CUDA/CUDNN Upgrade Runbook

So you wanna upgrade PyTorch to support a new CUDA? Follow these steps in order! They are adapted from previous CUDA upgrade processes.

0. Before starting upgrade process

A. Currently Supported CUDA and CUDNN Libraries

Here is the supported matrix for CUDA and CUDNN (versions can be looked up in https://github.com/pytorch/pytorch/blob/main/.ci/docker/common/install_cudnn.sh)

CUDA	CUDNN	additional details
11.8.0	9.1.0.70	Legacy CUDA Release
12.6.3	9.5.1.17	Stable CUDA Release
12.8.0	9.7.1.26	Latest CUDA Release
	9.8.0.87	Latest CUDA Nightly

B. Check the package availability

Package availability to validate before starting upgrade process :

CUDA and CUDNN is available for Linux and Windows: https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_570.86.10_linux.run (x86) https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_570.86.10_linux_sbsa.run (aarch64) https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/
CUDA is available on Docker hub images : https://hub.docker.com/r/nvidia/cuda Following example is for cuda 12.4: https://gitlab.com/nvidia/container-images/cuda/-/tree/master/dist/12.4.0/ubuntu2204/devel?ref_type=heads (TODO: Update this for 12.8) (Make sure to use version without CUDNN, it should be installed separately by install script)
Validate new driver availability: https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html. Check following table: Table 3. CUDA Toolkit and Corresponding Driver Versions

1. Maintain Progress and Updates

Make an issue to track the progress, for example #56721: Support 11.3. This is especially important as many PyTorch external users are interested in CUDA upgrades.

2. Modify scripts to install the new CUDA for Manywheel Docker Linux containers.

There are two types of Docker containers we maintain in order to build Linux binaries: libtorch, and manywheel. They all require installing CUDA and then updating code references in respective build scripts/Dockerfiles. This step is about manywheel.

Follow this PR 145567 for all steps in this section
Find the CUDA install link here
Get the cudnn link from NVIDIA on the PyTorch Slack
Modify install_cuda.sh and install_cuda_aarch64.sh
Run the install_128 chunk of code on your devbox to make sure it works.
Modify build-manywheel-images.yml with the latest CUDA version 12.8 in this case.
To test that your code works, from the root builder repo, run something similar to export CUDA_VERSION=12.8 && .ci/docker/manywheel/build_scripts/build_docker.sh for the manywheel images.
Once the PR in step1 is merged, validate manylinux docker hub manylinux2_28-builder:cuda12.8 and manylinuxaarch64-builder:cuda12.8 to see that images have been built and correctly tagged. These images are used in the next step to build Magma for linux.

3. Update Magma for Linux

Build Magma for Linux. Our Linux CUDA docker images require magma build, so we need to build magma-cuda and push it to the ossci-linux s3 bucket:

The code to build Magma is in the pytorch/pytorch repo
Currently, this is mainly copy-paste in magma/Makefile if there are no major code API changes/deprecations to the CUDA version. Previously, we've needed to add patches to MAGMA, so this may be something to check with NVIDIA about.
To push the package, please update build-magma-linux workflow
NOTE: This step relies on the pytorch/manylinux2_28-builder:cuda${DESIRED_CUDA}-main image (changes to .github/workflows/build-manywheel-images.yml), so make sure you have pushed the new manywheel-builder prior.

4. Modify scripts to install the new CUDA for Libtorch Docker Linux containers. Modify builder supporting scripts

There are two types of Docker containers we maintain in order to build Linux binaries: libtorch, and manywheel. They all require installing CUDA and then updating code references in respective build scripts/Dockerfiles. This step is about libtorch containers.

Add setup for our Docker libtorch:

Follow this PR PR 145789 for all steps in this section. For libtorch, the code changes are usually copy-paste.
Merge the above the PR, and it should automatically push the images to Docker Hub with GitHub Actions. Make sure to update the cuda_version to the version you're adding in respective YAMLs, such as .github/workflows/build-libtorch-images.yml.
Verify that the workflow that pushes the images succeed by selecting and verifying them in the Actions page. Furthermore, check https://hub.docker.com/r/pytorch/libtorch-cxx11-builder/tags to verify that the right tags exist for libtorch types of images.

5. Generate new Windows AMI, test and deploy to canary and prod.

Please note, since this step currently requires access to corporate AWS, this step should be performed by Meta employee. To be removed, once automated. Also note that Windows AMI takes about a week to build, so start this step early.

For Windows you will need to rebuild the test AMI, please refer to this PR. After this is done, run the release of Windows AMI using this proecedure. As time of this writing this is manual steps performed on dev machine. Please note that packer, aws cli needs to be installed and configured!
After step 1 is complete and new Windows AMI have been deployed to AWS. We need to deploy the new AMI to our canary environment (https://github.com/pytorch/pytorch-canary) through https://github.com/fairinternal/pytorch-gha-infra example : PR . After this is completed Submit the code for all windows workflows to https://github.com/pytorch/pytorch-canary and make sure all test are passing for all CUDA versions.
After that we can deploy the Windows AMI out to prod using the same pytorch-gha-infra repository.

