app.py

from llama_cpp import Llama
import gradio as gr
import os
import requests
import time

# Percorso locale del modello - Qwen2.5-0.5B-Instruct VELOCE
MODEL_PATH = "qwen2.5-0.5b-instruct-q4_k_m.gguf"
MODEL_URL = "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct-GGUF/resolve/main/qwen2.5-0.5b-instruct-q4_k_m.gguf"

def download_model():
    """Scarica il modello se non esiste già"""
    if not os.path.exists(MODEL_PATH):
        print("📥 Downloading Qwen2.5-0.5B-Instruct model...")
        try:
            response = requests.get(MODEL_URL, stream=True, timeout=300)
            response.raise_for_status()
            
            total_size = int(response.headers.get('content-length', 0))
            downloaded = 0
            
            with open(MODEL_PATH, 'wb') as f:
                for chunk in response.iter_content(chunk_size=8192):
                    if chunk:
                        f.write(chunk)
                        downloaded += len(chunk)
                        if total_size > 0:
                            progress = (downloaded / total_size) * 100
                            print(f"📥 Download progress: {progress:.1f}%")
            
            # Verifica che il file sia completo
            if os.path.getsize(MODEL_PATH) < 100000:  # Almeno 100KB
                print("❌ Downloaded file seems corrupted")
                os.remove(MODEL_PATH)
                return False
            
            print("✅ Model downloaded successfully!")
            return True
            
        except Exception as e:
            print(f"❌ Error downloading model: {e}")
            if os.path.exists(MODEL_PATH):
                os.remove(MODEL_PATH)  # Rimuovi file corrotto
            return False
    else:
        print("✅ Model already exists!")
        # Verifica che il file esistente sia valido
        if os.path.getsize(MODEL_PATH) < 100000:
            print("❌ Existing file seems corrupted, re-downloading...")
            os.remove(MODEL_PATH)
            return download_model()  # Riprova
        return True

# Scarica il modello
model_loaded = download_model()
llm = None  # Inizializza a None

if model_loaded:
    # Inizializza il modello SUPER OTTIMIZZATO con Qwen2.5-0.5B
    try:
        llm = Llama(
            model_path=MODEL_PATH,
            n_ctx=2048,  # Aumentato grazie al modello più piccolo
            n_threads=4,  # Più thread possibili con modello piccolo
            n_batch=256,  # Batch size ottimizzato
            use_mlock=False,  # Disabilitato per HF Free
            verbose=False,
            n_gpu_layers=0,
            use_mmap=True,  # Usa memory mapping per efficienza
            low_vram=True,  # Modalità low memory
            rope_scaling_type=1,  # Ottimizzazione RoPE
            rope_freq_base=10000.0
        )
        print("✅ Qwen2.5-0.5B Model loaded successfully!")
    except Exception as e:
        print(f"❌ Error loading model: {e}")
        llm = None
else:
    print("❌ Model not available, using fallback responses")

# System prompt OTTIMIZZATO per Qwen2.5
system_prompt = """<|im_start|>system
You are an expert D&D Dungeon Master. Create immersive, engaging adventures with vivid descriptions. Always end your responses with a question or choice for the player. Keep responses concise but atmospheric.
<|im_end|>"""

def generate_random_opening():
    """Genera un inizio casuale per l'avventura usando l'AI"""
    if llm is None:
        # Fallback solo se il modello non è disponibile
        import random
        openings = [
            "You enter a torch-lit dungeon. Water drips from ancient stones. A passage splits left and right. Which way?",
            "You're in a misty forest clearing. An old well sits in the center, rope disappearing into darkness. Investigate?",
            "The tavern door creaks open. Hooded figures look up from their ale. The barkeep waves you over. Approach?"
        ]
        return f"🌟 **New Adventure!** 🌟\n\n{random.choice(openings)}"
    
    try:
        # Prompt ottimizzato per Qwen2.5
        opening_prompt = f"""{system_prompt}
<|im_start|>user
Generate a creative D&D adventure opening in 2-3 sentences. Set an intriguing scene and end with a question for the player.
<|im_end|>
<|im_start|>assistant"""
        
        output = llm(
            opening_prompt,
            max_tokens=80,  # Leggermente più alto per qualità
            temperature=0.8,
            top_p=0.9,
            repeat_penalty=1.1,
            stop=["<|im_end|>", "<|im_start|>", "User:", "Player:"]
        )
        
        opening = output["choices"][0]["text"].strip()
        
        # Assicurati che finisca con una domanda
        if not opening.endswith('?'):
            opening += " What do you do?"
        
        return f"🌟 **New Adventure!** 🌟\n\n{opening}"
        
    except Exception as e:
        print(f"Error generating opening: {e}")
        return f"🌟 **New Adventure!** 🌟\n\nYou find yourself in a mysterious place. Strange things are happening. What do you do?"

