qiP0UddlmZddlZddlZddlmZmZmZddlm Z ddl Z ddl Z ddl mZddlZ ddlmZddlmZmZddlmZmZmZmZmZmZmZmZdd lm Z m!Z!dd l"m#Z#dd l$m%Z%m&Z&erdd l'm(Z(m)Z)d Z*da+de,d<ejZGddZ.ejZGddZ/ddZ0d dZ1ejdd!dZ3 d"dZ4GddeZ5d#dZ6GddZ7 d$dZ8 d%dZ9y)&) annotationsN)Any TYPE_CHECKINGUnion)patch) OrderedSet)configselect_algorithm)Buffer ChoiceCallerLayoutMultiTemplateBufferOperationBufferShapeAsConstantBuffer StorageBox TensorBox) KernelInputsMMKernelInputs) SchedulerNode) NullHandlerV) GeneratorSequencedistributed_autotunedist.ProcessGroup | None _AUTOTUNE_PGc.eZdZUdZdZded<dZded<y)_DistributedAutotuneStatezA State used to track autotuning during a graph_context() rintautotuned_indexautotuned_local_countN)__name__ __module__ __qualname____doc__r!__annotations__r"j/home/ubuntu/crypto_trading_bot/.venv/lib/python3.12/site-packages/torch/_inductor/distributed_autotune.pyrr(s  OS"#3"r)rc"eZdZUded<ded<y)_DistributedAutotuneInfor indexboollocalN)r#r$r%r'r(r)r*r,r,7s J Kr)r,ctjr@tjr,t tjj datSy)Npt2_distributed_autotune_pg)pg_tag)dist is_availableis_initializedrdistributed_c10d_new_group_with_tagr(r)r*get_autotune_pgr8=sI t224  00DD4EL r)cltjsJt|}t|}t ||y)z Finish the distributed autotuning by propagating the autotuning results between the ranks and then replacing the placeholder with the real Buffer. N)r distributed_max_autotune_gemm_autotune_local_nodes_sync_autotune_remote_nodes) schedulerautotune_resultschoices_by_indexs r*schedulerAIs5  / // /,Y7-.9&67r)c#KttjdtrJtjt dtjt y#tjt wxYww)zd Wrapped around processing a graph, sets up figuring out which ranks tune which shapes. F)check_poisonedN) isinstancerget_distributed_autotune_staterset_distributed_autotune_staterr(r)r* graph_contextrGTsg  ((>! $$%>%@A8  ((7((7sAB A) B )BB ctjsytx}syt|dkrytj }|j }|xj dz c_||jz|jk(}t||tjjt<|r|xjdz c_ ytjj j"j%t'|||S)z Used by an op (like `mm`) to determine if the op should be autotuned locally (returns None) or remotely (returns a placeholder Buffer). Nr )r r:r8lenrdistributed_autotune_stater!sizerankr, current_nodemeta_DISTRIBUTED_AUTOTUNE_KEYr"torch _inductorirrcreate_DistributedAutotuneBuffer)namechoicesinputslayout autotune_pgstater-r/s r*maybe_autotune_remoter[es  / /*, ,K , 7|q ( (E  ! !E Q K$$& &+*:*:*< N)r rnrobjectrgraphroriginal_inputsrDrXr get_choicer^nodesr)rf ser_choicern kernel_inputschoicebuffers r*autotunez#_DistributedAutotuneBuffer.autotunes @ \\!''; 5 *+BT-A-A+BCMdkk62 22**4;; FF.!!##% F fi0 00   s BCC)rgr]rW list[Buffer]rXrreturnNone)rlz int | Nonerzzdict[ChoiceCaller, float])rt_SerializedChoicerzr) r#r$r%r&r'rdrerx __classcell__)rhs@r*rTrTsU((( (  ( "(" ""r)rTcpt}|sJdg|jz}tjj |||t d|D}dg|z}d}|D]@}|D]9}t |tsJ||jJ|||j<|dz };B||k(s Jd|d||S)zT Perform the all_gather to collect the autotune results from all the ranks. N)groupc32K|]}t|ywrj)rI).0xs r* z_sync..s0SV0srr zcount mismatch:  != ) r8rKrP distributedall_gather_objectsumrDr|r-)r?rY all_states node_countr@ check_count other_resultsrvs r*r<r<s "#K ;269I9I9K0KJ '' 4DK'X0Z00J150CK# # Ff&78 88#FLL19 99-3 V\\ * 1 K    $V(8 D &VV $ r)cLeZdZdZddZd dZed dZed dZd dZ y) r|z This is a serializer for the autotune choice. KernelTemplateChoice can't be serialized directly (the template and inputs prevent this) so we need to serialize it by parts and reconstruct later on. c||_tj||_|j |j |_yrj)r-r|_template_uid_from_choice template_uid_compute_kwargs descriptionkwargs)rfr-rvs r*rdz_SerializedChoice.