espnet3.components.modeling.optimization_spec.OptimizerRuntimeState

source

class espnet3.components.modeling.optimization_spec.OptimizerRuntimeState(accum_counter: int = 0, update_step: int = 0)

Bases: object

Track custom runtime counters for one named optimizer.

Lightning already checkpoints optimizer and scheduler state_dict() values. This dataclass exists only for extra ESPnet3 runtime state that Lightning does not manage for named multi-optimizer training:

• accum_counter: how many backward passes have been accumulated since the

last optimizer step,

• update_step: how many times this optimizer has actually been updated.

These counters are saved and restored through on_save_checkpoint() and on_load_checkpoint() in the Lightning module.

accum_counter : int = 0

update_step : int = 0