Multi-Token Attention

This is another potential improvement to the transformer architecture from Facebook (the other one that comes to mind is this one from same authors: https://arxiv.org/abs/2405.18719), but note that it comes with a major problem that might not be obvious at first glance: it's just not usable in practice without a ton of work. It modifies the innards of the attention mechanism, so it is incompatible with Flash Attention (or any other optimized attention library), and you do not want to train anything beyond toy models without Flash Attention (the performance hit is just way too big).

There's pytorch's FlexAttention which could maybe make this practical, but currently it's just way too buggy.

Why do you say FlexAttention is too buggy? I have heard about a lot of successful usages of it, and never heard about any such problems.

Also note, depending on your model dimensions and sequence lengths, often the attention computation plays only a minor role (maybe 10% overall or so), and the MLP computation dominates.

Last time I tried it I encountered both showstopper bugs (it was completely obviously broken) and subtle correctness bugs (it looked like it was working, but since I'm paranoid I have unit tests for everything and numerically the errors were too big compared to what you'd get with eager attention or Flash Attention), and it was too slow for my taste compared to Flash Attention so I just dropped it. And I wasn't even doing anything super exotic with it.

Maybe it's better now, but I'd still consider using FlexAttention without a corresponding unit test checking its accuracy against an equivalent eager implementation completely irresponsible.

What unit tests do you use for nn modules and how do you come up with them?