Issue Station
Sign In
Tools hereToolbox

pytorch - ✅(Solved) Fix [RFC] CUDA support matrix for Release 2.12 - introduce CUDA 13.2 as experimental. [5 pull requests, 1 comments, 2 participants]

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…
GitHub stats
pytorch/pytorch#178665•Fetched 2026-04-08 01:40:26
View on GitHub
Comments
1
Participants
2
Timeline
47
Reactions
2
Author
atalmanatalman
Participants
atalmanatalmanjohnnynunezjohnnynunez
Timeline (top)
subscribed ×24mentioned ×10labeled ×5cross-referenced ×4

PR fix notes

PR #179072: [CD] Deprecate CUDA 12.8 builds in favor of CUDA 13.0

Description (problem / solution / changelog)

Remove CUDA 12.8 from the binary build matrix and regenerate nightly workflows. CUDA 13.0 is already the stable version, making 12.8 redundant.

https://github.com/pytorch/pytorch/issues/178665

Changed files

  • .github/scripts/generate_binary_build_matrix.py (modified, +2/-13)
  • .github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml (modified, +0/-518)
  • .github/workflows/generated-linux-binary-manywheel-nightly.yml (modified, +0/-551)
  • .github/workflows/generated-windows-binary-libtorch-debug-nightly.yml (modified, +0/-245)
  • .github/workflows/generated-windows-binary-wheel-nightly.yml (modified, +499/-2191)

PR #629: feat(ci): Build nightly for cu126, cu130, and cu132 for Linux

Description (problem / solution / changelog)

Ref: https://github.com/pytorch/pytorch/issues/178665

Changed files

  • .github/workflows/utils/full_matrix_linux.json (modified, +2/-2)
  • .github/workflows/utils/minimal_matrix_linux.json (modified, +3/-3)
  • .github/workflows/utils/minimal_matrix_macos.json (modified, +5/-0)
  • .github/workflows/utils/minimal_matrix_windows.json (modified, +5/-0)
  • CHANGELOG.md (modified, +1/-1)
  • README.md (modified, +3/-3)

PR #179516: [CI] Move 12.8 CI jobs to 13.0

Description (problem / solution / changelog)

Moving all the 12.8 CI jobs to 13.0 as 13.0 is the stable release now for 2.11 Step 1 in https://github.com/pytorch/pytorch/issues/178665

cc @atalman

Changed files

  • .github/workflows/attention_op_microbenchmark.yml (modified, +4/-4)
  • .github/workflows/b200-distributed.yml (modified, +10/-10)
  • .github/workflows/b200-symm-mem.yml (modified, +10/-10)
  • .github/workflows/docker-builds.yml (modified, +4/-4)
  • .github/workflows/h100-cutlass-backend.yml (modified, +10/-10)
  • .github/workflows/h100-distributed.yml (modified, +10/-10)
  • .github/workflows/h100-symm-mem.yml (modified, +10/-10)
  • .github/workflows/inductor-pallas.yml (modified, +2/-2)
  • .github/workflows/operator_microbenchmark.yml (modified, +4/-4)
  • .github/workflows/pull.yml (modified, +8/-8)
  • .github/workflows/quantization-periodic.yml (modified, +2/-2)
  • .github/workflows/slow.yml (modified, +0/-29)
  • .github/workflows/target-determination-indexer.yml (modified, +1/-1)
  • .github/workflows/test-b200.yml (modified, +11/-11)
  • .github/workflows/test-h100.yml (modified, +15/-15)
  • .github/workflows/torchtitan.yml (modified, +3/-3)
  • .github/workflows/trunk.yml (modified, +11/-63)

PR #180052: [CI] Migrate 12.8 CI jobs to 13.0

Description (problem / solution / changelog)

