[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

cuda support of src:pytorch

Hi team,

I reviewed cuDNN's license again, it finally explicitly declared that
the .h, and .so files are distributable (Chapter 8)
That means we can upload it to the non-free section, without modifying
the files (e.g. stripping). The packaging work should be done in nvidia
team, and I've already set up a repository[2] on salsa many years ago.
And we can copy some bits from Archlinux packaging:

Apart from cudnn, missing from the archive is nccl 
https://github.com/NVIDIA/nccl (BSD-3-Clause, can enter the contrib section).

As another cuda-related dependency, src:cub is already present in the
archive, thanks to the debian nvidia maintainers.

IIUC, B-D: nvidia-cuda-toolkit, libcudnn-dev, libcub-dev, libnccl-dev,
(optional: libmagma-dev) should be enough for us to enable the build
of pytorch-cuda.

Should we hurry up, we may still have chance to get a cuda version of
pytorch into the archive. Anybody willing to lend me a hand on the nccl
packaging [1]? Currently I'm still working on the unstable->testing
migration issue of pytorch, and plan to investigate src:nvidia-cudnn

[1] repo should be hosted in nvidia-team https://salsa.debian.org/nvidia-team/nvidia-nccl
[2] https://salsa.debian.org/nvidia-team/nvidia-cudnn

Reply to: