[Accelerator] Fix issue with 8bit models

[younesbelkada]

What does this PR do?

Fixes #1147

In theory it is not possible to fine-tune 8-bit models except if you use adapters that can be used only is a PeftModel is used for training (I also need to test the snippet below with a PeftModel to make sure this is relevant or not). EDIT: passing an 8bit PeftModel through accelerator.prepare seems to work fine.

But in some systems you can use the accelerator to load an 8bit model and use it out of the training scope (e.g. get the model's logits and use it in another model)

I am not sure if we should support 8-bit models using Accelerator, but if so, I propose the following changes in this PR
Happy also to revert the tests / bnb dependency

To reproduce:

import torch
import torch.nn.functional as F

from transformers import AutoModelForCausalLM
from datasets import load_dataset
from accelerate import Accelerator

model = AutoModelForCausalLM.from_pretrained(
    "EleutherAI/gpt-neo-125m",
    load_in_8bit=True,
    device_map="balanced",
)

accelerator = Accelerator()


model = accelerator.prepare(model)

cc @sgugger

Ran all the slow tests and got errors on DeepSpeed and FSDP tests but not sure if the failing is related to my PR

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[younesbelkada]

I think this PR is only necessary in case people want to design systems that uses 8-bit models in their training loop without backpropagating on the 8-bit model (for example in RLHF), as using adapters works already out of the box right now

[sgugger]

Thanks for the fix!

[younesbelkada]

@sgugger I think this will fail as the docker image is not using the main branch of transformers no?

[younesbelkada]

happy to skip it until the next release of transformers

[sgugger]

Yes we are not installing from source.

[sgugger]

Thanks again!