U
    yh                  	   @   s   d dl Z d dlZd dlZd dlmZ d dlZd dlmZ d dlm	Z	 d dl
mZmZmZmZ d dlmZ d dlmZmZmZmZ dd Zdejeeeejeej ed	d
dZejeeedddZeee ejdddZdS )    N)Optional)distributed_c10d)ShardShardedTensorShardedTensorMetadataTensorPropertiesShardMetadata)
DeviceMeshDTensor	Replicater   c                 C   s:   |  dkrd|  d| S d|  d| d| |  S d S )Ncpuzrank:/:)lower)rankdevice_typenum_devices_per_node r   U/var/www/html/venv/lib/python3.8/site-packages/torch/distributed/fsdp/_shard_utils.py_get_remote_device_str   s    r   )tensorr   
world_sizer   pgdevicereturnc              
      sz  | j |dd}t||krh||  }dd |  D t|  d | | d< t||g}ng }dd |D }	dgtt	
dd |	D dd  }
dgt|	d d	  fd
d|
D }|dkrtjn|j  fddtt|	D }t|	t|  kr$t|ks*n tdd t||	|D }t||  t| j| jdtj|  dd}tj||dS )z
    Shard a tensor to chunks along the first dimension. The local rank will gets its
    corresponding chunk as the local shard to create a ShardedTensor.
    r   )dimc                 S   s   g | ]}d qS r   r   .0_r   r   r   
<listcomp>*   s     z0_create_chunk_sharded_tensor.<locals>.<listcomp>c                 S   s   g | ]}t | qS r   )listsize)r   chunkr   r   r   r!   1   s     c                 S   s   g | ]}|d  qS r   r   )r   
chunk_sizer   r   r   r!   3   s     N   c                    s   g | ]}|g  qS r   r   )r   Zd0)offsetsr   r   r!   6   s     c                    s    g | ]}t t| qS r   )r   distZget_global_rank)r   r)r   r   r   r   r   r!   <   s   
c                 S   s   g | ]\}}}t |||qS r   r   )r   offsetr#   Z	placementr   r   r   r!   E   s   F)dtypelayoutZrequires_gradZmemory_formatZ
pin_memory)Zshards_metadatar#   Ztensor_properties)sharded_tensor_metadataZprocess_group)r$   lencloner#   mathceilr   Zfrom_tensor_and_offsetsr"   	itertools
accumulater   Z_get_pg_default_devicetyperangeAssertionErrorzipr   r   r,   r-   torchZcontiguous_format	is_pinnedr   Z+_init_from_local_shards_and_global_metadata)r   r   r   r   r   r   chunksZlocal_shardZlocal_shardsZchunk_sizesZdim0_offsetsZchunk_offsets
placementsZshard_metadatar.   r   )r   r   r(   r   r   _create_chunk_sharded_tensor   sT    
*
  r=   )r   r   device_meshr   c                 C   sZ   |    } dd t|jD }dd t|jD }td|d< tj| ||ddj|dS )	z
    Shard a tensor to chunks along the first dimension. The local rank will gets its
    corresponding chunk as the local tensor to create a DTensor.
    c                 S   s   g | ]
}t  qS r   r   r   r   r   r   r!   g   s     z)_create_chunk_dtensor.<locals>.<listcomp>c                 S   s   g | ]
}t  qS r   r?   r   r   r   r   r!   h   s     r   r&   F)Z	run_check)r<   )r0   detachr6   ndimDShardr   Z
from_localredistribute)r   r   r>   Zreplicate_placementsZshard_placementsr   r   r   _create_chunk_dtensorY   s    
   rD   )r   parent_meshr   c                 C   s>   |dkst tt| j}t |d< | j| j|d} |  S )zT
    All gather a DTensor in its sharded dimension and return the local tensor.
    Nr&   )r>   r<   )	r7   r"   copydeepcopyr<   r   rC   r>   Zto_local)r   rE   r<   r   r   r   _all_gather_dtensorr   s    
rH   )N)rF   r3   r1   typingr   r9   Ztorch.distributeddistributedr)   r   Z'torch.distributed._shard.sharded_tensorr   r   r   r   Z&torch.distributed._shard.sharding_specr	   Ztorch.distributed._tensorr
   r   r   rB   r   ZTensorintZProcessGroupr   r=   rD   rH   r   r   r   r   <module>   s:    ?