및 "언제" 절)을 적용합니다. 읽기 전용 Skill, 스크립트 기반 Skill, 다중 파일 Skill 등 일반적인 패턴에 대한 템플릿을 제공합니다.
triaging-issues
pytorch
GitHub 이슈를 분류하여 온콜 팀에 라우팅하고, 레이블을 적용하며, 질문을 종료합니다. 새로운 PyTorch 이슈를 처리하거나 이슈 분류를 요청받았을 때 사용하세요.
wheel-size-analyzer
pytorch
PyTorch nightly wheel 크기를 GitHub Actions 아티팩트 API를 사용하여 날짜 범위에 걸쳐 분석합니다. 바이너리 크기 변경 추적, wheel 크기 식별에 사용합니다…
release-go-live-binary-build-matrix
pytorch
tools/scripts/generate_binary_build_matrix.py를 PyTorch 릴리스가 라이브될 때 업데이트합니다. CURRENT_STABLE_VERSION을 새로운 안정 버전으로 올리고, 해당…
pr-review
pytorch
PyTorch 풀 리퀘스트의 코드 품질, 테스트 커버리지, 보안 및 하위 호환성을 검토합니다. PR을 검토할 때, 코드 변경 사항을 검토하도록 요청받았을 때 사용합니다.
qualcomm
pytorch
QNN(Qualcomm AI Engine Direct) 백엔드를 빌드, 테스트 또는 개발합니다. backends/qualcomm/에서 작업하거나 QNN을 빌드할 때 사용합니다(계속…).