Fix: support PEFT/LoRA with added tokens

[mapmeld]

PEFT supports adding tokens/embeddings: https://github.com/huggingface/peft/blob/main/examples/causal_language_modeling/peft_lora_clm_with_additional_tokens.ipynb

When loading a pretrained model, tokenizer, and PEFT including new tokens, the model fails because the model and adapter have a different number of embeddings:

! lm_eval --model hf \
    --model_args pretrained=gradientai/Llama-3-8B-Instruct-262k,
       tokenizer=monsoon-nlp/llama3-biotokenpretrain-kaniwa,
       load_in_4bit=True,
       peft=monsoon-nlp/llama3-biotokenpretrain-kaniwa \
    --tasks hellaswag \
    --device cuda:0 \
    --batch_size 1

  File "/content/lm-evaluation-harness/lm_eval/models/huggingface.py", line 204, in __init__
    self._create_model(
  File "/content/lm-evaluation-harness/lm_eval/models/huggingface.py", line 582, in _create_model
    self._model = PeftModel.from_pretrained(
  File "/usr/local/lib/python3.10/dist-packages/peft/peft_model.py", line 356, in from_pretrained
    model.load_adapter(model_id, adapter_name, is_trainable=is_trainable, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/peft/peft_model.py", line 730, in load_adapter
    load_result = set_peft_model_state_dict(self, adapters_weights, adapter_name=adapter_name)
  File "/usr/local/lib/python3.10/dist-packages/peft/utils/save_and_load.py", line 249, in set_peft_model_state_dict
    load_result = model.load_state_dict(peft_model_state_dict, strict=False)
  File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 2153, in load_state_dict
    raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for PeftModelForCausalLM:
	size mismatch for base_model.model.model.embed_tokens.modules_to_save.default.weight: copying a param with shape torch.Size([128260, 4096]) from checkpoint, the shape in current model is torch.Size([128256, 4096]).
	size mismatch for base_model.model.lm_head.weight: copying a param with shape torch.Size([128260, 4096]) from checkpoint, the shape in current model is torch.Size([128256, 4096]).

In this case we want to run self._model.resize_token_embeddings(len(self.tokenizer))

The core fix is:

if peft:
  if self._model.config.vocab_size != len(self.tokenizer):
    # resize model for LoRAs with added tokens
    self._model.resize_token_embeddings(len(self.tokenizer))
  self._model = PeftModel.from_pretrained( ...)

The other change is moving initialization of self.tokenizer earlier in the script, so we have access to the tokenizer at this point in time.

CoLab proof of concept:
https://colab.research.google.com/drive/12NaNeDHRCMVhyIHL4_Z6AZQq6q_E2cP8

[haileyschoelkopf]

Thanks very much for the fix @mapmeld !

Added a log message for when this resizing happens, but otherwise LGTM!