CUMM-GPU

This repo is just a fork from cumm — CUDA Matrix Multiply library.

⬇️ Install

Before you install, you must uninstall all existing versions to avoid conflicts:

pip uninstall -y cumm cumm-cu130 cumm-cu128 cumm-cu126 cumm-cu121 

For easy installation using prebuilt wheels (example for CUDA 13.0):

pip install --extra-index-url https://ratharog.github.io/cumm-spconv/ cumm-cu130

See ratharog.github.io/cumm-spconv for prebuilt wheel info and supported CUDA versions.

🛠️ Build and install from source

Building from source requires:

• A C++ compiler
• CUDA Toolkit installed (for CUDA support)

On Linux with CUDA 13.0:

export CUMM_CUDA_VERSION="13.0"
export CUMM_CUDA_ARCH_LIST="all"
export CUMM_DISABLE_JIT="1"
pip install git+https://github.com/rathaROG/cumm-gpu.git

On Windows with CUDA 13.0 + MSVC 2019 (cmd terminal):

git clone https://github.com/rathaROG/cumm-gpu.git
cumm-gpu\tools\msvc_setup.bat
RMDIR /s /q cumm-gpu
SET CUMM_CUDA_VERSION=13.0
SET CUMM_CUDA_ARCH_LIST=all
SET CUMM_DISABLE_JIT=1
pip install git+https://github.com/rathaROG/cumm-gpu.git

On any platform with CPU only:

pip install git+https://github.com/rathaROG/cumm-gpu.git

Note:

• For detailed build logs, add --verbose to the pip install command.
• On Windows, you can also use the scripts below to build and install from source; feel free to modify:
  • For CUDA 12.8: install_cumm_cu128.bat
  • For CUDA 13.0: install_cumm_cu130.bat

📝 License