Source code for lightning.pytorch.strategies.fsdp

# Copyright The Lightning AI team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import logging
import shutil
from contextlib import contextmanager, nullcontext
from datetime import timedelta
from pathlib import Path
from typing import (
    TYPE_CHECKING,
    Any,
    Callable,
    Dict,
    Generator,
    List,
    Literal,
    Mapping,
    Optional,
    Set,
    Tuple,
    Type,
    Union,
)

import torch
from lightning_utilities.core.rank_zero import rank_zero_only as utils_rank_zero_only
from torch import Tensor
from torch.nn import Module
from torch.optim import Optimizer
from typing_extensions import override

import lightning.pytorch as pl
from lightning.fabric.plugins import CheckpointIO, ClusterEnvironment
from lightning.fabric.plugins.collectives.torch_collective import default_pg_timeout
from lightning.fabric.strategies import _StrategyRegistry
from lightning.fabric.strategies.fsdp import (
    _METADATA_FILENAME,
    _activation_checkpointing_kwargs,
    _auto_wrap_policy_kwargs,
    _distributed_checkpoint_load,
    _distributed_checkpoint_save,
    _get_full_state_dict_context,
    _get_sharded_state_dict_context,
    _init_cpu_offload,
    _init_sharding_strategy,
    _is_full_checkpoint,
    _is_sharded_checkpoint,
    _move_torchmetrics_to_device,
    _optimizer_has_flat_params,
    _setup_activation_checkpointing,
)
from lightning.fabric.strategies.model_parallel import _load_raw_module_state
from lightning.fabric.utilities.distributed import (
    _distributed_is_initialized,
    _get_default_process_group_backend_for_device,
    _init_dist_connection,
    _sync_ddp_if_available,
)
from lightning.fabric.utilities.distributed import group as _group
from lightning.fabric.utilities.imports import _TORCH_GREATER_EQUAL_2_2
from lightning.fabric.utilities.init import _has_meta_device_parameters_or_buffers
from lightning.fabric.utilities.load import _lazy_load, _materialize_tensors
from lightning.fabric.utilities.optimizer import _optimizers_to_device
from lightning.fabric.utilities.seed import reset_seed
from lightning.fabric.utilities.types import _PATH, ReduceOp
from lightning.pytorch.core.optimizer import LightningOptimizer
from lightning.pytorch.plugins.precision import Precision
from lightning.pytorch.plugins.precision.fsdp import FSDPPrecision
from lightning.pytorch.strategies.launchers.subprocess_script import _SubprocessScriptLauncher
from lightning.pytorch.strategies.parallel import ParallelStrategy
from lightning.pytorch.strategies.strategy import TBroadcast
from lightning.pytorch.trainer.states import TrainerFn
from lightning.pytorch.utilities.model_helpers import is_overridden
from lightning.pytorch.utilities.rank_zero import rank_zero_info, rank_zero_only, rank_zero_warn

if TYPE_CHECKING:
    from torch.distributed.device_mesh import DeviceMesh
    from torch.distributed.fsdp.fully_sharded_data_parallel import CPUOffload, MixedPrecision, ShardingStrategy
    from torch.distributed.fsdp.wrap import ModuleWrapPolicy

    _POLICY = Union[Set[Type[Module]], Callable[[Module, bool, int], bool], ModuleWrapPolicy]
    _SHARDING_STRATEGY = Union[ShardingStrategy, Literal["FULL_SHARD", "SHARD_GRAD_OP", "NO_SHARD", "HYBRID_SHARD"]]


log = logging.getLogger(__name__)


[docs]class FSDPStrategy(ParallelStrategy): r"""Strategy for Fully Sharded Data Parallel provided by torch.distributed. Fully Sharded Training shards the entire model across all available GPUs, allowing you to scale model size, whilst using efficient communication to reduce overhead. In practice, this means we can remain at parity with PyTorch DDP, whilst scaling our model sizes dramatically. The technique is similar to ZeRO-Stage 3. For more information check out `this blogpost <https://pytorch.org/blog/introducing-pytorch-fully-sharded-data-parallel-api>`__. Defaults have been set and options have been exposed, but may require configuration based on your level of memory/speed efficiency. We suggest having a look at `this tutorial <https://pytorch.org/tutorials/intermediate/FSDP_tutorial.html>`__ for more information. Arguments: cpu_offload: See ``cpu_offload`` parameter in :class:`torch.distributed.fsdp.FullyShardedDataParallel`. mixed_precision: See ``mixed_precision`` parameter in :class:`torch.distributed.fsdp.FullyShardedDataParallel`. auto_wrap_policy: Same as ``auto_wrap_policy`` parameter in :class:`torch.distributed.fsdp.FullyShardedDataParallel`. For convenience, this also accepts a set of the layer classes to wrap. activation_checkpointing: Deprecated. Use ``activation_checkpointing_policy``. activation_checkpointing_policy: Same as ``auto_wrap_policy`` parameter in :class:`torch.distributed.fsdp.FullyShardedDataParallel` but used when selecting the modules for which you want to enable activation checkpointing. Enabling this can free up a significant amount of memory at the cost of speed since activations in these layers need to be recomputed during backpropagation. For convenience, this also accepts a set of the layer classes to wrap. sharding_strategy: Select whether to shard model parameters, gradients, optimizer states, or a combination of them. Available values are: - ``"FULL_SHARD"``: Shards model parameters, gradients, and optimizer states (default). - ``"SHARD_GRAD_OP"``: Shards gradients and optimizer states only. Model parameters get replicated. - ``"NO_SHARD"``: No sharding (identical to regular DDP). - ``"HYBRID_SHARD"``: Shards model parameters, gradients, and optimizer states within a single machine, but replicates across machines. See also the `device_mesh` parameter below. Also accepts a :class:`torch.distributed.fsdp.ShardingStrategy` enum value. device_mesh: A tuple `(replication size, sharding size)` that defines over how many devices to shard and replicate the model. The product of the two numbers must equal the world size. Only valid in combination with the `HYBRID_SHARD` sharding strategy. state_dict_type: The format in which the state of the model and optimizers gets saved into the checkpoint. - ``"full"``: The full weights and optimizer states get assembled on rank 0 and saved to a single file. - ``"sharded"``: Each rank saves its shard of weights and optimizer states to a file. The checkpoint is a folder with as many files as the world size. \**kwargs: See available parameters in :class:`torch.distributed.fsdp.FullyShardedDataParallel`. """ strategy_name = "fsdp" _registered_strategies: List[str] = [] def __init__( self, accelerator: Optional["pl.accelerators.Accelerator"] = None, parallel_devices: Optional[List[torch.device]] = None, cluster_environment: Optional[ClusterEnvironment] = None, checkpoint_io: Optional[CheckpointIO] = None, precision_plugin: Optional[Precision] = None, process_group_backend: Optional[str] = None, timeout: Optional[timedelta] = default_pg_timeout, cpu_offload: Union[bool, "CPUOffload", None] = None, mixed_precision: Optional["MixedPrecision"] = None, auto_wrap_policy: Optional["_POLICY"] = None, activation_checkpointing: Optional[Union[Type[Module], List[Type[Module]]]] = None, activation_checkpointing_policy: Optional["_POLICY"] = None, sharding_strategy: "_SHARDING_STRATEGY" = "FULL_SHARD", state_dict_type: Literal["full", "sharded"] = "full", device_mesh: Optional[Union[Tuple[int], "DeviceMesh"]] = None, **kwargs: Any, ) -> None: super().__init__( accelerator=accelerator, parallel_devices=parallel_devices, cluster_environment=cluster_environment, checkpoint_io=checkpoint_io, precision_plugin=precision_plugin, ) self.num_nodes = 1 self._process_group_backend = process_group_backend self._timeout: Optional[timedelta] = timeout self.cpu_offload = _init_cpu_offload(cpu_offload) self.mixed_precision = mixed_precision self.kwargs = _auto_wrap_policy_kwargs(auto_wrap_policy, kwargs) if device_mesh is not None: if not _TORCH_GREATER_EQUAL_2_2: raise ValueError("The `device_mesh` argument is only supported in torch >= 2.2.") self.kwargs["device_mesh"] = device_mesh self.sharding_strategy = _init_sharding_strategy(sharding_strategy, self.kwargs) # Avoids the need for user to reference params in `configure_optimizers` via # `self.trainer.model.parameters()` and enables support for multiple parameter groups. self.kwargs.setdefault("use_orig_params", True) self._activation_checkpointing_kwargs = _activation_checkpointing_kwargs( activation_checkpointing, activation_checkpointing_policy ) self._state_dict_type = state_dict_type @property @override def root_device(self) -> torch.device: assert self.parallel_devices is not None return self.parallel_devices[self.local_rank] @property def num_processes(self) -> int: return len(self.parallel_devices) if self.parallel_devices is not None else 0 @property def process_group_backend(self) -> Optional[str]: return self._process_group_backend @property def mixed_precision_config(self) -> Optional["MixedPrecision"]: if self.mixed_precision: return self.mixed_precision plugin = self.precision_plugin if isinstance(plugin, FSDPPrecision): return plugin.mixed_precision_config return None @property @override def precision_plugin(self) -> FSDPPrecision: plugin = self._precision_plugin if plugin is not None: assert isinstance(plugin, FSDPPrecision) return plugin return FSDPPrecision("32-true") @precision_plugin.setter @override def precision_plugin(self, precision_plugin: Optional[FSDPPrecision]) -> None: if precision_plugin is not None and not isinstance(precision_plugin, FSDPPrecision): raise TypeError( f"The FSDP strategy can only work with the `FSDPPrecision` plugin, found {precision_plugin}" ) self._precision_plugin = precision_plugin @property @override def distributed_sampler_kwargs(self) -> Dict: return {"num_replicas": (self.num_nodes * self.num_processes), "rank": self.global_rank} @property @override def restore_checkpoint_after_setup(self) -> bool: return True @property @override def lightning_restore_optimizer(self) -> bool: return False
[docs] @override def setup_environment(self) -> None: super().setup_environment() log.debug(f"{self.__class__.__name__}: setting up distributed...") reset_seed() # determine which process we are and world size self.set_world_ranks() self._process_group_backend = self._get_process_group_backend() assert self.cluster_environment is not None _init_dist_connection(self.cluster_environment, self._process_group_backend, timeout=self._timeout) # if 'device_mesh' in the `kwargs` is provided as a tuple, update it into the `DeviceMesh` object here if isinstance(self.kwargs.get("device_mesh"), tuple): from torch.distributed.device_mesh import init_device_mesh self.kwargs["device_mesh"] = init_device_mesh("cuda", self.kwargs["device_mesh"])
def _get_process_group_backend(self) -> str: return self._process_group_backend or _get_default_process_group_backend_for_device(self.root_device) def set_world_ranks(self) -> None: if self.cluster_environment is not None: self.cluster_environment.set_global_rank(self.node_rank * self.num_processes + self.local_rank) self.cluster_environment.set_world_size(self.num_nodes * self.num_processes) # `LightningEnvironment.set_global_rank` will do this too, but we cannot rely on that implementation detail # additionally, for some implementations, the setter is a no-op, so it's safer to access the getter rank_zero_only.rank = utils_rank_zero_only.rank = self.global_rank @override def _configure_launcher(self) -> None: assert self.cluster_environment is not None if not self.cluster_environment.creates_processes_externally: self._launcher = _SubprocessScriptLauncher(self.cluster_environment, self.num_processes, self.num_nodes) @override def _setup_model(self, model: Module) -> Module: """Wraps the model into a :class:`~torch.distributed.fsdp.fully_sharded_data_parallel.FullyShardedDataParallel` module.""" from torch.distributed.fsdp import FullyShardedDataParallel if any(isinstance(mod, FullyShardedDataParallel) for mod in model.modules()): if _has_meta_device_parameters_or_buffers(model): rank_zero_warn( "The model is already wrapped in `FSDP` but there are still parameters on the meta device." ) if "auto_wrap_policy" in self.kwargs: # The user has wrapped their submodules manually, don't apply the auto wrap policy. rank_zero_warn( "A FSDP `auto_wrap_policy` is set, but the model is already wrapped. The policy will be ignored." ) del self.kwargs["auto_wrap_policy"] else: log.debug(f"setting up FSDP model with device id: {self.root_device.index}, kwargs: {self.kwargs}") model = FullyShardedDataParallel( module=model, cpu_offload=self.cpu_offload, mixed_precision=self.mixed_precision_config, sharding_strategy=self.sharding_strategy, device_id=self.root_device.index, **self.kwargs, ) _move_torchmetrics_to_device(model, self.root_device) # activation checkpointing needs to be set up after wrapping the model _setup_activation_checkpointing(model, self._activation_checkpointing_kwargs) return model
