PyTorch CUDA 13 packages + RAPIDS

[jameslamb]

Description

When we first rolled out CUDA 13 support across RAPIDS (#208), there was not yet a PyTorch release with CUDA 13 support, so some workarounds like skipping tests and relying on nightlies were used.

This tracks the follow-up work to remove those workarounds.

Benefits of this work

• ensures users have a good experience with RAPIDS and PyTorch using CUDA 13

Acceptance Criteria

• RAPIDS CUDA 13 CI builds, project dependencies, etc. rely on actual releases of PyTorch, not nightlies

Approach

Subscribe to notifications on the main tracking issue: pytorch/pytorch#159779

Wheels

There are torch 2.9.x and 2.10.x nightly wheels with CUDA 13 support at: https://download.pytorch.org/whl/nightly/cu130

We are already using those nightlies in wheel tests for cugraph and cugraph-pyg . Once there's a new official release of torch with CUDA 13 wheels, update RAPIDS environments to use those releases and stop using nightlies.

Switch uses of nightly torch wheels to actual releases (GitHub search)

• cugraph (CUDA 13: use releases, not nightlies, for PyTorch wheels cugraph#5356)
• cugraph-gnn (Switch to release channel for PyTorch + CUDA 13 cugraph-gnn#355)

Conda

Recall that the pytorch channel is no longer supported (pytorch/pytorch#138506), so we need conda-forge::pytorch packages with CUDA 13 support.

For that to happen, all of pytorch's dependencies will also need packages on conda-forge with CUDA 13 support. You can see the dependency graph and status at https://conda-forge.org/status/migration/?name=cuda130. I think we need these to happen:

• libmagma (Upgrade to CUDA 13.0 conda-forge/libmagma-feedstock#30)
• libcudss (Upgrade to CUDA 13.0 update to v0.7.0 conda-forge/libcudss-feedstock#20)
• triton (Upgrade to CUDA 13.0 conda-forge/triton-feedstock#58)
• pytorch (Upgrade to CUDA 13.0 conda-forge/pytorch-cpu-feedstock#457 Triton bump, py3.14 + CUDA 13.0 conda-forge/pytorch-cpu-feedstock#477)

Once there are CUDA 13 conda-forge::pytorch packages:

• restore cugraph-pyg CUDA 13 conda test jobs (CI: add CUDA 13 conda python tests cugraph-gnn#296)
• move cudf integration tests back to using rapidsai/ci-conda:{rapids}-latest (see Revert "Pin to CUDA 12 image for integration tests (#19903)" NVIDIA/cudf#19948 (comment) ... fixed in use CUDA 13 for third-party integration tests NVIDIA/cudf#20748)

Notes

N/A

[jameslamb]

All of pytorch's upstream dependencies in conda-forge now have CUDA 13 support, so the migrator created a PR 😁

conda-forge/pytorch-cpu-feedstock#457

I'm watching that, once it lands and there are CUDA 13 packages I can handle the testing changes in cugraph-gnn and elsewhere in RAPIDS.

[jameslamb]

Updating for anyone following along.

I've moved the target release to v26.04. To close this we'll need PyTorch conda packages with CUDA 13 support, trying to help out with that on conda-forge/pytorch-cpu-feedstock#457

[jameslamb]

Work on CUDA 13 packages for conda-forge` has moved here: conda-forge/pytorch-cpu-feedstock#477

[bdice]

Now that we have PyTorch conda-forge packages for CUDA 13, I am working on adding CI testing for RMM's PyTorch integration here: rapidsai/rmm#2240

[jameslamb]

Now that cugraph-gnn support is done, I think we can close this! That RMM PR and related issue (rapidsai/rmm#2239) track the work in RMM.

[jakirkham]

Thanks James! 🙏