-
Notifications
You must be signed in to change notification settings - Fork 25.7k
Closed
Labels
enhancementNot as big of a feature, but technically not a bug. Should be easy to fixNot as big of a feature, but technically not a bug. Should be easy to fixmodule: binariesAnything related to official binaries that we release to usersAnything related to official binaries that we release to usersmodule: cudaRelated to torch.cuda, and CUDA support in generalRelated to torch.cuda, and CUDA support in generaloncall: relengIn support of CI and Release EngineeringIn support of CI and Release EngineeringtriagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate moduleThis issue has been looked at a team member, and triaged and prioritized into an appropriate module
Milestone
Description
🚀 The feature, motivation and pitch
CUDA 13.0 is released on 8/4, creating issue tracker for CUDA 13.0 binaries enablement.
CUDA 13.0 is a major upgrade over CUDA 12, benefits from upgrading in the nightlies binaries are mainly:
- CUDA 13.0 supports all NVIDIA architectures from Turing through Blackwell. Featuring sm_110 (Jetson Thor), sm_103 (B300/GB300), sm_121(DGX SPARK/DIGITS). Note: CUDA 12.9 first introduced sm_12x and sm_10x arch support, which will be compatible with sm_120 and sm_100 binaries respectively. Thor and SPARK now both on SBSA platforms.
- Binary size reduction from compression method change: Supporting --compress-mode flag for nvcc across all drivers of CUDA 13.X toolkits, enabling the possibility to use --compress-mode=size for significant size reduction (~71% less for CUDA Math APIs for example). Default will be --compress-mode=default which gives smaller binaries than LZ4 speed mode used in previous CUDA versions.
- New features to Math libraries: cuBLAS improvements, including implicit autotuning mode
- Arm platforms are now unified in CUDA Toolkit, enabling single-install and consistent builds across all Arm architectures.
- General CUDA code compilation performance https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html#id2
See https://developer.nvidia.com/blog/whats-new-and-important-in-cuda-toolkit-13-0/ for more comprehensive notes.
Update Plan
| Feature | Target Date |
|---|---|
| Linux and Linux aarch64 CD Nightly | Aug 22, 2025 |
| Enablement of CUDA 13.0 in PyTorch CI | Aug 29, 2025 |
| Windows CD Nightly | Aug 29, 2025 |
| Enable CUDA 13.0 in Ecosystem Libraries | Aug 29, 2025 |
| Removal of CUDA 12.9 Builds | Aug 29, 2025 |
PRs
Linux builds CD nightly (x86, sbsa, libtorch)
- Almalinux docker --> Magma build
- Add almalinux docker for magma and CUDA installation for CUDA 13 #160201
- Add Magma build for CUDA 13.0 #160770
- Build addition
- x86: [CD] Add CUDA 13.0 x86 nightly builds #160956
- sbsa: [CD] [aarch64] Add CUDA 13.0 sbsa nightly build #161257
- torchvision/torchaudio x86: Add x86 domain builds for CUDA 13.0 test-infra#7046
- torchvision/torchaudio sbsa: Add sbsa CUDA 13.0 domain builds test-infra#7059
- libtorch build: [CD] Add cuda 13.0 libtorch builds, remove CUDA 12.9 builds #161916
- Build size reduction
- 13.0.U2 upgrade
CUDA 13 CI
Windows CD nightly
- Add Windows AMI
- Add Windows CUDA 13 AMI test-infra#7003
- Update CUDA 13 CUDNN to 9.12.0.46 test-infra#7020
- Fix CUDA 13.0 Windows AMI building test-infra#7035
- Add Windows build script
Alternatives
No response
Additional context
No response
cc @seemethere @malfet @atalman @ptrblck @msaroufim @eqy @jerryzh168 @nWEIdia
atalman, greyhound3, levicki, johnnynunez, Margen67 and 1 moredonhuvy, jameslamb and Margen67eqy, Athlmyn, Puiching-Memory, johnnynunez, Margen67 and 1 more
Metadata
Metadata
Assignees
Labels
enhancementNot as big of a feature, but technically not a bug. Should be easy to fixNot as big of a feature, but technically not a bug. Should be easy to fixmodule: binariesAnything related to official binaries that we release to usersAnything related to official binaries that we release to usersmodule: cudaRelated to torch.cuda, and CUDA support in generalRelated to torch.cuda, and CUDA support in generaloncall: relengIn support of CI and Release EngineeringIn support of CI and Release EngineeringtriagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate moduleThis issue has been looked at a team member, and triaged and prioritized into an appropriate module
Type
Projects
Status
Done