shSSKJr SSKrSSKrSSKrSSKrSSKrSSKrSSKrSSK r SSK r SSK J r J r SSKJr SSKJrJrJrJr SSKJrJrJrJrJr SSKrSSKrSSKJr SSKJr SS KJ r SS K!J"r" SS K#J$r$J%r%J&r&J'r'J(r( SS K)J*r*J+r+ SS K,J-r- SSK.J/r/ \(aSSK0J1r1 SSK2J3r3 SSK4J5r5 SSK6J7r7 SSK8J9r9 SSK:J;r; SSK8Jr> SSK?J@r@ SrASqB\-"\CS5rD\R"\C5rF"SS5rG"SS5rH"S S!\I5rJ\RS) annotationsN)IterableSequence)ThreadPoolExecutor)byrefc_size_tc_void_pCDLL)AnyCallableOptional TYPE_CHECKINGUnion)multiprocessing)get_interface_for_device) rand_strided)ir) CppCodeCache CUDACodeCache DLLWrapperget_hash PyCodeCache) get_gpu_typeis_gpu)getArtifactLogger) OrderedSet) BaseProcess)Queue) ModuleType)TritonTemplateCaller) WorkspaceArg)config)WorkspaceZeroMode) benchmarker)VCUDA_VISIBLE_DEVICESF autotuningc\rSrSrSrg)Ping;N__name__ __module__ __qualname____firstlineno____static_attributes__r,t/Users/tiagomarins/Projetos/claudeai/copy_bank/venv/lib/python3.13/site-packages/torch/_inductor/autotune_process.pyr*r*;r3r*c\rSrSrSrg)Pong?r,Nr-r,r3r4r7r7?r5r3r7c\rSrSrSrg)!NonzeroWorkspaceNotSupportedErrorCr,Nr-r,r3r4r:r:Cr5r3r:c#v# UcSv g[RR[5n[ U5[R['Sv Uc[R[ gU[R['g!Uc[R[ fU[R['f=f7f)z Context manager to set the CUDA_VISIBLE_DEVICES environment variable to the specified single device. If device is None, don't manipulate the environment. N)osenvirongetr'str)devicecurrents r4set_cuda_visible_devicerCGs ~ jjnn12G'*6{BJJ#$7 ? /0/6BJJ+ , ? /0/6BJJ+ ,sA B9B1B92B66B9c\rSrSr%SrSrS\S'SrS\S'SrS\S 'Sr S\S '\ SS j5r \ SS j5r SS jr SSjrSSjrSSjrSSSjjrSSjrSSjrSSSjjrSrg) TuningProcess\z Abstraction for launching a helper process to benchmark kernels. Spawns the parent process and uses multiprocessing queues to send benchmark requests and return results. N Optional[int]rAzOptional[BaseProcess]processzOptional[Queue[Any]] request_queueresponse_queuec[RS[RR [ 55 [ RX5 g![a [RS5 gf=f)z$ Entry point for the child process. z2Entering TuningProcess child. Visible devices = %szException in TuningProcessN) autotuning_logdebugr=r>r?r'rEworkloop Exception exception)rIrJs r4 process_mainTuningProcess.process_mainisZ  @ JJNN/ 0  C  " "= A C  $ $%A B CsAA10A1c&UR5nUcg[U[5(aUR[ 55 OL[U[ 5(a URUR 55 O[S[U535eM)z, Work loop for the benchmarking subprocess. NzInvalid request type ) r? isinstancer*putr7BenchmarkRequest benchmark RuntimeErrortype)rIrJobjs r4rNTuningProcess.workloopzsx ##%C{C&&""46*C!122""3==?3"%:49+#FGGr3crURSL=(a# URSL=(a URSL$)z/ True if the sub-process has been initialized. NrHrIrJselfs r4validTuningProcess.valids; LL $ 0""$. 0##4/ r3c.S=Ul=UlUlg)z" Reset to an uninitialized state. Nr]r^s r4clearTuningProcess.clearsCGF Ft)D,?r3cUR5(ag[R"S5nUR5UlUR5UlUR URURUR 4S9UlURce[UR5 URR5 SSS5 g!,(df  g=f)z Create child process, request/response queues, and do the warm up. Set the environment to make only the provided GPU device visible to the process. Nspawn)targetargs) r`r get_contextrrIrJProcessrQrHrCrAstart)r_ctxs r4 initializeTuningProcess.initializes ::<< ))'2 YY[!iik{{$$""###  ||''' $T[[ 1 LL   2 1 1s 0C C"cxUR5 URceURRU5 g)z( Push a work item to the child process. N)rmrIrU)r_rZs r4rUTuningProcess.puts4 !!