AttributeError on import in 8.0.14 when using PyTorch<1.6.0 (?)

[lumi-tworthey]

As of the v8.0.14, the following traceback is occurring at import time in a project that has PyTorch pinned to v1.4.0:

.venv/lib/python3.6/site-packages/thinc/util.py:34: in <module>
    has_torch_amp = not torch.cuda.amp.common.amp_definitely_not_available()
AttributeError: module 'torch.cuda' has no attribute 'amp'

It seems torch.cuda.amp did not exist in torch until version 1.6.0. The offending line (at util.py:34) is guarded by an except ImportError, but not AttributeError for the cases in which the installed torch version does not know about the module being accessed in the first place

    has_torch_amp = not torch.cuda.amp.common.amp_definitely_not_available()

[adrianeboyd]

Hi, we're working on a fix in #610. Hopefully we'll have a new release very soon. In the meanwhile v8.0.13 should still be compatible with older versions of pytorch.

[adrianeboyd]

As a note, the current recommended lower version for torch is v1.5.0, so there may be other issues with v1.4.0. (But if you've been using it without issues, then it may be totally fine. This was set a while ago and I'd have to track down the details.)

[adrianeboyd]

Please try upgrading to v8.0.15, which should hopefully resolve the AttributeError on import.

However, after reviewing the changes and running more thorough tests of all recent torch releases, we've revised the lower recommended torch version to v1.6.0, so I'm not sure whether you still might run into other errors with v1.4.0.