Support finetuning with LoRA

[katalinic-gc]

What does this PR do?

To enable it, apply below, which is basically identical to upstream.

config = LoraConfig(
    r=16, 
    lora_alpha=32, 
    target_modules=["q_proj", "v_proj"], 
    lora_dropout=0.05, 
    bias="none"
)
model = get_peft_model(model, config)
model.print_trainable_parameters()

Some online finetuning walkthroughs also include

    remove_unused_columns=False,  # required as the PeftModel forward doesn't have the signature of the wrapped model's forward
    label_names=["labels"],  # same reason as above

to the training args, e.g. (IPU)Seq2SeqTrainingArguments.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[rrva]

@katalinic-gc I would like to use this to finetune whisper-large-v2 on a fairly large dataset (700k examples). Is this usable as is (if I fix the merge conflicts)? What steps from https://www.graphcore.ai/posts/fine-tune-openais-whisper-automatic-speech-recognition-asr-model would need to be different with this approach?

[katalinic-gc]

@katalinic-gc I would like to use this to finetune whisper-large-v2 on a fairly large dataset (700k examples). Is this usable as is (if I fix the merge conflicts)? What steps from https://www.graphcore.ai/posts/fine-tune-openais-whisper-automatic-speech-recognition-asr-model would need to be different with this approach?

It won't be usable for large on IPUs due to OOM. We are internally working on supporting that; if and when available, we'll announce it.