Build instructions for ROCM

Matthias_Moller · July 3, 2024, 5:38am

I wonder if it would be possible to share the build instructions/configurations for the official ROCM releases of LibTorch (the C++ API not PyTorch). Also, and that is the main reason why I need to compile LibTorch from source, would it be possible to compile future releases with MPI/Distributed enabled?

Thanks

fluidnumerics_joe · March 5, 2025, 7:31pm

Hey @Matthias_Moller ,
I know I’m a bit late to the game here, but I found this on the ROCm fork of Pytorch that should address the first part of your question

github.com/rocm/pytorch

docs/libtorch.rst

main

libtorch (C++-only)
===================

The core of pytorch does not depend on Python. A
CMake-based build system compiles the C++ source code into a shared
object, libtorch.so.

AMD ROCm Support
------------------------------
If you're compiling for AMD ROCm then first run this command:
::
   cd <pytorch_root>

   # Only run this if you're compiling for ROCm
   python tools/amd_build/build_amd.py

Additional information about ROCm support can be found in the top-level
`README <https://github.com/pytorch/pytorch/blob/main/README.md>`_.

Building libtorch using Python

This file has been truncated. show original

This covers building libtorch for AMD GPUs in the top section. I’m planning on giving this a try.

For your second question, are you wanting the official ROCm release built with MPI enabled ? That might be a question best opened on an issue at GitHub · Where software is built

Matthias_Moller · March 17, 2025, 7:59pm

Thanks very much @fluidnumerics_joe. I’ll make an issue about building the official ROCm release with MPI enabled.