from transformers import AutoTokenizer, T5ForConditionalGeneration
import torch

def load_model():
    model_name = "Salesforce/codet5-base-multi-sum"
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = T5ForConditionalGeneration.from_pretrained(model_name)
    model.eval()
    model.to(torch.device("cuda" if torch.cuda.is_available() else "cpu"))
    return tokenizer, model

def generate_explanation(code, tokenizer, model):
    device = model.device
    # Final prompt style: generate docstring
    input_text = f"generate docstring: {code.strip()}"
    input_ids = tokenizer.encode(input_text, return_tensors="pt", truncation=True).to(device)
    output = model.generate(input_ids, max_new_tokens=150, early_stopping=True)
    return tokenizer.decode(output[0], skip_special_tokens=True)