Brioche · In service
The same recipe gives the same brioche.
Brioche trains deep learning models with PyTorch. A run restarted with the same settings gives the same result, and an interruption loses nothing.
What it does
Brioche adapts to your model.
It depends on no particular model: a task is declared once, then Brioche runs the training.
01
Resume at the exact step
After an interruption, training restarts at the step where it stopped, and Brioche checks it is the same run.
02
Checked settings
Settings are written in TOML, inherited from run to run and overridden on the fly. A typo is refused before the start.
03
Driven remotely
From a laptop you send a run to the compute machine, follow it live and queue the next ones.
04
Runs compared
Statistics at every step, best state kept, reports and comparisons in Markdown, JSON and SVG.
The principle
Every result can be reproduced.
In research, a figure that cannot be reproduced is worthless. Brioche pins down everything that affects a run and keeps track of each one.
- A registry of every run
- Setting sweeps in a queue
- Background checkpoints
Questions
What we are asked most.
Does Brioche work with my model?
Yes, if your model is written with PyTorch. Brioche is not tied to any model: you declare the task, it runs the training.
What if the machine stops?
Training restarts at the exact step where it stopped, with the same data in the same order.
What hardware does it run on?
A machine with a CUDA-capable graphics card, driven from another computer.
Brioche serves our Labs every day.
It trains our world model and will serve our language models. If you want it for your own training runs, write to us.
Write to us