In reply to @mcwitt:matrix.org
Just to close the loop, the fix in my case was to set cmakeFlags = [ "-DCMAKE_CUDA_HOST_COMPILER=${cudaPackages.cudatoolkit.cc}/bin/cc" ] (and eventually found many examples of this in nixpkgs). Thanks for pointing me in the right direction!

Three more things that popped into my head (sorry, I am actively consuming coffee):

1. When we do override the C/C++ compilers by setting the CC/CXX environment variables, that doesn't change binutils, so (in my case) I still see ar/ranlib/ld and friends from gcc12 being used. Is that a problem? I don't know if version bumps to those tools can cause as much damage as libraries compiled with different language standards.
2. If a package needs to link against libcuda.so specifically, what's the best way to make the linker aware of those stubs? I set LIBRARY_PATH and that seemed to do the trick: https://github.com/NixOS/nixpkgs/pull/218166/files#diff-ab3fb67b115c350953951c7c5aa868e8dd9694460710d2a99b845e7704ce0cf5R76
3. Is it better to set environment variables as env.BLAG = "blarg" (I saw a tree-wide change about using env because of "structuredAttrs") in the derivation or to export them in the shell, in something like preConfigure?

Three more things that popped into my head (sorry, I am actively consuming coffee):

1. When we do override the C/C++ compilers by setting the CC/CXX environment variables, that doesn't change binutils, so (in my case) I still see ar/ranlib/ld and friends from gcc12 being used. Is that a problem? I don't know if version bumps to those tools can cause as much damage as libraries compiled with different language standards.
2. If a package needs to link against libcuda.so specifically, what's the best way to make the linker aware of those stubs? I set LIBRARY_PATH and that seemed to do the trick: https://github.com/NixOS/nixpkgs/pull/218166/files#diff-ab3fb67b115c350953951c7c5aa868e8dd9694460710d2a99b845e7704ce0cf5R76
3. Is it better to set environment variables as env.BLAG = "blarg" (I saw a tree-wide change about using env because of "structuredAttrs") in the derivation or to export them in the shell, in something like preConfigure?

EDIT: I should put these in my issue about docs for CUDA packaging...

RE: linker

Idk but I think it's kind of the point of the separate linkage phase that we can have some flexibility in mix-n-matching languages and compilers? I'm rn waiting for a build of faiss where I set CUDAHOSTCXX (as you suggested) but don't override stdenv (which means that gcc12 would still be used for .cpp/.cc files). I expect that it's going to succeed

RE: libcuda.so
I don't really know any uses we have for the stubs, we usually want apps to load libcuda.so from /run/opengl-driver/lib. So we add that to the runpaths and if we use autopatchelf we tell it to ignore the missing libcuda.so

Runpath looks kind of weird (it has but gcc11 and gcc12 lib/):

❯ patchelf --print-rpath /nix/store/2gs5jpwka7604pi9wqab4bp2hsxjkzjx-faiss-1.7.2/lib/python3.10/site-packages/faiss/_swigfaiss.so
/run/opengl-driver/lib:/nix/store/2gs5jpwka7604pi9wqab4bp2hsxjkzjx-faiss-1.7.2/lib64:/nix/store/2gs5jpwka7604pi9wqab4bp2hsxjkzjx-faiss-1.7.2/lib:/nix/store/713pzpgy2yhmnh3vs8cfdpv4j8pmqsmm-cudatoolkit-11.7.0/lib64/stubs:/nix/store/ps7an26cirhh0xy1wrlc2icvfhrd39cj-gcc-11.3.0-lib/lib:/nix/store/713pzpgy2yhmnh3vs8cfdpv4j8pmqsmm-cudatoolkit-11.7.0/lib:/nix/store/li1fg5xf6rzmpm7zlcnsymy8wfpmx0vj-cudatoolkit-11.7.0-lib/lib:/nix/store/m39wyb50jz4mqj22459nz397ascvmgiv-blas-3/lib:/nix/store/lqz6hmd86viw83f9qll2ip87jhb7p1ah-glibc-2.35-224/lib:/nix/store/k88zxp7cvd5gpharprhg9ah0vhz2asq7-gcc-12.2.0-lib/lib

