Skip to content

Callbacks

Lightning owns the training loop. graphids only ships callbacks that encode policy Lightning's stock callbacks don't:

  • Sha256ModelCheckpoint — lightning.pytorch.callbacks.ModelCheckpoint
  • a <ckpt>.sha256 sidecar so graphids._fs.atomic_load can verify bytes on read (GPFS truncates surprise us; sidecar is the load-time integrity check).

graphids._mlflow.MLflowSystemMetricsCallback records phase throughput and native MLflow system metrics. graphids._mlflow.MLflowSystemMetricsSummaryCallback summarizes native metric histories for offline diagnostic runs.

Lightning owns the training loop and callback lifecycle.

graphids.callbacks

callbacks

graphids-specific Lightning callbacks.

Sha256ModelCheckpoint

Bases: ModelCheckpoint

ModelCheckpoint + <ckpt>.sha256 sidecar after every save.

GPFS truncation surprises happen on OSC; the sidecar is the load-time integrity check used by _fs.atomic_load.