WIP: python3Packages.torch: migrate to setupCudaHook#245516
WIP: python3Packages.torch: migrate to setupCudaHook#245516SomeoneSerge wants to merge 3 commits intoNixOS:masterfrom
Conversation
| # confirm that cudatoolkits are sync'd across dependencies | ||
| assert !(MPISupport && cudaSupport) || mpi.cudatoolkit == cudatoolkit; | ||
| assert !cudaSupport || magma.cudaPackages.cudatoolkit == cudatoolkit; | ||
| with cudaPackages; |
There was a problem hiding this comment.
I probably shouldn't use with
There was a problem hiding this comment.
At least not with that big of a scope.
|
Result of 24 packages marked as broken and skipped:
174 packages failed to build:
246 packages built:
|
Failed derivationsDetails
|
|
I think a lot of these are new failures on master due to |
|
RE: torchvision Unbelievavble. This actually resulted in a WITH_CUDA=OFF build. The setup py even seems to ignore USE_CUDA=1. We really should invoke cmake directly |
This silently resets use_cuda to false. There's a bunch of other occasions where pytorch does that. The behaviour we'd like whenwe've set with_cuda but a dependency is missing is immediate failure |
|
Looking at the logs, this torch revision seems to ignore cmake_cudatoolkit_root. Maybe the release is still using the legacy findcuda. I guess I'll just come back to this in 2 weeks |
|
Superseded by the much more thorough implementation in #249259! |
Description of changes
Things done
sandbox = trueset innix.conf? (See Nix manual)nix-shell -p nixpkgs-review --run "nixpkgs-review rev HEAD". Note: all changes have to be committed, also see nixpkgs-review usage./result/bin/)