TOPH PHASE 0 | Split Checklist

☐ PART 1: LAPTOP

ALIENWARE LAPTOP • RTX 4050 • 6GB • DO THIS FIRST

1. Get your laptop's IP address

Open Command Prompt. Run: ipconfig Find IPv4 Address under your active adapter. Write it down. You'll need it on the desktop later.

2. Download llama.cpp (CUDA build for Windows)

Go to:
https://github.com/ggerganov/llama.cpp/releases

Look for the latest release. Download the file named something like:
llama-b____-bin-win-cuda-cu12.4-x64.zip

Your laptop has CUDA 13.0 / Driver 581.95. The cu12.4 build should work (backward compatible). If it doesn't, grab the Vulkan build instead as fallback:
llama-b____-bin-win-vulkan-x64.zip

3. Extract to C:\llama

Create the folder and extract: mkdir C:\llama Extract the ZIP contents into C:\llama
Verify rpc-server.exe exists in the folder. That's the only file that matters on this machine.

4. Open firewall port 50052

Open Command Prompt as Administrator (right-click → Run as administrator). Run: netsh advfirewall firewall add rule name="llama-rpc" dir=in action=allow protocol=TCP localport=50052 This lets the desktop connect to the laptop over your home network. Only affects local network, not exposed to internet.

5. Start the RPC server

Open a regular Command Prompt. Run: cd C:\llama rpc-server.exe --host 0.0.0.0 --port 50052 You should see:
main: server is listening on 0.0.0.0:50052

Leave this window open. Don't close it.
The laptop is now a remote GPU waiting for the desktop to connect.

LAPTOP IS DONE. No model download needed. No Python. No Hugging Face account. Just llama.cpp extracted, firewall opened, RPC server running. Walk to the desktop.

☐ PART 2: DESKTOP

ALIENWARE AURORA • RTX 5070 • 9GB • DO THIS AFTER LAPTOP IS RUNNING

1. Get your desktop's IP address

Open Command Prompt. Run: ipconfig Write down the IPv4 Address. You'll also need the laptop IP you wrote down earlier.

2. Verify you can reach the laptop

ping LAPTOP_IP_HERE Replace with the actual IP you wrote down from the laptop. If you get replies, they're connected. If "Request timed out" — check that both machines are on the same WiFi/router.

3. Download llama.cpp (same CUDA build)

Same file as laptop:
https://github.com/ggerganov/llama.cpp/releases

Download the same cuda-cu12.4-x64.zip (or Vulkan). Extract to C:\llama

4. Download the model

Go to:
https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF

Scroll down to find: Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf (~4.92 GB)
Click the filename to download. Save to: C:\llama\models\

Create the models folder first if needed: mkdir C:\llama\models
If it asks for a Hugging Face login, create a free account:
https://huggingface.co/join

5. Start distributed inference

This is the moment. Open Command Prompt. Run: cd C:\llama llama-server.exe --model C:\llama\models\Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf --host 0.0.0.0 --port 8080 --n-gpu-layers 33 --rpc LAPTOP_IP_HERE:50052 --ctx-size 4096 --chat-template llama3 Replace LAPTOP_IP_HERE with your laptop's actual IP address.

What happens:
→ Desktop loads the model
→ Detects laptop RPC server as remote GPU
→ Splits layers across BOTH GPUs
→ RTX 5070 takes majority
→ RTX 4050 takes the rest
→ Web server starts on port 8080

6. Open the chat

Open a browser on the desktop: http://localhost:8080 Type a prompt. Hit enter. Watch it run.

Or from the laptop browser (while RPC is still running): http://DESKTOP_IP_HERE:8080

7. Verify both GPUs are working

While it's generating a response, run on BOTH machines: nvidia-smi Desktop RTX 5070: VRAM in use ✓
Laptop RTX 4050: VRAM in use ✓

Both showing memory usage = distributed inference is live.
Phase 0 complete. Two machines. One model. No cloud. No kill switch.

DESKTOP IS DONE. You now have a distributed AI running on hardware you own, on internet you pay for, with a model that's free and open. No API key. No subscription. No platform. Nobody can turn this off except you.