app.py

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

app = FastAPI()

# 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)

class GenerationRequest(BaseModel):
    prompt: str
    max_length: int = 2048
    num_beams: int = 3
    early_stopping: bool = True
    no_repeat_ngram_size: int = 3

@app.post("/generate")
async def generate_text(request: GenerationRequest):
    inputs = tokenizer(request.prompt, return_tensors="pt").to(device)
    outputs = model.generate(
        **inputs,
        max_length=request.max_length,
        num_beams=request.num_beams,
        early_stopping=request.early_stopping,
        no_repeat_ngram_size=request.no_repeat_ngram_size,
        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 {"generated_text": output_text}

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=8080)