Get ava.py running with nixified builds

I had to redo some of my python overrides and crib the bitsandbytes.nix
from upstream.

Ava is failing because:

    ValueError: Tokenizer class LlamaTokenizer does not exist or is not
    currently imported.

I think this means I need to update my nixpkgs pin, so I'm gonna
snapshot my work in git, do the update, and that might obviate the local
bitsandbytes.nix anyway.

1 files changed, 74 insertions, 70 deletions

diff --git a/Biz/Bild/Deps/bitsandbytes.nix b/Biz/Bild/Deps/bitsandbytes.nix
index 8d4fa53..b87c0c4 100644
--- a/Biz/Bild/Deps/bitsandbytes.nix
+++ b/Biz/Bild/Deps/bitsandbytes.nix
@@ -1,94 +1,98 @@
 { lib
-, fetchgit
 , buildPythonPackage
-, pytorch
+, fetchFromGitHub
+, python
+, pythonOlder
+, pytestCheckHook
 , setuptools
-, typer
-, linuxPackages
-, pytest
-
-# CUDA
-, cudaSupport ? true
-, cudaPackages ? {}
-, addOpenGLRunpath ? null
-, gcc
+, torch
+, einops
+, lion-pytorch
+, scipy
+, symlinkJoin
 }:
 
 let
-  inherit (linuxPackages) nvidia_x11;
-  inherit (cudaPackages) cudatoolkit;
-  cudaVersion =
-    lib.strings.removeSuffix "0"
-      (lib.strings.concatStrings
-        (lib.strings.splitString "." cudatoolkit.version));
-  libraryPath = lib.strings.makeLibraryPath [
-      cudatoolkit
-      cudatoolkit.lib
-      nvidia_x11
-  ];
-in buildPythonPackage rec {
   pname = "bitsandbytes";
-  version = "unstable-2022-12-21";
+  version = "0.38.0";
+
+  inherit (torch) cudaPackages cudaSupport;
+  inherit (cudaPackages) cudaVersion;
+
+  # NOTE: torchvision doesn't use cudnn; torch does!
+  #   For this reason it is not included.
+  cuda-common-redist = with cudaPackages; [
+    cuda_cccl # <thrust/*>
+    libcublas # cublas_v2.h
+    libcurand
+    libcusolver # cusolverDn.h
+    libcusparse # cusparse.h
+  ];
 
-  src = fetchgit {
-    url = "https://simatime.com/git/ben/bitsandbytes.git";
-    rev = "31ef751bea48eeee2e0e95aca79df8e59b4c25c4";
-    sha256 = "sha256-/a2NFVuMSvSDELeXyfDdU9FZEJcDxCLa4VbMWBUCRI0=";
+  cuda-native-redist = symlinkJoin {
+    name = "cuda-native-redist-${cudaVersion}";
+    paths = with cudaPackages; [
+      cuda_cudart # cuda_runtime.h cuda_runtime_api.h
+      cuda_nvcc
+    ] ++ cuda-common-redist;
   };
 
-  # any exe must be wrapped with this, it doesn't effect the build but it does
-  # show up on the logs so keep it here for documentation purposes
-  LD_LIBRARY_PATH = libraryPath;
+  cuda-redist = symlinkJoin {
+    name = "cuda-redist-${cudaVersion}";
+    paths = cuda-common-redist;
+  };
+
+in
+buildPythonPackage {
+  inherit pname version;
+  format = "pyproject";
 
-  # this allows the build system to find nvcc
-  CUDA_HOME = "${cudatoolkit}";
+  disabled = pythonOlder "3.7";
+
+  src = fetchFromGitHub {
+    owner = "TimDettmers";
+    repo = pname;
+    rev = "refs/tags/${version}";
+    hash = "sha256-gGlbzTDvZNo4MhcYzLvWuB2ec7q+Qt5/LtTbJ0Rc+Kk=";
+  };
 
-  preBuild = ''
-    make cuda11x CUDA_VERSION=${cudaVersion} GPP=${gcc}/bin/g++ -j3
-    ## this python doesn't know where to get libcuda, so explicitly tell it
-    ## this is probably really bad practice, fix this
-    substituteInPlace bitsandbytes/cuda_setup/main.py \
-      --replace "libcuda.so" "${nvidia_x11}/lib/libcuda.so"
+  postPatch = ''
+    substituteInPlace Makefile --replace "/usr/bin/g++" "g++" --replace "lib64" "lib"
+    substituteInPlace bitsandbytes/cuda_setup/main.py  \
+      --replace "binary_path = package_dir / binary_name"  \
+                "binary_path = Path('$out/${python.sitePackages}/${pname}')/binary_name"
+  '' + lib.optionalString torch.cudaSupport ''
+    substituteInPlace bitsandbytes/cuda_setup/main.py  \
+      --replace "/usr/local/cuda/lib64" "${cuda-native-redist}/lib"
   '';
 
-  propagatedBuildInputs = [
-    (pytorch.override({ inherit cudaSupport;}))
-    setuptools
-  ] ++ lib.optionals cudaSupport [
-    typer
-    cudatoolkit
-    cudatoolkit.lib
-    nvidia_x11
-  ];
+  CUDA_HOME = "${cuda-native-redist}";
 
-  nativeBuildInputs = lib.optionals cudaSupport [
-    gcc
-    addOpenGLRunpath
-  ];
+  preBuild = if torch.cudaSupport then
+    with torch.cudaPackages;
+    let cudaVersion = lib.concatStrings (lib.splitVersion torch.cudaPackages.cudaMajorMinorVersion); in
+    ''make CUDA_VERSION=${cudaVersion} cuda${cudaMajorVersion}x''
+  else
+    ''make CUDA_VERSION=CPU cpuonly'';
 
-  preFixup = lib.optionalString cudaSupport ''
-    find $out -type f \( -name '*.so' -or -name '*.so.*' \) | while read lib
-    do
-      addOpenGLRunpath "$lib"
-      patchelf \
-        --set-rpath "${libraryPath}" \
-        "$lib"
-    done
-  '';
+  nativeBuildInputs = [ setuptools ] ++ lib.optionals torch.cudaSupport [ cuda-native-redist ];
+  buildInputs = lib.optionals torch.cudaSupport [ cuda-redist ];
 
-  checkInputs = [
-    pytest
+  propagatedBuildInputs = [
+    torch
   ];
 
-  # disabled because the test suite cannot find any GPUs in the nix sandbox
-  doCheck = false;
+  doCheck = false;  # tests require CUDA and also GPU access
+  nativeCheckInputs = [ pytestCheckHook einops lion-pytorch scipy ];
 
   pythonImportsCheck = [
-    "torch"
-    # this tries to load and verify the cuda device on import, since this runs
-    # in the nix sandbox it will fail with a bunch of errors, but curiously the
-    # build still succeeds...
     "bitsandbytes"
   ];
-}
 
+  meta = with lib; {
+    homepage = "https://github.com/TimDettmers/bitsandbytes";
+    description = "8-bit CUDA functions for PyTorch";
+    license = licenses.mit;
+    maintainers = with maintainers; [ bcdarwin ];
+  };
+}