- Package:
- python3-torch
- Source:
- pytorch
- Description:
- Tensors and Dynamic neural networks in Python (Python Interface)
- Submitter:
- Dima Kogan
- Date:
- 2022-07-06 23:21:03 UTC
- Severity:
- grave
Hi. Thanks for maintaining pytorch. I can't imagine the pain it took to
get this packaged. Currently it doesn't work, unfortunately:
$ python3 -mtorch
Traceback (most recent call last):
File "/usr/lib/python3.10/runpy.py", line 187, in _run_module_as_main
mod_name, mod_spec, code = _get_module_details(mod_name, _Error)
File "/usr/lib/python3.10/runpy.py", line 146, in _get_module_details
return _get_module_details(pkg_main_name, error)
File "/usr/lib/python3.10/runpy.py", line 110, in _get_module_details
__import__(pkg_name)
File "/usr/lib/python3/dist-packages/torch/__init__.py", line 196, in <module>
from torch._C import *
ImportError: /usr/lib/x86_64-linux-gnu/libtorch_cpu.so.1.8: undefined symbol: _ZN4onnx12optimization8OptimizeERKNS_10ModelProtoERKSt6vectorINSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEESaISA_EE
This symbol is missing. I looked around, and I can't figure out which
package was supposed to provide it. Without it the linking fails, and
the package is unusable. Am I missing some dependency?
Hi, Thanks for the bug report. I'm aware of the break, and other users have reported this issue some time before: https://lists.debian.org/debian-ai/2022/06/msg00060.html The break is due to onnx 1.12 upgrade. The pytorch version in the new queue works fine with onnx 1.12, as mentioned in the above mailing list post. If you would like to continue using pytorch 1.8 for a while, you may need to pin onnx to the previous version, or rollback using our snapshot archive. When dealing with the pytorch 1.12 upgrade, I lost (to be honest I would like to stick to 1.8 but we have to go through python 3.10 transition) access to build machines due to complicated reasons, and the access will not be recovered. So in order to continue the pytorch upgrade, I have to upload reverse dependencies to unstable early, so that I can build pytorch and upload to the NEW queue. Theoretically this bug can only be resolved when pytorch 1.12 passes the new queue.
Thanks for the description. If the fix is already in NEW, then there's nothing much to do here, so I'm closing this report. Thanks again for working on this.