--- s#r3cURceURceUnSnUb&US:a US-nURRSS9nUcURRUS9nU$![Ra# URR 5(deOf=fUcM]US:aMNf![Ra< URR nUcURUUS9 eUR5 ef=f)a Get a response from the child process. Raises queue.Empty on timeout or if the process dies. This method is (so far) only used by TuningProcessPool, where torch._inductor.config entries are being used to populate the timeouts: Arguments: @param result_timeout: Timeout in seconds, defaults to 120.0 or to config.max_autotune_subproc_result_timeout_seconds when called by TuningProcessPool @param graceful_timeout: Timeout in seconds to allow graceful shutdown (SIGTERM is sent after this time). Defaults to 3.0 or to config.max_autotune_subproc_graceful_timeout_seconds @param terminate_timeout: Timeout in seconds after SIGTERM, until we send SIGKILL if the process remains alive. Defaults to 1.0 or to config.max_autotune_subproc_terminate_timeout_seconds. Returns: A response from the child process (Any type) N?g?timeout)graceful_timeoutterminate_timeout) rHrJr?queueEmptyis_aliveexitcodekillrc)r_result_timeoutrurvremaining_timeoutresstatuss r4r?TuningProcess.gets,,||'''""... $2!'38IS8P%,%""1155c5B;--11:K1LC !;;"#||4466! 7" (38IS8P;; ..>II)9*;JJL s5B4A, B4,4B# B4"B##B4+B44ADcUR5(a:URceURceURRS5 gg)z( Signal the child process to terminate. N)r`rHrIrUr^s r4 terminateTuningProcess.terminatesJ ::<<<<+ ++%%1 11    " "4 ( r3ctURb+URR5 UR5 gg)z% Wait for the child process to exit. N)rHjoinrcr^s r4waitTuningProcess.waits, << # LL    JJL $r3cRURGbUR5 URRUS9 URR5(a[R SURR 5 URR5 URRUS9 URR5(aD[RSURR 5 URR5 UR5 gg)Nrsz&Sending SIGTERM to process with PID %dz&Sending SIGKILL to process with PID %d) rHrrryrLwarningpiderrorr{rc)r_rurvs r4r{TuningProcess.kills << # NN  LL  &6  7||$$&&&&<LL$$ &&( !!*;!<<<((**"((@ ((LL%%' JJL! $r3r])rI Queue[Any]rJrreturnNone)rboolrr)rZr rr)g^@g@rr)rr )g@rr)r.r/r0r1__doc__rA__annotations__rHrIrJ staticmethodrQrNr`rcrmrUr?rrr{r2r,r3r4rErE\s !FM %)G ")*.M'.+/N(/C!C"C CC HH  G !2$MP1 1f)r3rEcp\rSrSr%SrSrS\S'SrS\S'SSjrSS jr SS jr SS jr SS jr S r g)TuningProcessPooliz Maintains a pool of TuningProcesses to benchmark kernels in parallel across devices. By default, we create one TuningProcess per device and set the sub-process environment to make only that device visible. Nz$Optional[queue.Queue[TuningProcess]] processeszOptional[ThreadPoolExecutor]executorcURSLURSL:XdeURbgUR5n[R SU5 [ R "5UlUHPn[US9nUR5 UR[55 URRU5 MR URR H'n[URSS9[5(aM'e [[U5S9Ul[ (d"SqSSKnUR%UR&5 gg)z Start the child processes. Nz$Sub-process autotune device list: %s)rA)r|) max_workersTr)rrget_device_listlogrMrwrrErmrUr*rTr?r7rlenEXIT_HANDLER_REGISTEREDatexitregisterr)r_devicesrAprs r4rmTuningProcessPool.initialize$s$&DMMT,ABBB >> % &&( 8'BFV,A LLN EE$&M NN  q ! %%Aaee4e8$?? ??