Distributed Checkpoint - torch.distributed.checkpoint

Distributed Checkpoint (DCP) support loading and saving models from multiple ranks in parallel. It handles load-time resharding which enables saving in one cluster topology and loading into another.

DCP is different than torch.save and torch.load in a few significant ways:

• It produces multiple files per checkpoint, with at least one per rank.

• It operates in place, meaning that the model should allocate its data first and DCP uses that storage instead.

The entrypoints to load and save a checkpoint are the following:

torch.distributed.checkpoint.load_state_dict(state_dict, storage_reader, process_group=None, coordinator_rank=0, no_dist=False, planner=None)

Loads a distributed state_dict in SPMD style.

Each rank will try to read the least amount of data necessary to fullfill the requested state_dict. When loading ShardedTensor instances, each rank only reads data for their local shards.

Warning

All tensors in state_dict must be allocated on their destination device prior to calling this function.

All non-tensor data is loaded using torch.load() and modified in place on state_dict.

Warning

Users must call load_state_dict on the root module to ensure load pos-processing and non-tensor data properly propagates.

Parameters:

• state_dict (Dict[str, Any]) – The state_dict to load. Note that this state dict will updated in place.

• storage_reader (StorageReader) – StorageReader used to load data from.

• process_group (ProcessGroup) – ProcessGroup to be used for cross-rank synchronization.

• coordinator_rank (int) – Rank to use to coordinate the checkpoint. rank0 is used by default.

• no_dist (bool) – If True, distributed checkpoint will not save in SPMD style. (Default: False)

Returns:

None.

Return type:

None

Examples

>>> my_model = MyModule()
>>> optimizer = Adagrad(my_model.parameters())
>>> model_state_dict = my_model.state_dict()
>>> fs_storage_loader = torch.distributed.checkpoint.FileSystemLoader("/checkpoint/1")

>>> torch.distributed.checkpoint.load_state_dict(
>>>     state_dict=model_state_dict,
>>>     storage_reader=fs_storage_loader,
>>> )

>>> # module.load_state_dict() function might have customized steps
>>> # to flush the state_dict, must call it to
>>> # ensure correct behavior.
>>> my_model.load_state_dict(model_state_dict)

Note

load_state_dict uses collectives to coordinate reads across ranks. For NCCL-based process groups, internal tensor representations of objects must be moved to the GPU device before communication takes place. In this case, the device used is given by torch.cuda.current_device() and it is the user’s responsibility to ensure that this is set so that each rank has an individual GPU, via torch.cuda.set_device().

torch.distributed.checkpoint.save_state_dict(state_dict, storage_writer, process_group=None, coordinator_rank=0, no_dist=False, planner=None)

Saves a distributed model in SPMD style.

This function is different from torch.save() as it handles ShardedTensor by having each rank only save their local shards.

Warning

There is no guarantees of Backwards Compatibility across PyTorch versions for saved state_dicts.

Warning

If using the process_group argument, make sure that only its ranks call save_state_dict and that all data in state_dict belong to it.

Note

This function can be used to save a state_dict with an intialized process group by passing no_dist=True. This can be used to produce a checkpoint that can consumed by load_state_dict is a SPMD fashion.

Parameters:

• state_dict (Dict[str, Any]) – A state_dict

• storage_writer (StorageWriter) – Instance of StorageWrite use to perform writes.

• process_group (ProcessGroup) – ProcessGroup to be used for cross-rank synchronization.

• coordinator_rank (int) – Rank to use to coordinate the checkpoint. rank0 is used by default.

• no_dist (bool) – If True, distributed checkpoint will not save in SPMD style. (Default: False)

Returns:

Metadata object for the saved checkpoint.

Return type:

Metadata

Example

>>> my_model = MyModule()

>>> model_state_dict = my_model.state_dict()

>>> fs_storage_writer = torch.distributed.checkpoint.FileSystemWriter("/checkpoint/1")
>>> torch.distributed.checkpoint.save_state_dict(
>>>     state_dict=model_state_dict,
>>>     storage_writer=fs_stroage_writer,
>>> )

Note

save_state_dict uses collectives to coordinate writes across ranks. For NCCL-based process groups, internal tensor representations of objects must be moved to the GPU device before communication takes place. In this case, the device used is given by torch.cuda.current_device() and it is the user’s responsibility to ensure that this is set so that each rank has an individual GPU, via torch.cuda.set_device().

