Windows Installation Issue about ftorch

[jatkinson1000]

Hi @vbalmer

Thanks for bringing this to our attention and trying those simple suggestions.

I have now spun up a Windows VM and successfully replicated the problem you are having (and also found an opportunity to improve some of our documentation in #123) which is the first step.

I'll work on finding a solution and get back to you once we have more news.
If you make any progress in the meantime please do let us know!

[jatkinson1000]

Adding notes:

It seems to be that there is an issue locating mkl correctly (math kernel library used for mathematical operations).

This source suggests that it could be because calling the setvars.bat sets a number of environment variables, including CMAKE_PREFIX_PATH required for linking, but we are perhaps the overwriting this when we set -DCMAKE_PREFIX_PATH in the cmake command.

To check this I have tried running CMake with -DTorch_DIR=<path> instead of -DCMAKE_PREFIX_PATH=<path>, but still get the same error.

I tried setting the MKL_ROOT environment variable, but no luck as @vbalmer suggested above:

set MKL_ROOT="C:/Program Files (x86)/Intel/oneAPI/mkl/latest"

I also tried setting it to C:/Program Files (x86)/Intel/oneAPI/mkl which is the actual location on my machine. Still no luck.
Interestingly, however CMake does output a warning saying it is ignoring this variable for compatibility with policy CMP0074:

CMake Warning (dev) at C:/Users/Test/Downloads/libtorch-win-shared-with-deps-2.3.0+cpu/libtorch/share/cmake/Caffe2/public/mkl.cmake:1 (find_package):
  Policy CMP0074 is not set: find_package uses <PackageName>_ROOT variables.
  Run "cmake --help-policy CMP0074" for policy details.  Use the cmake_policy
  command to set the policy and suppress this warning.

  Environment variable MKL_ROOT is set to:

    "C:/Program Files (x86)/Intel/oneAPI/mkl"

  For compatibility, CMake is ignoring the variable.

It seems that despite setting CMake is ignoring the location to link from, so next steps might be to try hadr-coding to see if we can force a link for success at all.

I am running cmake with the --trace flag for more output, but this is a bit too much.
Running cmake build with verbose is helpful, however:

cmake --build . --verbose

@jwallwork23 these may be useful as a starting point if you get chance to look.

[jatkinson1000]

Related issues andf PRs on the PyTorch repo:

• pytorch/pytorch#24334
• pytorch/pytorch#124869
• pytorch/pytorch#124925

Some suggest that it is an issue with the FindMKL file supplied with pytorch/cmake/
Modules/FindMKL.cmake
Tried with the nightly build of libtorch to see if there has been a recent fix, but still the same issue.

Tried setting -DCMAKE_FIND_PACKAGE_PREFER_CONFIG variable to try and force CMake to use the native FindMKL.cmake file rather than the one provided by PyTorch but the issue persists.

It may be that we consider opening an issue on the PyTorch repository as this seems to perhaps be an issue with libtorch.