Skip to content

Cloud & GPU Commands

Vllama integrates with Kaggle to let you offload compute-heavy tasks to their free GPU — no payment required, just a Kaggle account.

vllama login — Authenticate with Kaggle

Save your Kaggle credentials so Vllama can submit kernels on your behalf.

Syntax

vllama login --service kaggle [--username <user>] [--key <api_key>]

How to Get Your Kaggle API Key

  1. Go to kaggle.com and log in
  2. Click your profile picture → Settings
  3. Scroll to API section → Create New Token
  4. A kaggle.json file downloads containing your username and key

Examples

# Provide credentials directly
vllama login --service kaggle --username manvith --key abcdef1234567890

# Use existing kaggle.json at ~/.kaggle/kaggle.json
vllama login --service kaggle

vllama logout — Remove Credentials

vllama logout

Removes saved Kaggle credentials from your machine.

vllama init gpu — Initialize a GPU Session

Prepares a GPU-enabled Kaggle kernel for use.

vllama init gpu --service kaggle

This is typically called automatically when you run commands with --service kaggle. You can run it manually to verify connectivity.

Using Kaggle GPU with Any Command

Add --service kaggle to these commands to offload to Kaggle's free T4 GPU:

# Image generation on Kaggle GPU
vllama run stabilityai/sd-turbo --service kaggle --prompt "..."

# Video generation on Kaggle GPU
vllama run_video damo-vilab/text-to-video-ms-1.7b --service kaggle --prompt "..."

# Image to 3D on Kaggle GPU
vllama image3d --path photo.jpg --service kaggle

# Video to 3D on Kaggle GPU
vllama video3d --path clip.mp4 --service kaggle

How It Works

When you add --service kaggle:

  1. Vllama creates a Kaggle notebook kernel via the Kaggle API
  2. The kernel runs with a free T4 GPU enabled
  3. Vllama installs the required dependencies inside the kernel
  4. The model runs and generates the output
  5. The output file is downloaded to your local machine automatically

Kaggle GPU limits

Kaggle provides approximately 30 GPU hours per week for free accounts. Each image generation takes a few minutes.

Environment Variables

Alternatively, set credentials as environment variables instead of using vllama login:

export KAGGLE_USERNAME=your_username
export KAGGLE_KEY=your_api_key

Or create a .env file in your project directory:

KAGGLE_USERNAME=your_username
KAGGLE_KEY=your_api_key
HF_HOME=/path/to/model/cache        # optional: custom model cache dir
HF_TOKEN=your_huggingface_token     # optional: for gated models like Llama 2

For a detailed step-by-step walkthrough, see the No GPU Guide.