The following types define the IO interface used during checkpoint:

class torch.distributed.checkpoint.StorageReader

Interface used by load_state_dict to read from storage.

One StorageReader instance acts as both the coordinator and the follower in a distributed checkpoint. As part of initialization, each instance is told its role.

A subclass should expected the following sequence of calls by load_state_dict:

1. (all ranks) read_metadata()

2. (all ranks) set_up_storage_reader()

3. (all ranks) prepare_local_plan()

4. (coordinator) prepare_global_plan()

5. (all ranks) read_data()

abstract prepare_global_plan(plans)

Perform centralized planning of storage loading.

This method is only called on the coordinator instance.

While this method can produce a completely different plan, the prefered way is to store storage specific data in LoadPlan::storage_data.

Parameters:

plans (List[LoadPlan]) – A list of LoadPlan instances, one for each rank.

Returns:

A list of transformed LoadPlan after storage global planning

Return type:

List[LoadPlan]

abstract prepare_local_plan(plan)

Perform storage-specific local planning.

While this method can produce a completely different plan, the recomended way is to store storage specific data in LoadPlan::storage_data.

Parameters:

plan (LoadPlan) – The local plan from the LoadPlan in use.

Returns:

A transformed LoadPlan after storage local planning

Return type:

LoadPlan

abstract read_data(plan, planner)

Reads all items from plan using planner to resolve the data.

A subclass should call LoadPlanner::load_bytes to deserialize a BytesIO object into the right place.

A subclass should call LoadPlanner::resolve_tensor to get access to the tensors that in should load data into.

It’s the StorageLayer responsibility to properly schedule any cross device copies required.

Parameters:

• plan (LoadPlan) – The local plan to execute on

• planner (LoadPlanner) – The planner object to use to resolve items.

Returns:

A future that completes once all reads are finished.

Return type:

Future[None]

abstract read_metadata()

Reads the checkpoint metadata.

Returns:

The metatada object associated with the checkpoint being loaded.

Return type:

Metadata

abstract set_up_storage_reader(metadata, is_coordinator)

Initialize this instance.

Parameters:

• metadata (Metadata) – The metadata schema to use.

• is_coordinator (bool) – Whether this instance is reponsible for coordinating the checkpoint.

class torch.distributed.checkpoint.StorageWriter

Interface used by save_state_dict to write to storage.

One StorageWriter instance acts as both the coordinator and the follower in a distributed checkpoint. As part of initialization, each instance is told its role.

A subclass should expect the following sequence of calls.

1. (all ranks) set_up_storage_writer()

2. (all ranks) prepare_local_plan()

3. (coordinator) prepare_global_plan()

4. (all ranks) write_data()

5. (coordinator) finish()

abstract finish(metadata, results)

Writes the metadata and marks the current checkpoint as sucessful.

The actual format/schema used for serializing metadata is an implemetation detail. The only requirement is that it’s recoverable in to the same object graph.

Parameters:

• metadata (Metadata) – metadata for the new checkpoint

• results (List[List[WriteResult]]) – A list of WriteResults from all ranks.

Returns:

None

Return type:

None

abstract prepare_global_plan(plans)

Perform centralized planning of storage.

This method is only called on the coordinator instance.

While this method can produce a completely different plan, the prefered way is to store storage specific data in SavePlan::storage_data.

Parameters:

plans (List[SavePlan]) – A list of SavePlan instances, one for each rank.

Returns:

A list of transformed SavePlan after storage global planning

Return type:

List[SavePlan]

abstract prepare_local_plan(plan)

Perform storage-specific local planning.

While this method can produce a completely different plan, the recomended way is to store storage specific data in SavePlan::storage_data.

