kagglex

Execute local Python code, modules, and experiments seamlessly on Kaggle GPUs and TPUs.

Overview

kagglex enables machine learning practitioners and researchers to transparently package and dispatch local Python code, standalone scripts, or complete packages to Kaggle's cloud GPU and TPU environments without tedious manual uploading or notebook maintenance.

Features

• Flexible project detection for standalone scripts, flat packages, or src/ layout projects
• Pre-flight validation with local AST syntax checking and credential verification
• Automatic ignore filtering powered by pathspec supporting .gitignore and .kaggleignore
• Dual-tier packaging preventing bloated uploads and base64 truncation
• Automated multi-GPU execution using torchrun
• Secure Kaggle Secrets integration for WandB, HuggingFace, and custom credentials
• Local experiment history repository for tracking past runs, statuses, and durations
• Interactive REPL execution on active Kaggle notebooks via Jupyter proxy URL
• Programmatic Python SDK alongside the kagglex CLI

Installation

Install via uv or pip:

uv pip install kagglex

CLI Usage

Run a Standalone Script

kagglex run --file train.py --gpu t4-2x --title "Pilot Training"

Run a Module with Multi-GPU

kagglex run "python -m mypkg.train --epochs 10" --gpu t4-2x --multi-gpu

Stream Remote Logs in Real-Time

kagglex run "python train.py" --stream

List Recent Runs

kagglex list

Check Status or Cancel a Run

kagglex status my-experiment
kagglex cancel my-experiment

Note: Kaggle's public API does not support programmatic session cancellation for running kernels. kagglex cancel marks the local run record as cancelled and provides the direct Kaggle URL to stop the session in the web interface.

Pull Downloaded Outputs

kagglex pull my-experiment --output-dir ./results --include-outputs "*.json" "checkpoints/*"

Push a Kaggle Dataset

kagglex dataset push --data-dir ./data/embeddings --title "Embeddings Dataset"

Auto-Dataset Payload Offloading

When your project files or local assets exceed the inline payload limit (5 MB), automatically stage and upload them as a private Kaggle dataset:

kagglex run --file train.py --auto-dataset

Inspect GPU and TPU Quota Usage

Track your rolling 7-day accelerator consumption against Kaggle's weekly quotas (30 GPU hours / 20 TPU hours):

kagglex quota --days 7

Because Kaggle's public API does not expose an endpoint to query remaining weekly quota balances, kagglex computes usage locally from run records in .kagglex/runs.json over a configurable rolling window (default: 7 days). Limit thresholds can be configured via --gpu-limit / --tpu-limit or in pyproject.toml.

Declarative Configuration

Define project defaults in pyproject.toml or kagglex.toml to avoid repetitive CLI arguments:

# pyproject.toml
[tool.kagglex]
gpu = "t4-2x"
multi_gpu = true
kaggle_secrets = ["WANDB_API_KEY", "HF_TOKEN"]
datasets = ["username/my-dataset"]
include_outputs = ["*.json", "checkpoints/*"]

[tool.kagglex.env]
WANDB_PROJECT = "my-experiment"

Then simply run:

kagglex run --file train.py

Interactive REPL on Running Kaggle Notebooks

When a notebook is already open in Kaggle, copy its proxy URL (Run -> Kaggle Jupyter Server -> Copy URL) and run commands with sub-second feedback:

# Verify connection
kagglex exec --url "https://kkb-production.jupyter-proxy.kaggle.net?token=..." --test

# Query remote GPU status
kagglex exec --url "https://kkb-production.jupyter-proxy.kaggle.net?token=..." --gpu-info

# Execute inline Python snippets
kagglex exec "import torch; print(torch.cuda.device_count())"

# Execute a local Python file remotely
kagglex exec --file evaluate.py

# List and transfer files
kagglex exec --list-files
kagglex exec --upload ./checkpoint.pt
kagglex exec --download run_results.json -o ./local_results.json

Alternatively, set the environment variable:

export KAGGLE_JUPYTER_URL="https://kkb-production.jupyter-proxy.kaggle.net?token=..."

Python SDK Usage

from kagglex import KaggleRunner, RunConfig

runner = KaggleRunner()
job = runner.run(
    command="python -m mypkg.train --batch-size 64",
    title="Fine Tuning Run",
    gpu="t4-2x",
    multi_gpu=True,
    wait=True,
)

print(f"Status: {job.status}")
job.pull_outputs(destination_dir="./results")