Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++
1.7 KB
Raw Blame
{
  description = "Nix environment for Strata";

  inputs = {
    nixpkgs.url = "github:nixos/nixpkgs/nixos-unstable";
  };

  outputs = { self, nixpkgs, ... }:
    let
      system = "x86_64-linux";
      pkgs = import nixpkgs {
        inherit system;
        # Allow NVIDIA's CUDA toolkit (nvcc, cudart, cuBLAS, ...).
        config.allowUnfree = true;
        config.cudaSupport = true;
      };

      # setup.py wants CUDA 13.0: the ready-made engine is built with 13.0.88, and 13.2.0/13.2.1 miscompile for sm_120.
      cudaPackages = pkgs.cudaPackages_13_0;
      # One merged toolkit root (bin/nvcc, include/, lib/ with cudart and cuBLAS) for CUDA_HOME / CUDAToolkit_ROOT.
      cuda = cudaPackages.cudatoolkit;
      # The gcc that this nvcc supports as its host compiler.
      stdenv = cudaPackages.backendStdenv;

    in
    {
      devShells."${system}" = {
        default = (pkgs.mkShell.override { inherit stdenv; }) {
          packages = with pkgs; [
            # ./setup.sh: makes .venv and runs setup.py
            python3

            # setup.py's engine build. cmake/ninja from PATH: the pip wheels' binaries do not run on NixOS.
            cuda
            cmake
            ninja
          ];

          shellHook = ''
            export CUDA_HOME=${cuda}
            export CUDA_PATH=${cuda}
            export CUDAHOSTCXX=${stdenv.cc}/bin/c++
            # The pip wheels in .venv (numpy, pillow, ...) need libstdc++ and zlib; the engine needs the driver's libcuda.
            export LD_LIBRARY_PATH=/run/opengl-driver/lib:${pkgs.lib.makeLibraryPath [ stdenv.cc.cc.lib pkgs.zlib ]}''${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}
          '';
        };
      };
    };
}