[docs] @override def setup(self, trainer: "pl.Trainer") -> None: assert self.accelerator is not None self.accelerator.setup(trainer) assert self.model is not None if trainer.state.fn == TrainerFn.FITTING and self._layer_sync: self.model = self._layer_sync.apply(self.model) self.model = self.precision_plugin.convert_module(self.model) if is_overridden("configure_sharded_model", self.lightning_module): # legacy: we don't skip setup with the `configure_model` alternative rank_zero_info( "You have overridden `LightningModule.configure_sharded_model` hook. It will assume that all the layers" " are already wrapped for sharding and won't wrap the entire model using `FullyShardedDataParallel`." ) else: self.model = self._setup_model(self.model) self.barrier() if trainer.state.fn == TrainerFn.FITTING: self.setup_optimizers(trainer) self.setup_precision_plugin() if trainer.state.fn == TrainerFn.FITTING: _optimizers_to_device(self.optimizers, self.root_device)
[docs] @override def setup_optimizers(self, trainer: "pl.Trainer") -> None: # If we're setting up for evaluation after fitting, we need to discard the optimizers # since we're rewrapping the model, otherwise optimizer param references are no longer valid # and subsequent checkpoint saving can fail self._reset_optimizers_and_schedulers() if self.kwargs.get("use_orig_params"): return super().setup_optimizers(trainer) invalid_params_error = False try: # If `use_orig_params=False` the user needs to do access `self.trainer.model.parameters()` in # `configure_optimizers()` super().setup_optimizers(trainer) except ValueError as ex: if "optimizer got an empty parameter list" not in str(ex): raise invalid_params_error = True if invalid_params_error or any(not _optimizer_has_flat_params(optimizer) for optimizer in self.optimizers): # We avoid this limitation by setting `use_orig_params=True` raise ValueError( "The optimizer does not seem to reference any FSDP parameters. HINT: Make sure to create the" " optimizer after setting up the model by referencing `self.trainer.model.parameters()` in the" " `configure_optimizers()` hook." ) return None
[docs] @override def model_to_device(self) -> None: # FSDP takes care of moving the model to device pass
[docs] @contextmanager @override def tensor_init_context(self, empty_init: Optional[bool] = None) -> Generator[None, None, None]: # Materialization happens in `setup`. When modules get wrapped by FSDP, the sequence of operations is: # 1) materialize module 2) call `reset_parameters()` 3) shard the module. # These operations are applied to each submodule 'bottom up' in the module hierarchy. empty_init_context = torch.device("meta") if empty_init else nullcontext() with empty_init_context, self.precision_plugin.tensor_init_context(): yield
[docs] @contextmanager @override def model_sharded_context(self) -> Generator[None, None, None]: log.debug(f"{self.__class__.__name__}: entered model_sharded_context.") from torch.distributed.fsdp.fully_sharded_data_parallel import FullyShardedDataParallel from torch.distributed.fsdp.wrap import enable_wrap with enable_wrap( wrapper_cls=FullyShardedDataParallel, cpu_offload=self.cpu_offload, mixed_precision=self.mixed_precision_config, sharding_strategy=self.sharding_strategy, device_id=self.root_device.index, **self.kwargs, ): yield
[docs] @override def barrier(self, name: Optional[str] = None) -> None: if not _distributed_is_initialized(): return if torch.distributed.get_backend() == "nccl": torch.distributed.barrier(device_ids=self._determine_device_ids()) else: torch.distributed.barrier()
[docs] @override def broadcast(self, obj: TBroadcast, src: int = 0) -> TBroadcast: if not _distributed_is_initialized(): return obj obj = [obj] torch.distributed.broadcast_object_list(obj, src, group=_group.WORLD) return obj[0]
