bitsandbytes_8bit_model_loading_with_transformers.py

1import torch
2from transformers import AutoModelForCausalLM, AutoTokenizer
3
4model_id = "facebook/opt-125m"
5
6# To load a model in 8-bit, use the load_in_8bit=True argument
7model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", load_in_8bit=True)
8tokenizer = AutoTokenizer.from_pretrained(model_id)
9
10text = "Hello, my name is"
11inputs = tokenizer(text, return_tensors="pt").to("cuda")
12outputs = model.generate(**inputs)
13print(tokenizer.decode(outputs[0], skip_special_tokens=True))