app.py

import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel, PeftConfig

# Model configuration - Gemma-3n-E4B fine-tuned
MODEL_ID = "Laserhun/gemma-3n-E4B-luau-finetuned"
BASE_MODEL_ID = "google/gemma-3n-E4B"

print("Loading Gemma-3n-E4B fine-tuned model...")
try:
    # Try loading as PEFT model
    peft_config = PeftConfig.from_pretrained(MODEL_ID)
    
    # Load base model
    base_model = AutoModelForCausalLM.from_pretrained(
        BASE_MODEL_ID,
        torch_dtype=torch.float16,
        device_map="auto",
        trust_remote_code=True,
        ignore_mismatched_sizes=True
    )
    
    # Load PEFT adapters
    model = PeftModel.from_pretrained(base_model, MODEL_ID)
    print("Loaded Gemma-3n-E4B as PEFT model")
except:
    # Load as regular model
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_ID,
        torch_dtype=torch.float16,
        device_map="auto",
        trust_remote_code=True
    )
    print("Loaded as regular model")

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
if not tokenizer.pad_token:
    tokenizer.pad_token = tokenizer.eos_token

def generate_luau_code(prompt, max_length=512, temperature=0.7, top_p=0.95):
    """Generate Luau code using Gemma-3n-E4B model"""
    
    # Format for Gemma-3n
    formatted_prompt = f"<start_of_turn>user\n{prompt}<end_of_turn>\n<start_of_turn>model\n"
    
    # Tokenize
    inputs = tokenizer(formatted_prompt, return_tensors="pt", truncation=True, max_length=512)
    
    # Move to device
    inputs = {k: v.to(model.device) for k, v in inputs.items()}
    
    # Generate
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=max_length,
            temperature=temperature,
            top_p=top_p,
            do_sample=True,
            pad_token_id=tokenizer.pad_token_id,
            eos_token_id=tokenizer.eos_token_id
        )
    
    # Decode
    generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    
    # Extract response
    if "<start_of_turn>model" in generated_text:
        response = generated_text.split("<start_of_turn>model")[-1].strip()
    else:
        response = generated_text[len(formatted_prompt):].strip()
    
    return response

# Create Gradio interface
iface = gr.Interface(
    fn=generate_luau_code,
    inputs=[
        gr.Textbox(
            lines=4,
            placeholder="Describe the Luau code you want to generate...",
            label="Enter your Luau code request"
        ),
        gr.Slider(minimum=100, maximum=1000, value=512, step=50, label="Max Length"),
        gr.Slider(minimum=0.1, maximum=1.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top P")
    ],
    outputs=gr.Code(language="lua", label="Generated Luau Code"),
    title="🎮 Gemma-3n-E4B Luau Code Generator",
    description="Generate Roblox Luau code using Gemma-3n-E4B model (8B parameters, 4B runtime) fine-tuned on Luau corpus.",
    examples=[
        ["Create a smooth part movement function with easing", 512, 0.7, 0.95],
        ["Write a door script with click interaction and smooth animation", 512, 0.7, 0.95],
        ["Generate a complete inventory system with add, remove, and display functions", 700, 0.7, 0.95],
        ["Create a spawning system for objects at random positions", 400, 0.7, 0.95],
        ["Write a leaderboard system that saves player scores", 600, 0.7, 0.95]
    ],
    theme=gr.themes.Soft()
)

if __name__ == "__main__":
    iface.launch()