[docs] @override def reduce( self, tensor: Union[Tensor, Any], group: Optional[Any] = None, reduce_op: Optional[Union[ReduceOp, str]] = "mean", ) -> Tensor: """Reduces a tensor from several distributed processes to one aggregated tensor. Args: tensor: the tensor to sync and reduce group: the process group to gather results from. Defaults to all processes (world) reduce_op: the reduction operation. Defaults to 'mean'/'avg'. Can also be a string 'sum' to calculate the sum during reduction. Return: reduced value, except when the input was not a tensor the output remains is unchanged """ if isinstance(tensor, Tensor): return _sync_ddp_if_available(tensor, group, reduce_op=reduce_op) return tensor
def _determine_device_ids(self) -> List[int]: return [self.root_device.index]
[docs] @override def teardown(self) -> None: log.debug(f"{self.__class__.__name__}: tearing down strategy...") pl_module = self.lightning_module if ( pl_module is not None # `self.lightning_module._trainer` can be None if teardown gets called on an exception before # the trainer gets set on the LightningModule and pl_module._trainer is not None and pl_module._trainer.state.fn == TrainerFn.FITTING and self._layer_sync ): assert self.model is not None self.model = self._layer_sync.revert(self.model) assert self.cluster_environment is not None assert self.accelerator is not None self.cluster_environment.teardown() self.precision_plugin.teardown() self.accelerator.teardown()
@classmethod def get_registered_strategies(cls) -> List[str]: return cls._registered_strategies @classmethod @override def register_strategies(cls, strategy_registry: _StrategyRegistry) -> None: if not torch.distributed.is_available(): return strategy_registry.register( "fsdp", cls, description="Fully Sharded Data Parallel (FSDP) training", ) cls._registered_strategies.append("fsdp") strategy_registry.register( "fsdp_cpu_offload", cls, description="Fully Sharded Data Parallel (FSDP) training with Full Sharding and CPU Offloading", cpu_offload=True, ) cls._registered_strategies.append("fsdp_cpu_offload")
[docs] @override def lightning_module_state_dict(self) -> Dict[str, Any]: assert self.model is not None if self._state_dict_type == "sharded": state_dict_ctx = _get_sharded_state_dict_context(self.model) elif self._state_dict_type == "full": state_dict_ctx = _get_full_state_dict_context(self.model, world_size=self.world_size) else: raise ValueError(f"Unknown state_dict_type: {self._state_dict_type}") with state_dict_ctx: return self.model.state_dict()
@override def load_model_state_dict(self, checkpoint: Mapping[str, Any], strict: bool = True) -> None: # Override to do nothing, FSDP already loaded the states in `load_checkpoint()` pass
[docs] @override def optimizer_state(self, optimizer: Optimizer) -> Dict[str, Tensor]: from torch.distributed.fsdp import FullyShardedDataParallel as FSDP from torch.distributed.fsdp import OptimStateKeyType if isinstance(optimizer, LightningOptimizer): optimizer = optimizer._optimizer assert self.model is not None if self._state_dict_type == "sharded": with _get_sharded_state_dict_context(self.model): return FSDP.optim_state_dict(self.model, optimizer) elif self._state_dict_type == "full": with _get_full_state_dict_context(self.model, world_size=self.world_size): state_dict = FSDP.optim_state_dict(self.model, optimizer) if self.global_rank == 0: # Store the optimizer state dict in standard format state_dict = FSDP.rekey_optim_state_dict(state_dict, OptimStateKeyType.PARAM_ID, self.model) return state_dict raise ValueError(f"Unknown state_dict_type: {self._state_dict_type}")
@override def load_optimizer_state_dict(self, checkpoint: Mapping[str, Any]) -> None: # Override to do nothing, the FSDP already loaded the states in `load_checkpoint()` pass
