Initializing libiomp5.dylib, but found libomp.dylib already initialized. #78490
Description
🐛 Describe the bug
The issue appears on MacOS py3.8, it started after updating to the latest nightly 1.13.0.dev20220525-py3.8_0 from core (previously I was at 1.12.0.dev20220309-py3.8_0, so the issue could have been introduced earlier than May 25th). I'm receiving the following after importing numpy and pytorch together:
$ python -c "import numpy;import torch"
OMP: Error pytorch/vision#15: Initializing libiomp5.dylib, but found libomp.dylib already initialized.
OMP: Hint This means that multiple copies of the OpenMP runtime have been linked into the program. That is dangerous, since it can degrade performance or cause incorrect results. The best thing to do is to ensure that only a single OpenMP runtime is linked into the process, e.g. by avoiding static linking of the OpenMP runtime in any library. As an unsafe, unsupported, undocumented workaround you can set the environment variable KMP_DUPLICATE_LIB_OK=TRUE to allow the program to continue to execute, but that may cause crashes or silently produce incorrect results. For more information, please see http://www.intel.com/software/products/support/.
python3 -mtorch.utils.collect_env
Collecting environment information...
PyTorch version: 1.13.0.dev20220525
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A
OS: macOS 12.3.1 (x86_64)
GCC version: Could not collect
Clang version: 12.0.0 (clang-1200.0.32.21)
CMake version: version 3.18.4
Libc version: N/A
Python version: 3.8.12 | packaged by conda-forge | (default, Oct 12 2021, 21:50:38) [Clang 11.1.0 ] (64-bit runtime)
Python platform: macOS-10.16-x86_64-i386-64bit
Is CUDA available: False
CUDA runtime version: No CUDA
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
Versions of relevant libraries:
[pip3] efficientnet-pytorch==0.7.1
[pip3] mypy==0.931
[pip3] mypy-extensions==0.4.3
[pip3] numpy==1.22.4
[pip3] torch==1.13.0.dev20220525
[pip3] torchdata==0.4.0a0+652986b
[pip3] torchvision==0.14.0a0+9a72fd6
[pip3] torchviz==0.0.2
[conda] blas 2.112 mkl conda-forge
[conda] blas-devel 3.9.0 12_osx64_mkl conda-forge
[conda] efficientnet-pytorch 0.7.1 pypi_0 pypi
[conda] libblas 3.9.0 12_osx64_mkl conda-forge
[conda] libcblas 3.9.0 12_osx64_mkl conda-forge
[conda] liblapack 3.9.0 12_osx64_mkl conda-forge
[conda] liblapacke 3.9.0 12_osx64_mkl conda-forge
[conda] mkl 2021.4.0 h89fa619_689 conda-forge
[conda] mkl-devel 2021.4.0 h694c41f_690 conda-forge
[conda] mkl-include 2021.4.0 hf224eb6_689 conda-forge
[conda] numpy 1.22.4 py38h3ad0702_0 conda-forge
[conda] pytorch 1.13.0.dev20220525 py3.8_0 pytorch-nightly
[conda] torchdata 0.4.0a0+652986b pypi_0 pypi
[conda] torchvision 0.14.0a0+9a72fd6 dev_0 <develop>
[conda] torchviz 0.0.2 pypi_0 pypi
Strangely, importing first torch works:
python -c "import torch;import numpy;print('works')"
works
Setting KMP_DUPLICATE_LIB_OK=TRUE as env var solves the issue while invoking from console:
KMP_DUPLICATE_LIB_OK=TRUE python -c "import numpy;import torch;print('works')"
works
Sometimes I get segfaults thought, this doesn't seem like a stable solution.
Versions
Latest Core nightly (20220525).