OpenEquivariance fails on wheel pip package installation on ARM systems.

[achacond]

OpenEquivariance fails on wheel pip package installation on ARM systems.
This is likely due to using an x86 pre-compiled libtorch_cpu.so library (see details below)

 [17/19] Linking CXX shared module oeq_stable_cuda.cpython-312-aarch64-linux-gnu.so
      FAILED: [code=1] oeq_stable_cuda.cpython-312-aarch64-linux-gnu.so
      : && /usr/bin/aarch64-linux-gnu-g++ -fPIC -O3 -DNDEBUG  -shared -Wl,-s -Wl,--gc-sections -Wl,--dependency-file=CMakeFiles/oeq_stable_cuda.dir/link.d  -o oeq_stable_cuda.cpython-312-aarch64-linux-gnu.so CMakeFiles/oeq_stable_cuda.dir/openequivariance/extension/libtorch_tp_jit_stable.cpp.o CMakeFiles/oeq_stable_cuda.dir/openequivariance/extension/json11/json11.cpp.o  -Wl,-rpath,/tmp/tmp977gblf5/build/_deps/libtorch-src/lib:/usr/local/cuda-13.1/targets/sbsa-linux/lib:/tmp/tmp977gblf5/build:  libnanobind-static.a  _deps/libtorch-src/lib/libtorch_cpu.so  _deps/libtorch-src/lib/libc10.so  /usr/local/cuda-13.1/targets/sbsa-linux/lib/libcudart.so  /usr/local/cuda/targets/sbsa-linux/lib/stubs/libcuda.so  /usr/local/cuda-13.1/targets/sbsa-linux/lib/libnvrtc.so  libtorch_cuda.so  -ldl  /usr/lib/aarch64-linux-gnu/librt.a && :
      /usr/bin/ld: _deps/libtorch-src/lib/libtorch_cpu.so: error adding symbols: file in wrong format
      collect2: error: ld returned 1 exit status
      [18/19] Building CXX object CMakeFiles/oeq_stable_cuda_aoti.dir/openequivariance/extension/libtorch_tp_jit_stable.cpp.o
      ninja: build stopped: subcommand failed.
      
      *** CMake build failed
      [end of output]
  
  note: This error originates from a subprocess, and is likely not a problem with pip.
  ERROR: Failed building wheel for openequivariance
Failed to build openequivariance
error: failed-wheel-build-for-install

× Failed to build install

Steps for the reproduction:

Download the wheel package and decompress:
pip download --no-clean openequivariance

File CMakeLists.txt shows the source of the pre-compiled torch libraries:

# Download LibTorch 
include(FetchContent)
FetchContent_Declare(
    libtorch
    URL "https://download.pytorch.org/libtorch/cpu/libtorch-shared-with-deps-2.10.0%2Bcpu.zip"
)
message(STATUS "Downloading LibTorch...")
FetchContent_MakeAvailable(libtorch)

By download them and check the format of the *.so lib with file we can identify that uses x86 binary format:
wget https://download.pytorch.org/libtorch/cpu/libtorch-shared-with-deps-2.10.0%2Bcpu.zip

# file libtorch/lib/libtorch_cpu.so 
libtorch/lib/libtorch_cpu.so: ELF 64-bit LSB shared object, x86-64, version 1 (GNU/Linux), dynamically linked, BuildID[sha1]=e59a73f8ca02c19463bd923dc568ce765b2a1106, not stripped

[vbharadwaj-bk]

Thanks for the detailed report! I merged a PR that should fix this by downloading the aarch64 libtorch. Because I don't have access to an ARM system, would you mind testing with

pip install "git+https://github.com/PASSIONLab/OpenEquivariance#subdirectory=openequivariance"

If this works, I will create a new release.

[achacond]

Many thanks! I can confirm that now the installing issue is gone on ARM systems:

# pip install "git+https://github.com/PASSIONLab/OpenEquivariance#subdirectory=openequivariance"
Collecting git+https://github.com/PASSIONLab/OpenEquivariance#subdirectory=openequivariance
  Cloning https://github.com/PASSIONLab/OpenEquivariance to /tmp/pip-req-build-jrqcy74i
  Running command git clone --filter=blob:none --quiet https://github.com/PASSIONLab/OpenEquivariance /tmp/pip-req-build-jrqcy74i
  Resolved https://github.com/PASSIONLab/OpenEquivariance to commit 61da2dfc02150622e5a726890b15c564ec3439db
  Installing build dependencies ... done
  Getting requirements to build wheel ... done
  Installing backend dependencies ... done
  Preparing metadata (pyproject.toml) ... done
Requirement already satisfied: setuptools in /usr/local/lib/python3.12/dist-packages (from openequivariance==0.6.6) (80.10.1)
Requirement already satisfied: ninja in /usr/local/lib/python3.12/dist-packages (from openequivariance==0.6.6) (1.13.0)
Requirement already satisfied: jinja2 in /usr/local/lib/python3.12/dist-packages (from openequivariance==0.6.6) (3.1.6)
Requirement already satisfied: numpy in /usr/local/lib/python3.12/dist-packages (from openequivariance==0.6.6) (2.1.0)
Requirement already satisfied: MarkupSafe>=2.0 in /usr/local/lib/python3.12/dist-packages (from jinja2->openequivariance==0.6.6) (3.0.3)
Building wheels for collected packages: openequivariance
  Building wheel for openequivariance (pyproject.toml) ... done
  Created wheel for openequivariance: filename=openequivariance-0.6.6-cp312-cp312-linux_aarch64.whl size=312671 sha256=3bb8530eec433a62a135863c230bbad4c5d07dc6ce33e200e6ca63552bf55b51
  Stored in directory: /tmp/pip-ephem-wheel-cache-w2bu55uv/wheels/e1/06/14/b16acfa1feebd2f16e75c0d43d673add03c592bfb36101daeb
Successfully built openequivariance
Installing collected packages: openequivariance
Successfully installed openequivariance-0.6.6

[vbharadwaj-bk]

v0.6.6 released and uploaded to PyPI; we recommend PT2.10 or higher to get the benefits of the stable-ABI C++ extension, although <=2.9 will still work as well; happy hunting :)