[docs] @override def save_checkpoint( self, checkpoint: Dict[str, Any], filepath: _PATH, storage_options: Optional[Any] = None ) -> None: if storage_options is not None: raise TypeError( "`FSDPStrategy.save_checkpoint(..., storage_options=...)` is not supported because" " `FSDPStrategy` does not use the `CheckpointIO`." ) path = Path(self.broadcast(filepath)) if path.is_dir() and self._state_dict_type == "full" and not _is_sharded_checkpoint(path): raise IsADirectoryError(f"The checkpoint path exists and is a directory: {path}") if self._state_dict_type == "sharded": if path.is_file(): path.unlink() path.mkdir(parents=True, exist_ok=True) converted_state = {"model": checkpoint.pop("state_dict")} converted_state.update({ f"optimizer_{idx}": optim_state for idx, optim_state in enumerate(checkpoint.pop("optimizer_states", [])) }) _distributed_checkpoint_save(converted_state, path) if self.global_rank == 0: torch.save(checkpoint, path / _METADATA_FILENAME) elif self._state_dict_type == "full": if _is_sharded_checkpoint(path): shutil.rmtree(path) return super().save_checkpoint(checkpoint=checkpoint, filepath=path) else: raise ValueError(f"Unknown state_dict_type: {self._state_dict_type}")
@override def load_checkpoint(self, checkpoint_path: _PATH) -> Dict[str, Any]: # broadcast the path from rank 0 to ensure all the states are loaded from a common path path = Path(self.broadcast(checkpoint_path)) from torch.distributed.fsdp import FullyShardedDataParallel as FSDP assert self.model is not None assert self.lightning_module is not None if _is_sharded_checkpoint(path): from torch.distributed.checkpoint.optimizer import load_sharded_optimizer_state_dict state_dict_ctx = _get_sharded_state_dict_context(self.model) with state_dict_ctx: module_state = {"model": self.model.state_dict()} _distributed_checkpoint_load(module_state, path) self.model.load_state_dict(module_state["model"], strict=self.lightning_module.strict_loading) if self.lightning_module.trainer.state.fn == TrainerFn.FITTING and self.optimizers: from torch.distributed.checkpoint import FileSystemReader # TODO: replace with newer APIs # https://github.com/pytorch/pytorch/issues/119800#issuecomment-1942156271 reader = FileSystemReader(path=path) # the optimizer states must be loaded separately for idx, optim in enumerate(self.optimizers): optim_key = f"optimizer_{idx}" optim_state = load_sharded_optimizer_state_dict( model_state_dict=module_state["model"], optimizer_key=optim_key, storage_reader=reader, ) flattened_osd = FSDP.optim_state_dict_to_load( optim_state_dict=optim_state[optim_key], model=self.model, optim=optim, ) optim.load_state_dict(flattened_osd) # Load metadata (anything not a module or optimizer) metadata = torch.load(path / _METADATA_FILENAME) return metadata if _is_full_checkpoint(path): checkpoint = _lazy_load(path) _load_raw_module_state( checkpoint.pop("state_dict"), module=self.model, world_size=self.world_size, strict=self.lightning_module.strict_loading, ) # Materialize lazy tensors if there are any left in the checkpoint # The `torch.Optimizer.load_state_dict` method can't load lazy tensors because of deepcopy pickle issues checkpoint = _materialize_tensors(checkpoint) from torch.distributed.fsdp import FullyShardedDataParallel as FSDP from torch.distributed.fsdp import OptimStateKeyType optimizer_states = checkpoint.get("optimizer_states") if optimizer_states is None or self.lightning_module.trainer.state.fn != TrainerFn.FITTING: # If the optimizer states are not present, we don't need to do anything (backward compatibility) return checkpoint if len(self.optimizers) != len(optimizer_states): raise RuntimeError( f"You have configured {len(self.optimizers)} optimizers but the checkpoint contains" f" {len(optimizer_states)} optimizers to load. Please resume training with the same number" " of optimizers or edit the checkpoint manually to remove states." ) # rank0_only should be false because we need to load the optimizer state on all ranks with _get_full_state_dict_context(self.model, world_size=self.world_size, rank0_only=False): for optimizer, opt_state in zip(self.optimizers, optimizer_states): if isinstance(list(opt_state["state"].keys())[0], int): # Handling the case where the optimizer state is saved from a normal optimizer opt_state = FSDP.rekey_optim_state_dict(opt_state, OptimStateKeyType.PARAM_NAME, self.model) opt_state = FSDP.optim_state_dict_to_load( optim_state_dict=opt_state, model=self.model, optim=optimizer, ) optimizer.load_state_dict(opt_state) return checkpoint raise ValueError( f"The path {str(path)!r} does not point to a valid checkpoint. Make sure the path points to either a" " directory with FSDP checkpoint shards, or a single file with a full checkpoint." )