Moving all the 12.8 CI jobs to 13.0 as 13.0 is the stable release now for 2.11 Step 1 in https://github.com/pytorch/pytorch/issues/178665

cc @atalman @malfet @eqy @nWEIdia @ptrblck

Changed files

  • .github/workflows/attention_op_microbenchmark.yml (modified, +4/-4)
  • .github/workflows/b200-distributed.yml (modified, +10/-10)
  • .github/workflows/b200-symm-mem.yml (modified, +10/-10)
  • .github/workflows/docker-builds.yml (modified, +3/-4)
  • .github/workflows/h100-cutlass-backend.yml (modified, +10/-10)
  • .github/workflows/h100-distributed.yml (modified, +10/-10)
  • .github/workflows/h100-symm-mem.yml (modified, +10/-10)
  • .github/workflows/inductor-pallas.yml (modified, +2/-2)
  • .github/workflows/operator_microbenchmark.yml (modified, +4/-4)
  • .github/workflows/pull.yml (modified, +9/-9)
  • .github/workflows/quantization-periodic.yml (modified, +2/-2)
  • .github/workflows/slow.yml (modified, +0/-29)
  • .github/workflows/target-determination-indexer.yml (modified, +1/-1)
  • .github/workflows/test-b200.yml (modified, +11/-11)
  • .github/workflows/test-h100.yml (modified, +15/-15)
  • .github/workflows/torchtitan.yml (modified, +3/-3)
  • .github/workflows/trunk.yml (modified, +5/-6)

PR #1925: ci: add CUDA 13.2 build and nightly test support

Description (problem / solution / changelog)

Summary

  • Add CUDA 13.2.0 to the build matrix (Linux x64, aarch64, Windows)
  • Bump Jimver/cuda-toolkit from v0.2.29 to v0.2.35 for Windows CUDA 13.2 support
  • Add CUDA 13.2.0 to the nightly test matrix with torch 2.12 nightly (+cu132)
  • Add torch_nightly input to test-runner.yml for pre-release torch installs (uses --pre, no version pin)

Not added to PR tests — nightly only for now, per discussion.

Using CUDA 13.2.0 (not 13.2.1) since Docker images and Jimver action don't have 13.2.1 yet — can bump once NVIDIA publishes images.

No changes needed to build-cuda.sh, CMakeLists.txt, or pyproject.toml — existing 13.*.* patterns handle 13.2 already.

Ref: https://github.com/pytorch/pytorch/issues/178665

Test plan

  • CI builds pass for CUDA 13.2.0 on Linux (x64 + aarch64) and Windows
  • Existing CUDA versions still build correctly (no regression from Jimver action bump)
  • Nightly test picks up torch 2.12 nightly with --pre install path

🤖 Generated with Claude Code

Changed files

  • .github/workflows/python-package.yml (modified, +2/-2)
  • .github/workflows/test-runner.yml (modified, +11/-2)
  • .github/workflows/tests-nightly.yml (modified, +6/-1)
  • CMakeLists.txt (modified, +5/-0)
RAW_BUFFERClick to expand / collapse

🚀 RFC

Similar to #172663 and #165111, this RFC proposes the CUDA support matrix for PyTorch Release 2.12.

Summary of changes proposed

  • Deprecate CUDA 12.8 — Remove from CI/CD
  • Keep CUDA 12.6 as Legacy — Maintained for backwards compatibility and support for older architecture
  • Keep CUDA 13.0 as Stable — Published to PyPI - Full Coverage in CI
  • Introduce CUDA 13.2 as Experimental — Covered by CI, nightly builds available

PyTorch 2.12 Version Support

ComponentSpecification
Legacy CUDACUDA 12.6 (CUDNN 9.10.2.21)
Stable CUDACUDA 13.0 (CUDNN TBD)
Experimental CUDACUDA 13.2 (CUDNN TBD)
DeprecatedCUDA 12.8 (removed from CI/CD)

Linux x86_64 and Windows Architecture Support

CUDA VersionArchitectures
CUDA 12.6.3Maxwell(5.0), Pascal(6.0), Volta(7.0), Turing(7.5), Ampere(8.0, 8.6), Hopper(9.0)
CUDA 13.0.xTuring(7.5), Ampere(8.0, 8.6), Hopper(9.0), Blackwell(10.0, 12.0)
CUDA 13.2.xTuring(7.5), Ampere(8.0, 8.6), Hopper(9.0), Blackwell(10.0, 12.0)

