tch-rs does not run on m1 mac

[antimora]

tch-rs does not run on m1 MacOS. To fix the problem I followed these steps adopted from https://github.com/ssoudan/tch-m1.

1. pip3 install torch
2. export paths:

export LIBTORCH=$(python3 -c 'import torch; from pathlib import Path; print(Path(torch.__file__).parent)')
export DYLD_LIBRARY_PATH=${LIBTORCH}/lib

[chrisvander]

What's the error message you're running into? How did you install torch?

I ended up symlinking the torch directory into the project directory.

[chrisvander]

Could also check out this repository, pretty much exactly what I did and seems to be a fairly minimal environment setup.

[antimora]

@chrisvander builds ok but when I tried running I got some linking errors. I think tch-rs downloads for mac but the version is not correct. I think it downloads for x86 but it needs ARM64 version which is provided by python3's pytorch.

I used the same solution as your suggested repository.

This is the script I created to help users to overcome. Unfortunately there is no way to set the variables by build.rs script.

Instead I updated our build.rs to error out and let the users know to run the script.

Also I updated our Cargo.toml to skip download for mac users.

Hope this temporary workaround helps others too. I looking forward when this will be fixed in tch-rs.

[kornelski]

The Mac download at https://download.pytorch.org/libtorch/cpu/libtorch-macos-1.13.0.zip contains only the Intel version.

Maybe disable downloads on aarch64 and build from source instead?

[chrisvander]

I made an environment in Conda and installed it that way. That version is aarch64.

[LaurentMazare]

Thanks, I just made some changes to the torch-sys build.rs to avoid the download for apple aarch64 and point to this issue + display the pip install commands from the first post, hopefully this will make it clearer to end users how to get this to work until pytorch folks provide pre-built aarch64 binaries.
(also I don't have an actual way to test the changes but hopefully they were straightforward enough)

[antimora]

@LaurentMazare I tested your changes on Mac and I can confirm it panics as intended:

   Compiling torch-sys v0.11.0 (/Users/user/Projects/tch-rs/torch-sys)
error: failed to run custom build command for `torch-sys v0.11.0 (/Users/user/Projects/tch-rs/torch-sys)`

Caused by:
  process didn't exit successfully: `/Users/user/Projects/tch-rs/target/debug/build/torch-sys-04e74d17bc8358a0/build-script-build` (exit status: 101)
  --- stdout
  cargo:rerun-if-env-changed=TORCH_CUDA_VERSION
  cargo:rerun-if-env-changed=LIBTORCH

  --- stderr
  thread 'main' panicked at 'Pre-built version of libtorch for apple silicon are not available.
                              You can install torch manually following the indications from https://github.com/LaurentMazare/tch-rs/issues/629
                              pip3 install torch==2.0.0

                              Then update the following environment variables:
                              export LIBTORCH=$(python3 -c 'import torch; from pathlib import Path; print(Path(torch.__file__).parent)')
                              export DYLD_LIBRARY_PATH=${LIBTORCH}/lib
                              ', torch-sys/build.rs:123:29
  note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
[tch-rs]$ uname -a
Darwin MacBook-Pro.local 22.3.0 Darwin Kernel Version 22.3.0: 
Mon Jan 30 20:38:37 PST 2023; root:xnu-8792.81.3~2/RELEASE_ARM64_T6000 arm64
[tch-rs]$ 

[LaurentMazare]

Great thanks a lot @antimora

[antimora]

Linking a Torch's issue ticket here: pytorch/pytorch#96046 Someone requested "Add arm64 builds for libtorch on MacOS with mps support"

[patientplatypus6]

Thanks for the python script antimora - that's what fixed the problem for me!

[anmolduainter]

I am getting error
symbol not found in flat namespace '__ZN3c1019UndefinedTensorImpl10_singletonE'

[xsa-dev]

I am getting error symbol not found in flat namespace '__ZN3c1019UndefinedTensorImpl10_singletonE'

i got:

dyld[49777]: symbol not found in flat namespace '__ZN3c1019UndefinedTensorImpl10_singletonE'

Indeed, I am experiencing a similar issue. Have you attempted to address the problems arising from colleagues' earlier writing?

Solved by this: #629 (comment)