devbox.py

import modal
import random
import platform

from images import (
    standard_devbox_image, cuda_devbox_image, doc_processing_image,
    assisted_coding_image, llm_playroom_image, llamacpp_cpu_image, rdp_devbox_image, forensic_analysis_image
)
from shared_runtime import run_devbox_shared, run_rdp_devbox_shared
from config import get_resource_config, LLAMACPP_DEVBOX_ARGS, LLAMACPP_GPU_DEVBOX_ARGS, LLAMACPP_IDLE_TIMEOUT


app = modal.App(
    name="personal-devbox-launcher",
)


dev_volume = modal.Volume.from_name("my-dev-volume", create_if_missing=True)

juicy_secrets = [modal.Secret.from_name("ssh-public-key"), modal.Secret.from_name("gemini-api-key"), modal.Secret.from_name("exa-api-key")] # Right now, all boxes will be filled with these secrets even if it may not directly be intented.

# Common arguments for the devbox functions loaded from config.py and None values such as secret and volume are filled with values defined above
cpu_devbox_args = get_resource_config("cpu", secrets= juicy_secrets, volume = dev_volume)

cpu_devbox_args_rdp = get_resource_config("cpu", secrets = juicy_secrets, volume = dev_volume, is_rdp = True)

gpu_devbox_args = get_resource_config("gpu", secrets = juicy_secrets, volume = dev_volume, is_rdp = True)

gpu_devbox_args_rdp = get_resource_config("gpu", secrets = juicy_secrets, volume = dev_volume, is_rdp = True)

# llama.cpp Research Center config - fill in secrets and volume
llamacpp_devbox_args = LLAMACPP_DEVBOX_ARGS.copy()
llamacpp_devbox_args["secrets"] = juicy_secrets
llamacpp_devbox_args["volumes"] = {"/data": dev_volume}

# GPU llama.cpp Research Center config - fill in secrets and volume
llamacpp_gpu_devbox_args = LLAMACPP_GPU_DEVBOX_ARGS.copy()
llamacpp_gpu_devbox_args["secrets"] = juicy_secrets
llamacpp_gpu_devbox_args["volumes"] = {"/data": dev_volume}

@app.function(image=standard_devbox_image, **cpu_devbox_args)
def launch_devbox(extra_packages: list[str] | None = None):
    """Launches a non-GPU personal development environment."""
    run_devbox_shared(extra_packages, devbox_type="standard_devbox")


@app.function(image=cuda_devbox_image, gpu="t4", **gpu_devbox_args)
def launch_devbox_t4(extra_packages: list[str] | None = None):
    """Launches a T4 GPU-powered personal development environment."""
    run_devbox_shared(extra_packages, devbox_type="cuda_devbox_t4")


@app.function(image=cuda_devbox_image, gpu="l4", **gpu_devbox_args)
def launch_devbox_l4(extra_packages: list[str] | None = None):
    """Launches an L4 GPU-powered personal development environment."""
    run_devbox_shared(extra_packages, devbox_type="cuda_devbox_l4")


@app.function(image=cuda_devbox_image, gpu="a10g", **gpu_devbox_args)
def launch_devbox_a10g(extra_packages: list[str] | None = None):
    """Launches an A10G GPU-powered personal development environment."""
    run_devbox_shared(extra_packages, devbox_type="cuda_devbox_a10g")

@app.function(image=rdp_devbox_image, **cpu_devbox_args_rdp)
def launch_rdp_devbox(extra_packages: list[str] = None):
    """Launches an RDP desktop development environment."""
    run_rdp_devbox_shared(extra_packages)

@app.function(image=rdp_devbox_image, gpu="t4", **gpu_devbox_args_rdp)
def launch_rdp_devbox_t4(extra_packages: list[str] = None):
    """Launches an RDP desktop with T4 GPU."""
    run_rdp_devbox_shared(extra_packages)

@app.function(image=rdp_devbox_image, gpu="l4", **gpu_devbox_args_rdp)
def launch_rdp_devbox_l4(extra_packages: list[str] = None):
    """Launches an RDP desktop with L4 GPU."""
    run_rdp_devbox_shared(extra_packages)

@app.function(image=rdp_devbox_image, gpu="a10g", **gpu_devbox_args_rdp)
def launch_rdp_devbox_a10g(extra_packages: list[str] = None):
    """Launches an RDP desktop with A10G GPU."""
    run_rdp_devbox_shared(extra_packages)

@app.function(image=doc_processing_image, **gpu_devbox_args)
def launch_doc_processor():
    """Launches a document processing environment with Pandoc and TeX Live."""
    run_devbox_shared(extra_packages=None, devbox_type="doc_processing")

@app.function(image=assisted_coding_image, **cpu_devbox_args)
def launch_assisted_coding():
    """Launches a development environment with Gemini CLI & OpenCode pre-installed."""
    run_devbox_shared(extra_packages=None, devbox_type="assisted_coding")

@app.function(image=llm_playroom_image, **gpu_devbox_args_rdp) # Needs review
def launch_llm_playroom():
    """Launches an LLM Playroom environment with Ollama and preloaded models."""
    run_devbox_shared(extra_packages=None, devbox_type="llm_playroom") # add model selection here too!

@app.function(image=forensic_analysis_image, **cpu_devbox_args)
def launch_forensics_image():
  """ Launches A Forensic Analysis Machine With Volatilty3 pre-installed. """
  run_devbox_shared(extra_packages=None, devbox_type="forensic_analysis")

# llama.cpp Research Center - Curated Model Catalog
LLAMACPP_MODELS = {
    "1": {
        "name": "Qwen3.5-9B-Instruct",
        "repo_id": "unsloth/Qwen3.5-9B-GGUF",
        "filename": "Qwen3.5-9B-Q4_K_M.gguf",
        "size": "5.68GB",
        "type": "text",
        "description": "Best overall - thinking mode, 256K context",
        "function_calling": True,
        "settings": {"temp": 0.7, "top_p": 0.8, "top_k": 20}
    },
    "2": {
        "name": "Bonsai-8B",
        "repo_id": "prism-ml/Bonsai-8B-gguf",
        "filename": "Bonsai-8B-Q1_0.gguf",
        "size": "1.15GB",
        "type": "text",
        "description": "1-bit ultra-lightweight - 14x smaller, 6x faster on GPU",
        "function_calling": True,
        "settings": {"temp": 0.5, "top_p": 0.85, "top_k": 20}
    },
    "3": {
        "name": "SmolVLM-500M",
        "repo_id": "ggml-org/SmolVLM-500M-Instruct-GGUF",
        "filename": "ggml-model-mmproj-f16.gguf",
        "size": "~1.5GB",
        "type": "vision",
        "description": "Fast image captioning",
        "function_calling": False,
        "settings": {"temp": 0.7, "top_p": 0.9, "top_k": 40}
    },
    "4": {
        "name": "Gemma4-26B-A4B",
        "repo_id": "unsloth/gemma-4-26b-a4b-it-GGUF",
        "filename": "gemma-4-26b-a4b-it-Q4_K_M.gguf",
        "size": "~17GB",
        "type": "vision",
        "description": "Google's latest - MoE (4B active), 256K context, multimodal",
        "function_calling": True,
        "settings": {"temp": 0.7, "top_p": 0.8, "top_k": 20}
    },
    "5": {
        "name": "Custom Model",
        "repo_id": None,
        "filename": None,
        "size": "User-defined",
        "type": "text",
        "description": "Any GGUF model from HuggingFace",
        "function_calling": True,
        "settings": {"temp": 0.7, "top_p": 0.9, "top_k": 40}
    },
}
#The models list will be moved to a better /dedicated file
@app.function(image=llamacpp_cpu_image, **llamacpp_devbox_args)
def launch_llamacpp_playroom():
    """llama.cpp Research Center with curated models, WebUI, and EXA Search."""
    import os
    import shutil
    import subprocess
    import sys
    import modal
    import time
    import json
    import httpx
    from utils import inject_ssh_key
    from exa_helper import get_exa_api_key, get_tools_for_model
    
    modal.interact()  # Enable user input
    
    print("\n🧠 Launching llama.cpp Research Center")
    print("💻 CPU-Optimized Inference")
    print("🔍 EXA Search Integration")
    print("🌐 Built-in WebUI")
    
