These things really do not have an actual reason. Take for example the Nic people who go to every project with a bash script that uses #!/bin/bash and tells them to use #!/usr/bin/env bash. There is no justification for trying to fix every bash script in existence over the operating system being able to resolve what the "current" bash version should be (/usr/bin/env already has to pick a version). This is pointless O(n) busy work Nix people have created for themselves when an O(1) solution exists.
Getting a patch into `nixpkgs` is not a reasonable bar for running some program you want to try. This is an (admittedly polite) way of saying more or less "you're holding it wrong". I respectfully disagree.
For expert users, creating a flake and/or standalone derivation is a little more reasonable of an ask. It is an unreasonable ask of people who are considering adopting Nix, and the friction is artificial, it serves no practical benefit (you'll notice that things like `nix-ld` are in `nix-community` now, long-time community contributors like Mic92 said a long time ago, we need to stop the bleeding on this).
And once you get out of the merely alienating to anyone used to being treated well, and into actually complex software builds, "possible once you learn the incantations" goes to "never happens, you will never run this without Docker on NixOS".
I won't settle for that, so I fix things:
```
b7r6 on ultraviolence ~
nix run github:sensenet-ai/nvidia-sdk#python
Python 3.12.12 (main, Oct 9 2025, 11:07:00) [GCC 15.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorrt_llm
W0619 15:07:35.110000 14798 /nix/store/xgk8wcis2avy1vsv66viw150jv78v6qs-ngc-python-packages-ngc-25.12-rootfs/lib/python3.12/site-packages/torch/utils/cpp_extension.py:2422] TORCH_CUDA_ARCH_LIST is not set, all archs for visible cards are included for compilation.
W0619 15:07:35.110000 14798 /nix/store/xgk8wcis2avy1vsv66viw150jv78v6qs-ngc-python-packages-ngc-25.12-rootfs/lib/python3.12/site-packages/torch/utils/cpp_extension.py:2422] If this is not desired, please set os.environ['TORCH_CUDA_ARCH_LIST'] to specific architectures.
[TensorRT-LLM] TensorRT LLM version: 1.1.0
>>> import torch
>>> torch.cuda.is_available()
True
>>>
```