Adding support for building for AMD on Windows

[rwfsmith]

Some pretty simple changes to get bitsandbytes to build for AMD ROCm on Windows. I've been running this for a few days for inference and training, and it seems to be working well. Running unit tests now.

How to build this:

Clone repo

# Install bitsandbytes from source
# Clone bitsandbytes repo
git clone https://github.com/bitsandbytes-foundation/bitsandbytes.git && cd bitsandbytes/

Build and install

Build doesn't work with the Visual Studio compiler, so make sure ROCM_HOME is set to your ROCm root folder and that Ninja is installed
This should work with ROCm 7.1.1 or newer, including the TheRock nightly builds

PowerShell

cmake -B build -G Ninja `
-DCOMPUTE_BACKEND=hip -DCMAKE_BUILD_TYPE=Release `
-DCMAKE_CXX_COMPILER="$($env:ROCM_HOME)/lib/llvm/bin/clang++.exe" `
-DCMAKE_HIP_COMPILER="$($env:ROCM_HOME)/lib/llvm/bin/clang++.exe" ` 
-DCMAKE_PREFIX_PATH="$($env:ROCM_HOME)" `
-DBNB_ROCM_ARCH="gfx1151"

cmake --build build --config Release

Cmd

cmake -B build -G Ninja ^
-DCOMPUTE_BACKEND=hip ^
-DCMAKE_BUILD_TYPE=Release ^
-DCMAKE_CXX_COMPILER="%ROCM_HOME%/lib/llvm/bin/clang++.exe" ^
-DCMAKE_HIP_COMPILER="%ROCM_HOME%/lib/llvm/bin/clang++.exe" ^
-DCMAKE_PREFIX_PATH="%ROCM_HOME%" ^
-DBNB_ROCM_ARCH="gfx1151"

cmake --build build --config Release

Then install

pip install -e .

or

pip wheel -e . -w dist

resolves #1844

[rwfsmith]

164 unit tests failed, seems to be accuracy issues with test_gemv_4bit