ContextQMD
Back to Pytorch

Releasing PyTorch

RELEASE.md

2.11.033.0 KB
Original Source

Releasing PyTorch

<!-- toc --> <!-- tocstop -->

Release Compatibility Matrix

Following is the Release Compatibility Matrix for PyTorch releases:

PyTorch versionPythonC++Stable CUDAExperimental CUDAStable ROCm
2.10>=3.10, <=(3.14, 3.14t experimental)C++17CUDA 12.6 (CUDNN 9.10.2.21), CUDA 12.8 (CUDNN 9.10.2.21)CUDA 13.0 (CUDNN 9.15.1.9)ROCm 7.1
2.9>=3.10, <=(3.14, 3.14t experimental)C++17CUDA 12.6 (CUDNN 9.10.2.21), CUDA 12.8 (CUDNN 9.10.2.21)CUDA 13.0 (CUDNN 9.13.0.50)ROCm 6.4
2.8>=3.9, <=3.13, (3.13t experimental)C++17CUDA 12.6 (CUDNN 9.10.2.21), CUDA 12.8 (CUDNN 9.10.2.21)CUDA 12.9 (CUDNN 9.10.2.21)ROCm 6.4
2.7>=3.9, <=3.13, (3.13t experimental)C++17CUDA 11.8 (CUDNN 9.1.0.70), CUDA 12.6 (CUDNN 9.5.1.17)CUDA 12.8 (CUDNN 9.7.1.26)ROCm 6.3
2.6>=3.9, <=3.13, (3.13t experimental)C++17CUDA 11.8, CUDA 12.4 (CUDNN 9.1.0.70)CUDA 12.6 (CUDNN 9.5.1.17)ROCm 6.2.4
2.5>=3.9, <=3.12, (3.13 experimental)C++17CUDA 11.8, CUDA 12.1, CUDA 12.4, CUDNN 9.1.0.70NoneROCm 6.2
2.4>=3.8, <=3.12C++17CUDA 11.8, CUDA 12.1, CUDNN 9.1.0.70CUDA 12.4, CUDNN 9.1.0.70ROCm 6.1
2.3>=3.8, <=3.11, (3.12 experimental)C++17CUDA 11.8, CUDNN 8.7.0.84CUDA 12.1, CUDNN 8.9.2.26ROCm 6.0
2.2>=3.8, <=3.11, (3.12 experimental)C++17CUDA 11.8, CUDNN 8.7.0.84CUDA 12.1, CUDNN 8.9.2.26ROCm 5.7
2.1>=3.8, <=3.11C++17CUDA 11.8, CUDNN 8.7.0.84CUDA 12.1, CUDNN 8.9.2.26ROCm 5.6
2.0>=3.8, <=3.11C++14CUDA 11.7, CUDNN 8.5.0.96CUDA 11.8, CUDNN 8.7.0.84ROCm 5.4
1.13>=3.7, <=3.10C++14CUDA 11.6, CUDNN 8.3.2.44CUDA 11.7, CUDNN 8.5.0.96ROCm 5.2
1.12>=3.7, <=3.10C++14CUDA 11.3, CUDNN 8.3.2.44CUDA 11.6, CUDNN 8.3.2.44ROCm 5.0

PyTorch CUDA Support Matrix

For Release 2.9 and 2.10 PyTorch Supports following CUDA Architectures:

CUDAarchitectures supported for Linux x86 and Windows buildsnotes
12.6.3Maxwell(5.0), Pascal(6.0), Volta(7.0), Turing(7.5), Ampere(8.0, 8.6), Hopper(9.0)
12.8.1Volta(7.0), Turing(7.5), Ampere(8.0, 8.6), Hopper(9.0), Blackwell(10.0, 12.0)
13.0.0Turing(7.5), Ampere(8.0, 8.6), Hopper(9.0), Blackwell(10.0, 12.0+PTX)+PTX available on linux builds only
CUDAarchitectures supported for Linux aarch64 builds
12.6.3Ampere(8.0), Hopper(9.0)
12.8.1Ampere(8.0), Hopper(9.0), Blackwell(10.0, 12.0)
13.0.0Ampere(8.0), Hopper(9.0), Blackwell(10.0, 11.0, 12.0+PTX)

Release Cadence

Following is the release cadence. All future dates below are tentative. For latest updates on the release schedule, please follow dev discuss. Please note: Patch Releases are optional.

