[AMD ROCm] `_validate_bnb_multi_backend_availability()` incorrectly tries to alter a frozenset.

[anadon]

System Info

When building for the ROCm/HIP backend, then importing the python module, the following traceback is printed:

Traceback (most recent call last):
  File "<frozen runpy>", line 198, in _run_module_as_main
  File "<frozen runpy>", line 88, in _run_code
  File "/home/anadon/Documents/code/Kiwi-LLaMA/kiwillama/__main__.py", line 20, in <module>
    train()
  File "/home/anadon/Documents/code/Kiwi-LLaMA/kiwillama/train.py", line 45, in train
    base_model = AutoModelForCausalLM.from_pretrained(
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/nix/store/6va0jj7wpj3xb3vkxkwz78apvsd2ckyx-python3.12-transformers-4.49.0/lib/python3.12/site-packages/transformers/models/auto/auto_factory.py", line 564, in from_pretrained
    return model_class.from_pretrained(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/nix/store/6va0jj7wpj3xb3vkxkwz78apvsd2ckyx-python3.12-transformers-4.49.0/lib/python3.12/site-packages/transformers/modeling_utils.py", line 262, in _wrapper
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/nix/store/6va0jj7wpj3xb3vkxkwz78apvsd2ckyx-python3.12-transformers-4.49.0/lib/python3.12/site-packages/transformers/modeling_utils.py", line 3698, in from_pretrained
    hf_quantizer.validate_environment(
  File "/nix/store/6va0jj7wpj3xb3vkxkwz78apvsd2ckyx-python3.12-transformers-4.49.0/lib/python3.12/site-packages/transformers/quantizers/quantizer_bnb_4bit.py", line 83, in validate_environment
    validate_bnb_backend_availability(raise_exception=True)
  File "/nix/store/6va0jj7wpj3xb3vkxkwz78apvsd2ckyx-python3.12-transformers-4.49.0/lib/python3.12/site-packages/transformers/integrations/bitsandbytes.py", line 558, in validate_bnb_backend_availability
    return _validate_bnb_multi_backend_availability(raise_exception)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/nix/store/6va0jj7wpj3xb3vkxkwz78apvsd2ckyx-python3.12-transformers-4.49.0/lib/python3.12/site-packages/transformers/integrations/bitsandbytes.py", line 499, in _validate_bnb_multi_backend_availability
    available_devices.discard("cpu")  # Only Intel CPU is supported by BNB at the moment
    ^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'frozenset' object has no attribute 'discard'

Reproduction

Using a System with an AMD CPU, and a RX 7900 XTX, build the following Nix Flake then import bitsandbytes.

{
  description = "Define development dependencies.";

  inputs = {
    # Which Nix upstream package branch to track
    nixpkgs.url = "nixpkgs/nixos-unstable";
    process-compose-flake.url = "github:Platonic-Systems/process-compose-flake";
    services-flake.url = "github:juspay/services-flake";
  };

  # What results we're going to expose
  outputs = { nixpkgs, process-compose-flake, services-flake, ... }:
    let

      supportedSystems = [ "x86_64-linux" "aarch64-linux" "aarch64-darwin" ];
      forAllSystems = f: nixpkgs.lib.genAttrs supportedSystems (system: f rec {

        # Configure package settings
        pkgs = import nixpkgs { 
          inherit system; 
          # Accept the following un-free licenses 
          config.allowUnfreePredicate = pkg: builtins.elem (nixpkgs.lib.getName pkg) [
            "cuda_nvcc"
            "cudnn"
            "libcublas"
            "cuda_cudart"
            "cuda_cccl"
            "libcufile"
            "libcurand" # because of bitsandbytes
            "libcusolver" # because of bitsandbytes
            "libnvjitlink" # because of bitsandbytes
            "libcusparse" # because of bitsandbytes
          ];

