FSDPMixedPrecisionPlugin¶

class lightning.pytorch.plugins.precision.FSDPMixedPrecisionPlugin(precision, device, scaler=None)[source]¶

AMP for Fully Sharded Data Parallel (FSDP) Training.

Warning

This is an experimental feature.

clip_grad_by_norm(*_, **__)[source]¶

Clip gradients by norm.

forward_context()[source]¶

For FSDP, this context manager is a no-op since conversion is already handled internally.

See: https://pytorch.org/docs/stable/fsdp.html for more details on mixed precision.