Tokenizers padding_side was not validate to be "right" in trainer_sft.py

from transformers import AutoTokenizer
AutoTokenizer.from_pretrained("OpenAssistant/llama2-13b-orca-8k-3319").padding_side
>> 'left'
AutoTokenizer.from_pretrained("TheBloke/Llama-2-13B-fp16")
>> 'left'
AutoTokenizer.from_pretrained("mosaicml/mpt-7b").padding_side
>> 'right'
AutoTokenizer.from_pretrained("huggyllama/llama-7b").padding_side
>> 'left'
AutoTokenizer.from_pretrained("OpenAssistant/llama-30b-sft-v8.2-2.4k-steps-system").padding_side
>> 'left'

Since llama models are using left padding, the supervised training dialoguecollator would cause the label_mask to pad in a different direction as the tokenizer.pad (input_ids, attention_mask), as torch.stack (label_mask) implements the right padding strategy.

Printing out the dataloader results in trainer_sft.py would also verify the issue

    train_dataloader = DataLoader(train, collate_fn=train_collate_fn, batch_size=9, shuffle=True)
    for batch in train_dataloader:
        for idx, question in enumerate(batch['input_ids']):
            print('-------')
            print(tokenizer.decode(question[batch['label_masks'][idx]]).replace('</s>', '')+'\n')

I think there's no padding_side assigned to right in the trainer_sft.py pipeline, so by default llama models we have trained are bit faulty