    # --- Persistence Setup ---
    persistent_storage_dir = "/data/.config_persistence"
    os.makedirs(persistent_storage_dir, exist_ok=True)
    
    items_to_persist = [
        ".bash_history", ".bashrc", ".profile", ".viminfo", ".vimrc",
        ".gitconfig", ".ssh/config", ".ssh/known_hosts",
    ]
    
    for item in items_to_persist:
        home_path = f"/root/{item}"
        volume_path = f"{persistent_storage_dir}/{item}"
        os.makedirs(os.path.dirname(home_path), exist_ok=True)
        os.makedirs(os.path.dirname(volume_path), exist_ok=True)
        
        if os.path.lexists(home_path):
            if os.path.isdir(home_path) and not os.path.islink(home_path):
                shutil.rmtree(home_path)
            else:
                os.remove(home_path)
        
        os.symlink(volume_path, home_path)
    
    # --- Model Selection ---
    print("\n📦 Model Selection:")
    for key, model in LLAMACPP_MODELS.items():
        print(f"{key}. {model['name']} - {model['description']}")
        print(f"   Size: {model['size']} | Type: {model['type']}")
    
    try:
        model_choice = input("\nEnter model number (1-7): ").strip()
    except EOFError:
        model_choice = "1"
    
    if model_choice not in LLAMACPP_MODELS:
        print("❌ Invalid choice. Defaulting to Qwen3.5-9B")
        model_choice = "1"
    
    selected_model = LLAMACPP_MODELS[model_choice].copy()
    
    # Handle custom model input
    if model_choice == "7":
        print("\n📝 Custom Model Details:")
        repo_id = input("Repo ID (e.g., unsloth/Qwen3.5-9B-GGUF): ").strip()
        filename = input("Filename (e.g., model-Q4_K_M.gguf): ").strip()
        selected_model["repo_id"] = repo_id
        selected_model["filename"] = filename
        selected_model["name"] = filename.replace(".gguf", "")
    
    # --- Download Model ---
    model_dir = "/data/models"
    os.makedirs(model_dir, exist_ok=True)
    model_path = f"{model_dir}/{selected_model['filename']}"
    
    if not os.path.exists(model_path):
        print(f"\n📦 Downloading {selected_model['repo_id']}/{selected_model['filename']}...")
        print("   This may take a few minutes on first run...")
        
        try:
            subprocess.run([
                "hf", "download",
                selected_model["repo_id"],
                selected_model["filename"],
                "--local-dir", model_dir
            ], check=True)
            print(f"✅ Model downloaded to {model_path}")
        except subprocess.CalledProcessError:
            print(f"❌ Download failed. Check repo_id and filename.")
            print(f"   Repo: {selected_model['repo_id']}")
            print(f"   File: {selected_model['filename']}")
            return
    else:
        print(f"✅ Model already cached: {model_path}")
    
    # --- EXA Search Check ---
    exa_key = get_exa_api_key()
    if exa_key:
        print("✅ EXA Search API key found")
        use_exa = True
    else:
        print("⚠️  EXA Search API key not found. Web search disabled.")
        print("   Add 'exa-api-key' secret to Modal to enable.")
        use_exa = False
    
    # --- Start llama-server ---
    model_settings = selected_model.get('settings', {})
    
    server_cmd = [
        "llama-server",
        "-m", model_path,
        "--host", "0.0.0.0",
        "--port", "8080",
        "--threads", "2",
        "--ctx-size", "4096",
        "--batch-size", "512",
        "--mlock",
        "--jinja",  # Enable function calling + WebUI
        "--temp", str(model_settings.get('temp', 0.7)),
        "--top-p", str(model_settings.get('top_p', 0.9)),
        "--top-k", str(model_settings.get('top_k', 40)),
    ]
    
    print(f"\n🚀 Starting llama-server for {selected_model['name']}...")
    print(f"   Settings: temp={model_settings.get('temp', 0.7)}, top_p={model_settings.get('top_p', 0.9)}")
    
    # Start server in background
    server_process = subprocess.Popen(
        server_cmd,
        stdout=subprocess.DEVNULL,
        stderr=subprocess.PIPE
    )
    
    # Wait for server to be ready
    time.sleep(3)
    for i in range(30):
        try:
            resp = httpx.get("http://localhost:8080/health", timeout=1.0)
            if resp.status_code == 200:
                print("✅ llama-server ready!")
                break
        except:
            pass
        time.sleep(1)
    else:
        print("❌ llama-server failed to start. Check model file.")
        server_process.terminate()
        return
    
