sgm0XddlmZddlmZmZmZmZddlmZm Z GddZ GddeZ Gdd eZ Gd d eZ Gd d eZGddeZdedeedefdZde deee fde fdZdee deee fdee e ffdZde de defdZdee dee e fdefdZy) )Enum) NamedTupleDictListSet)Nodemap_argc8eZdZdZdeddfdZdZdZdZd Z y) PartitionzPartition class contains all the information about an individual partition. It also provides necessary methods for manipulation the partition. partition_idreturnNct|_||_t|_t|_d|_d|_g|_y)Nr)setnodesr parentschildren bfs_levelused_mem_byteslogical_device_ids)selfr s Z/var/www/html/venv/lib/python3.12/site-packages/torch/fx/experimental/partitioner_utils.py__init__zPartition.__init__ s< # ('*u (+  #$-/c,t|jSN)strr )rs r__str__zPartition.__str__s4$$%%rcd|_|jD]+}|xjt||jz c_-y)Nr)rrget_extra_size_of)rnodes rrecalculate_mem_sizezPartition.recalculate_mem_sizes<JJ GD   #4T4::#F F  Grc@i}t|j|jt|j|j|D],}|jdvs|j j |.|j j ||jy)N>get_attr placeholder)r args setdefaultkwargsopraddr")rr! input_nodesns radd_nodezPartition.add_nodesy(*  ;112 [334 "Att22 q! " t !!#rc|jvrjj|i}t|j|jt|j |j|D]K}t fd|jDs"|jdvs1jj|Mjyy)Nc3:K|]}|jvywr)r).0r,rs r z(Partition.remove_node..5s,-ATZZ's>r$r%) rremover r&r'r(allusersr)r")rr!r+ input_nodes` r remove_nodezPartition.remove_node)s 4::  JJ  d #,.K DII{55 6 DKK!7!7 8* 2 1;1A1A mm'BBJJ%%j1  2  % % ' r) __name__ __module__ __qualname____doc__intrrr"r-r6rrr r s/0S0T0&G $(rr c,eZdZUeed<eed<eed<y)Devicenameavailable_mem_bytes logical_idN)r7r8r9r__annotations__r;r<rrr>r><s IOrr>c"eZdZUeed<eed<y) NodeLatencymem_latency_seccomputer_latency_secNr7r8r9floatrBr<rrrDrDBsrrDc,eZdZUeed<eed<eed<y)PartitionLatencyrErFoverall_latency_secNrGr<rrrJrJIsrrJc eZdZdZdZdZdZdZy) PartitionModerN)r7r8r9 size_based sparse_nn cost_awarekl_based aot_basedr<rrrMrMRsJIJHIrrMceZdZUeeed<ejZeed<dZ e ed<iZ e e efed<iZe e efed<iZe eeefed<dZeed <y ) PartitionerConfigdevicesmodetransfer_rate_bytes_per_secnode_to_latency_mappingnode_to_partition_mapping#partition_to_logical_device_mappingF saturate_hostN)r7r8r9rr>rBrMrRrZr\rHr]rrrDr^r;r_r`boolr<rrrXrXZsq &\'22D-2),,79T$ "34913tD#I3@B'c49n)=BM4rrXr!rr cJi}t|j|jt|j|jd}|D]0}||vst |dd}|r||j z }'t dt |dd}|r||jz }|St d)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)r r&r'r(getattr output_size RuntimeError total_size)r!rr+total_size_of_input_nodesr,rcs rr r es %'K DII{--. DKK//0 ! B E> L$7J)Z-C-CC)"#@AAB|T2J!Z%:%::! %$899r partitionr]c dtdttfd}dtdtffd |}tddd}|D]3}|tddd}|j|jkDs2|}5|S)zVGiven a partition and its nodes' latency, return a PartitionLatency for this partitionrir cg}jD]y}|jdvri}t|j|jt|j |jt fd|Dri|j|{|S)z>Given a partition, return a list of nodes on the top bfs level>r$r%c3ZK|]"}|jvxr|jdv$yw)>r$r%N)rr))r0r,ris rr1zFget_latency_of_one_partition..get_top_nodes..s4Y__$P5P)PPs(+)rr)r r&r'r(anyappend)ri top_nodesr!r+s` r get_top_nodesz3get_latency_of_one_partition..get_top_nodess " OO 'Dww55,.K DII{55 6 DKK!7!7 8(  & 'rr!c  |}|jt|j|jz}|j|jz}|j|jz}t |j j j}|rGtddd}|D]2} |t|||} | j|jkDs1| }4|St|||S)zyGiven a top node of a partition, this function returns the latency of the critical path in the partition r[rErFrK) rKmaxrFrErr4 intersectionrrJ) r!partition_latency node_latencyrKrErFr4 max_latencyr,new_partition_latency dfs_helperr]ris rryz0get_latency_of_one_partition..dfs_helpers/t4 /CCc  - -|/K/KG    - - 0L0L L   2 2\5V5V V DJJ,,Y__= * ##SVK 8(2$')=?R)%*==!556#8K 8  13F  rr[rr)r rrrJrK)rir]rprocritical_path_latencyr!rurys`` @rget_latency_of_one_partitionr{s tDz(( ( 5E( Xi(I,#3 6&   ##SV    1 1#77 8%6 ! 6 ! r partitionsc:i}|D]}t||}|||<|S)zGiven all the partitions and node_to_latency_mapping dictionary, return a mapping dictionary of each partition to its overall latency )r{)r|r]partition_to_latency_mappingrirus r get_partition_to_latency_mappingrsD GI D 8 . 3D$Y/ D ('rparent_partitionchild_partitionr\c|jgk7r)|jgk7r|j|jk(ryd}t}|jD]}i}t|j|j t|j |j |D]E}||jvs||vst|dd}|||jz }|j|G||z S)zfGiven two partitions (parent and child), calculate the communication latency between the two. r[rrcN) rrrr r&r'r(rdrer*) rrr\ comm_size visited_nodesr!r+r,rcs rget_comm_latency_betweenrs ++r1  . ." 4  / /?3U3U UIEM  %% %(*  ;112 [334 %A$***q /E$Q d; )!7!77I!!!$  % % 2 22rr~cdtdtdtffd dttdttfd}||}d}|D]}|d}||kDs|}|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 rilatency_so_far_secr c||jz }|j}|jr6d}|jD]#}t|| }|||z}||kDs"|}%|S|S)zJThis function helps to recursively get the latency of a path of partitionsr[)rKrr) rirrmax_latency_secchildcomm_latency_secnew_latency_secryr~r\s rryz4get_latency_of_partitioned_graph..dfs_helpers :    %%   !O"++ 6#;u&A$ #--0@@##_4&5O 6# "!!rr|clg}|D],}t|jdk(s|j|.|S)zvThis function is to return all the partitions without parents as the starting points of all the paths r)lenrrn)r|top_partitionsris rget_top_partitionsz.get_top_partitions-sA# 1I9$$%*%%i0 1rr[)r rHr) r|r~r\rrcritical_path_latency_secri latency_secrys `` @r get_latency_of_partitioned_graphrs|"i"U"u", tI 4 ? ( 3N ##4  C0 2 2(3 %4 %$rN)enumrtypingrrrr torch.fx.noderr r r>rDrJrMrXr;r r{rrHrrr<rrrs&..'1(1(hZ  * zD  %D%T%s%8V!V!37k8I3JV!V!r (Y (:>t[?P:Q ( )% %& ( !3!3!3"'!3H0%Y0%"&y2B'B"C0%"'0%r