Lptorch
An extention of pytorch for low precision training / inference
Install / Use
/learn @sunwoolee/LptorchREADME
lptorch
lptorch is CUDA extension library of pytorch for low precision neural network training / inference.
lptorch only supports $2^N$ scale factor which can be implemented with bit shift operation in hardware.
Install_requires
- torch (with cuda)
- onnx
Installation
git clone https://github.com/sunwoolee/lptorch.git
cd lptorch
make clean
make run
Installation with multiple CUDA architectures
Check your GPU compute capability at https://developer.nvidia.com/cuda-gpus#compute
export TORCH_CUDA_ARCH_LIST="{your GPU compute capability}"
# export TORCH_CUDA_ARCH_LIST="6.1 7.5 8.6"
Follow Installation in the latest cuda architecture environment.
See the wiki for more information about the package.