          # Some packages are reported broken but we need them to even build, so enable them anyways.
          config.allowBroken=true;

          overlays = [
            (final: prev: { 

              python312Packages = prev.python312Packages // {
                trl = pkgs.python312Packages.buildPythonPackage rec {
                  pname = "trl";
                  version = "v0.16.0";
                  # Declare repos which are then later used to build packages
                  # See https://ryantm.github.io/nixpkgs/builders/fetchers/ for more details.
                  src = pkgs.fetchFromGitHub {
                    owner = "huggingface";
                    repo = "trl";
                    rev = "${version}";
                    sha256 = "sha256-+ab952LXUM3nSpsil/xH2PrqTA9uNdt82m1dLN1iEQg=";
                  };
                  propagatedBuildInputs = [ (with pkgs.python312Packages; [ 
                    datasets
                    rich
                    accelerate
                    transformers
                  ])];
                };

                bitsandbytes-hip = pkgs.python312Packages.buildPythonPackage rec {
                  pname = "bitsandbytes";
                  version = "0.45.1";
                  format = "other";
                
                  # Directly fetch the wheel file instead of using pip during the build
                  wheel = pkgs.fetchurl {
                    url = "https://github.com/bitsandbytes-foundation/bitsandbytes/releases/download/continuous-release_multi-backend-refactor/bitsandbytes-0.45.1.dev0-py3-none-manylinux_2_24_x86_64.whl";
                    hash = "sha256-Z/7V+LU8XNXXh/WKwVKNHalSarRQLjjGijI+iGPY3K4="; # Your original hash looked correct
                  };

                  src = pkgs.fetchFromGitHub {
                    owner = "bitsandbytes-foundation";
                    repo = "${pname}";
                    rev = "multi-backend-refactor";
                    sha256 = "sha256-WWNhrhQYaauvhW2xylZ0ROoOfGxqpUUWoD2d9YLWFUE="; # Your updated hash
                  };
                
                  dontBuild = true;
                
                  # Dependencies
                  nativeBuildInputs = with pkgs; [
                    #python312Packages.pip
                    #python312Packages.wheel
                    #python312Packages.setuptools
                    unzip
                    #git
                    patchelf
                    makeWrapper
                  ];
                
                  buildInputs = with pkgs; [
                    rocmPackages.clr
                    rocmPackages.hipblas
                    rocmPackages.rocblas
                    rocmPackages.rocrand
                    rocmPackages.hipcub
                    rocmPackages.miopen
                  ];
                
                  propagatedBuildInputs = with pkgs.python312Packages; [
                    torch
                    scipy
                    numpy
                  ];
                
                  # Custom install phase
                  installPhase = ''
                    # Create Python package directory
                    mkdir -p $out/${pkgs.python312.sitePackages}
                    
                    # Extract the wheel directly to the site-packages directory
                    unzip ${wheel} -d $TMPDIR/wheel_extract
                    
                    # Copy the package content
                    cp -r $TMPDIR/wheel_extract/bitsandbytes $out/${pkgs.python312.sitePackages}/
                    cp -r $TMPDIR/wheel_extract/bitsandbytes-*.dist-info $out/${pkgs.python312.sitePackages}/
                    
                    # Fix RPATH in the shared libraries - specify all relevant ROCm .so files
                    find $out/${pkgs.python312.sitePackages}/bitsandbytes -name "*.so" | while read sofile; do
                      echo "Patching RPATH for $sofile"
                      patchelf --set-rpath "${pkgs.lib.makeLibraryPath buildInputs}" "$sofile"
                    done
                    
                    # Create bin directory and wrapper script
                    mkdir -p $out/bin
                    makeWrapper ${pkgs.python312}/bin/python3 $out/bin/python3-bnb \
                      --set PYTHONPATH $out/${pkgs.python312.sitePackages}:$PYTHONPATH \
                      --set BNB_COMPUTE_BACKEND "HIP" \
                      --set HIP_VISIBLE_DEVICES "0" \
                      --set LD_LIBRARY_PATH "${pkgs.lib.makeLibraryPath buildInputs}"
                  '';
                
