pytorch
Decomission CUDA 10.2 support
We keep cuda 10.2 only for pypi releases https://pypi.org/project/torch/#files, which has 850mb size limit. Here is the pypi support ticket we opened last year to increase the limit. Here is related discussion Issue 56055
We should discontinue cuda 10. It will likely require pypi wheel without cuda support, but we can't propagate this cuda 10 situation indefinitely. We could print an error message when users attempt to initialize cuda context with their pypi installs. The problem though is users who don't get pytorch directly, but get it through various dependency resolution methods, like torch listed in requirements.txt.
possible solution to explore: dynamically link with CUDA so pytorch is distribuded without CUDA on pypi.org Implement post install script to get CUDA using curl and install it after pypi install is complete:
Subtask for decomissioning:
- [x] #1134
- [x] #1133
- [x] #1132
- [x] #1154
We should proceed with the following schedule on dropping support for 10.2 :
- Release 1.12 - announce
- Release 1.13 - actual drop
One solution of dropping CUDA 10.2 that was discussed with @ptrblck is to link with CUDA dynamically rather then statically as done now. Then we download CUDA runtime on the install machine similar to this: https://docs.nvidia.com/cuda/cuda-quick-start-guide/index.html#pip-wheels-installation-linux
@malfet @ngimel @ezyang Please let me know if we have anything against us trying to implement dynamic linking of CUDA 11+ ? This way we can continue to distribute pytorch on pypi and decomission CUDA 10.2