WandbCacheCleanupCallback

class WandbCacheCleanupCallback(*args, **kwargs)

Bases: Callback

Periodically run wandb artifact cache cleanup to cap local cache size.

Parameters:

• max_cache_size (str) – Size cap passed to the W&B CLI, e.g., “10GB”, “5GB”.

• every_n_epochs (int) – Run cleanup when (current_epoch + 1) % N == 0.

• run_on_fit_start (bool) – If True, also run once at fit start.

• only_on_global_rank_zero (bool) – If True, only run on rank 0 in DDP.

• executable (str) – CLI executable name/path for wandb (default: “wandb”).

• extra_env (Mapping[str, str] | None) – Optional environment overrides for the subprocess.

• background (bool) – If True, run non-blocking via background thread + Popen.

• timeout (int | None) – Optional timeout (seconds) for blocking mode.

__init__(

max_cache_size='5GB',

every_n_epochs=1,

run_on_fit_start=False,

only_on_global_rank_zero=True,

executable='wandb',

extra_env=None,

background=True,

timeout=None,

)

Initializes the WandbCacheCleanupCallback.

Parameters:

• max_cache_size (str)

• every_n_epochs (int)

• run_on_fit_start (bool)

• only_on_global_rank_zero (bool)

• executable (str)

• extra_env (Mapping[str, str] | None)

• background (bool)

• timeout (int | None)

Return type:

None

on_fit_start(trainer, pl_module)

Runs cleanup at the start of fitting.

Parameters:

• trainer (pytorch_lightning.Trainer)

• pl_module (pytorch_lightning.LightningModule)

Return type:

None

on_train_epoch_end(trainer, pl_module)

Runs cleanup at the end of each training epoch.

Parameters:

• trainer (pytorch_lightning.Trainer)

• pl_module (pytorch_lightning.LightningModule)

Return type:

None