Minor VersionRelease branch cutRelease dateFirst patch release dateSecond patch release date
2.1Aug 2023Oct 2023Nov 2023Dec 2023
2.2Dec 2023Jan 2024Feb 2024Mar 2024
2.3Mar 2024Apr 2024Jun 2024Not planned
2.4Jun 2024Jul 2024Sept 2024Not planned
2.5Sep 2024Oct 2024Nov 2024Not planned
2.6Dec 2024Jan 2025Not plannedNot planned
2.7Mar 2025Apr 2025Jun 2025Not planned
2.8Jun 2025Jul 2025(Aug 2025)Not planned
2.9Sept 2025Oct 2025(Nov 2025)Not planned
2.10Dec 2025Jan 2026Not plannedNot planned
2.1116 Feb 202618 Mar 2026(Apr 2026)Not planned
2.1213 Apr 202613 May 2026(Jun 2026)Not planned
2.138 Jun 20268 Jul 2026(Aug 2026)Not planned
2.143 Aug 20262 Sept 2026(Oct 2026)Not planned
2.1528 Sept 202628 Oct 2026(Nov 2026)Not planned
2.1623 Nov 202622 Dec 2026(Jan 2027)Not planned

General Overview

Releasing a new version of PyTorch generally entails 3 major steps:

  1. Cutting a release branch preparations
  2. Cutting a release branch and making release branch specific changes
  3. Drafting RCs (Release Candidates), and merging cherry picks
  4. Preparing and Creating Final Release Candidate
  5. Promoting Final RC to stable and performing release day tasks

Frequently Asked Questions

  • Q: What is a release branch cut ?

    • A: When bulk of the tracked features merged into the main branch, the primary release engineer starts the release process of cutting the release branch by creating a new git branch based off of the current main development branch of PyTorch. This allows PyTorch development flow on main to continue uninterrupted, while the release engineering team focuses on stabilizing the release branch in order to release a series of release candidates (RC). The activities in the release branch include both regression and performance testing as well as polishing new features and fixing release-specific bugs. In general, new features are not added to the release branch after it was created.

  • Q: What is a cherry-pick ?

    • A: A cherry pick is a process of propagating commits from the main into the release branch, utilizing git's built in cherry-pick feature. These commits are typically limited to small fixes or documentation updates to ensure that the release engineering team has sufficient time to complete a thorough round of testing on the release branch. To nominate a fix for cherry-picking, a separate pull request must be created against the respective release branch and then mentioned in the Release Tracker issue (example: https://github.com/pytorch/pytorch/issues/94937) following the template from the issue description. The comment nominating a particular cherry-pick for inclusion in the release should include the committed PR against main branch, the newly created cherry-pick PR, as well as the acceptance criteria for why the cherry-pick is needed in the first place. This process can be automated by using entering a comment @pytorchbot cherry-pick -c [reason] on the PR you wish to cherry-pick.

Cutting a release branch preparations

Following requirements need to be met prior to cutting a release branch:

  • Triton release branch must be created (e.g., release/3.6.x) and the Triton pin update PR must be landed (e.g., #168096) at least 1 week before the branch cut
  • Resolve all outstanding issues in the milestones that are feature work and release blocking (for example release 2.10 milestone). A report of outstanding cherry-picks can be produced by running the github-analytics-daily workflow
  • Validate that all new workflows have been created in the PyTorch and domain libraries included in the release. Validate it against all dimensions of release matrix, including operating systems (Linux, MacOS, Windows), Python versions as well as CPU architectures (x86 and arm) and accelerator versions (CUDA, ROCm, XPU).
  • All viable/strict jobs are green, which requires the following jobs to pass: pull, trunk, lint, linux-aarch64
  • All the nightly jobs for pytorch and domain libraries should be green. Validate this using the following HUD links:

Cutting release branches

pytorch/pytorch

Release branches are typically cut from the branch viable/strict as to ensure that tests are passing on the release branch.

There's a convenience script to create release branches from current viable/strict. Perform following actions :

  • Perform a fresh clone of pytorch repo using
bash
git clone [email protected]:pytorch/pytorch.git
  • Execute following command from PyTorch repository root folder:
bash
DRY_RUN=disabled scripts/release/cut-release-branch.sh

This script should create 2 branches:

  • release/{MAJOR}.{MINOR}
  • orig/release/{MAJOR}.{MINOR}

PyTorch ecosystem libraries

Note: Release branches for individual ecosystem libraries should be created after first release candidate build of PyTorch is available in staging channels (which happens about a week after PyTorch release branch has been created). This is absolutely required to allow sufficient testing time for each of the domain library. Domain libraries branch cut is performed by Ecosystem Library POC. Test-Infra branch cut should be performed at the same time as PyTorch core branch cut. Convenience script can also be used for domains.

