Adds a multiply_grads akin to fairseq

[muellerzr]

What does this PR do?

Helps w/ DDP wrt huggingface/transformers#34283

fairseq came up with a technique to help also with countering gradient accumulation by countering grad update by multiplying the gradients by num_gpus / num_items_in_batch. This PR adds an interface for the Trainer to hook into.

Without it:

Fixes # (issue)

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.

@SunMarc

[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.

[SunMarc]

Nice ! It would be also nice to compare if it is better to update the optimizer or the loss directly in transformers.

[muellerzr]

I found in general both were needed.

[sbwww]

There is a minor concern, would this fix be suitable for all backends (PyTorch DDP/FDSP, DeepSpeed, Megatron)?

I notice that DeepSpeed optimizer has unique groups (bf16_groups and fp32_groups_gradients) in its wrapped optimizer. And, grads in bf16_groups are accumulated into fp32_groups_gradients and deallocated in each backward call, .

I'm not sure whether param_groups in accelerator is a wrap of bf16_groups or fp32_groups_gradients in DeepSpeed? Previously I had an attempt to manipulate grads in optimizer. However, grads in optimizer are all 0 after backward when I use accelerate + DeepSpeed.

So, I'm afraid multiply_grads would not work for current DeepSpeed integration.

---

DeepSpeed/deepspeed/runtime/bf16_optimizer.py :

https://github.com/microsoft/DeepSpeed/blob/b647fb2470f8f6fefe5cab0ea84a2d89696eb898/deepspeed/runtime/bf16_optimizer.py#L316-L329