U d|@sLddlmZddlmZmZddlmZddlmZeGddde Z dS)) dataclass)ListUnion)reduce)_remote_devicec@sTeZdZUdZdddgZeeed<eeed<ee e fed<ddZ ddZ d S) ShardMetadataaH Represents a shard of the overall Tensor including its offsets, lengths and device placement. Args: shard_offsets(List[int]): Offsets in the original tensor indicating the start offsets for this shard. Should have the same rank as the original tensor. shard_sizes(List[int]): Integers indicating the size of each dimension for this shard. Should have the same rank as the original tensor. placement(:class:`torch.distributed._remote_device`): Specifies the placement of this shard. shard_offsets shard_sizes placementcCs4dd}t||jd}t||j|}|||j}|S)NcSs|d>t|S)N)hash)abrE/tmp/pip-unpacked-wheel-ua33x9lu/torch/distributed/_shard/metadata.py _hash_reducesz,ShardMetadata.__hash__.._hash_reduce%)rrr r )selfrresrrr__hash__s  zShardMetadata.__hash__cCst|jtrt|j|_t|jt|jkrJtdt|jd|jdtt|jD]0}|j|dkrrtd|j|dkrXtdqXdS)NzMshard_offsets and shard_sizes should have the same number of elements, found z and z respectivelyrzshard_offsets should be >=0zshard_sizes should be >= 0) isinstancer strrlenrr ValueErrorrange)rirrr __post_init__'s  zShardMetadata.__post_init__N) __name__ __module__ __qualname____doc__ __slots__rint__annotations__rrrrrrrrrrs     rN) Z dataclassesrtypingrr functoolsrZtorch.distributed.remote_devicerobjectrrrrrs