app.py

import os
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import gradio as gr

# Configuration
MODEL_ID = "google/gemma-1.1-2b-it"  # Using smaller 2B version
HF_TOKEN = os.getenv("HF_TOKEN")
MAX_TOKENS = 80  # Conservative limit

def load_model():
    """Simplified model loading that works in Spaces"""
    print("🔄 Loading model...")
    
    # Load tokenizer
    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, token=HF_TOKEN)
    
    # Explicit CPU-only loading
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_ID,
        torch_dtype=torch.float32,  # Required for CPU
        token=HF_TOKEN
    ).to('cpu')  # Explicit CPU placement
    
    print("✅ Model loaded successfully!")
    return tokenizer, model

tokenizer, model = load_model()

def predict(topic):
    """Memory-safe generation"""
    try:
        prompt = f"Create a short script about {topic}:\n1) Hook\n2) Point\n3) CTA\n\nScript:"
        inputs = tokenizer(prompt, return_tensors="pt").to('cpu')
        
        outputs = model.generate(
            **inputs,
            max_new_tokens=MAX_TOKENS,
            temperature=0.7
        )
        
        return tokenizer.decode(outputs[0], skip_special_tokens=True)
    
    except Exception as e:
        return f"Error: {str(e)}"

# Minimal interface
gr.Interface(
    fn=predict,
    inputs=gr.Textbox(label="Topic"),
    outputs=gr.Textbox(label="Script", lines=4),
    api_name="predict"
).launch(server_name="0.0.0.0")