Parameters:

plan (SavePlan) – The local plan from the SavePlanner in use.

Returns:

A transformed SavePlan after storage local planning

Return type:

SavePlan

abstract set_up_storage_writer(is_coordinator)

Initialize this instance.

Parameters:

is_coordinator (bool) – Whether this instance is reponsible for coordinating the checkpoint.

abstract write_data(plan, planner)

Write all items from plan using planner to resolve the data.

A subclass should call SavePlanner::resolve_data on each item from the plan to get access to the underlying object to write.

Subclasses should lazily call resolve_data as it can allocate memory. In case of tensors, make following assuptions:

• They might be on any device, including not matching the one on WriteItem::tensor_data

• They might be views or not contiguous. Only the projection needs to be saved.

Parameters:

• plan (SavePlan) – The save plan to execute.

• planner (SavePlanner) – Planner object to be used to resolve items to data.

Returns:

A future that completes to a list of WriteResult

Return type:

Future[List[WriteResult]]

The following types define the planner interface used during checkpoint:

class torch.distributed.checkpoint.LoadPlanner

Abstract class defining the protocol used by load_state_dict to plan the load process.

LoadPlanner are stateful objects that can be used to customize the whole load process.

LoadPlanner acts as an access proxy to the state_dict, so any transfomation done to it will be visible to the whole process.

A planner subclass can expect the following sequence of calls during load_state_dict:

1. set_up_planner - called on all ranks.

   Signals the start of loading a checkpoint.

2. create_local_plan - called on all ranks.

   Process the state_dict and produces a LoadPlan that will be sent for global planning.

3. create_global_plan - called on the coordinator rank only.

   Takes the LoadPlan from all ranks and make any global decision.

4. load_bytes - called multiple times on each rank

   This is called once per non-tensor value in state_dict.

5. resolve_tensor and commit_tensor - called multiple times on each rank

   They are called in pair for each Tensor value in state_dict.

Users are recomended to extend DefaultLoadPlanner instead of this interface directly as most changes can be expressed by changes in a single method.

There are two usual patterns of extension:

Rewriting state_dict. This is the simplest way to extend the load process as it doesn’t requite understanding the intrincacies of how LoadPlan works. We need to keep a reference to the original state_dict as load happens in place so we need to be able to perform it in place

>>> class RenamePlanner(DefaultLoadPlanner):
>>>     def set_up_planner(self, state_dict, metadata, is_coordinator):
>>>         self.original_state_dict = state_dict
>>>         super().set_up_planner(self, {"foo_" + k: v for k, v in state_dict.items()}, is_coordinator)
>>>
>>>     def load_bytes(self, read_item, value):
>>>         # Remove the "foo_" prefix
>>>         self.original_state_dict[read_item.dest_index.fqn[4:]] = torch.load(value)

Modifying resolve_tensor and commit_tensor to handle load time transformation.

>>> class MetaModelMaterialize(DefaultSavePlanner):
>>>     def resolve_tensor(self, read_item):
>>>         tensor = super().resolve_tensor(read_item)
>>>         return torch.empty_like(tensor, device="cpu")
>>>
>>>     def commit_tensor(self, read_item, tensor):
>>>         self.state_dict[read_item.dest_index.fqn] = tensor

abstract commit_tensor(read_item, tensor)

This method is called once the StorageReader finished loading data into tensor.

The provided tensor is the same one returned by the call to resolve_tensor. This method is only needed if this LoadPlanner needs to post process tensor prior to copying it back to the one in the state_dict.

The contents of tensor will follow its device synchronization model.

abstract create_global_plan(global_plan)

Compute the global load plan and return plans for each rank.

. N.B. This is called on the coordinator rank only

Return type:

List[LoadPlan]

abstract create_local_plan()

Create a LoadPlan based on state_dict and metadata provided by set_up_planner.

. N.B. This is called on every rank.

Return type:

LoadPlan

abstract finish_plan(central_plan)

Accept the plan from coordinator and return final LoadPlan.

Return type:

LoadPlan

abstract load_bytes(read_item, value)

Load the item described by read_item``and ``value.

