llm_qa.py

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_name = "meta-llama/Llama-2-7b-chat-hf"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", torch_dtype=torch.float16)

def format_prompt(context, query):
    return f"""You are a helpful medical assistant. Answer only based on the following context. If you are unsure, say 'I don't know.'

Context:
{context}

Query: {query}
Answer:"""

def generate_answer(prompt):
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    output = model.generate(**inputs, max_new_tokens=150, do_sample=True, temperature=0.7, top_p=0.9)
    return tokenizer.decode(output[0], skip_special_tokens=True).split("Answer:")[-1].strip()