PyTorchWrapper causes error on deserializing

[polm]

How to reproduce the behaviour

Add a PyTorchWrapper component to a language pipeline and then do this:

nlp.to_disk("something")
nlp2 = spacy.load("something")

Motivating case is covered in https://github.com/explosion/spaCy/discussions/8291. This issue touches code in Thinc and spacy-transformers.

The issue is that the model at construction time has fewer layers than after initialize is called. When deserializing Thinc detects this as an issue and throws an error.

Your Environment

• Operating System:
• Python Version Used:
• spaCy Version Used:
• Environment Information:

[svlandeg]

Related issue, cf code in IO tests of https://github.com/explosion/spacy-transformers/pull/277:

nlp.to_disk(file_path)
nlp2 = util.load_model_from_config(orig_config, auto_fill=True, validate=True)
nlp2.initialize(lambda: train_examples)
nlp2.from_disk(file_path)

--> It shouldn't be necessary to call initialize ?

Can this be fixed by implementing an "empty shim" when constructing the Transformer to avoid the shape mismatch?

[svlandeg]

Update: the transformers bug at least is fixed by explosion/spacy-transformers#285

[oliviercwa]

Additional repro steps from #9317 to regress the fix