search

pytorch / builder

Continuous builder and binary build scripts for pytorch
BSD 2-Clause "Simplified" License
339 stars 227 forks source link

Update PyTorch-CUDA for CUDA 12.4.1 Bump #1947

Closed nWEIdia closed 3 months ago

nWEIdia commented 3 months ago

Update constraints to use cuda 12.4.1 as [ and cuda 12.5.0 as ) CUDA 12.4.1 page: https://docs.nvidia.com/cuda/archive/12.4.1/cuda-toolkit-release-notes/index.html CUDA 12.5.0 page: https://docs.nvidia.com/cuda/archive/12.5.0/cuda-toolkit-release-notes/index.html

Similar to: https://github.com/pytorch/builder/pull/1792

cc @atalman @malfet @ptrblck @eqy @tinglvv

nWEIdia commented 3 months ago

Maybe. there was a build number change to distinguish the pytorch-cuda at 2.4.0 release time vs. the new pytorch-cuda. But I guess there is no automatic way to let 2.4.0 pick up pytorch-cuda v6 ? This PR is v7.

nWEIdia commented 3 months ago

PyTorch-CUDA is like packaging index. The 2.4.0 binary's dependency should have already been hard-coded, therefore, it would know which cuda dependency conda package to pull? Agree it would be best to test it out first in a test channel.

nWEIdia commented 3 months ago

It looks like indeed the CUDA 12.4.1 anaconda pytorch-cuda package would be picked up, if this PR gets merged. Tested with anaconda/nweidia pytorch-cuda where both v6 and v7 were present and used the following anaconda command: conda install pytorch pytorch-cuda=12.4 -c nweidia -c pytorch -c nvidia

nWEIdia commented 3 months ago

What if we just upload the pytorch-cuda 12.4.1 packages to pytorch-nightly channel? At next release, if cuda is updated to 12.4.1 too, we can then upload the packages to pytorch channel? Would this avoid the problem of "one tries to install pytorch-2.4.0 using this new pytorch-cuda definition" ?

cc @malfet @atalman @ptrblck