[RFC] Cuda support matrix for Release 2.5 #134015
Description
🐛 Describe the bug
Similar to: #123456
Opening this RFC to discuss CUDA version support for future PyTorch releases:
Option 1 - CUDA 11.8 and CUDA 12.1 and 12.4.1
CUDA 11.8, CUDNN 9.1.0.70
CUDA 12.1, CUDNN 9.1.0.70 - Version hosted on pypi
CUDA 12.4.1, CUDNN 9.1.0.70
Pros: Support for older driver
Cons: Will not be able to land NCCL update. Large support matrix, hence more expensive to maintain
This is conservative approach. Will allow us to announce CUDA 11.8 deprecation for 2.5 release and deprecate for 2.6 release.
Option 2 - CUDA 12 only:
CUDA 12.1, CUDNN 9.1.0.70 - Version hosted on pypi
CUDA 12.4, CUDNN 9.1.0.70
Pros: Will be able to go ahead with NCCL update: #133593 . Going forward we support CUDA 12 versions only.
Cons: People with older driver will not be able to use latest pytorch
Option 3
CUDA 11.8, CUDNN 9.1.0.70
CUDA 12.4, CUDNN 9.1.0.70 - Version hosted on pypi, as stable
Pros: Support for older driver
Cons: Will not be able to land NCCL update. We would need to still support CUDA 11.8 untill we migrate to next version of CUDA 12.5 or 12.6.
CUDA Driver compatibility. Please refer to:
https://docs.nvidia.com/deploy/cuda-compatibility/index.html#minor-version-compatibility
| CUDA Toolkit | Linux x86_64 Minimum Required Driver Version | Windows Minimum Required Driver Version |
|---|---|---|
| CUDA 12.x | >=525.60.13 | >=527.41 |
| CUDA 11.x | >= 450.80.02* | >=452.39* |
cc @seemethere @malfet @osalpekar @svekars @brycebortree @ptrblck @msaroufim @albanD @ezyang @nWEIdia @Skylion007
Versions
2.5