__init__s4 -GGO**6+=+=> r)c&|j}i|j}d|vrF|jdjd}t j ||d|dk(|d<i}ddlm}m}||} ||| |||} | jS)z= Deserialize the ChoiceCaller and return it. BLOCK_Krr EVEN_K)DictKernelTemplateParamsKernelTemplateChoice) _template_from_uidrrsget_sizesympygcdkernel_template_choicerrrv) rfrXrWtemplaterk extra_kwargsrrparamsktcs r*rrz_SerializedChoice.get_choices **, DKK    q!**,Q/A$yyF9,=>&BSSF8 ')  *&1"8V\66Rzzr)cj|siSi}|jdD]}|jdd\}}|j|j}}|dk(rd||<C|dk(rd||<N|jrt|||<m|j dr|j dsJ|dd ||<|S) zI Given a template description turn it into input kwargs. ,=r TrueTFalseF')splitstripisdigitr startswithendswith)rrcfgkeyvals r*rz!_SerializedChoice._compute_kwargss I46$$S) (Cyya(HCyy{CIIKCf}"s #s !#hs ~~c*s||C/@@@!!Bis  ( r)ct|tjr<|jjdk(ryt d|jjt|tj ryt dt|)z Given a ChoiceCaller figure out which template represents it. This is reversed by _template_from_uid(). mmz!torch._inductor.kernel.mm.aten_mmz TODO: kernel z%torch._inductor.kernel.mm.mm_templatezTODO: )rDr ExternKernelCallerrvrU RuntimeErrorTritonTemplateCallertype)rvs r*rz+_SerializedChoice._template_uid_from_choicesq f.AA B}}!!T):"]6==3E3E2H#IJJ  0 E E F:V ~67 7r)c|jjd}t|d}|ddD]}t||}|S)z2 See _template_uid_from_choice(). .rr N)rrglobalsgetattr)rfpartsobjrs r*rz$_SerializedChoice._template_from_uid,sO!!'',ia!qr "A#q/C " r)N)r-r rvr rzr{)rXrrWrrzzChoiceCaller | None)rr]rzz dict[str, Union[int, str, bool]])rvr rzr])rzr) r#r$r%r&rdrr staticmethodrrrr(r)r*r|r|s> ? 4088$r)r|c>g}|jD]}t|ts|jx}#t|tr4t|t sE|j x}T|jx}c|jt}|{|jsJ|j\}}t|j|} |j| tj } t#|| j$k(s!Jdt#|d| j$d|S)zt Go through the nodes in the scheduler and autotune the kernels which should be autotuned by this rank. z'incorrect local autotuned nodes found (r))rsrDrnoderTr origin_noderNgetrOr/get_min_choicer|r-appendrrJrIr") r>r?r inner_noderrNinfo min_choice_rvrZs r*r;r;7s.13 ($ . )) #J ,  j"< = *&9: %11 1K : $$ $D - xx12 < zzz #113 A"4::z:'A (D  ( (E  E$?$? ? 1#6F2G1HUMhMhLiijk ? r)ct|jD]\}}t|tst|jx}t s4|j J|j jt}|j||j}|j}t|tsJ|j}t|tsJ|j|jk(sJ|j||||y)zo Go through the nodes in the scheduler and autotune the nodes that were autotuned on remote ranks. N) enumeratersrDrrrTrrNrOrxr-datarrrX _replace_node) r>r@ir dist_noder out_tensorbox out_storage out_buffers r*r=r=jsY__-D4 dM *z)) #Y&@0 ((4 44((--.GHD%../? /KLM',,Kk:6 66$))Jj/: ::$$ (8(88 88  # #J 1d CDr))rzr)r>#torch._inductor.scheduler.Schedulerrzr{)rzzGenerator[None, None, None]) rUr]rVzlist[ChoiceCaller]rWryrXrrzz(TensorBox | ShapeAsConstantBuffer | None)r?list[_SerializedChoice]rzSequence[_SerializedChoice])r>rrzr)r>rr@rrzr{): __future__r contextlib dataclassestypingrrr unittest.mockrrtorch._loggingrPtorch.distributedrr3torch.fxtorch.utils._ordered_setrr r rRr r rrrrrrrurrr>r virtualizedrrcollections.abcrrrOrr' dataclassrr,r8rAcontextmanagerrGr[rTr<r|r;r=r(r)r*rsG",,  /&   8$'33)- &-  # # #  8  8 8  *4@JP-B4!44p8ZZz0200fD2D1D Dr)