    # --- Start EXA Proxy (if enabled) ---
    if use_exa:
        from exa_proxy import start_exa_proxy
        print("  Starting EXA Search proxy...")
        start_exa_proxy(port=8081, target_port=8080)
        web_port = 8081
    else:
        web_port = 8080
    
    # --- SSH Setup ---
    inject_ssh_key()
    subprocess.run(["/usr/sbin/sshd"])
    
    # --- Port Forwarding ---
    with modal.forward(22, unencrypted=True) as ssh_tunnel:
        ssh_command = f"ssh root@{ssh_tunnel.host} -p {ssh_tunnel.unencrypted_port}"
        
    with modal.forward(web_port, unencrypted=True) as web_tunnel:
            web_url = f"http://{web_tunnel.host}:{web_tunnel.unencrypted_port}"
            
            print("\n" + "="*60)
            print("🚀 Your llama.cpp Research Center is ready!")
            print("="*60)
            print(f"\n🌐 WebUI: {web_url}")
            print(f"   (Open in browser for chat interface)")
            print(f"\n🔐 SSH: {ssh_command}")
            print(f"\n🧠 Model: {selected_model['name']} ({selected_model['size']})")
            if use_exa:
                print(f"🔍 EXA Search: Enabled (function calling)")
            print(f"\n⏰ Idle timeout: {LLAMACPP_IDLE_TIMEOUT // 60} minutes")
            print("\n" + "="*60)
            
            # --- Idle Monitor ---
            idle_time = 0
            check_interval = 30
            
            while idle_time < LLAMACPP_IDLE_TIMEOUT:
                time.sleep(check_interval)
                
                # Check for active SSH sessions
                ssh_result = subprocess.run(
                    "ps -ef | grep 'sshd: root@' | grep -v grep",
                    shell=True,
                    capture_output=True,
                )
                
                # Check for active WebUI connections
                web_active = False
                try:
                    stats = httpx.get("http://localhost:8080/health", timeout=1.0)
                    web_active = True  # Server is running
                except:
                    pass
                
                if ssh_result.stdout or web_active:
                    idle_time = 0
                else:
                    idle_time += check_interval
                    remaining = LLAMACPP_IDLE_TIMEOUT - idle_time
                    print(f"\r💤 No activity. Shutting down in {remaining // 60}m {remaining % 60}s...", end="", flush=True)
            
            print(f"\n\n⏰ Idle timeout reached ({LLAMACPP_IDLE_TIMEOUT}s). Shutting down.")
            server_process.terminate()

@app.function(image=llamacpp_cpu_image, **llamacpp_gpu_devbox_args)
def launch_llamacpp_playroom_gpu():
    """GPU-accelerated llama.cpp Research Center with curated models, WebUI, and EXA Search."""
    import os
    import shutil
    import subprocess
    import sys
    import modal
    import time
    import json
    import httpx
    from utils import inject_ssh_key
    from exa_helper import get_exa_api_key, get_tools_for_model
    
    modal.interact()
    
    print("\n🧠 Launching llama.cpp Research Center")
    print("🚀 GPU-Accelerated Inference (T4)")
    print("🔍 EXA Search Integration")
    print("🌐 Built-in WebUI")
    
    # --- Persistence Setup ---
    persistent_storage_dir = "/data/.config_persistence"
    os.makedirs(persistent_storage_dir, exist_ok=True)
    
    items_to_persist = [
        ".bash_history", ".bashrc", ".profile", ".viminfo", ".vimrc",
        ".gitconfig", ".ssh/config", ".ssh/known_hosts",
    ]
    
    for item in items_to_persist:
        home_path = f"/root/{item}"
        volume_path = f"{persistent_storage_dir}/{item}"
        os.makedirs(os.path.dirname(home_path), exist_ok=True)
        os.makedirs(os.path.dirname(volume_path), exist_ok=True)
        
        if os.path.lexists(home_path):
            if os.path.isdir(home_path) and not os.path.islink(home_path):
                shutil.rmtree(home_path)
            else:
                os.remove(home_path)
        
        os.symlink(volume_path, home_path)
    
