Hydra Personal GPU Cluster

This repo now has a clean Colab notebook that turns a Google Colab GPU runtime into a personal remote GPU cluster.

It does not install model repos. It only starts:

• a FastAPI control server
• an ngrok tunnel
• token-authenticated routes for command execution, package installs, jobs/logs, upload, and download

What To Upload

Upload/open this in Colab:

notebooks/hydra_personal_gpu_cluster.ipynb

notebooks/hydra_a100_cluster.ipynb is also the same cluster-only notebook for convenience.

Colab Setup

1. Open the notebook in Google Colab.
2. Select a GPU runtime.
3. Add a Colab Secret named NGROK_AUTHTOKEN.
4. Run the notebook top to bottom.
5. Copy the printed HYDRA_CLUSTER_URL and HYDRA_CLUSTER_TOKEN.

The notebook prints:

HYDRA_CLUSTER_URL=https://...
HYDRA_CLUSTER_TOKEN=...
HYDRA_CLUSTER_DOCS_URL=https://.../docs
HYDRA_CLUSTER_OPENAPI_URL=https://.../openapi.json

Local Control

From this repo:

export HYDRA_CLUSTER_URL="https://your-ngrok-url"
export HYDRA_CLUSTER_TOKEN="printed-token"

python3 scripts/hydra_personal_cluster_client.py health
python3 scripts/hydra_personal_cluster_client.py status
python3 scripts/hydra_personal_cluster_client.py exec -- nvidia-smi
python3 scripts/hydra_personal_cluster_client.py exec -- python -m pip --version

Install from here:

python3 scripts/hydra_personal_cluster_client.py pip-install -- numpy==1.26.4 opencv-python-headless
python3 scripts/hydra_personal_cluster_client.py apt-install -- ffmpeg git
python3 scripts/hydra_personal_cluster_client.py packages --raw

Move files:

python3 scripts/hydra_personal_cluster_client.py files /content
python3 scripts/hydra_personal_cluster_client.py upload local.file /content/local.file
python3 scripts/hydra_personal_cluster_client.py download /content/remote.file outputs/remote.file

Run long jobs:

python3 scripts/hydra_personal_cluster_client.py exec --no-wait --timeout 7200 -- python train_or_infer.py
python3 scripts/hydra_personal_cluster_client.py jobs
python3 scripts/hydra_personal_cluster_client.py logs <job_id>

Routes

The notebook server exposes:

• GET /health
• GET /status
• POST /exec
• GET /jobs
• GET /jobs/{job_id}
• GET /logs/{job_id}
• POST /files/upload
• GET /files/download
• GET /files/list

All routes require:

Authorization: Bearer <HYDRA_CLUSTER_TOKEN>

Treat the token like root access to the Colab runtime.