Can't build hasktorch-0.2.0.0 into external cabal project (linux, cuda)

[melendezd]

I was able to get the examples to work just fine (static typed MNIST CNN), but I'm having trouble building hasktorch-0.2.0.0 into a separate cabal project.
Steps:

• Clone hasktorch
• Run get-deps.h in hasktorch/deps/
• Copy hasktorch/deps/libtorch/include/* into /usr/include
• Copy hasktorch/deps/libtorch/lib/* into /lib (not sure if this is necessary with the setenv script)
• Source setenv in hasktorch
• Run hasktorch/setup-cabal.sh in hasktorch
• cabal build hasktorch
  Hasktorch builds successfully on its own.

I then create a new cabal project called ml in a separate folder. Here's my ml.cabal file:

cabal-version:      2.4  
name:               ml  
version:            0.1.0.0  
executable ml  
    main-is:          Main.hs  
    build-depends:    base ^>=4.14.2.0, hasktorch==0.2.0.0  
    hs-source-dirs:   app  
    default-language: Haskell2010

I also create a cabal.project file:

packages: .
          /home/mapa/git/hasktorch/libtorch-ffi-helper
          /home/mapa/git/hasktorch/libtorch-ffi
          /home/mapa/git/hasktorch/hasktorch

Then, I run cabal build and get the following error:

Resolving dependencies...
Build profile: -w ghc-8.10.5 -O1
In order, the following will be built (use -v for more details):
 - libtorch-ffi-1.9.0.0 (lib) (first run)
 - hasktorch-0.2.0.0 (lib:hasktorch) (first run)
 - ml-0.1.0.0 (exe:ml) (first run)
Preprocessing library for libtorch-ffi-1.9.0.0..
Building library for libtorch-ffi-1.9.0.0..
[ 45 of 105] Compiling Torch.Internal.Unmanaged.Optim ( src/Torch/Internal/Unmanaged/Optim.hs, /home/mapa/projects/ml/dist-newstyle/build/x86_64-linux/ghc-8.10.5/libtorch-ffi-1.9.0.0/build/Torch/Internal/Unmanaged/Optim.o, /home/mapa/projects/ml/dist-newstyle/build/x86_64-linux/ghc-8.10.5/libtorch-ffi-1.9.0.0/build/Torch/Internal/Unmanaged/Optim.dyn_o )

/tmp/ghc6075_0/ghc_5.cpp:4:10: error:
     fatal error: torch/types.h: No such file or directory
        4 | #include <torch/types.h>
          |          ^~~~~~~~~~~~~~~
  |
4 | #include <torch/types.h>
  |          ^
compilation terminated.
`gcc' failed in phase `C Compiler'. (Exit code: 1)
cabal: Failed to build libtorch-ffi-1.9.0.0 (which is required by exe:ml from
ml-0.1.0.0).

Am I approaching this wrong?

[austinvhuang]

@melendezd Is stack an option? If so the template here https://github.com/hasktorch/hasktorch-stack-skeleton takes care of the hasktorch git dependency using stack.yaml.

We'd like to have a cabal skeleton as well, but nobody has gotten to it yet (there's a nix one as well https://github.com/hasktorch/hasktorch-skeleton). If you need cabal maybe the troubleshooting here could be the basis for one.

[junjihashimoto]

In your use case, include path is not enough.
libtorch needs two include paths.
One is libtorch/include. Another is libtorch/include/torch/csrc/api/include.
This command may be helpful.
https://github.com/hasktorch/libtorch-binary-for-ci/blob/master/Makefile#L40

I recommend to use above skeltons.

[Reep236]

Heads up for hasktorch-skeleton, not sure why this is yet, but (in my experience) the .cabal file must be named hasktorch-skeleton.cabal or cabal will fail to locate it (even if the glob *.cabal is used). I haven't had the chance to figure out why this is yet. Also watch out for using Cuda 11, as the libtorch nix overlay has the incorrect version 11.0 instead of 11.1. I fixed the second issue in my local copy but never got to submitting a PR for it. Will do ASAP.

I'm really not sure about the first one. Maybe someone else has some more insight, but I'll hopefully have time to look at it some more later today.