U yh&@sddlZddlZddlmZddlmZmZmZmZddl m Z m Z m Z ddl mZmZmZmZddlmZddd gZed d Gd ddZGd ddejZGdd d ejZdS)N) dataclass)AnyListOptionalUnion)Metadata MetadataIndex StorageMeta)LoadPlan LoadPlannerSavePlan SavePlanner)Future WriteResult StorageWriter StorageReaderT)frozenc@s&eZdZUeed<eed<eed<dS)rindexZ size_in_bytesZ storage_dataN)__name__ __module__ __qualname__r__annotations__intrrrV/var/www/html/venv/lib/python3.8/site-packages/torch/distributed/checkpoint/storage.pyrs c@seZdZdZejdeeej dfddddZ eje ddddZ eje e d d d Zejee ee d d dZeje eeeedddZejeeeeddddZeejeeej fe dddZeedddZdS)ra8 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. 0) (all ranks) set checkpoint_id if users pass a valid checkpoint_id. 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() N checkpoint_idreturncCsdS)a Calls to indicates a brand new checkpoint write is going to happen. A checkpoint_id may be present if users set the checkpoint_id for this checkpoint write. The meaning of the checkpiont_id is storage-dependent. It can be a path to a folder/file or a key for a key-value storage. Args: checkpoint_id (Union[str, os.PathLike, None]): The ID of this checkpoint instance. The meaning of the checkpoint_id depends on the storage. It can be a path to a folder or to a file. It can also be a key if the storage is a key-value store. (Default: ``None``) Nrselfrrrrreset-szStorageWriter.reset)is_coordinatorrcCsdS)z Initialize this instance. Args: is_coordinator (bool): Whether this instance is responsible for coordinating the checkpoint. Nr)rr!rrrset_up_storage_writer?s z#StorageWriter.set_up_storage_writerplanrcCsdS)a Perform storage-specific local planning. While this method can produce a completely different plan, the recommended way is to store storage specific data in SavePlan::storage_data. Args: plan (SavePlan): The local plan from the ``SavePlanner`` in use. Returns: A transformed ``SavePlan`` after storage local planning Nrrr$rrrprepare_local_planJsz StorageWriter.prepare_local_planplansrcCsdS)a Perform centralized planning of storage. This method is only called on the coordinator instance. While this method can produce a completely different plan, the preferred way is to store storage specific data in SavePlan::storage_data. Args: plans: A list of ``SavePlan`` instances, one for each rank. Returns: A list of transformed ``SavePlan`` after storage global planning Nrrr(rrrprepare_global_planZsz!StorageWriter.prepare_global_planr$plannerrcCsdS)a 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 assumptions: - 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. Args: 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 Nrrr$r,rrr write_datalszStorageWriter.write_data)metadataresultsrcCsdS)a Write the metadata and marks the current checkpoint as successful. The actual format/schema used for serializing `metadata` is an implementation detail. The only requirement is that it's recoverable in to the same object graph. Args: metadata (Metadata): metadata for the new checkpoint results: A list of WriteResults from all ranks. Returns: None Nr)rr/r0rrrfinishszStorageWriter.finishcCsdSz Check if the given checkpoint_id is supported by the stroage. This allow us to enable automatic storage selection. Nrclsrrrrvalidate_checkpoint_idsz$StorageWriter.validate_checkpoint_idrcCsdS)a= Return the storage-specific metadata. This is used to store additional information in a checkpoint that can be useful for providing request-level observability. StorageMeta is passed to the ``SavePlanner`` during save calls. Returns None by default. TODO: provide an example Nrrrrr storage_metaszStorageWriter.storage_meta)N)rrr__doc__abcabstractmethodrstrosPathLiker boolr"r r&rr*r rrr.rr1 classmethodr5rr r8rrrrrs(   c@seZdZdZejdeeej dfddddZ eje dddZ eje e dd d d Zejeed d dZejeeeedddZejeeeddddZeejeeej fe dddZdS)raV 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``: 0) (all ranks) set checkpoint_id if users pass a valid checkpoint_id. 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() NrcCsdS)a Calls to indicates a brand new checkpoint read is going to happen. A checkpoint_id may be present if users set the checkpoint_id for this checkpoint read. The meaning of the checkpiont_id is storage-dependent. It can be a path to a folder/file or a key for a key-value storage. Args: checkpoint_id (Union[str, os.PathLike, None]): The ID of this checkpoint instance. The meaning of the checkpoint_id depends on the storage. It can be a path to a folder or to a file. It can also be a key if the storage is more like a key-value store. (Default: ``None``) Nrrrrrr szStorageReader.resetr6cCsdS)z Read the checkpoint metadata. Returns: The metadata object associated with the checkpoint being loaded. Nrr7rrr read_metadatas zStorageReader.read_metadata)r/r!rcCsdS)z Initialize this instance. Args: metadata (Metadata): The metadata schema to use. is_coordinator (bool): Whether this instance is responsible for coordinating the checkpoint. Nr)rr/r!rrrset_up_storage_readers z#StorageReader.set_up_storage_readerr#cCsdS)a Perform storage-specific local planning. While this method can produce a completely different plan, the recommended way is to store storage specific data in LoadPlan::storage_data. Args: plan (LoadPlan): The local plan from the ``LoadPlan`` in use. Returns: A transformed ``LoadPlan`` after storage local planning Nrr%rrrr&sz StorageReader.prepare_local_planr'cCsdS)a 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 preferred way is to store storage specific data in LoadPlan::storage_data. Args: plans: A list of ``LoadPlan`` instances, one for each rank. Returns: A list of transformed ``LoadPlan`` after storage global planning Nrr)rrrr*sz!StorageReader.prepare_global_planr+cCsdS)a Read 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. Args: 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. Nrr-rrr read_dataszStorageReader.read_datacCsdSr2rr3rrrr5sz$StorageReader.validate_checkpoint_id)N)rrrr9r:r;rr<r=r>r rrAr?rBr r&rr*r rrCr@r5rrrrrs    )r:r= dataclassesrtypingrrrrZ%torch.distributed.checkpoint.metadatarrr Z$torch.distributed.checkpoint.plannerr r r r Z torch.futuresr__all__rABCrrrrrrs