chat_history = []

def generate_dm_response_with_timeout(message, timeout=30):
    """Genera risposta con timeout ridotto per velocità"""
    if llm is None:
        # Fallback responses se il modello non è disponibile
        import random
        fallbacks = [
            "The path ahead is unclear. What's your next move?",
            "You hear footsteps approaching. How do you react?",
            "A mysterious door appears before you. Do you open it?",
            "The ground trembles slightly. What do you do?",
            "You find a strange artifact. Examine it closely?"
        ]
        return random.choice(fallbacks)
    
    try:
        # Prompt ottimizzato per Qwen2.5 con chat template
        prompt = f"{system_prompt}\n"
        
        # Mantieni più contesto grazie al modello efficiente
        context_turns = min(len(chat_history), 3)  # Ultimi 3 turni
        for turn in chat_history[-context_turns:]:
            prompt += f"<|im_start|>user\n{turn['user']}\n<|im_end|>\n"
            prompt += f"<|im_start|>assistant\n{turn['ai']}\n<|im_end|>\n"
        
        prompt += f"<|im_start|>user\n{message}\n<|im_end|>\n<|im_start|>assistant\n"
        
        # Parametri ottimizzati per Qwen2.5-0.5B
        start_time = time.time()
        output = llm(
            prompt, 
            max_tokens=100,  # Aumentato per qualità migliore
            stop=["<|im_end|>", "<|im_start|>", "User:", "Player:"],
            temperature=0.7,
            top_p=0.8,
            repeat_penalty=1.2,
            top_k=40,
            min_p=0.1  # Miglior controllo qualità
        )
        
        # Verifica se ha impiegato troppo tempo
        elapsed_time = time.time() - start_time
        if elapsed_time > timeout:
            print(f"Response took {elapsed_time:.1f}s (timeout: {timeout}s)")
            return "Time passes quickly. What do you do next?"
        
        text = output["choices"][0]["text"].strip()
        
        # Assicurati che ci sia sempre una domanda
        if not text.endswith(('?', '!', '.')):
            text += "?"
        
        print(f"✅ Response generated in {elapsed_time:.1f}s")
        return text
        
    except Exception as e:
        print(f"Error generating response: {e}")
        return "Something unexpected happens. What do you do next?"

def chat(message, history):
    global chat_history
    
    if not message.strip():
        return "You stand there, unsure. What would you like to do?"
    
    # Genera risposta del DM con timeout ridotto
    dm_response = generate_dm_response_with_timeout(message)
    
    # Aggiorna cronologia (mantieni più turni grazie al modello efficiente)
    chat_history.append({"user": message, "ai": dm_response})
    if len(chat_history) > 5:  # Mantieni 5 turni invece di 2
        chat_history = chat_history[-5:]
    
    return dm_response

def reset():
    global chat_history
    chat_history = []
    return generate_random_opening()

# Crea l'interfaccia SUPER OTTIMIZZATA
with gr.Blocks(title="Infinite Dungeon - Lightning Fast", theme=gr.themes.Soft()) as demo:
    gr.Markdown("# ⚡ Infinite Dungeon - Lightning Fast")
    gr.Markdown("*Powered by Qwen2.5-0.5B - Optimized for 5-15 second responses*")
    gr.Markdown("🚀 **Super fast AI D&D with perfect memory retention**")
    
    # Inizializza la chat
    chatbot = gr.Chatbot(
        value=[(None, "⚡ **Lightning Fast Adventure Ready!** ⚡\n\nPress 'New Adventure' to begin your quest!")],
        height=400,
        show_label=False
    )
    
    msg = gr.Textbox(
        label="Your action", 
        placeholder="What do you do? (e.g., 'I search the room', 'I attack the orc', 'I cast a spell')",
        max_lines=2
    )
    
    with gr.Row():
        submit = gr.Button("⚔️ Act", variant="primary", size="lg")
        reset_btn = gr.Button("🔄 New Adventure", variant="secondary")
    
    gr.Markdown("⚡ **Ultra-fast responses**: 5-15 seconds | 🧠 **Perfect memory**: Never forgets your adventure!")
    
    # Funzione per gestire la chat
    def respond(message, chat_history_ui):
        if not message.strip():
            return "", chat_history_ui
        
        # Mostra messaggio di caricamento
        chat_history_ui.append((message, "🎲 *The DM is thinking...*"))
        
        # Genera risposta
        bot_message = chat(message, chat_history_ui)
        chat_history_ui[-1] = (message, bot_message)
        
        return "", chat_history_ui
    
    # Funzione per il reset
    def reset_chat():
        new_opening = reset()
        return [(None, new_opening)]
    
    # Collegamenti eventi
    msg.submit(respond, [msg, chatbot], [msg, chatbot])
    submit.click(respond, [msg, chatbot], [msg, chatbot])
    reset_btn.click(reset_chat, outputs=[chatbot])

# Avvia l'app
if __name__ == "__main__":
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860,
        show_error=True
    )