import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
 
# Hugging Face model id
model_id = "philschmid/llama-2-7b-instruction-generator"
 
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False) # bug with fast tokenizer
model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True, torch_dtype=torch.float16) # we load the model in fp16 on purpose