NOTE: RELEASE_VERSION only needs to be specified if version.txt is not available in root directory

bash
DRY_RUN=disabled GIT_BRANCH_TO_CUT_FROM=main RELEASE_VERSION=1.11 scripts/release/cut-release-branch.sh

Making release branch specific changes for PyTorch

First you should cut a release branch for pytorch/test-infra:

  • Create a new branch using the naming convention release/[major].[minor], e.g. release/2.7
  • On that release branch, update branch pointers for any pytorch-managed reusable actions or workflows to point to the new release's branch (example).

Here are examples of changes that should be made to the pytorch/pytorch release branches so that CI / tooling can function normally on them:

These are examples of changes that should be made to the default branch after a release branch is cut

Making release branch specific changes for ecosystem libraries

Ecosystem libraries branch cut is done a few days after branch cut for the pytorch/pytorch. The branch cut is performed by the Ecosystem Library POC. After the branch cut is performed, the PyTorch Dev Infra member should be informed of the branch cut and Domain Library specific change is required before Drafting RC for this domain library.

Follow these examples of PR that updates the version and sets RC Candidate upload channel:

The CI workflow updating part of the above PRs can be automated by running: python release/apply-release-changes.py [version] (where version is something like '2.7'). That script lives in both pytorch/audio and pytorch/vision.

Running Launch Execution team Core XFN sync

The series of meetings for Core XFN sync should be organized. The goal of these meetings are the following:

  1. Establish release POC's from each of the workstreams
  2. Cover the tactical phase of releasing minor releases to the market
  3. Discuss possible release blockers

Following POC's should be assigned from each of the workstreams:

  • Core/Marketing
  • Release Eng
  • Doc Eng
  • Release notes
  • Partner

NOTE: The meetings should start after the release branch is created and should continue until the week of the release.

Drafting RCs (Release Candidates) for PyTorch and domain libraries

To draft RCs, a user with the necessary permissions can push a git tag to the main pytorch/pytorch git repository. Please note: exactly same process is used for each of the domain library

The git tag for a release candidate must follow the following format:

v{MAJOR}.{MINOR}.{PATCH}-rc{RC_NUMBER}

An example of this would look like:

v1.12.0-rc1

You can use following commands to perform tag from pytorch core repo (not fork):

  • Checkout and validate the repo history before tagging
git checkout release/1.12
git log --oneline
  • Perform tag and push it to github (this will trigger the binary release build)
git tag -f  v1.12.0-rc2
git push origin  v1.12.0-rc2

Pushing a release candidate tag should trigger the binary_build workflows. This trigger functionality is configured in [linux_binary_build_workflow.yml.j2]](https://github.com/pytorch/pytorch/blob/main/.github/pytorch-circleci-labels.yml) and in the matching templates for the other OSes.

To view the state of the release build, please navigate to HUD. And make sure all binary builds are successful.

Release Candidate Storage

Release candidates are currently stored in the following places:

Backups are stored in a non-public S3 bucket at s3://pytorch-backup

Release Candidate health validation

Validate that the release jobs for pytorch and domain libraries are green. Validate this using the following HUD links:

Validate that the documentation build has completed and generated an entry corresponding to the release in the docs repository.

Cherry Picking Fixes

Typically, within a release cycle fixes are necessary for regressions, test fixes, etc.

For fixes that are to go into a release after the release branch has been cut we typically employ the use of a cherry pick tracker.

An example of this would look like:

Please also make sure to add milestone target to the PR/issue, especially if it needs to be considered for inclusion into the dot release.

NOTE: The cherry pick process is not an invitation to add new features, it is mainly there to fix regressions

How to do Cherry Picking

You can now use pytorchbot to cherry pick a PyTorch PR that has been committed to the main branch using @pytorchbot cherry-pick command as follows (make sure that the cherry-pick tracker issue for the target release labelled as "release tracker" - this will allow the bot to find it and post comments).

usage: @pytorchbot cherry-pick --onto ONTO [--fixes FIXES] -c
                               {regression,critical,fixnewfeature,docs,release}

Cherry pick a pull request onto a release branch for inclusion in a release

optional arguments:
  --onto ONTO           Branch you would like to cherry pick onto (Example: release/2.2)
  --fixes FIXES         Link to the issue that your PR fixes (i.e. https://github.com/pytorch/pytorch/issues/110666)
  -c {regression,critical,fixnewfeature,docs,release}
                        A machine-friendly classification of the cherry-pick reason.