    # --- Model Selection ---
    print("\n📦 Model Selection:")
    for key, model in LLAMACPP_MODELS.items():
        print(f"{key}. {model['name']} - {model['description']}")
        print(f"   Size: {model['size']} | Type: {model['type']}")
    
    try:
        model_choice = input("\nEnter model number (1-7): ").strip()
    except EOFError:
        model_choice = "1"
    
    if model_choice not in LLAMACPP_MODELS:
        print("❌ Invalid choice. Defaulting to Qwen3.5-9B")
        model_choice = "1"
    
    selected_model = LLAMACPP_MODELS[model_choice].copy()
    
    # Handle custom model input
    if model_choice == "7":
        print("\n📝 Custom Model Details:")
        repo_id = input("Repo ID (e.g., unsloth/Qwen3.5-9B-GGUF): ").strip()
        filename = input("Filename (e.g., model-Q4_K_M.gguf): ").strip()
        selected_model["repo_id"] = repo_id
        selected_model["filename"] = filename
        selected_model["name"] = filename.replace(".gguf", "")
    
    # --- Download Model ---
    model_dir = "/data/models"
    os.makedirs(model_dir, exist_ok=True)
    model_path = f"{model_dir}/{selected_model['filename']}"
    
    if not os.path.exists(model_path):
        print(f"\n📦 Downloading {selected_model['repo_id']}/{selected_model['filename']}...")
        print("   This may take a few minutes on first run...")
        
        try:
            subprocess.run([
                "hf", "download",
                selected_model["repo_id"],
                selected_model["filename"],
                "--local-dir", model_dir
            ], check=True)
            print(f"✅ Model downloaded to {model_path}")
        except subprocess.CalledProcessError:
            print(f"❌ Download failed. Check repo_id and filename.")
            print(f"   Repo: {selected_model['repo_id']}")
            print(f"   File: {selected_model['filename']}")
            return
    else:
        print(f"✅ Model already cached: {model_path}")
    
    # --- EXA Search Check ---
    exa_key = get_exa_api_key()
    if exa_key:
        print("✅ EXA Search API key found")
        use_exa = True
    else:
        print("⚠️  EXA Search API key not found. Web search disabled.")
        print("   Add 'exa-api-key' secret to Modal to enable.")
        use_exa = False
    
    # --- Start llama-server with GPU optimizations ---
    model_settings = selected_model.get('settings', {})
    
    server_cmd = [
        "llama-server",
        "-m", model_path,
        "--host", "0.0.0.0",
        "--port", "8080",
        "--ctx-size", "32768",
        "--batch-size", "512",
        "-ngl", "99",         # Offload all layers to GPU
        "-fa", "on",          # Flash Attention
        "-ctk", "q8_0",       # Key cache quantization
        "-ctv", "q8_0",       # Value cache quantization
        "--jinja",
        "--temp", str(model_settings.get('temp', 0.7)),
        "--top-p", str(model_settings.get('top_p', 0.9)),
        "--top-k", str(model_settings.get('top_k', 40)),
    ]
    
    print(f"\n🚀 Starting GPU-accelerated llama-server for {selected_model['name']}...")
    print(f"   GPU: T4 (all layers offloaded)")
    print(f"   Settings: temp={model_settings.get('temp', 0.7)}, top_p={model_settings.get('top_p', 0.9)}")
    print(f"   Context: 32K with KV cache quantization")
    
    # Start server in background
    server_process = subprocess.Popen(
        server_cmd,
        stdout=subprocess.DEVNULL,
        stderr=subprocess.PIPE
    )
    
    # Wait for server to be ready
    time.sleep(3)
    for i in range(30):
        try:
            resp = httpx.get("http://localhost:8080/health", timeout=1.0)
            if resp.status_code == 200:
                print("✅ llama-server ready!")
                break
        except:
            pass
        time.sleep(1)
    else:
        print("❌ llama-server failed to start. Check model file.")
        server_process.terminate()
        return
    
    # --- Start EXA Proxy (if enabled) ---
    if use_exa:
        from exa_proxy import start_exa_proxy
        print("  Starting EXA Search proxy...")
        start_exa_proxy(port=8081, target_port=8080)
        web_port = 8081
    else:
        web_port = 8080
    