& +s7|D '&&* #  OODNN + 'r3c[R(dS/$[5n[U5nUR 5n[ [ R;aR[ R[ RS5Vs/sHn[U5PM nn[U5U::deU$[[U55$s snf)z4 Gather the list of devices to be used in the pool. N,) r#autotune_multi_devicerr device_countr'r=r>splitintrlistrange)r_gpu_typedevice_interfacecountdrs r4r!TuningProcessPool.get_device_listIs++6M>3H= --/ 2:: -')zz2F'G'M'Mc'RS'R!s1v'RGSw<5( ((NE%L!! Ts>B<c>URb!URR5 SUlURbbURRHnUR 5 M URRHnUR 5 M SUlgg)z* Signal all child processes to terminate. N)rshutdownrrwrr)r_rs r4rTuningProcessPool.terminate]su == $ MM " " $ DM >> %^^)) *^^))*!DN &r3cNURceURceURR5nURUR5 UR[R [R [R5URRU5 $![RaC [R"SUS35 [S5sURRU5 $f=f!URRU5 f=f)z Entry point for the thread-pool helper threads: Wait for an open TuningProcess, remove it from the queue, execute the benchmark in that subprocess, and return the TuningProcess to the queue. zFailed to benchmark choice 'z['. It will be ignored. Please debug the root cause in case the choice can bring perf gains.inf) bmreqrr?rUr#+max_autotune_subproc_result_timeout_seconds-max_autotune_subproc_graceful_timeout_seconds.max_autotune_subproc_terminate_timeout_secondsrwrxwarningswarnfloat)r_choicerHs r4rgTuningProcessPool.targetls ||'''~~)))..$$& FLL! (;;BBDDEE NN  w '{{ MM.vh7WW  <  NN  w '  NN  w 's$*TensorMeta.from_irnodes..sA&Qz!Z00&sfake)rlayout)fallback)rArrrrr)rTr from_irnodesallrLayoutBuffer get_dtype get_devicerr&graphsizevars size_hintsget_sizer#unbacked_symint_fallback get_stride size_hint get_layoutrget_name)clsirnodesrrnoderrAs r4rTensorMeta.from_irnodess] gx ( (>E Fg!1!1!!4gF FA&AAA AAM dBII & &99&6D    "!!!''""-- 88.GG$$//!88077##--!((88.  !GsFc[URURURURUR S9$)N)rAr extra_size)rrrrArrr^s r4 to_tensorTensorMeta.to_tensors2 JJ LL;;**{{   r3r,)rz/Union[LayoutOrBuffer, Sequence[LayoutOrBuffer]]r#Union[TensorMeta, list[TensorMeta]])r torch.Tensor) r.r/r0r1rr classmethodrrr2r,r3r4rrsQ   (( ++ KD-! E! ,! ! F r3rc\rSrSrSrS SjrS SjrS SjrSS.SSjjrSS.SS jjr S r g)rVia Only handle triton template benchmark for now. The extern kernel benchmark can be done inside the same process since they usually don't cause crash. Important: Instances of this class and subclasses have to be serializable across process boundaries. Do not put CUDA Tensors in here! c^Xl[U[5(aU/nX l[T[[ 45(a0[ T5S:a[U4SjT55(deTSmTUlX@l g)Nr!c3n># UH*nSH n[TSU5[X5:Hv M" M, g7f))rArrrrrN)getattr)rrattroutput_tensor_metas r4r,BenchmarkRequest.__init__..s=/ Q.q148GA.s!P9OA++--9Osrz0Skipping op due to nonzero workspace requirementrz6InChildProcess %s: load %f, create tensor %f, bench %f)rL isEnabledForloggingDEBUGtimerrrrrrr:inforrrMr@r ) r_rrrMstart_tscreate_tensor_elapser load_elapseout bench_elapses r4rWBenchmarkRequest.benchmark sM ++GMM: yy{H  }%* **!!P9O9O!PPM 33==?M #'99;#9 yy{H !!