Libtorch build for ROCM error: “aten/src/THH” not exist

[Fovini]

Issue description

I want to build libtorch for rocm from source and followed the instructions from here:Building libtorch using CMake. It occured this error:

CMake Error at aten/CMakeLists.txt:82 (add_subdirectory):
  add_subdirectory given source "src/THH" which is not an existing directory.
ROCm is enabled.

Directory “pytorch/aten/src/THH” does’t exist but it should be. According to pytorch/aten/CMakeLists.txt:

if(USE_ROCM)
  # TODO: AT_HIP_ENABLED (change this once we represent HIP as HIP in
  # ATen proper)
  set(AT_CUDA_ENABLED 1)
  add_subdirectory(src/THH)
  message("ROCm is enabled.")
elseif(USE_CUDA)
  set(AT_CUDA_ENABLED 1)
  add_subdirectory(src/THC)

If build for cuda, there exists the directory “pytorch/aten/src/THC”, but for rocm, the corresponding directory “THH” does not exist. I‘ve searched several versions pytorch source code, none of them have this directory.
So can someone tell me where can get this lost directory? Or have I missed something to install? I would appreciate for your help with this problem.

• How you installed PyTorch (conda, pip, source): source
• Build command you used (if compiling from source):
  cmake -DBUILD_SHARED_LIBS:BOOL=ON -DCMAKE_BUILD_TYPE:STRING=Release -DPYTHON_EXECUTABLE:PATH=which python3 -D CMAKE_PREFIX_PATH=/opt/rocm-6.1.0 -DCMAKE_INSTALL_PREFIX:PATH=../pytorch-install ../pytorch
• OS: Fedora 37 6.5.12-100.fc37.x86_64
• PyTorch version: 2.3
• Python version: 3.11.6
• ROCM version: 6.1.0
• GCC version (if compiling from source): 12.3.1
• CMake version: 3.27.7

cc @malfet @seemethere @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @dllehr-amd @jataylo @hongxiayang

[jeffdaily]

Looks like the doc you followed doesn't have the ROCm caveat for building from source, and that is you need to generate some of the sources before running the build. There is an extra step of converting the CUDA sources into HIP sources.

# hipify sources
python tools/amd_build/build_amd.py

[hongxiayang]

Thanks for @Fovini reporting the issue and thanks for @jeffdaily for the tips. Sounds like we should do an exercise of building libtorch from source successfully, and then update the documentation for ROCm.