Runpath looks kind of weird (it has both gcc11 and gcc12 lib/):

❯ patchelf --print-rpath /nix/store/2gs5jpwka7604pi9wqab4bp2hsxjkzjx-faiss-1.7.2/lib/python3.10/site-packages/faiss/_swigfaiss.so
/run/opengl-driver/lib:/nix/store/2gs5jpwka7604pi9wqab4bp2hsxjkzjx-faiss-1.7.2/lib64:/nix/store/2gs5jpwka7604pi9wqab4bp2hsxjkzjx-faiss-1.7.2/lib:/nix/store/713pzpgy2yhmnh3vs8cfdpv4j8pmqsmm-cudatoolkit-11.7.0/lib64/stubs:/nix/store/ps7an26cirhh0xy1wrlc2icvfhrd39cj-gcc-11.3.0-lib/lib:/nix/store/713pzpgy2yhmnh3vs8cfdpv4j8pmqsmm-cudatoolkit-11.7.0/lib:/nix/store/li1fg5xf6rzmpm7zlcnsymy8wfpmx0vj-cudatoolkit-11.7.0-lib/lib:/nix/store/m39wyb50jz4mqj22459nz397ascvmgiv-blas-3/lib:/nix/store/lqz6hmd86viw83f9qll2ip87jhb7p1ah-glibc-2.35-224/lib:/nix/store/k88zxp7cvd5gpharprhg9ah0vhz2asq7-gcc-12.2.0-lib/lib

In reply to @ss:someonex.net

RE: linker

Idk but I think it's kind of the point of the separate linkage phase that we can have some flexibility in mix-n-matching languages and compilers? I'm rn waiting for a build of faiss where I set CUDAHOSTCXX (as you suggested) but don't override stdenv (which means that gcc12 would still be used for .cpp/.cc files). I expect that it's going to succeed

RE: libcuda.so
I don't really know any uses we have for the stubs, we usually want apps to load libcuda.so from /run/opengl-driver/lib. So we add that to the runpaths and if we use autopatchelf we tell it to ignore the missing libcuda.so

If you have some time, could you take a crack at getting https://github.com/NixOS/nixpkgs/pull/218166 to build without the CUDA stub? Not sure if I’m missing something but it fails in the linking portion of the build phase, complaining about missing -lcuda if I don’t add the stub to the library path. Is there a better way to do that?

I don’t remember any of the other libraries failing like that, so I’m curious if there’s something weird going on.

In reply to @ss:someonex.net
#TODO there's no comment in versions.toml about where the gcc attribute comes from. I guess we take it from the release notes. Also there's no reason to hard-code gcc: looking at cudatoolkit12 release notes, they say they support clang as well

My approach to things like that has been to record something like minGcc, maxGcc, minClang, and maxClang and use them to find out valid compilers. From there, export the allowed stdenvs.

That has the added benefit of allowing new tool chains as they’re added to Nixpkgs so long as they fall in the supported range.

Although, at least personally speaking, I would also like a flag I could set globally to have NVCC just use whatever the stdenv does.

If a package requires a newer version of cudaPackages than is the default, how should that be handled? For example, I’m packaging NVIDIA’s Transformer Engine and that needs CUDA11.8+. It’s not enough to just pass it the newer, because every dependency the package has ALSO needs to use that newer version.

Any suggestions?

It’s not enough to just pass it the newer, because every dependency the package has ALSO needs to use that newer version.

Hm. What happens if they use different cuda?

Or someone mistakenly overrides one but not the other
I ended up doing this in python-modules:

  transformer-engine = callPackage ../development/python-modules/transformer-engine (
    let 
      cudaPackages = pkgs.cudaPackages_11_8;
      magma = pkgs.magma.override { inherit cudaPackages; };
      torch = self.torch.override { inherit cudaPackages magma; };
    in
    { 
      inherit cudaPackages torch;
    }
  );