Initial support for ppc64le

[mgiessing]

Initial support for PowerPC architecture following the design pattern introduced by the aarch64 PR

[matthewdouglas]

Thanks @mgiessing! I am not sure that we'll be able to commit to distributing binary wheels for ppc64le, but certainly welcome source compatibility!

I'll do some quick regression testing on this over the next couple days, but at first glance it looks good!

PowerPC support was deprecated in CUDA 12.4 and removed in 12.5. However my understanding is that there's still many AC922 systems with V100 GPUs out there, and maybe even some Power8 S822LC with P100s, so to add context and further advocate:

Example operational supercomputers:

• Summit / OLCF-4
• Sierra / ATS-2
• MareNostrum 4 / MN4 CTE-Power

Just a reference note: relates to #652

[mgiessing]

Thanks a lot @matthewdouglas!
Yeah - I do not expect to have binary wheels but as you said there are many people/organisations having P100/V100/T4 on Power Systems so source compatibility is desired :-)

Btw. I figured out I had to rename the libbitsandbytes_cuda122.so to libbitsandbytes_cuda122_nocublaslt.so otherwise it would crash during the test...not sure I've done something wrong during the build which was the following:

## System: AC922 // CUDA 12.2 // RHEL8.9

# Create bnb environment and install dependencies via mamba/conda and rocketce channel
micromamba create -n bnb \
    -c rocketce \
    -c defaults \
    python=3.10 \
    pytorch==2.1.2 \
    pandas \
    scipy \
    matplotlib && micromamba clean --all --yes

micromamba activate bnb

#Install remaining depenedencies via pypi
pip3 install lion-pytorch wheel einops pytest setuptools>=63 transformers accelerate

git clone https://github.com/mgiessing/bitsandbytes

export PATH=$PATH:/usr/local/cuda/bin

cmake -DCMAKE_CUDA_ARCHITECTURES=70 -DCOMPUTE_BACKEND=cuda -S .

make -j$(nproc)
pip install -e .

cp bitsandbytes/libbitsandbytes*.so bitsandbytes/libbitsandbytes_cuda122_nocublaslt.so

#Simple test to check if it works
python3 -m bitsandbytes

[matthewdouglas]

If you add -DNO_CUBLASLT=ON to the cmake step it will build libbitsandbytes_cuda122_nocublaslt.so.