I solved the problem about maturin (and am writing this response in case someone has my problem). In the edgecases.md document it was advised to use preferWheel and preferWheels, but I wasn't aware of this:

https://github.com/nix-community/poetry2nix/issues/1115#issuecomment-1611913900
One should prefer:

overrides = poetry2nix.overrides.withDefaults (<...>);

to:

overrides = poetry2nix.defaultPoetryOverrides.extend (<...>);

otherwise, preferWheel = true is not respected.

Now I have still a problem, and I'm wandering if someone encountered it before:

RuntimeError: Cannot find CUDA_HOME. CUDA must be available to build the package.

This happens when I try to build the wheel for the last package I need, llvm itself. As usual, one can try the failing build at the latest commit of the repo I linked in the first question. I wonder who should set CUDA_HOME, and how could I set it so that the build process of vllm sees it (I don't think a shellHook is early enough)

Now I have still a problem, and I'm wondering if someone encountered it before:

RuntimeError: Cannot find CUDA_HOME. CUDA must be available to build the package.

This happens when I try to build the wheel for the last package I need, llvm itself. As usual, one can try the failing build at the latest commit of the repo I linked in the first question. I wonder who should set CUDA_HOME, and how could I set it so that the build process of vllm sees it (I don't think a shellHook is early enough)

Now I have still a problem, and I'm wondering if someone encountered it before:

RuntimeError: Cannot find CUDA_HOME. CUDA must be available to build the package.

This happens when I try to build the wheel for the last package I need, vllm itself. As usual, one can try the failing build at the latest commit of the repo I linked in the first question. I wonder who should set CUDA_HOME, and how could I set it so that the build process of vllm sees it (I don't think a shellHook is early enough)

Now I have still a problem, and I'm wondering if someone encountered it before:

RuntimeError: Cannot find CUDA_HOME. CUDA must be available to build the package.

This happens when I try to build the wheel for the last package I need, vllm itself. As usual, one can try the failing build at the latest commit of the repo I linked in the first question. I wonder who should set CUDA_HOME, and how could I set it so that the build process of vllm sees it (I don't think a shellHook is early enough)

I solved this problem via:

  addCudaHome = { name, final, prev, pkg }@args:
    pkg.overridePythonAttrs (old: {
      preBuild = (old.preBuild or "")
        + ''export CUDA_HOME="${final.pkgs.cudatoolkit}"'';
    });

I am trying to get nvidia-cusparse-cu12, part of my dependencies, to build. It is currently failing with auto-patchelf not able to resolve libnvJitLink.so.12. So I tried this override which has no effect. Am I missing something obvious?

      python311.nvidia-cusparse-cu12 = super.python311.nvidia-cusparse-cu12.overridePythonAttrs (old: {
        preferWheel = true;
        nativeBuildInputs = (old.nativeBuildInputs or [ ]) ++ [ pkgs.cudaPackages.libnvjitlink ];
        autoPatchelfIgnoreMissingDeps = [ "libnvJitLink.so.12" ];
      });

In reply to @angryant:envs.net
What does your WIP nix look like?

somewhat like this: https://gist.github.com/paradx/af5de890145e70b5921e9169aecf1119

I had more of the override bits but in the end it just fails further down the line, there has to be an easier solution.