- installed Jupyter notebook
- installed
conda, e.g./opt/conda - built PyTorch
- from source
- libtorch
- installed via conda (but can have an issue with
_GLIBCXX_USE_CXX11_ABI)
There can be an issue if pytorch is built with _GLIBCXX_USE_CXX11_ABI=0, but xeus-cling assumes _GLIBCXX_USE_CXX11_ABI=1.
Please, make sure that
python -c "import torch; print(torch._C._GLIBCXX_USE_CXX11_ABI)"
> True
Otherwise, there will be linker errors with methods using std::string.
Install xeus-cling into conda env
(base) $ conda create -n cling
(base) $ conda activate cling
(cling) $ conda install xeus-cling -c conda-forge(cling) $ conda deactivate
(base) $ jupyter kernelspec install /opt/conda/envs/cling/share/jupyter/kernels/xcpp11 --sys-prefix
(base) $ jupyter kernelspec install /opt/conda/envs/cling/share/jupyter/kernels/xcpp14 --sys-prefix
(base) $ jupyter kernelspec install /opt/conda/envs/cling/share/jupyter/kernels/xcpp17 --sys-prefixRefresh Jupyter notebook page
Let's assume that PyTorch is built from source:
!ls /workspace/pytorch/torch/lib
> libtorch.so
> ...
Add the following cells to your C++ notebook
// If you want to add include path
#pragma cling add_include_path("/workspace/pytorch/torch/include")
#pragma cling add_include_path("/workspace/pytorch/torch/include/torch/csrc/api/include")
// If you want to add library path
#pragma cling add_library_path("/workspace/pytorch/torch/lib")
// If you want to load library
#pragma cling load("libtorch")
#pragma cling load("libtorch_cpu")
#pragma cling load("libc10")
// #pragma cling load("libtorch_cuda")
// #pragma cling load("libc10_cuda")If everything is set up correctly, the following code should work
#include <iostream>
#include <ATen/ATen.h>
auto p = at::CPU(at::kFloat);
std::cout << p << "\n";
auto t = at::ones({3, 4}, p);
std::cout << t << "\n";