experiment tracking, checkpoints and GPUs

Train on your cloud. Lose a machine, not the run.

Resume from the last checkpoint, fork from any other: weights, optimizer and data come back. GPUs run in your cloud only while there is work. Free: you pay your cloud, not us.

train.py

run resumes from step 48,000

A cluster of GPUs at work
pip install "ravex[moonclip]"
import ravex

@ravex.train_loop()
def train():
    ...  # unchanged
python train.py

the console

Every run, in one place.

The GPU Zero console: two runs compared, their training and eval loss on the same charts
Two runs of the quickstart, compared on the console's charts.

resume and fork

Any checkpoint is a new beginning.

Checkpoints as often as every step, only what changed. Run the script again and it carries on from the newest; fork any of them with new hyperparameters.

A GPU at work

your cloud

Your GPUs. Your bill. Your bucket.

Your providers, on your own accounts, the cheapest offer first; a machine is given back, by default, about 5 minutes after its last job. Checkpoints in your own bucket, or in one we keep for you until you connect yours.

Checkpoints rising into a cloud

open source

The parts that touch your training are open

Runtime, checkpoint engine, node agent: Apache-2.0, at github.com/gpuzero-dev. Read them, run them without us, see exactly what runs beside your model.

Measured on a 50M-parameter model on a 4-core CPU machine, a Moonclip save holds the training loop for 226 ms where torch.save holds it for 2.7 s (its README, with the script to measure your own).

Your next run can survive its machine.