For example, #120567 created a cherry pick PR #121232 onto release/2.2 branch to fix a regression issue. You can then refer to the original and the cherry-picked PRs on the release tracker issue. Please note that the cherry-picked PR will still need to be reviewed by PyTorch RelEng team before it can go into the release branch. This feature requires pytorchbot, so it's only available in PyTorch atm.

Cherry Picking Reverts

If a PR that has been cherry-picked into the release branch has been reverted, its cherry-pick must be reverted as well.

Reverts for changes that were committed into the main branch prior to the branch cut must be propagated into the release branch as well.

Preparing and Creating Final Release Candidate

The following requirements need to be met prior to creating the final Release Candidate:

  • Resolve all outstanding open issues in the milestone. There should be no open issues/PRs (for example 2.1.2). Each issue should either be closed or de-milestoned.

  • Validate that all closed milestone PRs are present in the release branch. Confirm this by running: python github_analyze.py --repo-path ~/local/pytorch --remote upstream --branch release/2.2 --milestone-id 40 --missing-in-branch

  • No outstanding cherry-picks that need to be reviewed in the issue tracker: https://github.com/pytorch/pytorch/issues/115300

  • Perform Release Candidate health validation. CI should have the green signal.

After the final RC is created, the following tasks should be performed:

Promoting RCs to Stable

Promotion of RCs to stable is done with this script: pytorch/test-infra:release/promote.sh

Users of that script should take care to update the versions necessary for the specific packages you are attempting to promote.

Promotion should occur in two steps:

  • Promote S3 artifacts (wheels, libtorch) and Conda packages
  • Promote S3 wheels to PyPI

NOTE: The promotion of wheels to PyPI can only be done once so take caution when attempting to promote wheels to PyPI, (see https://github.com/pypi/warehouse/issues/726 for a discussion on potential draft releases within PyPI)

Additional Steps to prepare for release day

The following should be prepared for the release day:

Modify release matrix

Modify the release matrix for the get started page. See the following PR as reference.

The PR to update published_versions.json and quick-start-module.js is auto generated. See the following PR as reference.

Please note: This PR needs to be merged on the release day and hence it should be absolutely free of any failures. To test this PR, open another test PR pointing to the Release Candidate location as described in the Release Candidate Storage section.

Open Google Colab issue

This is normally done right after the release is completed. We need to create a Google Colab issue. See the following example issue

Patch Releases

A patch release is a maintenance release of PyTorch that includes fixes for regressions found in a previous minor release. Patch releases typically will bump the patch version from semver (i.e. [major].[minor].[patch]).

Please note: Starting from 2.1, one can expect up to 2 patch releases after every minor release. Patch releases are only published for the latest minor release.

Patch Release Criteria

Patch releases should be considered if a regression meets the following criteria:

  1. Does the regression break core functionality (stable / beta features) including functionality in first party domain libraries?
  2. Is there not a viable workaround?
    • Can the regression be solved simply or is it not overcomable?

NOTE: Patch releases should only be considered when functionality is broken, documentation does not typically fall within this category

Patch Release Process

Patch Release Process Description

Main POC: Patch Release Managers, Triage Reviewers

Patch releases should follow these high-level phases. This process starts immediately after the previous release has completed. The patch release process takes around 4-5 weeks to complete.

  1. Triage is a process where issues are identified, graded, compared to Patch Release Criteria and added to Patch Release milestone. This process normally takes 2 weeks after the release completion.
  2. Go/No Go meeting between PyTorch Releng, PyTorch Core and Project Managers where potential issues triggering a release in milestones are reviewed, and following decisions are made:
  • Should the new patch release be created?
  • Timeline execution for the patch release
  1. Cherry picking phase starts after the decision is made to create a patch release. At this point, a new release tracker for the patch release is created, and an announcement will be made on official channels example announcement. The authors of the fixes to regressions will be asked to create their own cherry picks. This process normally takes 2 weeks.
  2. Updating version.txt in the release branch to match expected patch release version, see https://github.com/pytorch/pytorch/commit/f77213d3dae5d103a39cdaf93f21863843571e8d as an example
  3. Building Binaries, Promotion to Stable and testing. After all cherry picks have been merged, Release Managers trigger a new build and produce a new release candidate. An announcement is made on the official channel about the RC availability at this point. This process normally takes 2 weeks.
  4. General Availability

Triage

Main POC: Triage Reviewers

  1. Tag issues/pull requests that are candidates for a potential patch release with triage review
  2. Triage reviewers will then check if the regression/fix identified fits within the above mentioned Patch Release Criteria
  3. Triage reviewers will then add the issue/pull request to the related milestone (i.e. 1.9.1) if the regression is found to be within the Patch Release Criteria