    # --- SSH Setup ---
    inject_ssh_key()
    subprocess.run(["/usr/sbin/sshd"])
    
    # --- Port Forwarding ---
    with modal.forward(22, unencrypted=True) as ssh_tunnel:
        ssh_command = f"ssh root@{ssh_tunnel.host} -p {ssh_tunnel.unencrypted_port}"
    
    with modal.forward(web_port, unencrypted=True) as web_tunnel:
            web_url = f"http://{web_tunnel.host}:{web_tunnel.unencrypted_port}"
            
            print("\n" + "="*60)
            print("🚀 Your GPU llama.cpp Research Center is ready!")
            print("="*60)
            print(f"\n🌐 WebUI: {web_url}")
            print(f"   (Open in browser for chat interface)")
            print(f"\n🔐 SSH: {ssh_command}")
            print(f"\n🧠 Model: {selected_model['name']} ({selected_model['size']})")
            print(f"🎮 GPU: T4 (all layers offloaded)")
            if use_exa:
                print(f"🔍 EXA Search: Enabled (function calling)")
            print(f"\n⏰ Idle timeout: {LLAMACPP_IDLE_TIMEOUT // 60} minutes")
            print("\n" + "="*60)
            
            # --- Idle Monitor ---
            idle_time = 0
            check_interval = 30
            
            while idle_time < LLAMACPP_IDLE_TIMEOUT:
                time.sleep(check_interval)
                
                # Check for active SSH sessions
                ssh_result = subprocess.run(
                    "ps -ef | grep 'sshd: root@' | grep -v grep",
                    shell=True,
                    capture_output=True,
                )
                
                # Check for active WebUI connections
                web_active = False
                try:
                    stats = httpx.get("http://localhost:8080/health", timeout=1.0)
                    web_active = True
                except:
                    pass
                
                if ssh_result.stdout or web_active:
                    idle_time = 0
                else:
                    idle_time += check_interval
                    remaining = LLAMACPP_IDLE_TIMEOUT - idle_time
                    print(f"\r💤 No activity. Shutting down in {remaining // 60}m {remaining % 60}s...", end="", flush=True)
            
            print(f"\n\n⏰ Idle timeout reached ({LLAMACPP_IDLE_TIMEOUT}s). Shutting down.")
            server_process.terminate()
# Menu-driven local entrypoint.
@app.local_entrypoint()
def main():
  
    from ui_utils import create_box, show_spinner
    from quotes_loader import get_random_quote, format_quote
    from utils import inject_ssh_key, display_system_info
    from config import IDLE_TIMEOUT_SECONDS
    from gpu_utils import get_gpu_config, get_available_gpus

    banner = """
╔══════════════════════════════════════════════════════════╗
║                                                          ║
║            🚀  DEVBOX LAUNCHER  🚀                ║
║                                                          ║
║              " Potatoes Taste Nice 🤔 "               ║
║                                                          ║
╚══════════════════════════════════════════════════════════╝
"""
    print(banner)

    quote = get_random_quote()
    quote_box = format_quote(quote)
    create_box(quote_box, "💭 Programming Wisdom")
    display_system_info()

    menu_box = """
    🎯 Choose your DevBox Config:
    1. 🛠️  Standard DevBox general purpose development environment with optional extra packages
    
    2. 📄 Document Processing Box
    Pandoc + Full TeX Live for document work
    
    3. 🤖 AI Assistants Box
    Includes OpenCode and Gemini CLI
    
    4. 🧠 LLM Playroom 
    Ollama for AI experimentation
    
    5. 🖥️  RDP Desktop Box Full graphical desktop with XFCE and RDP access
    
    6. 🔬 llama.cpp Research Center (CPU)
    Curated models + WebUI + EXA Search + Multimodal
    
    7. 🔍 Forensics Analysis
    Analysis Machine using Volatilty3 and other tools

    8. 🚀 llama.cpp Research Center (GPU)
    GPU-accelerated inference + 32K context + T4 GPU
    """
    create_box(menu_box, "🚀 LAUNCH OPTIONS")
    
    try:
      choice = input("Enter your choice (1-8): ").strip()
    except EOFError:
      print("\nNo input received. Exiting.")
      return
      
    print(choice)
    if choice == "1": #remember to adjust subsequent indents at package_selction
      package_box = """
      📦 Want to install additional tools?
      Examples: htop tmux git neovim curl wget
      (leave empty for default setup)
      """
      create_box(package_box, "🛠️  EXTRA PACKAGES")
      try:
        tools_input = input("Enter tools (space-separated): ").strip()
      except EOFError:
        tools_input = ""

      package_list = tools_input.split() if tools_input else []
   
      if package_list:
        print(f"✅ Requesting with additional tools: {', '.join(package_list)}")
      else:
        print("✅ No extra tools requested.")