6. Modify code to install the new CUDA for Windows and update MAGMA for Windows

Follow this windows Magma and cuda build for cu128 for all steps in this section
To get the CUDA install link, just like with Linux, go here and upload that .exe file to our S3 bucket ossci-windows.
Review "Table 3. Possible Subpackage Names" of CUDA installation guide for windows link to make sure the Subpackage Names have not changed. These are specified in cuda_install.bat file
To get the cuDNN install link, you could ask NVIDIA, but you could also just sign up for an NVIDIA account and access the needed .zip file at this link. First click on cuDNN Library for Windows (x86) and then upload that zip file to our S3 bucket.
NOTE: When you upload files to S3, make sure to make these objects publicly readable so that our CI can access them!
If you have to upgrade the driver install for newer versions, which would look like updating the windows/internal/driver_update.bat file
1. Please check the CUDA Toolkit and Minimum Required Driver Version for CUDA minor version compatibility table in the release notes to see if a driver update is necessary.
Compile MAGMA with the new CUDA version. Update .github/workflows/build-magma-windows.yml to include new version.
Validate Magma builds by going to S3 ossci-windows. And querying for magma_

7. Add the new CUDA version to the nightly binaries matrix.

Adding the new version to nightlies allows PyTorch binaries compiled with the new CUDA version to be available to users through pip or just raw libtorch.

If the new CUDA version requires a new driver (see #1 sub-bullet), the CI and binaries would also need the new driver. Find the driver download here and update the link like so.
1. Please check the Driver Version table in the release notes to see if a driver update is necessary.
Follow this Add CUDA 12.8 manywheel x86 Builds to Binaries Matrix for steps 2-4 in this section.
Once PR 145792 is created make sure to attach ciflow/binaries, ciflow/nightly labels to this PR. And make sure all the new workflow with new CUDA version terminate successfully.
Testing nightly builds is done as follows:
- Make sure your commit to master passed all the test and there are no failures, otherwise the next step will not work
- Make sure your changes are promoted to viable/strict branch: https://github.com/pytorch/pytorch/tree/viable/strict . Run viable/strict promotion job to promote from master to viable/strict
- After your changes are promoted to viable/strict. Run nighly build job.
- Make sure your changes made to nightly branch https://github.com/pytorch/pytorch/tree/nightly
- Make sure all nightly build succeeded before continuing to Step #6
If Stable CUDA version changes, update latest tag for ghcr.io like so: https://github.com/pytorch/pytorch/blob/main/.github/scripts/generate_binary_build_matrix.py#L20

8. Add the new CUDA version to OSS CI.

Testing the new version in CI is crucial for finding regressions and should be done ASAP along with the next step (I am simply putting this one first as it is usually easier).

The configuration files will be subject to change, but usually you just have to replace an older CUDA version with the new version you're adding. Code reference for 12.6: PR 140793.
IMPORTANT NOTE: the CI is not always automatically triggered when you edit the workflow files! Ensure that the new CI job for the new CUDA version is showing up in the PR signal box. If it is not there, make sure you add the correct ciflow label (ciflow/periodic, for example) to trigger the test. Just because the CI is green on your pull request does NOT mean the test has been run and is green.
It is likely that there will be tests that no longer pass with the new CUDA version or GPU driver. Disable them for the time being, notify people who can help, and make issues to track them (like so).
After merging the CI PR, Please open temporary issues for new builds and tests marking them unstable, example issue. These issues should be closed after few days of opening, when newly added CI jobs are constantly green.

9. Update Linux Nvidia driver used during runner provisioning

If linux driver update is required. The driver should be updated during the runner provisioning otherwise nightly workflows will fail with multiple Nova workflows.

Post and merge PR 5243
Run workflow lambda-release-tag-runners workflow this worklow will create new release here
Post and merge PR 394
Deploy this change by running following workflow runners-on-dispatch-release

10. Add the new version to torchvision and torchaudio CI.

Torchvision and torchaudio is usually a dependency for installing PyTorch for most of our users. This is why it is important to also propagate the CI changes so that torchvision and torchaudio can be packaged for the new CUDA version as well.

Add a change to a binary build matrix in test-infra repo here
A code sample for torchvision: PR 7533
A code sample for torchaudio: PR 3284 You can combine all above three steps in one PR: [PR 6244] (https://github.com/pytorch/test-infra/pull/6244/files)
Almost every change in the above sample is copy-pasted from either itself or other existing parts of code in the builder repo. The difficulty again is not changing the config but rather verifying and debugging any failing builds.

This completes CUDA and CUDNN upgrade. Congrats! PyTorch now has support for a new CUDA version and you made it happen!

Upgrade CUDNN version only

If you require to update CUDNN version for already existing CUDA version, please perform the followin modifications.

Add new cudnn vesion to windows AMI: pytorch/test-infra#6290. Rebuild and retest the AMI. Follow step 6 Generate new Windows AMI, test and deploy to canary and prod.
Add new cudnn version to linux builds: https://github.com/pytorch/pytorch/pull/148963/files (including installation script and small wheel update)

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!