tj<0ddlmZddlmZddlmZmZGddZGddeZGdd eZ Gd d eZ Gd d eZ GddeZ dede edefdZdedeee fde fdZdeedeee fdeee ffdZdedededefdZdeedeee fdedefdZdS) )Enum) NamedTuple)map_argNodecVeZdZdZdeddfdZdefdZd dZde ddfd Z de ddfd Z dS) PartitionzPartition class contains all the information about an individual partition. It also provides necessary methods for manipulating the partition. partition_idreturnNct|_||_t|_t|_d|_d|_g|_dS)Nr)setnodesr parentschildren bfs_levelused_mem_byteslogical_device_ids)selfr s e/var/www/html/CA-Chatbot/venv/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 Grrci}t|j|jt|j|j|D]%}|jdvr|j|&|j||dS)N>get_attr placeholder)rargs 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..4s;,-ATZ'r>r!r") rremoverr#r$r%allusersr&r)rrr( 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)r N) __name__ __module__ __qualname____doc__intrrrrrr*r3rrrrs0S0T0000&&&&&GGGG $T $d $ $ $ $((((((((rrc.eZdZUeed<eed<eed<dS)Devicenameavailable_mem_bytes logical_idN)r4r5r6r__annotations__r8r9rrr;r;;s. IIIOOOOOrr;c$eZdZUeed<eed<dS) NodeLatencymem_latency_seccomputer_latency_secNr4r5r6floatr?r9rrrArAAs*rrAc.eZdZUeed<eed<eed<dS)PartitionLatencyrBrCoverall_latency_secNrDr9rrrGrGHs6rrGc"eZdZdZdZdZdZdZdS) PartitionModerN)r4r5r6 size_based sparse_nn cost_awarekl_based aot_basedr9rrrJrJQs'JIJHIIIrrJceZdZUeeed<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)r4r5r6listr;r?rJrOrWrYrErZdictrrAr[r8r\r]boolr9rrrUrUYs &\'2D-222),,,,79T$ "3499913tD#I333@B'c49n)=BBBM4rrUrrr c<i}t|j|jt|j|jd}|D]3}||vr-t |dd}|r ||jz }$t d4t |dd}|r ||jz }nt d|S)zGiven a node and a set of nodes, this function return the extra size that needed if this node is included in this set. r size_bytesNznode has no size_bytes attr)rr#r$r%getattr output_size RuntimeError total_size)rrr(total_size_of_input_nodesr)rbs rrrds %'K DI{-... DK/000 ! BB E>> L$77J B)Z-CC))"#@AAA |T22J:!Z%::!!8999 $$r partitionrZc  dtdttfd}dtdtdtffd |}tddd}|D]0}|tddd}|j|jkr|}1|S) zVGiven a partition and its nodes' latency, return a PartitionLatency for this partitionrhr cg}jD]r}|jdvr i}t|j|jt|j|jt fd|Ds||s|S)z>Given a partition, return a list of nodes on the top bfs level>r!r"c3<K|]}|jvo|jdvVdS)>r!r"N)rr&)r-r)rhs rr.zFget_latency_of_one_partition..get_top_nodes..sKY_$P5P)Pr)rr&rr#r$r%anyappend)rh top_nodesrr(s` r get_top_nodesz3get_latency_of_one_partition..get_top_nodess " O ' 'Dw555,.K DI{5 6 6 6 DK!7 8 8 8$ '  &&&rrpartition_latencyc  |}|jt|j|jz}|j|jz}|j|jz}t |j j}|rFtddd}|D]/} |t|||} | j|jkr| }0|St|||S)zyGiven a top node of a partition, this function returns the latency of the critical path in the partition rXrBrCrH) rHmaxrCrBr r1 intersectionrrG) rrp node_latencyrHrBrCr1 max_latencyr)new_partition_latency dfs_helperrZrhs rrxz0get_latency_of_one_partition..dfs_helpers /t4 /Cc  -|/KG G    - 0L L   2\5V V DJ,,Y_==  * ##SVK 8 8(2 $')=?R))%*=!566#8K  13F   rrXrr)rr^rrGrH)rhrZrorncritical_path_latencyrrprxs`` @rget_latency_of_one_partitionrzs tDz(( ( 2B( GW( ( ( ( ( ( ( ( X i((I,#3 6 6&J   ##SV       1#7 8 8%6 !  r partitionsc>i}|D]}t||}|||<|S)zGiven all the partitions and node_to_latency_mapping dictionary, return a mapping dictionary of each partition to its overall latency )rz)r{rZpartition_to_latency_mappingrhrps r get_partition_to_latency_mappingr~sJ GI DD 8 .  3D$Y// ''rparent_partitionchild_partitionrYc|jgkr|jgkr|j|jkrdSd}t}|jD]|}i}t|j|jt|j|j|D]A}||jvr6||vr2t|dd}| ||jz }| |B}||z S)zfGiven two partitions (parent and child), calculate the communication latency between the two. rXrrbN) rr rrr#r$r%rcrdr') rrrY comm_size visited_nodesrr(r)rbs rget_comm_latency_betweenrs +r11  ." 4 4  /?3U U UsIEEM  % % %(*  ;1222 [3444 % %A$***q /E/E$Q d;; )!77I!!!$$$  % 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 rhlatency_so_far_secr c||jz }|jr6d}|jD]*}t||}|||z}||kr|}+|S|S)zJThis function helps to recursively get the latency of a path of partitionsrX)rHrr) rhrmax_latency_secchildcomm_latency_secnew_latency_secrxr}rYs rrxz4get_latency_of_partitioned_graph..dfs_helpers :      #!O"+ 6 6#;u&A$$ #-*-0@@###_44&5O" "!!rr{cd|D}|S)zvThis function is to return all the partitions without parents as the starting points of all the paths cBg|]}t|jdk|S)r)lenr)r-rhs r zPget_latency_of_partitioned_graph..get_top_partitions..1s4   #S9J5K5Kq5P5PI5P5P5Prr9)r{top_partitionss rget_top_partitionsz.get_top_partitions,s(   '1   rrX)rrEr^) r{r}rYrrcritical_path_latency_secrh latency_secrxs `` @r get_latency_of_partitioned_graphr s"i"U"u"""""""",tI4 ?(' 33N ##44  jC00 2 2 2(3 % $$rN)enumrtypingr torch.fx.noderrrr;rArGrJrUr r8rr_rzr^r~rErrr9rrrsv''''''''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!3!3H/%Y/%"&y2B'B"C/%"'/% /%/%/%/%/%/%r