Error when loading extension and the corresponding Python module is already loaded

[frostedoyster]

I got this error while loading extensions for a featomic-based model when featomic.torch was already imported. I haven't been able to reproduce it outside of GitHub CI

terminate called after throwing an instance of 'c10::Error'
  what():  Only a single TORCH_LIBRARY can be used to register the namespace featomic; please put all of your definitions in a single TORCH_LIBRARY block.  If you were trying to specify implementations, consider using TORCH_LIBRARY_IMPL (which can be duplicated).  If you really intended to define operators for a single namespace in a distributed way, you can use TORCH_LIBRARY_FRAGMENT to explicitly indicate this.  Previous registration of TORCH_LIBRARY was registered at /project/featomic-torch/src/register.cpp:6; latest registration was registered at /project/featomic-torch/src/register.cpp:6
Exception raised from registerLibrary at /pytorch/aten/src/ATen/core/dispatch/Dispatcher.cpp:210 (most recent call first):
frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x96 (0x7f23dcf83056 in /home/runner/work/metatrain/metatrain/.tox/docs/lib/python3.13/site-packages/torch/lib/libc10.so)
frame metatensor/metatensor#1: c10::detail::torchCheckFail(char const*, char const*, unsigned int, std::string const&) + 0x64 (0x7f23dcf2c776 in /home/runner/work/metatrain/metatrain/.tox/docs/lib/python3.13/site-packages/torch/lib/libc10.so)
frame metatensor/metatensor#2: c10::Dispatcher::registerLibrary(std::string, std::string) + 0x3c5 (0x7f23debc81c5 in /home/runner/work/metatrain/metatrain/.tox/docs/lib/python3.13/site-packages/torch/lib/libtorch_cpu.so)
frame metatensor/metatensor#3: torch::Library::Library(torch::Library::Kind, std::string, std::optional<c10::DispatchKey>, char const*, unsigned int) + 0x416 (0x7f23dec09bc6 in /home/runner/work/metatrain/metatrain/.tox/docs/lib/python3.13/site-packages/torch/lib/libtorch_cpu.so)
frame metatensor/metatensor#4: <unknown function> + 0x18d1c (0x7f23d6bb2d1c in /home/runner/work/metatrain/metatrain/.tox/docs/lib/python3.13/site-packages/featomic/torch/torch-2.6/lib/libfeatomic_torch.so)
frame metatensor/metatensor#5: <unknown function> + 0x647e (0x7f2419a7f47e in /lib64/ld-linux-x86-64.so.2)
frame metatensor/metatensor#6: <unknown function> + 0x6568 (0x7f2419a7f568 in /lib64/ld-linux-x86-64.so.2)
frame metatensor/metatensor#7: _dl_catch_exception + 0xe5 (0x7f2419174af5 in /lib/x86_64-linux-gnu/libc.so.6)
frame metatensor/metatensor#8: <unknown function> + 0xdff6 (0x7f2419a86ff6 in /lib64/ld-linux-x86-64.so.2)
frame metatensor/metatensor#9: _dl_catch_exception + 0x88 (0x7f2419174a98 in /lib/x86_64-linux-gnu/libc.so.6)
frame metatensor/metatensor#10: <unknown function> + 0xe34e (0x7f2419a8734e in /lib64/ld-linux-x86-64.so.2)
frame metatensor/metatensor#11: <unknown function> + 0x9063c (0x7f241909063c in /lib/x86_64-linux-gnu/libc.so.6)
frame metatensor/metatensor#12: _dl_catch_exception + 0x88 (0x7f2419174a98 in /lib/x86_64-linux-gnu/libc.so.6)
frame metatensor/metatensor#13: _dl_catch_error + 0x33 (0x7f2419174b63 in /lib/x86_64-linux-gnu/libc.so.6)
frame metatensor/metatensor#14: <unknown function> + 0x9012e (0x7f241909012e in /lib/x86_64-linux-gnu/libc.so.6)
frame metatensor/metatensor#15: dlopen + 0x48 (0x7f24190906c8 in /lib/x86_64-linux-gnu/libc.so.6)
frame metatensor/metatensor#16: <unknown function> + 0x1a7bd (0x7f241601e7bd in /opt/hostedtoolcache/Python/3.13.2/x64/lib/python3.13/lib-dynload/_ctypes.cpython-313-x86_64-linux-gnu.so)
<omitting python frames>

[Luthaf]

This is weird: we are trying to prevent exactly this and not load extensions if they are already loaded: https://github.com/metatensor/metatensor/blob/23740d9b1618e33969ad3c5323fbd78c7821f22d/metatensor-torch/src/atomistic/model.cpp#L854

This was trying to load library from an extension_directory, while simultaneously having the extension loaded in Python, right?