      gpu_box = """
      🎮 Add GPU acceleration?
      • T4: Cost-effective, good for inference
      • L4: More Performant than T4
      • A10G: Higher performance, more VRAM
      (Enter 'y' To Attach a GPU, anything else for CPU-only)

      ⚠️  GPU usage incurs charges on your Modal account.
         See modal.com/pricing for current rates.
      """
      create_box(gpu_box, "⚡ GPU ACCELERATION")
      
      try:
        gpu_choice = input("Attach GPU? (y/n): ").lower().strip()
      except EOFError:
        gpu_choice = "n"

      if gpu_choice == "y":
        gpu_menu = """
        1. 🎯 T4 GPU (Cost-effective, good for inference)
        2. 🚀 L4 GPU (More Performant than T4)
        3. 💪 A10G GPU (Higher performance, more VRAM)
        """
        create_box(gpu_menu, "🎮 SELECT GPU TYPE")
        try:
          gpu_type_choice = input("Choose GPU (1-3): ").strip()
        except EOFError:
          print("\nNo input received. Exiting.")
          return

          gpu_types = {
            "1": ("T4", launch_devbox_t4),
            "2": ("L4", launch_devbox_l4),
            "3": ("A10G", launch_devbox_a10g),
            }

          if gpu_type_choice in gpu_types:
            gpu_name, launch_func = gpu_types[gpu_type_choice]
            
            gpu_launch_box = f"🎯 Launching with {gpu_name} GPU... Silicon Dust 🔥"
            create_box(gpu_launch_box, f"🚀 {gpu_name} POWERED")
            show_spinner("Initializing GPU environment", 2)
            launch_func.remote(extra_packages=package_list)
          else:
            print("❌ Invalid GPU choice. Please run again.")
            return
      else:
        cpu_box = """🖥️  Launching CPU-only environment...
        Potatoes Have High Carbohydrates Content 😉
        """
        create_box(cpu_box, "🚀 STANDARD DEVBOX")
        show_spinner("Preparing your DevBox", 2)
        launch_devbox.remote(extra_packages=package_list)

    elif choice == "2":
        doc_box = """
        📄 Launching Document Processing Box...
        📚 Pandoc + Full TeX Live Distribution
        """
        create_box(doc_box, "📄 DOCUMENT PROCESSING")
        show_spinner("Setting up document tools", 2)
        launch_doc_processor.remote()

    elif choice == "3":
        assistants_box = """
        🤖 Launching AI Assistants Box...
        🧠 Includes OpenCode and Gemini CLI
        🚀 Let's build something amazing together!
        """
        create_box(assistants_box, "🤖 AI ASSISTANTS")
        show_spinner("Initializing AI assistants", 2)
        launch_assisted_coding.remote()

    elif choice == "4":
        llm_box = """
        🧠 Launching LLM Playroom...
        🤖 Ollama with DeepSeek R1 (8B distilled)
        🚀 Ready for AI experimentation!
        """
        create_box(llm_box, "🧠 LLM PLAYROOM")
        show_spinner("Initializing LLM environment", 2)
        launch_llm_playroom.remote()

    elif choice == "5":
        rdp_box = """
        🖥️  Launching RDP Desktop Box...
        🖼️  XFCE Desktop Environment + RDP access
        """
        create_box(rdp_box, "🖥️  RDP DESKTOP")
        show_spinner("Setting up desktop environment", 2)
        
        package_box = """
        📦 Want to install additional desktop tools?
        Examples: firefox gedit vscode libreoffice(leave empty for default XFCE setup)
        """
        create_box(package_box, "🖥️  EXTRA DESKTOP PACKAGES")
        try:
          tools_input = input("Enter desktop tools(space-separated): ").strip()
        except EOFError:
          tools_input = ""

        package_list = tools_input.split() if tools_input else []

        if package_list:
          print(f"✅ Requesting with additional tools: {', '.join(package_list)}")
        else:
          print("✅ No extra desktop tools requested.")

