o hh@sddlZddlmZddlmZmZddlmZddlm Z m Z ddl m Z m Z mZmZgdZed d Gd d d ZGd ddejZGdddejZdS)N) dataclass)ListAny)Future)Metadata MetadataIndex)LoadPlanSavePlan SavePlanner LoadPlanner) WriteResult StorageWriter StorageReaderT)frozenc@s&eZdZUeed<eed<eed<dS)r index size_in_bytes storage_dataN)__name__ __module__ __qualname__r__annotations__intrrrZ/var/www/html/ai/venv/lib/python3.10/site-packages/torch/distributed/checkpoint/storage.pyr s  r c @seZdZdZejdeddfddZejdedefdd Z ejd e ede efd d Z ejded e de e efddZejdede e eddfddZdS)ra 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() is_coordinatorreturnNcCdS)z Initialize this instance. Args: is_coordinator (bool): Whether this instance is responsible for coordinating the checkpoint. Nr)selfrrrrset_up_storage_writer/ z#StorageWriter.set_up_storage_writerplancCr)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_plan:z StorageWriter.prepare_local_planplanscCr)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_planJz!StorageWriter.prepare_global_planplannercCr)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_data\szStorageWriter.write_datametadataresultscCr)a Writes 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,r-rrrfinishuszStorageWriter.finish)rrr__doc__abcabstractmethodboolrr r#rr'r rr r+rr.rrrrrs0    rc@seZdZdZejdefddZejdededdfdd Z ejd e de fd d Z ejd e e de e fddZ ejd e dededfddZdS)ra  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() rcCr)z Reads the checkpoint metadata. Returns: The metadata object associated with the checkpoint being loaded. Nr)rrrr read_metadatar zStorageReader.read_metadatar,rNcCr)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,rrrrset_up_storage_readers z#StorageReader.set_up_storage_readerr!cCr)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#r$z StorageReader.prepare_local_planr%cCr)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'r(z!StorageReader.prepare_global_planr)cCr)a 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. 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_data)rrrr/r0r1rr3r2r4r r#rr'r rr5rrrrrs   r)r0 dataclassesrtypingrr torch.futuresrr,rrr)r r r r __all__r ABCrrrrrrs  l