o H h0@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)setnodesrparentschildren bfs_levelused_mem_byteslogical_device_ids)selfrr[/var/www/vscode/kcb/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$rrr 5s  z(Partition.remove_node..>rr) r removerrrr allusersr!r)rrr# input_noderrr remove_node)s      zPartition.remove_node) __name__ __module__ __qualname____doc__intrrrr%r-rrrrrs  rc@s&eZdZUeed<eed<eed<dS)Devicenameavailable_mem_bytes logical_idN)r.r/r0r__annotations__r2rrrrr3<s  r3c@seZdZUeed<eed<dS) NodeLatencymem_latency_seccomputer_latency_secNr.r/r0floatr7rrrrr8Bs  r8c@s&eZdZUeed<eed<eed<dS)PartitionLatencyr9r:overall_latency_secNr;rrrrr=Is  r=c@s eZdZdZdZdZdZdZdS) PartitionModerN)r.r/r0 size_based sparse_nn cost_awarekl_based aot_basedrrrrr?Rs r?c@seZdZUeeed<ejZeed<dZ e ed<iZ e e efed<iZe e efed<iZe eeefed<dZeed <d S) PartitionerConfigdevicesmodetransfer_rate_bytes_per_secnode_to_latency_mappingnode_to_partition_mapping#partition_to_logical_device_mappingF saturate_hostN)r.r/r0listr3r7r?rDrKrMr<rNdictrr8rOr2rPrQboolrrrrrIZs  rIrr rcCsi}t|j|jt|j|jd}|D]}||vr,t|dd}|r(||j7}qtdqt|dd}|r<||j7}|Std)zGiven a node and a set of nodes, this function return the extra size that needed if this node is included in this set. r size_bytesNznode has no size_bytes attr)rrrr getattr output_size RuntimeError total_size)rr r#total_size_of_input_nodesr$rUrrrres     r partitionrNcsxdtdttfdd}dtdtffdd |}tdddd }|D]}|tdddd }|j|jkr9|}q%|S) zVGiven a partition and its nodes' latency, return a PartitionLatency for this partitionr[rcs`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$|] }|jvo |jdvVqdS)>rrN)r r!r&r[rrr(s  zFget_latency_of_one_partition..get_top_nodes..)r r!rrrr anyappend)r[ top_nodesrr#rr\r get_top_nodess    z3get_latency_of_one_partition..get_top_nodesrc s|}|jt|j|j}|j|j}|j|j}t|jj}|rDtdddd}|D]}|t|||} | j|jkrA| }q.|St|||S)zyGiven a top node of a partition, this function returns the latency of the critical path in the partition rLr9r:r>) r>maxr:r9r r+ intersectionr r=) rpartition_latency node_latencyr>r9r:r+ max_latencyr$new_partition_latency dfs_helperrNr[rrris:  z0get_latency_of_one_partition..dfs_helperrLra)rrRrr=r>)r[rNr`r_critical_path_latencyrrdrrhrget_latency_of_one_partitions&,rk partitionscCs$i}|D] }t||}|||<q|S)zGiven all the partitions and node_to_latency_mapping dictionary, return a mapping dictionary of each partition to its overall latency )rk)rlrNpartition_to_latency_mappingr[rdrrr get_partition_to_latency_mappings rnparent_partitionchild_partitionrMc Cs|jgkr|jgkr|j|jkrdSd}t}|jD]4}i}t|j|jt|j|j|D]}||jvrM||vrMt|dd}|durH||j7}| |q.q||S)zfGiven two partitions (parent and child), calculate the communication latency between the two. rLrrUN) rr r rrrr rVrWr") rorprM comm_size visited_nodesrr#r$rUrrrget_comm_latency_betweens&      rsrmcshdtdtdtffdd dttdttfdd}||}d }|D] }|d }||kr1|}q$|S) zGiven all partitions in a graph, find the critical path among all partitions and return its latency as the latency of the whole graph r[latency_so_far_secrcsR||j7}|jr'd}|jD]}t||}|||}||kr$|}q|S|S)zJThis function helps to recursively get the latency of a path of partitionsrL)r>r rs)r[rtmax_latency_secchildcomm_latency_secnew_latency_secrirmrMrrris& z4get_latency_of_partitioned_graph..dfs_helperrlcSsdd|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'r[rrr 2szPget_latency_of_partitioned_graph..get_top_partitions..r)rltop_partitionsrrrget_top_partitions-sz.get_top_partitionsrL)rr<rR)rlrmrMr}r|critical_path_latency_secr[ latency_secrryr get_latency_of_partitioned_graphs   rN)enumrtypingr torch.fx.noderrrr3r8r=r?rIr r2rrSrkrRrnr<rsrrrrrsL  4    Y    $