app.py

# app.py

import os
import gradio as gr
import torch
from transformers import (
    AutoTokenizer, 
    AutoModelForCausalLM, 
    BitsAndBytesConfig
)

MODEL_NAME = "NCAIR1/N-ATLaS"

# Read Hugging Face token from Space Secrets
HF_TOKEN = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACE_HUB_TOKEN")
if not HF_TOKEN:
    raise RuntimeError("HF_TOKEN missing. Add it in Space → Settings → Secrets.")

# 4-bit quantization config
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_use_double_quant=True,
)

print("Loading tokenizer...")
tokenizer = AutoTokenizer.from_pretrained(
    MODEL_NAME,
    use_auth_token=HF_TOKEN
)

print("Loading quantized N-ATLaS model (this may take a few minutes)...")
model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME,
    quantization_config=bnb_config,
    device_map="auto",  # automatically uses CPU or GPU if available
    use_auth_token=HF_TOKEN
)

model.eval()
print("✅ N-ATLaS loaded successfully with 4-bit quantization")

# Translation function
def translate(text, target_language):
    if not text.strip():
        return ""

    prompt = (
        f"You are a Nigerian language translation expert.\n"
        f"Translate the following English text to {target_language}.\n"
        f"Only output the translation.\n\n"
        f"Text:\n{text}\n\nTranslation:"
    )

    inputs = tokenizer(prompt, return_tensors="pt")

    # Make sure model outputs on the same device as inputs
    device = model.device
    inputs = {k: v.to(device) for k, v in inputs.items()}

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=128,  # reduce for CPU
            temperature=0.2,
            repetition_penalty=1.1,
            do_sample=False,
            eos_token_id=tokenizer.eos_token_id
        )

    decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)

    # Remove prompt part
    if "Translation:" in decoded:
        decoded = decoded.split("Translation:")[-1]

    return decoded.strip()

# Gradio interface
with gr.Blocks() as demo:
    gr.Markdown("## 🌍 N-ATLaS Translation (CPU / 4-bit Quantized)")
    
    inp = gr.Textbox(label="English text", lines=4)
    lang = gr.Dropdown(
        choices=["Hausa", "Igbo", "Yoruba", "English"],
        value="Hausa",
        label="Target language"
    )
    out = gr.Textbox(label="Translated text", lines=4)
    btn = gr.Button("Translate")
    
    btn.click(translate, inputs=[inp, lang], outputs=out)

demo.launch()