=N NB ))+0Kyy{HmmB>> > 99;1L  HD $    +1    R S<  s?E--*FF)rrrr) rr@rrrrr Iterable[Any]rrrrrrrzCallable[[], None]rrrrzOptional[torch.Tensor]rr) r.r/r0r1rrrr rrWr2r,r3r4rVrVs%%?%@ % " %  %6"*";G" "  15 "%". "  "15)$).)  ))r3rVcB\rSrSrSrSS SjjrSS.S SjjrSrg) TestBenchmarkRequesti9zx Supports unit testing. Defined in this file so that the TuningProcess sub-process knows how to unpickle these objects. NcXlgrvalue)r_r&s r4rTestBenchmarkRequest.__init__?s r3r cJURc [S5eUR$)Nz Failed to run)r&rOrs r4rWTestBenchmarkRequest.benchmarkBs# :: O, ,zzr3r%r)r&zOptional[float]rrr!)r.r/r0r1rrrWr2r,r3r4r#r#9s5 UY*;Q r3r#c0\rSrSrSS.SSjjrSrg)GPUDeviceBenchmarkMixiniJNr c[S/UQUP55n[U5S::d SU35e[SU5S5n[U5n[U5S:Xa[[ U55nOUR 5nUR U5 [R"U5nUR5 SSS5 U$!,(df  W$=f)Nc3# UH{n[U[R5(dM$[URR 5(dMJURR cMcURR v M} g7fr)rTtorchTensorrrArYindexrtensors r4r3GPUDeviceBenchmarkMixin.do_bench..Qs^$ 9&%,,/ v}}))*  ## FMM  9s#B"BB(Br!zCan not mix devices c3# UHAn[URR5(dM)URRv MC g7fr)rrArYr1s r4rr3Zs5 +F&--,,-# ""+s (A A cuda) rrnextritercurrent_devicerAr% benchmark_gpu synchronize) r_rrrdevice_idx_set device_typer device_idxrs r4r GPUDeviceBenchmarkMixin.do_benchKs $$ 9M9=9$  >"a'P+??O)PP' +    4K@ ~ ! #d>23J)88:J  $ $Z 0++B/C  ( ( *1 1 0 s 'C Cr,r!r.r/r0r1rr2r,r3r4r+r+Js/ 15 %.   r3r+c0\rSrSrSS.SSjjrSrg)CPUDeviceBenchmarkMixinimNr c.[R"U5$r)r% benchmark_cpur s r4r CPUDeviceBenchmarkMixin.do_benchns ((,,r3r,r!r?r,r3r4rArAms/ 15 -%-. -  --r3rAc^\rSrSrSSU4SjjjrS SjrSrS SjrSrU=r $) TritonBenchmarkRequestiwc >[T U]XX45 XPlX`lXplXlXlXlXlXl gr) superr module_pathmodule_cache_key num_stages num_warpsmatrix_instr_nonkdim waves_per_eukpack workspace_arg)r_rrrrrIrJrKrLrMrNrOrP __class__s r4rTritonBenchmarkRequest.__init__zsB 9KX& 0$"$8!( *r3c^^^^ ^ ^ ^ [R"URUR5n[R SURUR5 [ X0R5Rm [UR5mST Rl 0m SSK nSURT 5R;aST S'TR R"S:XaSm OPTR R"n[%U5nUR'UR(R R*5m UR,bUR,m UUUU U U U 4SjnU$[/[ X0R5[0R2R4R6R85(a"[:R<"T /TQTPTQ70T DST 0D6$[:R<"T /TQTPTQ70T DT SS .D6$) Nz"benchmark module key: %s, path: %sFrwarmupcpuc>TRn[R"U4S[RTRS9nTR [ R:waUR5 T"/TQTPUPTQ70TDTSS.D6 g)N)r!rrATstream benchmark_run) rr. empty_strideduint8rA zero_moder$ UNINITIALIZEDzero_) workspace_sizeworkspace_tensorrrr run_methodrY warmup_argrPs r4run_with_workspace>TritonBenchmarkRequest.make_run_fn..run_with_workspaces!.!4!4#(#6#6#%++(// $ !**.?.M.MM$**,"!%  !  ""&r3rYTrX)rload_by_key_pathrJrIrLrMrrrunrr__self__with_bandwidth_infoinspect signature parametersrArYrget_raw_streamrr0rPrTr. _inductorruntimetriton_heuristicsDebugAutotuner functoolspartial) r_rrmodrjr<rrdrrbrYrcrPs `` @@@@@r4r"TritonBenchmarkRequest.make_run_fns**4+@+@$BRBRS 0  ! !    S"2"2377 $//* 27 /  w((4?? ?#(Jx    $ $ -F'..33K7 D %44''..44F    ) ..M  2& %  C)) * OO # # 5 5 D D  $$      $$    " r3c[R"URUR5n[ XR 5R 5 gr)rrfrJrIrr precompile)r_rts r4rw!TritonBenchmarkRequest.precompiles7**4+@+@$BRBRS%%&113r3cZSUR<SUR<SUR<3$)Nself.kernel_name=z, self.module_path=z, self.module_cache_key=)rrIrJr^s r4__str__TritonBenchmarkRequest.