If you are experiencing problems building or using FTorch please see below for guidance on common problems.
If possible we recommend using the Windows Subsystem for Linux (WSL) to build the library. In this case the build process is the same as for a Linux environment.
If you need to build in native Windows please read the following information:
It is possible to build using Visual Studio and the Intel Fortran Compiler. In this case you must install the following:
You will then need to load the intel Fortran compilers using setvars.bat which is found in the Intel compiler install
directory (see the intel
docs)
for more details.
From cmd this can be done with:
call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat"
Finally you will need to add -G "NMake Makefiles" to the cmake command in the
regular install instructions.
So the basic command to build from cmd becomes:
cmake -G "NMake Makefiles" -DCMAKE_PREFIX_PATH="C:\Users\<path-to-libtorch-download>\libtorch" -DCMAKE_BUILD_TYPE=Release ..
cmake --build .
cmake --install .
The following is an example cmd script that installs FTorch and runs the integration tests. It assumes you have already
install cmake, git, the intel compilers and visual studio.
rem disable output for now
ECHO ON
rem load intel compilers
call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat"
rem download ftorch
git clone https://github.com/Cambridge-ICCS/FTorch.git
cd FTorch
rem make venv
python -m venv .ftorch
rem activate the environment
call .ftorch\Scripts\activate
rem install torch
pip install torch torchvision torchaudio
rem enable output
ECHO ON
rem run cmake to generate build scripts
rem (update CMAKE_PREFIX_PATH depending on location of ftorch venv)
cd src
cmake -Bbuild -G "NMake Makefiles" -DCMAKE_Fortran_FLAGS="/fpscomp:logicals" ^
-DCMAKE_PREFIX_PATH="C:\Users\Quickemu\Downloads\FTorch\.ftorch\Lib\site-packages" ^
-DCMAKE_BUILD_TYPE=Release ^
-DCMAKE_BUILD_TESTS=True ^
-DCMAKE_Fortran_COMPILER=ifx -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx
rem build and install ftorch
cmake --build build
cmake --install build
rem quit if this raises an error
if %errorlevel% neq 0 exit /b %errorlevel%
ECHO OFF
rem add ftorch and pytorch libs to path
rem (update these depending on where you installed ftorch and where you created the venv)
set PATH=C:\Users\Quickemu\Downloads\FTorch\.ftorch\Lib\site-packages;%PATH%
set PATH=C:\Program Files (x86)\FTorch\bin;%PATH%
set PATH=C:\Users\Quickemu\Downloads\FTorch\.ftorch\Lib\site-packages\torch\lib;%PATH%
cd ..
rem run integration tests
ECHO ON
run_integration_tests.bat
if %errorlevel% neq 0 exit /b %errorlevel%
We would also recommend Windows users to review the Windows CI workflow (.github/workflows/test_suite_windows.yml) for more
information, as this provides another example of how to build and run FTorch and its integration tests.
If using powershell the setvars and build commands become:
cmd /k '"C:\Program Files (x86)\Intel\oneAPI\setvars.bat" && powershell'
cmake -G "NMake Makefiles" -DCMAKE_PREFIX_PATH="C:\Users\<path-to-libtorch-download>\libtorch" -DCMAKE_BUILD_TYPE=Release ..
cmake --build .
cmake --install .
It may be tempting to build on Windows using MinGW. However, libtorch does not currently support MinGW. Instead please build using Visual Studio and the intel Fortran compiler (ifort) as detailed in the project README.
At the time of writing, libtorch is currently only officially available for x86
architectures (according to pytorch.org).
However, the version of PyTorch provided by pip install provides an ARM binary
for libtorch which works on Apple Silicon.
Therefore you should pip install torch in this situation and follow the guidance
on locating Torch within a virtual environment (venv) for CMake.
The reason input and output tensors to torch_model_forward are
contained in arrays is because it is possible to pass multiple input tensors to
the forward() method of a torch net, and it is possible for the net to return
multiple output arrays.
The nature of Fortran means that it is not possible to set an arbitrary number
of inputs to the torch_model_forward subroutine, so instead we use a single
array of input tensors which can have an arbitrary length. Similarly, a single
array of output tensors is used.
Note that this does not refer to batching data. This should be done in the same way as in Torch; by extending the dimensionality of the input tensors.
By default we disable gradient calculations for tensors and models and place models in
evaluation mode for efficiency.
These can be adjusted using the requires_grad and is_training optional arguments
in the Fortran interface. See the API procedures documentation
for details.
ftorch for large tensors?Currently FTorch represents the number of elements in an array dimension using
32-bit integers. For most users this will be more than enough, but if your code
uses large tensors (where large means more than 2,147,483,647 elements
in any one dimension (the maximum value of a 32-bit integer)), you may you may
need to compile ftorch with 64-bit integers. If you do not, you may receive a
compile time error like the following:
39 | call torch_tensor_from_array(tensor, in_data, tensor_layout, torch_kCPU)
| 1
Error: There is no specific subroutine for the generic ‘torch_tensor_from_array’ at (1)
To fix this, we can build ftorch with 64-bit integers. We need to modify this
line in src/ftorch.fypp
integer, parameter :: ftorch_int = int32 ! set integer size for FTorch library
We can use 64-bit integers by changing the above line to this
integer, parameter :: ftorch_int = int64 ! set integer size for FTorch library
Finally, you will need to run fypp (fypp is not a core dependency, so you
may need to install this separately) e.g.,
fypp src/ftorch.fypp src/ftorch.F90