hasktorch-stack-skeleton

Creating a hasktorch project

To use hasktorch, you need to include the libtorch precompiled libraries and reference a hasktorch commit in stack.yaml extra deps.

If you encounter linker errors you may need to specify a path to libtorch in the environment variable LD_LIBRARY_PATH. In this case you may want to download libtorch to ~/.libtorch and then run cp -rs $HOME/.libtorch libtorch inside your project. This will allow you to write

export LD_LIBRARY_PATH=$HOME/.libtorch/lib:$LD_LIBRARY_PATH

in one of your shell configuration files, instead of the relative path $PWD/libtorch mentioned below.

1) libtorch

Download the libtorch dependency in ./libtorch by downloading an archive here e.g. cpu-libtorch-cxx11-abi-shared-with-deps-latest.zip

$ curl -L https://github.com/hasktorch/libtorch-binary-for-ci/releases/download/1.8.0/cpu-libtorch-cxx11-abi-shared-with-deps-latest.zip >> libtorch.zip

$ unzip libtorch.zip

This will extract the libtorch dependencies inside ./libtorch/. In some cases, you may encounter linking errors when using hasktorch inside stack ghci, and you may need to do:

$ export LD_LIBRARY_PATH=$(pwd)/libtorch/lib:$LD_LIBRARY_PATH

before running stack commands.

Q: Can we put libtorch in a single location instead and replace by /path/to/libtorch/lib in bashrc instead? => yes, use symlink to libtorch inside the project and absolute path for the environment

2) hasktorch

Reference hasktorch 2 by git commit +libtorch in stack.yaml extra-deps:

# ./stack.yaml

resolver: lts-17.4

packages:
- . 

extra-deps:
  - git: https://github.com/hasktorch/hasktorch.git
    commit: 4e846fdcd89df5c7c6991cb9d6142007a6bb0a58
    subdirs:
      - hasktorch
      - libtorch-ffi
      - libtorch-ffi-helper

extra-include-dirs:
  - libtorch/include/torch/csrc/api/include
  - libtorch/include

extra-lib-dirs:
  - ./libtorch/lib

Reference hasktorch as a dependency in package.yaml too:

# ./package.yaml 
# ...
dependencies:
- base >= 4.7 && < 5
- hasktorch

Then you should be good to go and load the library:

$ stack build
$ stack ghci
ghci> import Torch
ghci> :t asTensor [1, 2, 3]

Troubleshooting

In case you can load the library but get linker errors like

λ> Torch.exp t
INTEL MKL ERROR: /opt/intel/mkl/lib/intel64/libmkl_vml_avx2.so: undefined symbol: mkl_blas4vml_dptrmm.
Intel MKL FATAL ERROR: cannot load libmkl_vml_avx2.so or libmkl_vml_def.so.

You may have missing libtorch links and should try fixing $LD_LIBRARY_PRELOAD as mentioned above.

See also the contents of the hasktorch setenv script here.

Otherwise contact the team on their slack channel.

Documentation

Once you're good to go, follow on to the tutorial here :)