This method is expected to modify in-place the underlying state_dict.

The contents of value are defined by the SavePlanner used to produce the checkpoint being loaded.

abstract resolve_tensor(read_item)

Return the tensor described by read_item to be used by the StorageReader to load read_item.

The tensor should alias with one on the underlying state_dict as StorageReader will replace its contents. If, for any reason, that’s not possible, the planner can use the commit_tensor method to copy the data back to the one in state_dict.

Return type:

Tensor

abstract set_up_planner(state_dict, metadata, is_coordinator)

Initialize this instance to load data into state_dict

. N.B. This is called on every rank.

class torch.distributed.checkpoint.LoadPlan(items: List[torch.distributed.checkpoint.planner.ReadItem], storage_data: Any = None, planner_data: Any = None)

class torch.distributed.checkpoint.ReadItem(type: torch.distributed.checkpoint.planner.LoadItemType, dest_index: torch.distributed.checkpoint.metadata.MetadataIndex, dest_offsets: torch.Size, storage_index: torch.distributed.checkpoint.metadata.MetadataIndex, storage_offsets: torch.Size, lengths: torch.Size)

class torch.distributed.checkpoint.SavePlanner

Abstract class defining the protocol used by save_state_dict to plan the save process.

SavePlanners are stateful objects that can be used to customize the whole save process.

SavePlanner acts as an access proxy to the state_dict, so any transfomation done to it will be visible to the whole process.

A planner subclass can expect the following sequence of calls during save_state_dict:

1. set_up_planner - called on all ranks.

   Signals the start of a checkpoint save.

2. create_local_plan - called on all ranks.

   Process the state_dict and produces a SavePlan that will be sent for global planning.

3. create_global_plan - called on the coordinator rank only.

   Takes the SavePlan from all ranks and make any global decision.

4. finish_plan - called on all ranks.

   This gives each rank a chance to adjust to global planning decisions.

5. resolve_data - called multiple times on each rank

   Lookups a value on the state_dict for the storage layer to write.

Users are recomended to extend DefaultSavePlanner instead of this interface directly as most changes can be expressed by changes in a single method.

There are 3 usual patterns of extension:

Rewriting state_dict. This is the simplest way to extend the save process as it doesn’t requite understanding the intrincacies of how SavePlan works:

>>> class RenamePlanner(DefaultSavePlanner):
>>>     def set_up_planner(self, state_dict, is_coordinator):
>>>         # prefix all keys with `foo_``
>>>         super().set_up_planner(self, {"foo_" + k: v for k, v in state_dict.items()}, is_coordinator)

Modifying local plan and lookup in tandem. This is useful when fine control of how data is persisted

>>> class FP16Planner(DefaultSavePlanner):
>>>     def create_local_plan(self):
>>>         plan = super().create_local_plan()
>>>         for p in plan:
>>>             if p.tensor_data is not None:
>>>                 p.tensor_data.properties.dtype = torch.float16
>>>
>>>     def resolve_data(self, write_item):
>>>         item = super().resolve_data(write_item)
>>>         return item if write_item.type == WriteItemType.BYTE_IO else item.to(torch.float16)

Using the global planning step to make central decisions that can’t be made individually by each rank

>>> from itertools import islice
>>> from dataclasses import replace
>>> class DDPLoadBalancingPlanner(DefaultSavePlanner):
>>>     # This uses the default local plan behavior of having all non-sharded writes in rank 0
>>>     # This sample doesn't handle ShardedTensors
>>>     def create_global_plan(self, all_plans):
>>>         def chunk(it, size):
>>>             it = iter(it)
>>>         return list(iter(lambda: tuple(islice(it, size)), ()))
>>>         all_plans = [
>>>             replace(plan, items=items) for plan, items in
>>>                 zip(all_plans, chunk(all_plans[0].items, len(all_plans)))
>>>         ]
>>>         return super().create_global_plan(all_plans)

Finally, some planners need to save additional metadata in the checkpoint, this is accomplished by having each rank contribute their data items in the local plan and the global planner aggregate them:

