stuck attempting to package vllm 0.3.3

[meditans]

Hi! I'm attempting to package the latest version of vllm using poetry2nix.

I believe I'm mostly there: I've patched numerous dependencies that either lacked build tools or, like safetensors, relied on maturin. Thus, I used preferWheel. Currently, I'm addressing the last package, vllm itself, but I'm encountering a few unresolved errors.

Please find the work I've done thus far in this autocontained flake: packaging-vllm.
I would appreciate ideas on how to finish packaging this, or directions on how to approach these problems. Then I could start upstreaming the patches that are needed to poetry2nix.

• In the master branch, I've tried to minimize the reliance on wheels. Here, I'm experiencing an issue with the CUDA_HOME variable possibly not being correctly set, and an incorrect version of stdenv (although I tried to fix the version).

• In the only-wheels branch, I've focused on exclusively using wheels. However, I’ve encountered collision issues that I'm unsure how to resolve.

The errors are immediately visible if issuing nix develop .# in the respective branch. I'm on NixOS and I have some cuda GPUs.
Thank you :)

ps. Also including @charles-dyfis-net because my starting point was the override he created for lmql (I learned a lot!)