{
  lib,
  buildPythonPackage,
  fetchurl,
  # dependencies
  numpy,
  numba,
  # torch-bin (CUDA wheel build) and triton-bin are passed by the caller;
  # the python scope defaults are the source-built torch (CPU-only) and
  # triton — using them would duplicate torch/triton in the closure.
  torch,
  triton,
  tqdm,
}:

buildPythonPackage rec {
  pname = "flashlib";
  version = "0.3.0";
  format = "wheel";

  src = fetchurl {
    url = "https://files.pythonhosted.org/packages/e5/b8/4c085892462e521bb9f2d943ff34fa219e3a5a206798f3c96a983538039f/flashlib-0.3.0-py3-none-any.whl";
    hash = "sha256-kDeRHzFf7zyfRFMmFZc2Ory4OGiBOtGyWKsTCLNP0Ro=";
  };

  # freetoken pins flashlib==0.3.0 (not in nixpkgs) and only uses its Triton
  # kernels (flashlib.kernels.slot_cache). nvidia-cutlass-dsl — needed only by
  # the CuTeDSL GEMM backends in flashlib.linalg, which drag in cuda-python and
  # nvdisasm binary wheels — is deliberately not packaged (and removed from
  # the wheel metadata so the runtime deps check passes).
  pythonRemoveDeps = [ "nvidia-cutlass-dsl" ];

  dependencies = [
    numpy
    numba
    torch
    triton
    tqdm
  ];

  # flashlib is CPU-safe at import time: CuteDSL imports are lazy and hardware
  # detection (flashlib._hw) is cuda-optional.
  pythonImportsCheck = [ "flashlib" ];

  meta = {
    description = "High-performance ML primitives — Triton and CuteDSL kernels for NVIDIA GPUs";
    homepage = "https://pypi.org/project/flashlib/";
    license = lib.licenses.asl20;
    platforms = lib.platforms.linux;
  };
}