Skip to content
Nodus

nodus.integrations

View Markdown

Framework callbacks that report training progress to Nodus from inside a container.

nodus.integrations.transformers.NodusCallback plugs into a Hugging Face Trainer (and every TRL trainer); nodus.integrations.lightning.NodusCallback into a Lightning Trainer. Both send metrics to nodus.log.metrics, report progress, save when the node asks for a checkpoint, and do nothing outside Nodus. Only the global rank 0 process reports, so a multi-node gang reports once.