Arch Linux User Repository

Search Criteria

Enter search criteria

Search by

Keywords

Out of Date

Sort by

Sort order

Per page

Package Details: python-bitsandbytes-git 0.44.1.11.g9264f02-1

Package Actions

Git Clone URL:	https://aur.archlinux.org/python-bitsandbytes-git.git (read-only, click to copy)
Package Base:	python-bitsandbytes-git
Description:	Lightweight wrapper around CUDA custom functions, in particular 8-bit optimizers, matrix multiplication (LLM.int8()), and quantization functions.
Upstream URL:	https://github.com/TimDettmers/bitsandbytes
Licenses:	MIT
Conflicts:	python-bitsandbytes
Provides:	python-bitsandbytes
Submitter:	Premik
Maintainer:	Premik
Last Packager:	Premik
Votes:	0
Popularity:	0.000000
First Submitted:	2023-12-26 15:48 (UTC)
Last Updated:	2024-11-15 22:19 (UTC)

Dependencies (12)

cuda (cuda11.1^AUR, cuda-12.2^AUR, cuda12.0^AUR, cuda11.4^AUR, cuda11.4-versioned^AUR, cuda12.0-versioned^AUR, cuda-12.5^AUR, cuda-12.9^AUR)
python
python-accelerate^AUR
python-einops^AUR
python-lion-pytorch^AUR
python-pytorch (python-pytorch-cxx11abi^AUR, python-pytorch-cxx11abi-opt^AUR, python-pytorch-cxx11abi-cuda^AUR, python-pytorch-cxx11abi-opt-cuda^AUR, python-pytorch-cxx11abi-rocm^AUR, python-pytorch-cxx11abi-opt-rocm^AUR, python-pytorch-cuda12.9^AUR, python-pytorch-opt-cuda12.9^AUR, python-pytorch-cuda, python-pytorch-opt, python-pytorch-opt-cuda, python-pytorch-opt-rocm, python-pytorch-rocm)
python-scipy (python-scipy-git^AUR, python-scipy-mkl-bin^AUR, python-scipy-mkl-tbb^AUR, python-scipy-mkl^AUR)
python-transformers^AUR
git (git-git^AUR, git-gl^AUR) (make)
python-build (make)
python-installer (make)
python-pytest (check)

Required by (3)

python-pytorch-lightning (requires python-bitsandbytes) (optional)
python-transformers (requires python-bitsandbytes) (optional)
python-trl (requires python-bitsandbytes) (optional)

Sources (1)

python-bitsandbytes-git

Latest Comments

arzeth commented on 2025-07-22 14:35 (UTC) (edited on 2025-07-22 14:36 (UTC) by arzeth)

First of all, python-scikit-build-core must be added now to makedepends=(...) to avoid ERROR Backend 'scikit_build_core.setuptools.build_meta' is not available. after the compilation.

And with the latest CUDA (12.9.0 and 12.9.1), there's now:

error: macro "_CCCL_PP_SPLICE_WITH_IMPL1" passed 3 arguments, but takes just 2

because of a bug in nvcc: https://github.com/NVIDIA/cccl/issues/4967

Solutions:

1) add -DCOMPUTE_CAPABILITY="61;62;70;72;75;80;86;87;89;90;100;101;120;103;121" to the cmake line (I delisted the five earliest archs to avoid hitting the unintended limit of archs in nvcc).

2) or fix nvcc itself:

curl -s https://github.com/NVIDIA/cccl/commit/3878bea049b71142ccd6041836efbd70ab2b3b2c.patch | sudo patch /opt/cuda/targets/x86_64-linux/include/cuda/std/__cccl/preprocessor.h

arzeth commented on 2024-04-18 19:32 (UTC) (edited on 2024-04-18 21:36 (UTC) by arzeth)

I had to replace the make line with

cmake -G Ninja -DCUDA_VERSION=$BNB_CUDA_VERSION -DCOMPUTE_BACKEND=cuda -DCOMPUTE_CAPABILITY="50;52;60;61;70;75;80;86;89;90" -DNO_CUBLASLT=0 -DCMAKE_BUILD_TYPE=Release -S .
cmake --build . --config Release

(50;52;... means compile code for each of the specified CUDA Capability versions, which can be found by running /opt/cuda/extras/demo_suite/deviceQuery, e.g. GTX 1660 Super is 7.5)

For non-CUDA users there are less args:

cmake -G Ninja -DCOMPUTE_BACKEND=cpu -DCMAKE_BUILD_TYPE=Release -S .
cmake --build . --config Release