Describe the bug
nccl is only ever built with the default version of cudaPackages because the line which adds the derivation never changes the version of cudaPackages provided:
https://github.com/NixOS/nixpkgs/blob/master/pkgs/top-level/cuda-packages.nix#L48
I confirmed this by making this change to the nccl derivation
cuda-redist = throw "cudaPackages.cudaVersion is ${cudaPackages.cudaVersion}";and trying to build it with two different versions of cudaPackages. Notice that the error is the same:
Details
[connorbaker@fedora nixpkgs]$ nix build --impure -L .#cudaPackages.nccl
warning: Git tree '/home/connorbaker/Documents/nixpkgs' is dirty
error:
… while calling the 'derivationStrict' builtin
at //builtin/derivation.nix:9:12: (source not available)
… while evaluating derivation 'nccl-2.16.5-1-cuda-11.7'
whose name attribute is located at /nix/store/bh247jakm9gvjrif38ssqz90zfgqxy05-source/pkgs/stdenv/generic/make-derivation.nix:302:7
… while evaluating attribute 'nativeBuildInputs' of derivation 'nccl-2.16.5-1-cuda-11.7'
at /nix/store/bh247jakm9gvjrif38ssqz90zfgqxy05-source/pkgs/stdenv/generic/make-derivation.nix:342:7:
341| depsBuildBuild = lib.elemAt (lib.elemAt dependencies 0) 0;
342| nativeBuildInputs = lib.elemAt (lib.elemAt dependencies 0) 1;
| ^
343| depsBuildTarget = lib.elemAt (lib.elemAt dependencies 0) 2;
(stack trace truncated; use '--show-trace' to show the full trace)
error: cudaPackages.cudaVersion is 11.7
[connorbaker@fedora nixpkgs]$ nix build --impure -L .#cudaPackages_11_8.nccl
warning: Git tree '/home/connorbaker/Documents/nixpkgs' is dirty
error:
… while calling the 'derivationStrict' builtin
at //builtin/derivation.nix:9:12: (source not available)
… while evaluating derivation 'nccl-2.16.5-1-cuda-11.7'
whose name attribute is located at /nix/store/bh247jakm9gvjrif38ssqz90zfgqxy05-source/pkgs/stdenv/generic/make-derivation.nix:302:7
… while evaluating attribute 'nativeBuildInputs' of derivation 'nccl-2.16.5-1-cuda-11.7'
at /nix/store/bh247jakm9gvjrif38ssqz90zfgqxy05-source/pkgs/stdenv/generic/make-derivation.nix:342:7:
341| depsBuildBuild = lib.elemAt (lib.elemAt dependencies 0) 0;
342| nativeBuildInputs = lib.elemAt (lib.elemAt dependencies 0) 1;
| ^
343| depsBuildTarget = lib.elemAt (lib.elemAt dependencies 0) 2;
(stack trace truncated; use '--show-trace' to show the full trace)
error: cudaPackages.cudaVersion is 11.7In python-packages.nix the pattern seems to be passing pythonPackages = self. However, it is relatively rare that the derivation takes pythonPackages as an argument. Instead, it seems to be preferred to have the derivation take, as arguments, the actual python packages required.
To bring the NCCL derivation in line with this pattern, it should take the CUDA packages it requires as arguments.
Notify maintainers
@NixOS/cuda-maintainers
Describe the bug
ncclis only ever built with the default version ofcudaPackagesbecause the line which adds the derivation never changes the version ofcudaPackagesprovided:https://github.com/NixOS/nixpkgs/blob/master/pkgs/top-level/cuda-packages.nix#L48
I confirmed this by making this change to the
ncclderivationand trying to build it with two different versions of
cudaPackages. Notice that the error is the same:Details
In
python-packages.nixthe pattern seems to be passingpythonPackages = self. However, it is relatively rare that the derivation takespythonPackagesas an argument. Instead, it seems to be preferred to have the derivation take, as arguments, the actual python packages required.To bring the NCCL derivation in line with this pattern, it should take the CUDA packages it requires as arguments.
Notify maintainers
@NixOS/cuda-maintainers