app.py

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

# Daftar model yang dipakai
MODELS = {
    "in2bg": "rahmanansah/t5-id-bugis",
    "bg2id": "rahmanansah/t5-bugis-id"
}

# Simpan tokenizer & model yang sudah diload
loaded_models = {}

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

def load_model(model_id):
    tokenizer = AutoTokenizer.from_pretrained(model_id)
    model = AutoModelForSeq2SeqLM.from_pretrained(model_id).to(device)
    return tokenizer, model

# Preload semua model
for key, model_id in MODELS.items():
    print(f"🔄 Loading {key} -> {model_id}")
    loaded_models[key] = load_model(model_id)
print("✅ Semua model sudah diload")

app = FastAPI()

class InputText(BaseModel):
    text: str
    model: str  # "in2bg" atau "bg2id"

@app.post("/translate")
def translate(input: InputText):
    if input.model not in loaded_models:
        return {"error": f"Model '{input.model}' tidak tersedia. Pilihan: {list(loaded_models.keys())}"}

    tokenizer, model = loaded_models[input.model]

    if not input.text.strip():
        return {"result": ""}

    text = input.text.strip()

    # Tambahkan prefix sesuai arah model
    if input.model == "in2bg":
        prefixed_text = f"translate id2bg: {text}"
    elif input.model == "bg2id":
        prefixed_text = f"translate bg2id: {text}"
    else:
        prefixed_text = text

    inputs = tokenizer(prefixed_text, return_tensors="pt").to(device)
    outputs = model.generate(**inputs, max_length=64)
    decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)

    return {"result": decoded}

if __name__ == "__main__":
    import uvicorn
    uvicorn.run("app:app", host="0.0.0.0", port=7860)