Linux aarch64 Architecture Support

CUDA VersionArchitectures
CUDA 12.6.3Ampere(8.0), Hopper(9.0)
CUDA 13.0.xAmpere(8.0), Hopper(9.0), Blackwell(10.0, 11.0, 12.0)
CUDA 13.2.xAmpere(8.0), Hopper(9.0), Blackwell(10.0, 11.0, 12.0)

PyPI Release

CUDA 13.0 will be the version published to PyPI for Release 2.12. CUDA 12.6 and 13.2 will be available via download.pytorch.org

Rationale

  • CUDA 12.8 deprecation: With CUDA 13.0 now stable and CUDA 12.6 retained as legacy, CUDA 12.8 no longer serves a distinct role in the support matrix. Removing it reduces CI/CD burden and binary build complexity.
  • CUDA 12.6 as legacy: Provides continued support for older GPU architectures (Maxwell, Pascal, Volta) that are not supported by CUDA 13.x.
  • CUDA 13.0 as stable: Already validated through 2.11 as stable. Continues as the primary CUDA version for PyPI releases.
  • CUDA 13.2 as experimental: Introduces the latest CUDA toolkit for early adopters and forward compatibility testing.

Action items

  • Remove CUDA 12.8 from CI/CD workflows
  • Remove CUDA 12.8 from binary build matrix (wheels, libtorch, conda)
  • Add CUDA 13.2 to Linux CI/CD workflows
  • Add CUDA 13.2 to Windows CI/CD workflows
  • Add CUDA 13.2 to Linux binary builds (nightly)
  • Add CUDA 13.2 to Windows binary builds (nightly)
  • Update release compatibility matrix documentation
  • Validate CUDA 13.2 nightly builds on supported architectures

cc @seemethere @malfet @tinglvv @nWEIdia @ptrblck @msaroufim @eqy @jerryzh168 @pytorch/pytorch-dev-infra @ngimel

extent analysis

Fix Plan

To implement the proposed CUDA support matrix for PyTorch Release 2.12, follow these steps:

  • Remove CUDA 12.8 from CI/CD workflows:
    • Update .github/workflows/ci.yml to exclude CUDA 12.8
    • Remove CUDA 12.8 from ci.py and ci.sh scripts
  • Remove CUDA 12.8 from binary build matrix:
    • Update setup.py to exclude CUDA 12.8
    • Remove CUDA 12.8 from conda/meta.yaml and conda/build.sh
  • Add CUDA 13.2 to CI/CD workflows:
    • Update .github/workflows/ci.yml to include CUDA 13.2
    • Add CUDA 13.2 to ci.py and ci.sh scripts
  • Add CUDA 13.2 to binary builds:
    • Update setup.py to include CUDA 13.2
    • Add CUDA 13.2 to conda/meta.yaml and conda/build.sh

Example code changes:

# setup.py
cuda_versions = ['12.6', '13.0', '13.2']  # Add CUDA 13.2

# ci.py
cuda_versions = ['12.6', '13.0', '13.2']  # Add CUDA 13.2

# conda/meta.yaml
dependencies:
  - cuda ==12.6
  - cuda ==13.0
  - cuda ==13.2  # Add CUDA 13.2

Verification

To verify the changes, run the following commands:

  • git checkout <branch-name>: Switch to the branch with the changes
  • python setup.py build: Build the package with the updated CUDA versions
  • python ci.py: Run the CI script with the updated CUDA versions
  • conda build.: Build the conda package with the updated CUDA versions

Extra Tips

  • Make sure to update the release compatibility matrix documentation to reflect the changes
  • Validate CUDA 13.2 nightly builds on supported architectures to ensure compatibility
  • Test the package with the updated CUDA versions to ensure functionality and performance are not affected.

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…
#task chaining#parallel task#integration issue#index setup#retrieval issue

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING