Significant performance improvement on MoE block of SwitchTransformer

[ranggihwang]

What does this PR do?

This PR includes a performant implementation of SwitchTransformersSparseMLP in the Google SwitchTransformer.
In the current implementation of the SwitchTransformer, it spans all possible experts, including the inactive ones.

for idx, expert in enumerate(self.experts.values()):
            token_indices = router_mask[:, :, idx].bool()
            next_states[token_indices] = expert(hidden_states[token_indices]).to(next_states.dtype)

This results in serious performance degradation of the SwitchTransformer.

As shown in this figure, the current implementation of the SwitchTransformer spans inactive experts, unnecessarily increasing latency. This issue can be particularly severe in models with a larger number of experts, as it needlessly spans more experts.

However, in my custom implementation of SwitchTransformersSparseMLP, it only accesses and computes the active experts.

Advantages

• This can significantly reduce the latency of the SwitchTransformer and make the model more accessible to a broader range of users.
• This change achieves greater latency reductions when expert parameters are offloaded to the CPU or SSD.
• This change addresses the problem of increasing latency proportional to the number of experts.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[amyeroberts]

cc @ArthurZucker @younesbelkada

[younesbelkada]

Nice work and great investigation ! Thanks for this ! Can you confirm the slow SwitchTransformers test pass?

[ranggihwang]

Is there anything else that I need to do for this PR?

[younesbelkada]

Hi @ranggihwang
Yes, please make sure to run make fixup so that the styling checks pass in our CI. For running the slow tests, can you run:

RUN_SLOW=1 pytest tests/models/switch_transformers/test_modeling_switch_transformers.py

[ranggihwang]

@younesbelkada
I've changed the coding style using make fixup, but I'm encountering an error when running this command:

RUN_SLOW=1 pytest tests/models/switch_transformers/test_modeling_switch_transformers.py

I've noticed the same error with the original switch transformer code, so I assume it's not due to my changes. How can I resolve this issue?

If you need the error log, I can attach it here.

[ArthurZucker]

Hey! Though we do span on all expert, the way the model is trained emphasize an even loading, meaning on average the experts should all be used no?
Could you share a bit more which model you are using, are you pretraining?

Let's revert changes related to linting that are not supposed to be here as well

[ranggihwang]

@ArthurZucker
That's a great point.
In the SwitchTransformer model, all experts can be used in the training phase, but in the inference, only some of them are utilized and that's the point where inefficiency is raised that the original code missed.
I'm mainly talking about inference, however, my custom code also can be used for training because it is mathematically equivalent to the original code.

For linting, is it okay for me to revert it to the previous version? Actually, this is the first time for me to contribute codes to HuggingFace, so I'm a bit confused about what to do next to get an acceptance for merging.