app.py

from fastapi import FastAPI
from pydantic import BaseModel
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch

# Load model and tokenizer once on startup
tokenizer = AutoTokenizer.from_pretrained("Salesforce/codet5p-220m")
model = AutoModelForSeq2SeqLM.from_pretrained("Salesforce/codet5p-220m")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)

app = FastAPI()

class Input(BaseModel):
    prompt: str

@app.post("/generate")
async def generate(input: Input):
    inputs = tokenizer(input.prompt, return_tensors="pt").to(device)
    outputs = model.generate(
        **inputs,
        max_length=2048,
        num_beams=3,
        early_stopping=True,
        no_repeat_ngram_size=3,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.pad_token_id,
    )
    output_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return {"output": output_text}