[DEL] Remove .from_inner2

[f-dangel]

Resolves #37.

[runame]

LGTM. One question: are there not some cases where this can be more efficient, i.e. when batch_size * seq_len >> feature_dim?

[f-dangel]

Yes, you are right that there are different pros and cons. I tried to sum them up in the following note:

from_inner.pdf

Actually, having done the analysis I think we need to discuss whether we should aim for simplicity by supporting only from_inner, or should dynamically switch between from_inner versus from_inner2 depending on batch_size, sequence_length, feature_dim. In the scenario you mention above, using the from_inner2 approach might save a lot of memory.

For now, I will not merge and label this for 'needs discussion'.

[f-dangel]

@yorkerlin please add this to the agenda for our next in-person discussion