CUda Matrix Multiply library.
cumm is developed during learning of CUTLASS, which use too much c++ template and make code unmaintainable. So I develop pccm, use python as meta programming language, to replace c++ template meta programming.
Now pccm become a foundational framework of cumm and my other c++ project such as spconv.
cumm also contains a python asyncio-based gemm simulator that share same meta program with CUDA code, enable gemm visualization and easy debug experience.
- 0.3.1: tv::DType enum value changed, this will affect all binary code of tv::Tensor user. you must recompile all code if upgrade to cumm >= 0.3.1.
- Ampere feature support (by EvernightAurora)
We offer python 3.9-3.13 and cuda 11.4/11.8/12.1/12.4/12.6 prebuilt binaries for linux (manylinux_2_28).
We offer python 3.9-3.13 and cuda 11.4/11.8/12.1/12.4/12.6 prebuilt binaries for windows 10/11.
We offer python 3.9-3.13 prebuilt binaries for Mac OS X >= 14.0 (Apple Silicon Only).
pip install cumm for CPU-only
pip install cumm-cu114 for CUDA 11.4
pip install cumm-cu126 for CUDA 12.6
WARNING Use code in tags!!! code in main branch may contain bugs.
The c++ code will be built automatically when you change c++ code in project.
- uninstall cumm installed by pip. you must ensure no "cumm" exists in
pip list | grep cumm - install build-essential, install CUDA
git clone https://github.com/FindDefinition/cumm,cd ./cumm,git checkout tags/<tag_name>,pip install -e .- in python,
import cummand wait for build finish.
- uninstall spconv and cumm installed by pip. you must ensure no "cumm" exists in
pip list | grep cumm - install visual studio 2019 or newer. make sure C++ development component is installed. install CUDA
- set powershell script execution policy
- start a new powershell, run
tools/msvc_setup.ps1 git clone https://github.com/FindDefinition/cumm,cd ./cumm,git checkout tags/<tag_name>,pip install -e .- in python,
import cummand wait for build finish.
WARNING Use code in tags!!! code in main branch may contain bugs.
WARNING: If CUMM_CUDA_VERSION is set with a CUDA version, following steps will create a wheel named "cumm-cuxxx", not "cumm", this means you must use cumm-cuxxx in dependency of your project which depend on cumm, not cumm. If CUMM_CUDA_VERSION isn't set, cumm will always built with CUDA, so the CUDA must exists in your system. The wheel name will be cumm even if it is built with cuda.
It's recommend to build Linux packages in official build docker. Build with CUDA support don't need a real GPU.
- select a cuda version. available: CUDA 11.1, 11.3, 11.4, 11.5, 12.0
- (Example for CUDA 11.4)
git clone https://github.com/FindDefinition/cumm,cd ./cumm,docker run --rm -e PLAT=manylinux2014_x86_64 -e CUMM_CUDA_VERSION=114 -v `pwd`:/io scrin/manylinux2014-cuda:cu114-devel-1.0.0 bash -c "source /etc/bashrc && /io/tools/build-wheels.sh"
- install build-essential, install CUDA
- set env for installed cuda version. for example,
export CUMM_CUDA_VERSION="11.4". If you want to build CPU-only, runexport CUMM_CUDA_VERSION="". IfCUMM_CUDA_VERSIONisn't set, you need to ensure cuda libraries are inside OS search path, and the built wheel name will becumm, otherwisecumm-cuxxx - run
export CUMM_DISABLE_JIT="1" - run
python setup.py bdist_wheel+pip install dists/xxx.whl
- install visual studio 2019 or newer. make sure C++ development package is installed. install CUDA
- set powershell script execution policy
- start a new powershell, run
tools/msvc_setup.ps1 - set env for installed cuda version. for example,
$Env:CUMM_CUDA_VERSION = "11.4". If you want to build CPU-only, run$Env:CUMM_CUDA_VERSION = "". . IfCUMM_CUDA_VERSIONisn't set, you need to ensure cuda libraries are inside OS search path, and the built wheel name will becumm, otherwisecumm-cuxxx - run
$Env:CUMM_DISABLE_JIT = "1" - run
python setup.py bdist_wheel+pip install dists/xxx.whl
- EvernightAurora: add ampere feature.
The work is done when the author is an employee at Tusimple.
Apache 2.0