        gpu_box = """
        🎮 Add GPU acceleration for desktop?
        • T4: Cost-effective, good for graphics
        • L4: Newer, more performant than T4
        • A10G: Higher performance, more VRAM
        (Enter 'y' for GPU options, anything else for CPU-only)

        ⚠️  GPU usage incurs charges on your Modal account.
           See modal.com/pricing for current rates.
        """
        create_box(gpu_box, "⚡ GPU ACCELERATION")

        try:
            gpu_choice = input("Attach GPU? (y/n): ").lower().strip()
        except EOFError:
            gpu_choice = "n"

        if gpu_choice == "y":
            gpu_menu = """
            1. 🎯 T4 GPU (Cost-effective, good for graphics)
            2. 🚀 L4 GPU (Newer, more performant than T4)
            3. 💪 A10G GPU (Higher performance, more VRAM)
            """
            create_box(gpu_menu, "🎮 SELECT GPU TYPE")
            try:
                gpu_type_choice = input("Choose GPU (1-3): ").strip()
            except EOFError:
                print("\nNo input received. Exiting.")
                return

            gpu_types = {
                "1": ("T4", launch_rdp_devbox_t4),
                "2": ("L4", launch_rdp_devbox_l4),
                "3": ("A10G", launch_rdp_devbox_a10g),
            }

            if gpu_type_choice in gpu_types:
                gpu_name, launch_func = gpu_types[gpu_type_choice]
                gpu_launch_box = f"""
                🎯 Launching RDP Desktop with {gpu_name} GPU...
                """
                create_box(gpu_launch_box, f"🚀 {gpu_name} POWERED RDP")
                show_spinner("Initializing GPU RDP environment", 2)
                launch_func.remote(extra_packages=package_list)
            else:
                print("❌ Invalid GPU choice. Please run again.")
                return
        else:
            cpu_box = """
            🖥️  Launching RDP Desktop (CPU-only)...
            """
            create_box(cpu_box, "🚀 RDP DESKTOP")
            show_spinner("Preparing your RDP Desktop", 2)
            launch_rdp_devbox.remote(extra_packages=package_list)

    elif choice == "6":
        research_box = """
        🔬 Launching llama.cpp Research Center...
        
        Models available:
        • Qwen3.5-9B (General reasoning)
        • Bonsai-8B (1-bit ultra-lightweight)
        • SmolVLM-500M (Image captioning)
        • Gemma4-26B-A4B (Multimodal)
        • Custom (Any HuggingFace GGUF)
        
        Features:
        • Built-in WebUI (browser chat)
        • EXA Search integration
        • Function calling support
        """
        create_box(research_box, "🔬 LLAMA.CPP RESEARCH CENTER")
        show_spinner("Preparing research environment", 3)
        launch_llamacpp_playroom.remote()
 
    elif choice == "7":
        forensics_box = """
        Immerse Yourself Into The Bits.... 01101010101010100101
        """
        create_box(forensics_box, "Forensics Machine")
        show_spinner("Interleaving Bits... Hold On Tight!", 3)
        launch_forensics_image.remote()

    elif choice == "8":
        research_gpu_box = """
        🚀 Launching GPU llama.cpp Research Center...
        
        Models available:
        • Qwen3.5-9B (General reasoning)
        • Bonsai-8B (1-bit ultra-lightweight)
        • SmolVLM-500M (Image captioning)
        • Gemma4-26B-A4B (Multimodal)
        • Custom (Any HuggingFace GGUF)
        
        GPU Features:
        • T4 GPU with all layers offloaded
        • 32K context with KV cache quantization
        • Flash Attention for faster inference
        • Built-in WebUI (browser chat)
        • EXA Search integration
        """
        create_box(research_gpu_box, "🚀 GPU LLAMA.CPP RESEARCH CENTER")
        show_spinner("Preparing GPU-accelerated environment", 3)
        launch_llamacpp_playroom_gpu.remote()

    else:
        error_box = """
        ❌ Invalid choice selected.
        Please run the launcher again and choose:
        • 1 for Standard DevBox
        • 2 for Document Processing
        • 3 for AI Assistants Box
        • 4 for LLM Playroom
        • 5 for RDP Desktop Box
        • 6 for llama.cpp Research Center (CPU)
        • 7 for Forensics Machine
        • 8 for llama.cpp Research Center (GPU)
        """
        create_box(error_box, "❌ ERROR")