קgm0ddlmZddlmZmZmZmZddlmZm Z GddZ GddeZ Gdd eZ Gd d eZ Gd d eZGddeZdedeedefdZde deee fde fdZdee deee fdee e ffdZde de defdZdee dee e fdefdZdS) )Enum) NamedTupleDictListSet)Nodemap_argc:eZdZdZdeddfdZdZdZdZd Z dS) 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|_dS)Nr)setnodesr parentschildren bfs_levelused_mem_byteslogical_device_ids)selfr s c/var/www/html/ai-engine/env/lib/python3.11/site-packages/torch/fx/experimental/partitioner_utils.py__init__zPartition.__init__ sH # ('*uu (+  #$-/c*t|jSN)strr )rs r__str__zPartition.__str__s4$%%%rcnd|_|jD]%}|xjt||jz c_&dS)Nr)rrget_extra_size_of)rnodes rrecalculate_mem_sizezPartition.recalculate_mem_sizesMJ G GD   #4T4:#F#F F    G Grci}t|j|jt|j|j|D]%}|jdvr|j|&|j||dS)N>get_attr placeholder)r args setdefaultkwargsopraddr")rr! input_nodesns radd_nodezPartition.add_nodes(*  ;1222 [3444 " "At222 q!!! t !!#####rcv|jvrj|i}t|j|jt|j|j|D]E}t fd|jDr#|jdvrj|F dSdS)Nc3*K|] }|jvVdSr)r).0r,rs r z(Partition.remove_node..5s;,-ATZ'r>r$r%) rremover r&r'r(allusersr)r")rr!r+ input_nodes` r remove_nodezPartition.remove_node)s 4:   J  d # # #,.K DI{5 6 6 6 DK!7 8 8 8* 2 2 1;1A2 m'BBBJ%%j111  % % ' ' ' ' '  r) __name__ __module__ __qualname____doc__intrrr"r-r6rrr r s0S0T0000&&&GGG $ $ $(((((rr c.eZdZUeed<eed<eed<dS)Devicenameavailable_mem_bytes logical_idN)r7r8r9r__annotations__r;r<rrr>r><s. IIIOOOOOrr>c$eZdZUeed<eed<dS) NodeLatencymem_latency_seccomputer_latency_secNr7r8r9floatrBr<rrrDrDBs*rrDc.eZdZUeed<eed<eed<dS)PartitionLatencyrErFoverall_latency_secNrGr<rrrJrJIs6rrJc"eZdZdZdZdZdZdZdS) PartitionModerN)r7r8r9 size_based sparse_nn cost_awarekl_based aot_basedr<rrrMrMRs'JIJHIIIrrMceZdZUeeed<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)r7r8r9rr>rBrMrRrZr\rHr]rrrDr^r;r_r`boolr<rrrXrXZs &\'2D-222),,,,79T$ "3499913tD#I333@B'c49n)=BBBM4rrXr!rr c<i}t|j|jt|j|jd}|D]3}||vr-t |dd}|r ||jz }$t d4t |dd}|r ||jz }nt d|S)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 DI{-... DK/000 ! BB E>> L$77J B)Z-CC))"#@AAA |T22J:!Z%::!!8999 $$r partitionr]c dtdttfd}dtdtffd |}tddd}|D]0}|tddd}|j|jkr|}1|S)zVGiven a partition and its nodes' latency, return a PartitionLatency for this partitionrir cg}jD]r}|jdvr i}t|j|jt|j|jt fd|Ds||s|S)z>Given a partition, return a list of nodes on the top bfs level>r$r%c3<K|]}|jvo|jdvVdS)>r$r%N)rr))r0r,ris rr1zFget_latency_of_one_partition..get_top_nodes..sKY_$P5P)Pr)rr)r r&r'r(anyappend)ri top_nodesr!r+s` r get_top_nodesz3get_latency_of_one_partition..get_top_nodess " O ' 'Dw555,.K DI{5 6 6 6 DK!7 8 8 8( '  &&&rr!c  |}|jt|j|jz}|j|jz}|j|jz}t |j j}|rFtddd}|D]/} |t|||} | j|jkr| }0|St|||S)zyGiven a top node of a partition, this function returns the latency of the critical path in the partition r[rErFrK) rKmaxrFrErr4 intersectionrrJ) r!partition_latency node_latencyrKrErFr4 max_latencyr,new_partition_latency dfs_helperr]ris rryz0get_latency_of_one_partition..dfs_helpers /t4 /Cc  -|/KG G    - 0L L   2\5V V DJ,,Y_==  * ##SVK 8 8(2 $')=?R))%*=!566#8K  13F   rr[rr)r rrrJrK)rir]rprocritical_path_latencyr!rurys`` @rget_latency_of_one_partitionr{s tDz(( ( 5E( ( ( ( ( ( ( ( X i((I,#3 6 6&J   ##SV       1#7 8 8%6 !  r partitionsc>i}|D]}t||}|||<|S)zGiven all the partitions and node_to_latency_mapping dictionary, return a mapping dictionary of each partition to its overall latency )r{)r|r]partition_to_latency_mappingrirus r get_partition_to_latency_mappingrsJ GI DD 8 .  3D$Y// ''rparent_partitionchild_partitionr\c|jgkr|jgkr|j|jkrdSd}t}|jD]|}i}t|j|jt|j|j|D]A}||jvr6||vr2t|dd}| ||jz }| |B}||z S)zfGiven two partitions (parent and child), calculate the communication latency between the two. r[rrcN) rrrr r&r'r(rdrer*) rrr\ comm_size visited_nodesr!r+r,rcs rget_comm_latency_betweenrs +r11  ." 4 4  /?3U U UsIEEM  % % %(*  ;1222 [3444 % %A$***q /E/E$Q d;; )!77I!!!$$$  % 2 22rr~cdtdtdtffd dttdttfd}||}d}|D]}|d}||kr|}|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}||kr|}+|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 6#;u&A$$ #-*-0@@###_44&5O" "!!rr|cng}|D]/}t|jdkr||0|S)zvThis function is to return all the partitions without parents as the starting points of all the paths r)lenrrn)r|top_partitionsris rget_top_partitionsz.get_top_partitions-sI# 1 1I9$%%**%%i000rr[)r rHr) r|r~r\rrcritical_path_latency_secri latency_secrys `` @r get_latency_of_partitioned_graphrs"i"U"u"""""""", tI 4 ?    (' 33N ##44  jC00 2 2 2(3 % $$rN)enumrtypingrrrr torch.fx.noderr r r>rDrJrMrXr;r r{rrHrrr<rrrst............''''''''1(1(1(1(1(1(1(1(hZ      *   zD        %D%T%s%%%%8V!V!37k8I3JV!V!V!V!V!r (Y (:>t[?P:Q ( )% %& ( ( ( ( !3!3!3"'!3!3!3!3H0%Y0%"&y2B'B"C0%"'0%0%0%0%0%0%r