Help & Reference

Everything you need to use Ozekai — CLI commands, Jupyter, file uploads, and tips.

🚀 Jupyter starts automatically

When you run connect-gpu, a JupyterLab session is launched inside your GPU machine automatically. Just open http://localhost:8888 in your browser — no setup required. You'll see the URL printed every time you connect.

CLI Commands

Ozekai has three commands. No flags to memorize — just run them and follow the prompts.

connect-gpu

Connect to your GPU machine. Prompts for your username and password, then drops you into a shell with full GPU access. Your Jupyter URL is printed on connect.

# Run it — that's all
$ connect-gpu
Username: alice
Password: ****
Container ready: d4f523a74c4d
Jupyter: http://localhost:8888
Connecting...

Session caching: After logging in, your credentials are cached for 25 minutes on that terminal. Subsequent commands won't prompt you again.

disconnect-gpu

Stop your GPU machine and billing. Your files in /workspace are preserved — they'll be there when you reconnect. Uses your cached session if still valid. You can also disconnect directly from your Dashboard — just click Stop Session & End Billing.

$ disconnect-gpu
Authenticating as alice...
Usage tracked: 3842 seconds.
Container user-alice stopped.

Important: Typing exit in the shell disconnects your terminal but does not stop billing. Run disconnect-gpu or use the Dashboard stop button.

upload-gpu

Upload files from your local machine into your GPU container's /workspace. Uses scp-like syntax — supports single files, directories, and glob patterns.

# Upload a single file
$ upload-gpu train.py

# Upload a directory
$ upload-gpu my_project/

# Upload multiple files using glob patterns
$ upload-gpu *.py

# Upload to a specific subdirectory
$ upload-gpu train.py config.yaml models/

# Upload all Python files to a subdirectory
$ upload-gpu *.py src/ experiments/

Glob patterns work: Your shell expands *.py, src/**/*.py, and other patterns before upload. Use them the same way you would with scp or cp.

Using Jupyter

Auto-started session

JupyterLab launches automatically when your container starts. Open your browser to http://localhost:8888 — no token or password required.

Jupyter is available on every connect-gpu — whether it's your first connection, a reconnect after disconnect-gpu, or a reconnect after exit.

Getting Files Into Your Container

You have multiple options for getting your code and data into your GPU machine:

upload-gpu — Upload files from your local machine (see above).
Jupyter upload — Use the upload button in JupyterLab's file browser.
git clone — Clone repos directly inside your container shell.
HuggingFace Hub — Pull models and datasets with huggingface_hub or transformers.
wget / curl — Download anything from the internet directly.

# Clone a repo
$ git clone https://github.com/your-org/your-repo.git

# Download a HuggingFace model
$ python -c "from transformers import AutoModel; AutoModel.from_pretrained('gpt2')"

# Download a dataset
$ wget https://example.com/dataset.tar.gz

Pre-installed Software

Every GPU machine comes with a production-ready ML stack pre-installed:

Python 3 with pip and conda (Miniforge)
PyTorch with CUDA support
JupyterLab (auto-started)
CUDA & cuDNN — full development toolkit
Common tools — git, vim, nano, htop, tmux, screen, wget, curl
Build tools — gcc, g++, cmake, ninja-build

Need something else? Install it with pip install or conda install. Your installed packages persist across reconnections.

Tips & Best Practices

Your /workspace directory persists across sessions. Keep your work there.
Always disconnect when you're done — run disconnect-gpu in the terminal or click Stop Session on your Dashboard. Note: exit alone doesn't stop billing.
Large datasets? Download them inside the container with wget or HuggingFace Hub rather than uploading from your laptop.
Use tmux or screen inside the container for long-running jobs that might outlast your terminal session.
Your Jupyter session survives terminal disconnects — just reopen the browser tab.

Contact & Support

Need help or have a question? Here's how to reach us:

Email: [email protected] — we respond within 24 hours.

Contact Support