[실습] merge base model and LoRA adapter_model, Upload huggingface-hub

Topic. Fine-tuning 후 기존 base 모델과 생성된 adapter_model 을 merge합니다.

---

1.  huggingface login

$ pip install huggingface_hub
$ huggingface-cli login

 

1.1  허깅페이스에서 발급받은 토큰 입력

2. Base Model 가져오기

from transformers import AutoModelForCausalLM, AutoTokenizer

model_pretrained_name = {base model id}
path_to_lora_adapters = {adapter model path}

#모델 불러오기
model = AutoModelForCausalLM.from_pretrained(
    model_pretrained_name,
    device_map="cuda:0",
    quantization_config=bnb_config,
)

#토크나이저 불러오기
tokenizer = AutoTokenizer.from_pretrained(
	model_pretrained_name,
	add_eos_token=True
)

#어뎁터 모델 불러오기
model = PeftModel.from_pretrained(
    model,
    path_to_lora_adapters,
    torch_dtype=torch.float16
)

---

3. Merge

model = model.merge_and_unload()

---

4. Huggingface-hub upload

hub_model_id = {hub_model_id}

# model 업로드
model.push_to_hub(hub_model_id, token=True)

# tokenizer 업로드
tokenizer.push_to_hub(hub_model_id, token=True)