[BUG] Avoid serializing Callable kwargs during Completion

[bahtman]

Fixes #1759

[okhat]

Wow super cool, thanks @bahtman ! I'll review this more carefully but I think we can merge soon

[okhat]

Hmm, will the lru_cache work for callables? Do we need to add "ignore" parameter for the lru_cache to ignore callable_kwargs?

[bahtman]

I was able to run the LM with a callable token provider. I’m unfamiliar with how the lru cache works, so unsure whether it actually caches the parameter.

This is a workaround so we don’t serialize it. But since we don’t serialize in the new Embed, it might be a relic? If that is the case we should just revert and remove the ujson part.

[bahtman]

I did some digging. lru cache works as long as parameters are hashable (has hash() method)
https://docs.python.org/3/library/functools.html#functools.lru_cache

This is the case for callables, but not mutable containers like lists and dicts. This is probably the reason for the serialization.