Issue Tracker for Patch releases

For patch releases, an issue tracker needs to be created. For a patch release, we require all cherry-pick changes to have links to either a high-priority GitHub issue or a CI failure from previous RC. An example of this would look like:

Only following issues are accepted:

  1. Fixes to regressions against previous major version (e.g. regressions introduced in 1.13.0 from 1.12.0 are pickable for 1.13.1)
  2. Low risk critical fixes for: silent correctness, backwards compatibility, crashes, deadlocks, (large) memory leaks
  3. Fixes to new features being introduced in this release
  4. Documentation improvements
  5. Release branch specific changes (e.g. blocking ci fixes, change version identifiers)

Building a release schedule / cherry picking

Main POC: Patch Release Managers

  1. After regressions / fixes have been triaged Patch Release Managers will work together and build /announce a schedule for the patch release
    • NOTE: Ideally this should be ~2-3 weeks after a regression has been identified to allow other regressions to be identified
  2. Patch Release Managers will work with the authors of the regressions / fixes to cherry pick their change into the related release branch (i.e. release/1.9 for 1.9.1)
    • NOTE: Patch release managers should notify authors of the regressions to post a cherry picks for their changes. It is up to authors of the regressions to post a cherry pick. If cherry pick is not posted the issue will not be included in the release.
  3. If cherry picking deadline is missed by cherry pick author, patch release managers will not accept any requests after the fact.

Building Binaries / Promotion to Stable

Main POC: Patch Release managers

  1. Patch Release Managers will follow the process of Drafting RCs (Release Candidates)
  2. Patch Release Managers will follow the process of Promoting RCs to Stable

Hardware / Software Support in Binary Build Matrix

PyTorch has a support matrix across a couple of different axis. This section should be used as a decision making framework to drive hardware / software support decisions

Python

PyTorch supports all minor versions of CPython that are not EOL: https://devguide.python.org/versions/

For each minor release independently, we only support patch releases as follows:

  • If the latest patch release is a bugfix release, we only support this one.
  • Otherwise, we support all the non-bugfix patch releases.

See https://github.com/pytorch/rfcs/blob/master/RFC-0038-cpython-support.md for details on the rules and process for upgrade and sunset of each version.

Accelerator Software

For accelerator software like CUDA and ROCm we will typically use the following criteria:

  • Support latest 2 minor versions

Special support cases

In some instances support for a particular version of software will continue if a need is found. For example, our CUDA 11 binaries do not currently meet the size restrictions for publishing on PyPI so the default version that is published to PyPI is CUDA 10.2.

These special support cases will be handled on a case by case basis and support may be continued if current PyTorch maintainers feel as though there may still be a need to support these particular versions of software.

Operating Systems

Supported OS flavors are summarized in the table below:

Operating System familyArchitectureNotes
Linuxaarch64, x86_64Wheels are manylinux2014 compatible, i.e. they should be runnable on any Linux system with glibc-2.17 or above.
MacOSarm64Builds should be compatible with MacOS 11 (Big Sur) or newer, but are actively tested against MacOS 14 (Sonoma). MPS support is enabled on MacOS 13 (Ventura) or later.
Windowsx86_64Builds are compatible with Windows-10 or newer.

Submitting Tutorials

Tutorials in support of a release feature must be submitted to the pytorch/tutorials repo at least two weeks before the release date to allow for editorial and technical review. There is no cherry-pick process for tutorials. All tutorials will be merged around the release day and published at pytorch.org/tutorials.

Special Topics

Updating submodules for a release

In the event a submodule cannot be fast forwarded, and a patch must be applied we can take two different approaches:

  • (preferred) Fork the said repository under the pytorch GitHub organization, apply the patches we need there, and then switch our submodule to accept our fork.
  • Get the dependencies maintainers to support a release branch for us

Editing submodule remotes can be easily done with: (running from the root of the git repository)

git config --file=.gitmodules -e

An example of this process can be found here:

Triton dependency for the release

In nightly builds for conda and wheels pytorch depend on Triton build by this workflow: https://hud.pytorch.org/hud/pytorch/pytorch/nightly/1?per_page=50&name_filter=Build%20Triton%20Wheel. The pinned version of triton used by this workflow is specified here: https://github.com/pytorch/pytorch/blob/main/.ci/docker/ci_commit_pins/triton.txt .

In Nightly builds we have following configuration:

However for release we have following :

Important: The release of https://pypi.org/project/triton/ needs to be requested from OpenAI once branch cut is completed. Please include the release PIN hash in the request: https://github.com/pytorch/pytorch/blob/release/2.1/.ci/docker/ci_commit_pins/triton.txt .