AutoAWQ Kernels is a new package that is split up from the main repository in order to avoid compilation times.
-
Windows: Must use WSL2.
-
NVIDIA:
- GPU: Must be compute capability 7.5 or higher.
- CUDA Toolkit: Must be 11.8 or higher.
-
AMD:
- ROCm: Must be 5.6 or higher. Build from source
The package is available on PyPi with CUDA 12.4.1 wheels:
pip install autoawq-kernels
To build the kernels from source, you first need to setup an environment containing the necessary dependencies.
- Python>=3.8.0
- Numpy
- Wheel
- PyTorch
- ROCm: You need to install the following packages
rocsparse-dev hipsparse-dev rocthrust-dev rocblas-dev hipblas-dev
.
pip install git+https://github.com/casper-hansen/AutoAWQ_kernels.git
Notes on environment variables:
-
TORCH_VERSION
: By default, we build using the current version of torch bytorch.__version__
. You can override it withTORCH_VERSION
.-
CUDA_VERSION
orROCM_VERSION
can also be used to build for a specific version of CUDA or ROCm.
-
-
CC
andCXX
: You can specify which build system to use for the C code, e.g.CC=g++-13 CXX=g++-13 pip install -e .
-
COMPUTE_CAPABILITIES
: You can specify specific compute capabilities to compile for:COMPUTE_CAPABILITIES="75,80,86,87,89,90" pip install -e .