libTorch M1 build - duplicate symbols and other errors

Aaron_Leese · March 9, 2023, 3:21pm

Hi all!

I’m trying to compile libTorch, but finding errors no matter how I build it.
This is on an M1 mac.

I’ve tried all the various recommend methods from the official docs, here:

pytorch/pytorch/blob/master/docs/libtorch.rst

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.

Building libtorch using Python
------------------------------

You can use a python script/module located in tools package to build libtorch
::
   cd <pytorch_root>

   # Make a new folder to build in to avoid polluting the source directories
   mkdir build_libtorch && cd build_libtorch

   # You might need to export some required environment variables here.
   Normally setup.py sets good default env variables, but you'll have to do
   that manually.

This file has been truncated. show original

:::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
Trying the first method:
python ../tools/build_libtorch.py

I get a bunch of undefined symbols error:


Undefined symbols for architecture arm64:
  "_MTLCopyAllDevices", referenced from:
Undefined symbols for architecture arm64:
  "_MTLCopyAllDevices", referenced from:
      at::mps::MPSDevice::MPSDevice() in libtorch_cpu.a(MPSDevice.mm.o)
Undefined symbols for architecture arm64:
  "_MTLCopyAllDevices", referenced from:
      at::mps::MPSDevice::MPSDevice() in libtorch_cpu.a(MPSDevice.mm.o)
      at::mps::MPSDevice::MPSDevice() in libtorch_cpu.a(MPSDevice.mm.o)

Guessing that maybe some subcomponent is compiling for x86_64 instead of arm? Maybe there is a flag necessary to force arm/M1 compilation?

:::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::

Using either other method (CMAKE or Setup.py, both outlined in the docs) it seems to get further (92%) but then dies in a linking phase with a bunch of duplicate symbols errors like this:

**duplicate symbol 'at::native::DispatchStub<void (*)(at::TensorIteratorBase&, c10::Scalar const&), at::native::leaky_relu_backward_stub>::DEFAULT' in:
    CMakeFiles/torch_cpu.dir/__/aten/src/ATen/native/cpu/Activation.cpp.DEFAULT.cpp.DEFAULT.cpp.o
    CMakeFiles/torch_cpu.dir/__/aten/src/ATen/native/cpu/Activation.cpp.DEFAULT.cpp.o
**

Looking at those files, they both just include the “Activation.cpp”. Why exactly you would have 2 files both including some other file is not something I understand, so maybe someone else knows how to make sense of this?