DDPShardedStrategy

class pytorch_lightning.strategies.DDPShardedStrategy(accelerator=None, parallel_devices=None, cluster_environment=None, checkpoint_io=None, precision_plugin=None, ddp_comm_state=None, ddp_comm_hook=None, ddp_comm_wrapper=None, model_averaging_period=None, process_group_backend=None, timeout=datetime.timedelta(seconds=1800), **kwargs)

Bases: pytorch_lightning.strategies.ddp.DDPStrategy

Optimizer and gradient sharded training provided by FairScale.

block_backward_sync()

Blocks syncing gradients behaviour on backwards pass.

This is useful for skipping sync when accumulating gradients, reducing communication overhead Returns: context manager with sync behaviour off

Return type:

Generator

optimizer_state(optimizer)

Returns state of an optimizer.

Allows for syncing/collating optimizer state from processes in custom plugins.

Return type:

Optional[dict]

pre_backward(closure_loss)

Run before precision plugin executes backward.

Return type:

None

setup(trainer)

Setup plugins for the trainer fit and creates optimizers.

Parameters:

trainer (Trainer) – the trainer instance

Return type:

None

property lightning_module: Optional[pytorch_lightning.core.module.LightningModule]

Returns the pure LightningModule without potential wrappers.

Return type:

Optional[LightningModule]