Make sure that nnabla and nnabla-ext-cuda
are cloned on the same level directory.
You needs NVIDIA Docker and GNU make
Note: workdir in "nnabla-ext-cuda/"
$ cd nnabla-ext-cuda
$ make all
make all
is as same as make bwd-cpplib bwd-wheel
After this you can find .whl file in ../nnabla/build_wheel/dist/
and build_wheel/dist/
$ cd nnabla-ext-cuda
$ make bwd-cpplib
Prepare to specify CUDA, cuDNN, and python version.
$ export PYTHON_VERSION_MAJOR=3
$ export PYTHON_VERSION_MINOR=10
$ export CUDA_VERSION_MAJOR=11
$ export CUDA_VERSION_MINOR=6.2
$ export CUDNN_VERSION=8
Then you can get with,
$ cd nnabla-ext-cuda
$ make all
Or you can specify every time.
$ cd nnabla-ext-cuda
$ make PYTHON_VERSION_MAJOR=3 PYTHON_VERSION_MINOR=10 CUDA_VERSION_MAJOR=11 CUDA_VERSION_MINOR=6.2 CUDNN_VERSION=8 all
Now, this section is same as Build CUDA extension on Windows