Does not work on Apple Silicon

[zzeee]

1.I have PyTorch 1.10 that works natively on my M1 Mac
I've downloaded libtorch from the website
2. I've added tch = "0.6.1" to cargo.toml
and when run cargo build I see:

note: ld: warning: ignoring file /Users/andrey/prj/nntest/target/debug/build/torch-sys-3c7217b1a617897c/out/libtorch/libtorch/lib/libtorch.dylib, building for macOS-arm64 but attempting to link with file built for macOS-x86_64
ld: warning: ignoring file /Users/andrey/prj/nntest/target/debug/build/torch-sys-3c7217b1a617897c/out/libtorch/libtorch/lib/libtorch_cpu.dylib, building for macOS-arm64 but attempting to link with file built for macOS-x86_64
ld: warning: ignoring file /Users/andrey/prj/nntest/target/debug/build/torch-sys-3c7217b1a617897c/out/libtorch/libtorch/lib/libc10.dylib, building for macOS-arm64 but attempting to link with file built for macOS-x86_64
Undefined symbols for architecture arm64:
"c10::impl::ExcludeDispatchKeyGuard::~ExcludeDispatchKeyGuard()", referenced from:
at::AutoDispatchBelowADInplaceOrView::~AutoDispatchBelowADInplaceOrView() in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
"c10::impl::ExcludeDispatchKeyGuard::ExcludeDispatchKeyGuard(c10::DispatchKeySet)", referenced from:
at::AutoDispatchBelowADInplaceOrView::AutoDispatchBelowADInplaceOrView() in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
"at::_ops::zeros::call(c10::ArrayRef, c10::optionalc10::ScalarType, c10::optionalc10::Layout, c10::optionalc10::Device, c10::optional)", referenced from:
at::zeros(c10::ArrayRef, c10::TensorOptions) in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
"vtable for torch::autograd::AutogradMeta", referenced from:
torch::autograd::AutogradMeta::AutogradMeta(c10::TensorImpl*, bool, torch::autograd::Edge) in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
NOTE: a missing vtable usually means the first non-inline virtual member function has no definition.
"c10::AutogradMetaInterface::~AutogradMetaInterface()", referenced from:
torch::autograd::AutogradMeta::AutogradMeta(c10::TensorImpl*, bool, torch::autograd::Edge) in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
"caffe2::TypeMeta::typeMetaDatas()", referenced from:
caffe2::TypeMeta::data() const in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
"c10::TensorImpl::set_autograd_meta(std::__1::unique_ptr<c10::AutogradMetaInterface, std::__1::default_deletec10::AutogradMetaInterface >)", referenced from:
torch::autograd::make_variable(at::Tensor, bool, bool) in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
"c10::UndefinedTensorImpl::_singleton", referenced from:
c10::UndefinedTensorImpl::singleton() in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
"c10::detail::torchInternalAssertFail(char const*, char const*, unsigned int, char const*, std::__1::basic_string<char, std::__1::char_traits, std::__1::allocator > const&)", referenced from:
c10::Device::validate() in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
caffe2::TypeMeta::fromScalarType(c10::ScalarType) in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
"at::ops::mul_Scalar::call(at::Tensor const&, c10::Scalar const&)", referenced from:
at::mul(at::Tensor const&, c10::Scalar const&) in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
"c10::detail::torchInternalAssertFail(char const*, char const*, unsigned int, char const*, char const*)", referenced from:
c10::intrusive_ptr<c10::VariableVersion::VersionCounter, c10::detail::intrusive_target_default_null_typec10::VariableVersion::VersionCounter >::intrusive_ptr(c10::VariableVersion::VersionCounter*) in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
c10::intrusive_ptr_target::~intrusive_ptr_target() in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
c10::intrusive_ptr<c10::TensorImpl, c10::UndefinedTensorImpl>::retain() in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
c10::ArrayRef::debugCheckNullptrInvariant() in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
"vtable for c10::AutogradMetaInterface", referenced from:
c10::AutogradMetaInterface::AutogradMetaInterface() in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
NOTE: a missing vtable usually means the first non-inline virtual member function has no definition.
"caffe2::TypeMeta::error_unsupported_typemeta(caffe2::TypeMeta)", referenced from:
caffe2::TypeMeta::toScalarType() in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
"c10::detail::torchCheckFail(char const*, char const*, unsigned int, char const*)", referenced from:
torch::autograd::AutogradMeta::AutogradMeta(c10::TensorImpl*, bool, torch::autograd::Edge) in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
c10::detail::torchInternalAssertFail(char const*, char const*, unsigned int, char const*, c10::detail::CompileTimeEmptyString) in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
c10::TensorImpl::itemsize() const in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
c10::TensorImpl::data() const in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
"at::print(std::__1::basic_ostream<char, std::__1::char_traits >&, at::Tensor const&, long long)", referenced from:
at::operator<<(std::__1::basic_ostream<char, std::__1::char_traits >&, at::Tensor const&) in libtorch_sys-2bfcaab8f2fcc6a6.rlib(torch_api.o)
ld: symbol(s) not found for architecture arm64
clang: error: linker command failed with exit code 1 (use -v to see invocation)

[FVershinin]

Any news?

[l-monninger]

This should tell you all you need to know: macOS-arm64 but attempting to link with file built for macOS-x86_64.

I don't think this an issue with this library at all, apart from a shortage of documentation. You simply don't have the right LibTorch build installed. How to do that, I don't know (#488).

[l-monninger]

@zzeee Btw, in the meantime, if you just want to experiment with the library, you should be able to run in Docker.

[zzeee]

I just wanted to have experiments 5 months ago :) Now I use Pytorch because I can not wait for such a long time :)

[kodieg]

I managed to install pytorch from pip and set LIBTORCH variable and DYLD_LIBRARY_PATH accordingly.

pip install pytorch
export LIBTORCH=/Users/<username>/opt/miniconda3/lib//python3.9/site-packages/torch/
export DYLD_LIBRARY_PATH=${LIBTORCH}/lib:$DYLD_LIBRARY_PAT

Note, I used miniconda, so your library path may differ.

[LaurentMazare]

Closing in favor of #488