app.py

import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
from threading import Thread
import random

# Configuration 🛠️
model_name = "microsoft/phi-3-mini-4k-instruct"
device = "cuda" if torch.cuda.is_available() else "cpu"

# Load model with memory optimizations
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16 if device == "cuda" else torch.float32,
    device_map="auto",
    low_cpu_mem_usage=True
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Safety tools 🛡️ (simplified)
BLOCKED_WORDS = ["violence", "hate", "gun", "personal"]
SAFE_IDEAS = ["Design a robot to clean parks 🌳", "Code a recycling game ♻️"]

def is_safe(text):
    text = text.lower()
    return not any(bad_word in text for bad_word in BLOCKED_WORDS)

def respond(message, history, system_message, max_tokens, temperature, top_p):
    if not is_safe(message):
        return f"🚫 Let's focus on positive projects! Try: {random.choice(SAFE_IDEAS)}"
    
    # Create prompt with limited history
    prompt = f"System: {system_message}\n"
    for user, bot in history[-2:]:  # Keep only last 2 exchanges
        prompt += f"User: {user}\nAssistant: {bot}\n"
    prompt += f"User: {message}\nAssistant:"
    
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    
    # Generation settings
    generation_kwargs = dict(
        inputs.input_ids,
        max_new_tokens=min(max_tokens, 256),
        temperature=min(temperature, 0.7),
        top_p=top_p,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )
    
    # Generate response
    outputs = model.generate(**generation_kwargs)
    response = tokenizer.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True)
    
    yield response

with gr.Blocks() as demo:
    gr.Markdown("# 🤖 REACT Ethical AI Lab")
    gr.ChatInterface(
        respond,
        additional_inputs=[
            gr.Textbox("Help students create ethical AI projects", label="Guidelines"),
            gr.Slider(64, 256, value=128, label="Max Length"),
            gr.Slider(0.1, 0.7, value=0.3, label="Creativity"),
            gr.Slider(0.5, 1.0, value=0.9, label="Focus")
        ],
        examples=[
            ["How to make a solar-powered robot?"],
            ["Simple air quality sensor code"]
        ]
    )

if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0")