>>> class SaveExtraDataPlanner(DefaultSavePlanner):
>>>     def create_local_plan(self) -> SavePlan:
>>>         plan = super().create_local_plan()
>>>         return replace(plan, planner_data="per-rank-data")
>>>
>>>     def create_global_plan(self, all_plans: List[SavePlan]) -> Tuple[List[SavePlan], Metadata]:
>>>         global_plan, metadata = super().create_global_plan(all_plans)
>>>         merged_data = [p.planner_data for p in global_plan]
>>>         metadata = replace(metadata, planner_data=merged_data)
>>>         return global_plan, metadata

abstract create_global_plan(all_plans)

Compute the global checkpoint plan and return the local plan of each rank.

This is called on the coordinator rank only.

Return type:

Tuple[List[SavePlan], Metadata]

abstract create_local_plan()

Compute the save plan for the current rank. This will be aggregated and passed to create_global_plan. Planner specific data can be passed through SavePlan::planner_data.

This is called on all ranks.

Return type:

SavePlan

abstract finish_plan(new_plan)

Merge the plan created by create_local_plan and the result of create_global_plan.

This is called on all ranks.

Return type:

SavePlan

abstract resolve_data(write_item)

Lookup the object associated with write_item in state_dict and apply any transformation (such as serialization) prior to the storage layer consuming it.

Called on each rank multiple times, at least once per WriteItem in the final SavePlan.

This method should be idepotent and thread-save. StorageWriter implementations are free to call it as frequently as they need.

Any transformation that allocates memory should be lazily done when his method is called in order to reduce peak memory required by checkpointing.

When returning tensors, they can be on any device or format, they can be views too. It’s the storage layer responsibility to figure out how to save them.

Return type:

Union[Tensor, BytesIO]

abstract set_up_planner(state_dict, is_coordinator)

Intialize this planner to save state_dict.

Implementations should save those values as they won’t be provided lated in the save process.

This is called on all ranks.

class torch.distributed.checkpoint.SavePlan(items: List[torch.distributed.checkpoint.planner.WriteItem], storage_data: Any = None, planner_data: Any = None)

class torch.distributed.checkpoint.WriteItem(index: torch.distributed.checkpoint.metadata.MetadataIndex, type: torch.distributed.checkpoint.planner.WriteItemType, tensor_data: Union[torch.distributed.checkpoint.planner.TensorWriteData, NoneType] = None)

We provide a filesystem based storage layer:

class torch.distributed.checkpoint.FileSystemReader(path)

class torch.distributed.checkpoint.FileSystemWriter(path, single_file_per_rank=True, sync_files=True, thread_count=1, per_thread_copy_ahead=10000000)

Basic implementation of StorageWriter using file IO.

This implementation makes the following assumptions and simplifications:

• The checkpoint path is an empty or non-existing directory.

• File creation is atomic

The checkpoint consist of one file per write request plus a .metadata file with the serialized metadata.

We provide default implementations of LoadPlanner and SavePlanner that can handle all of torch.distributed constructs such as FSDP, DDP, ShardedTensor and DistributedTensor.

class torch.distributed.checkpoint.DefaultSavePlanner(flatten_state_dict=True, flatten_sharded_tensors=True, dedup_replicated_tensors=True)

lookup_object(index)

This is an extension from the planner interface to make it easy to extend the default planner

Return type:

Any

transform_object(write_item, object)

This is an extension from the planner interface to make it easy to extend the default planner

class torch.distributed.checkpoint.DefaultLoadPlanner(flatten_state_dict=True, flatten_sharded_tensors=True)

DefaultLoadPlanner that adds multiple features on top of LoadPlanner.

In particular it adds the following:

flatten_state_dict: Handle state_dict with nested dicts flatten_sharded_tensors: For FSDP in 2D parallel mode

lookup_tensor(index)

This is an extension from the planner interface to make it easy to extend the default planner

Return type:

Tensor

transform_tensor(read_item, tensor)

This is an extension from the planner interface to make it easy to extend the default planner