Colab GPU Acceleration & Remote Development

You can connect free or rented cloud GPUs (Google Colab, Lambda Labs, RunPod) to your local development environment to offload model training, heavy benchmark runs, or LLM inference without heating up your workstation.

Bundled Colab Notebooks

AI Friend ships three production-ready Colab notebooks in notebooks/, each self-contained — no Docker, Postgres, Neo4j, or NATS required:

Notebook	Purpose	GPU Tier
`ai_friend_voice_training.ipynb`	Fine-tunes a GPT-SoVITS voice clone from your own recordings.	T4 (hard requirement), 30-90+ min
`ai_friend_eval_harness.ipynb`	Runs the behavioral eval gate (`backend/evals/`) against real Ollama models.	Helps, not required, 5-20 min/model
`ai_friend_llm_benchmark.ipynb`	Measures raw Ollama generation throughput/latency/VRAM across model sizes.	T4 (that's the point), 5-15 min

Connecting Local VS Code to Colab GPU (Remote SSH)

You can edit files and run terminal commands directly on a Colab GPU from your local VS Code.

Step 1: In Google Colab

Run this snippet in a new Colab cell to create a secure Cloudflare SSH tunnel:

python

!pip install colab_ssh --upgrade
from colab_ssh import launch_ssh_cloudflared
launch_ssh_cloudflared(password="my_secure_password")

Colab will output your SSH config block:

text

Host colab
    HostName <tunnel-id>.trycloudflare.com
    User root
    Port 22

Step 2: In Local VS Code

Install the official Remote - SSH extension.
Open ~/.ssh/config and paste the snippet from Colab.
Press Cmd+Shift+P $\rightarrow$ Remote-SSH: Connect to Host... $\rightarrow$ colab.
Enter your password.

You now have a live terminal and full filesystem access inside the Colab GPU runtime.

← Voice Training on GPU Disaster Recovery & Backup →