[SwitchTransformer] Significant performance improvement on MoE blocks

[ranggihwang]

What does this PR do?

This is an edited version of the previously closed PR (#30490)

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.

@ArthurZucker and @younesbelkada

[younesbelkada]

Thanks a lot ! Looks very good ! Can you make sure the styling checks pas make fixup && make fix-copies

[ranggihwang]

Thanks @younesbelkada
I've done make fixup && make fix-copies before in #30490 (review)
But I was requested to revert in the end.

How do I need to this correctly? Would you please let me know how to do this?

[younesbelkada]

Hi @ranggihwang
Thanks ! Hmm I think there was a misunderstanding on my side at that time, if you could run the styling checks and push the results here (it should only change 2 files, switch transformers & gpt_san_japanese file), it would be great !

[ranggihwang]

Shouldn't the styling check be done for the src/transformers/models/switch_transformers/modeling_switch_transformers.py?
I haven't changed anything except the file and it seems like there's no file named gpt_san_japanese_file in my repo.

[younesbelkada]

since gpt_san_japanese uses blocks that are copied from switch transformers, running make fix-copies will propagate the changes you introduced in that file as well, see: https://app.circleci.com/pipelines/github/huggingface/transformers/94615/workflows/e1e9d110-614a-411a-a0f5-b7d4146e4db8/jobs/1241937

[amyeroberts]

@younesbelkada @ranggihwang gpt san has been deprecated, so we don't really want these changes to be propogated. I've just merged in #31153 which removes the # Copied from headers for this model. Rebasing on main will include this and remove the need to run make fix-copies here. Thanks!

[younesbelkada]

Perfect thanks for the heads up @amyeroberts !
@ranggihwang feel free to proceed as suggested by amy 🙏

[ranggihwang]

@amyeroberts @younesbelkada
Thank you for your advice, Amy and Younes.

I've just rebase it to main and commit it. Would you please check if it is correct?

[younesbelkada]

Thanks @ranggihwang ! Now styling checks are failing, can you run make fixup and commit the changes ?

[ranggihwang]

Okay, now make fixup is done!

[younesbelkada]

Thanks !

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ArthurZucker]

I'll review this as I reviewed the previous PR, want to make sure the suggestions are all applied!

[ArthurZucker]

Could you apply the suggestion I did in the previous PR

[ArthurZucker]

Suggested change

	router_mask = router_mask.bool()
	idx_mask = router_mask.transpose(1, 2) # Batch * experts * tokens
	idx_mask = torch.cat(torch.split(idx_mask, 1, dim=0), dim=2) # 1 * experts * (batch * tokens)
	idx_mask = idx_mask.sum(dim=2)
	idx_mask = idx_mask.squeeze() # length: number of experts / value: number of tokens
	idx_mask = torch.nonzero(idx_mask, as_tuple=True)[
	0
	].tolist() # length: number of "activated" expert / value: index
	idx_mask = router_mask.reshape(batch*seq_len, num_experts).transpose(0,1).sum(dim=1)
	idx_mask = torch.nonzero(idx_mask, as_tuple=True)[0].tolist()

• the comment about shapes! 🤗

[ranggihwang]

The batch_size, seq_len, and num_experts are not defined in the funciton.
So, I've defined it with the router_mask and reflected your suggestions.

Thank you @ArthurZucker !

[ArthurZucker]

thanks!

[ranggihwang]

@ArthurZucker @younesbelkada
Please let me know if I need to rebase it again :)

[younesbelkada]

Still LGTM ! Let's wait for @ArthurZucker 's final review!

[ArthurZucker]

Thanks a lot! 🤗

[ArthurZucker]

Could this be propagated to the qwen code @ranggihwang ? I know that they have some variants with lots of experts!

[ranggihwang]

@ArthurZucker I think it can be adopted for many MoE models in HuggingFace not only qwen-moe but also for NLLB-MoE, Mixtral, etc.

[ArthurZucker]

awesome! Then if you are interested feel free to open a PR and ping me! 🤗
Some models need compile support which might be a little bit tricky we'll see