                  # Skip tests for now
                  doCheck = false;
                
                  # Simple import check to verify installation
                  pythonImportsCheck = [ "bitsandbytes" ];
                
                  meta = with pkgs.lib; {
                    description = "8-bit optimizers and matrix multiplication with ROCm support";
                    homepage = "https://github.com/bitsandbytes-foundation/bitsandbytes";
                    license = licenses.mit;
                    platforms = platforms.linux;
                  };
                };
              };
            })
          ];
        };

        # Specify service processes which should be made available to run via `nix run ...`.
        servicesMod = (import process-compose-flake.lib { inherit pkgs; }).evalModules {
          modules = [
            services-flake.processComposeModules.default
            {
              services.ollama."ollama1" = {
                enable = true;
                acceleration = "rocm";
              };
            }
          ];
        };
      });

    in {
      packages = forAllSystems ({ servicesMod, ... }: {
        default = servicesMod.config.outputs.package;
      });

      # Declare what packages we need as a record. The use as a record is
      # needed because, without it, the data contained within can't be
      # referenced in other parts of this file.
      devShells = forAllSystems ({pkgs, servicesMod}: {
        default = pkgs.mkShell rec {
          packages = with pkgs; [
            python312Full 
            python312Packages.distlib
            python312Packages.cython
            python312Packages.setuptools
            python312Packages.setuptoolsBuildHook
            python312Packages.wheel
            python312Packages.vllm
            python312Packages.beautifulsoup4
            python312Packages.types-beautifulsoup4
            python312Packages.keyring
            python312Packages.peft
            python312Packages.trl
            python312Packages.bitsandbytes-hip
            cudaPackages.cudnn
            cudaPackages.libcublas
            cudaPackages.cuda_cudart
            python312Packages.torchWithoutCuda
            direnv
            pkg-config
            cmake
            blas
            lapack
            gcc_multi 
            gccMultiStdenv
            gcc-unwrapped
            ruff
            ninja
            gfortran
            meson
            glibc_multi
            ollama-rocm
            openblas
            cudaPackages.cuda_nvcc
            zlib
            # niv
            # NOTE: Put additional packages you need in this array. Packages may be found by looking them up in
            # https://search.nixos.org/packages
          ];

          # Getting the library paths needed for Python to be put into
          # LD_LIBRARY_PATH
          pythonldlibpath = "${pkgs.stdenv.cc.cc.lib}/lib:${pkgs.stdenv.cc.cc.lib.outPath}/lib:${pkgs.lib.makeLibraryPath packages}:$NIX_LD_LIBRARY_PATH";

          shellHook = ''
            export LD_LIBRARY_PATH="${pythonldlibpath}"
            export BNB_COMPUTE_BACKEND="HIP"
            export HIP_VISIBLE_DEVICES="0"
            export ROCM_PATH=${pkgs.rocmPackages.clr}
            export HIP_PATH=${pkgs.rocmPackages.clr}
          '';
        };
      });
    };
}

Expected behavior

When imported, no error should be encountered under these operating conditions.

[anadon]

This is another issue related in #1271

[anadon]

This is caused by the transformers module, but impacts usage here as well. I'm not sure how you guys want to treat this.

[anadon]

Related bug created at transformers

[anadon]

I need someone with all the appropriate hardware to test anadon/transformers#1 . I have the sneaking suspicion that the reason this wasn't caught earlier was because available_devices wasn't a frozenset.

[anadon]

huggingface/transformers#36975

[matthewdouglas]

@anadon Thanks for looking into this! I've commented and given my approval on the transformers PR.

As an FYI, now that we have merged #1544 to main, we're going to start to move the AMD integration out of the multiplatform branch. In the very near future we're going to push toward mainlining with a new interface. Stay tuned!