o Zh0@s$ddlmZddlmZddlmZmZGdddZGdddeZGdd d eZ Gd d d eZ Gd d d eZ GdddeZ dede edefddZdedeee fde fddZdeedeee fdeee ffddZdededefdd Zdeed!eee fdefd"d#Zd$S)%)Enum) NamedTuple)map_argNodec@sBeZdZdZdeddfddZddZd d Zd d Zd dZ dS) PartitionzPartition class contains all the information about an individual partition. It also provides necessary methods for manipulation the partition. partition_idreturnNcCs4t|_||_t|_t|_d|_d|_g|_dS)Nr)setnodesrparentschildrenZ bfs_levelused_mem_byteslogical_device_ids)selfrrV/var/www/auris/lib/python3.10/site-packages/torch/fx/experimental/partitioner_utils.py__init__ s zPartition.__init__cCs t|jSN)strrrrrr__str__s zPartition.__str__cCs,d|_|jD] }|jt||j7_qdS)Nr)rr get_extra_size_of)rnoderrrrecalculate_mem_sizes zPartition.recalculate_mem_sizecCsXi}t|j|jt|j|j|D] }|jdvr|j|q|j||dS)N>get_attr placeholder)rargs setdefaultkwargsopr addr)rr input_nodesnrrradd_nodes    zPartition.add_nodecs~|jvr=j|i}t|j|jt|j|j|D]}tfdd|jDr6|jdvr6j|q dSdS)Nc3s|]}|jvVqdSr)r .0r#rrr 5s  z(Partition.remove_node..>rr) r removerrrrallusersr r)rrr"Z input_noderrr remove_node)s      zPartition.remove_node) __name__ __module__ __qualname____doc__intrrrr$r+rrrrrs  rc@s&eZdZUeed<eed<eed<dS)DevicenameZavailable_mem_bytesZ logical_idN)r,r-r.r__annotations__r0rrrrr1<s  r1c@seZdZUeed<eed<dS) NodeLatencymem_latency_seccomputer_latency_secNr,r-r.floatr3rrrrr4Bs  r4c@s&eZdZUeed<eed<eed<dS)PartitionLatencyr5r6overall_latency_secNr7rrrrr9Is  r9c@s eZdZdZdZdZdZdZdS) PartitionModerN)r,r-r. size_basedZ sparse_nnZ cost_awareZkl_basedZ aot_basedrrrrr;Rs r;c@seZdZUeeed<ejZeed<dZ e ed<iZ e e efed<iZe e efed<iZe eeefed<dZeed <d S) PartitionerConfigZdevicesmodetransfer_rate_bytes_per_secnode_to_latency_mappingnode_to_partition_mapping#partition_to_logical_device_mappingF saturate_hostN)r,r-r.listr1r3r;r@rBrDr8rEdictrr4rFr0rGrHboolrrrrrAZs  rArr rcCsi}t|j|jt|j|jd}|D]}||vr,t|dd}|r(||j7}qtdqt|dd}|r<||j7}|Std)zGiven a node and a set of nodes, this function return the extra size that needed if this node is included in this set. r size_bytesNznode has no size_bytes attr)rrrrgetattr output_size RuntimeError total_size)rr r"Ztotal_size_of_input_nodesr#rLrrrres     r partitionrEcsxdtdttfdd}dtdtffdd |}tdddd }|D]}|tdddd }|j|jkr9|}q%|S) zVGiven a partition and its nodes' latency, return a PartitionLatency for this partitionrQrcs`g}jD](}|jdvr qi}t|j|jt|j|jtfdd|Ds-||q|S)z>Given a partition, return a list of nodes on the top bfs level>rrc3s$|] }|jvo |jdvVqdS)>rrN)r r r%rQrrr's  zFget_latency_of_one_partition..get_top_nodes..)r r rrrranyappend)rQ top_nodesrr"rrRr get_top_nodess    z3get_latency_of_one_partition..get_top_nodesrc s|}|jt|j|j}|j|j}|j|j}t|jj}|rDtdddd}|D]}|t|||} | j|jkrA| }q.|St|||S)zyGiven a top node of a partition, this function returns the latency of the critical path in the partition rCr5r6r:) r:maxr6r5r r* intersectionr r9) rpartition_latencyZ node_latencyr:r5r6r*Z max_latencyr#Znew_partition_latency dfs_helperrErQrrr\s:  z0get_latency_of_one_partition..dfs_helperrCrW)rrIrr9r:)rQrErVrUZcritical_path_latencyrrZrr[rget_latency_of_one_partitions&,r] partitionscCs$i}|D] }t||}|||<q|S)zGiven all the partitions and node_to_latency_mapping dictionary, return a mapping dictionary of each partition to its overall latency )r])r^rEpartition_to_latency_mappingrQrZrrr get_partition_to_latency_mappings r`parent_partitionchild_partitionrDc Cs|jgkr|jgkr|j|jkrdSd}t}|jD]4}i}t|j|jt|j|j|D]}||jvrM||vrMt|dd}|durH||j7}| |q.q||S)zfGiven two partitions (parent and child), calculate the communication latency between the two. rCrrLN) rr r rrrrrMrNr!) rarbrDZ comm_sizeZ visited_nodesrr"r#rLrrrget_comm_latency_betweens&      rcr_cshdtdtdtffdd dttdttfdd}||}d }|D] }|d }||kr1|}q$|S) zGiven all partitions in a graph, find the critical path among all partitions and return its latency as the latency of the whole graph rQlatency_so_far_secrcsR||j7}|jr'd}|jD]}t||}|||}||kr$|}q|S|S)zJThis function helps to recursively get the latency of a path of partitionsrC)r:r rc)rQrdZmax_latency_secchildZcomm_latency_secZnew_latency_secr\r_rDrrr\s& z4get_latency_of_partitioned_graph..dfs_helperr^cSsdd|D}|S)zvThis function is to return all the partitions without parents as the starting points of all the paths cSsg|] }t|jdkr|qS)r)lenr )r&rQrrr 2szPget_latency_of_partitioned_graph..get_top_partitions..r)r^top_partitionsrrrget_top_partitions-sz.get_top_partitionsrC)rr8rI)r^r_rDrjriZcritical_path_latency_secrQZ latency_secrrfr get_latency_of_partitioned_graphs   rkN)enumrtypingrZ torch.fx.noderrrr1r4r9r;rAr r0rrJr]rIr`r8rcrkrrrrsL  4    Y    $