__str__s2#$""$$8t'7'7&99RD[TU]XX45 XPlSUlSUlSUlSUlSUlSUl[R"URS5uUlUlg)NrFso) rHr source_coder` workspaceDLL_workspace_size_updatedhash_key source_filerwriter_rrrrrrQs r4rCUDABenchmarkRequest.__init__sk 9KX&#$15)-',$ "*7*=*=d>N>NPT*U' t'r3c[RSU5 [R"URS5 [RSU5 g)NPrecompiling %srDone precompiling %s)rLrMrcompilerr^s r4rwCUDABenchmarkRequest.precompile s< .5d..53T:r3c HUR5 UR5 [U5U/-Vs/sHn[UR 55PM nn[ R SURURURURUUR5 [[RR5R5n[!URUR5n[S5nUR"S:af[R$"UR"S-S-[R&UR(S9Ul[UR*R 55n[,R."U/UQURQSPUPUP76$s snf)Nzqmake_run_fn: self.kernel_name=%s, self.source_file=%s, self.hash_key=%s, self.DLL=%s, args=%s, self.extra_args=%srrW)ensure_dll_loadedupdate_workspace_sizerr data_ptrrLrMrrrrrr.r5current_stream cuda_streamrr`zerosfloat64rArrrrs)r_rrr2rh stream_ptrrb workspace_ptrs r4r CUDABenchmarkRequest.make_run_fns|   ""$}-? ? V__& '?           MM HH  OO ejj779EEF TXXt'7'78      ""[[$$q(Q.mm$++DN %T^^%<%<%>?M    __         3 s#Fc  UR(agUR5 [URVs1sHoRiM sn5n[ US-5Vs/sHn[ S5PM nn[ [RR5R5n[URUR5n[5nU"/UQURQ[!U5PSPUP76 [RR#5 UR$Ul[(R+SUR&URUR,UR.URUUR5 SUlgs snfs snf)Nr!zupdate_workspace_size called: new workspace size=%d, self.kernel_name=%s, self.source_file=%s, self.hash_key=%s, self.DLL=%s, args=%s, self.extra_args=%sT)rrrrrrr r.r5rrrrrrrrr:r&r`rLrMrr)r_metaunique_input_count_rhrrbc_workspace_sizes r4r*CUDABenchmarkRequest.update_workspace_size;s^  ' '    #'#9#9 :#94YY#9 : )..@1.D(EF(E1(EFejj779EEF TXXt'7'78 #:  __            .44 h          MM HH  OO (,$; ;Fs F"F cURc5[R"URS5uUlUlUlgg)Nr)rrloadrrrr^s r4r&CUDABenchmarkRequest.ensure_dll_loaded_s: 88 8E8J8J  $9 5DHdmT%5 r3c`URbURR5 SUlgr)rcloserr^s r4r #CUDABenchmarkRequest.cleanup_run_fnes! 88  HHNN r3cZSUR<SUR<SUR<3$)Nrzz, self.source_file=z, self.hash_key=)rrrr^s r4r{CUDABenchmarkRequest.__str__js0#$""$$8t'7'7&99JDMM;KLLr3)rrrrrrr` rr@rrrrrrrr@rrr rr}) r.r/r0r1rrwrrrr r{r2r~rs@r4rrsVV?V@ V " V  V V$;% *% ;G% % N",H  MMr3rct^\rSrSrSU4SjjrSrS SjrS SjrS SjrSr U=r $) CppBenchmarkRequestinc`>[TU]XX45 XPl[U5UlSUlgr)rHrrrrrrs r4rCppBenchmarkRequest.__init__rs. 9KX& - 6:r3c[RSU5 [R"URSS9 [RSU5 g)NrrUr<r)rLrMrrrr^s r4rwCppBenchmarkRequest.precompiles< .5$**>3T:r3ch[R"URSS9Ul[ U5U/-Vs/sHo3R 5PM nn[ RSURURUUR5 [URUR5n[SUR55(de[R/[U5[[ UR55--Ul[ R""U/UQURQ76$s snf)NrUrzJmake_run_fn: self.kernel_name=%s, self.DLL=%s, args=%s, self.extra_args=%sc3V# UHn[U[R5v M! g7fr)rTctypes c_ulonglong)rargs r4r2CppBenchmarkRequest.make_run_fn..s R/3:c6#5#566/s'))rrrrrrrLrMrrrrrrrargtypesrrrs)r_rrr2rhrbs r4rCppBenchmarkRequest.make_run_fns $$T%5%55I04]0C}o0UV0Uf!0UV X    HH  OO  TXXt'7'78 R$//RRRRR%112 ID12 2      __  !WsD/cURb8[URS5(aURR5 ggg)Nr)rhasattrrr^s r4r "CppBenchmarkRequest.cleanup_run_fns< 88  txx)) * r3c"SUR<3$)Nrz)rr^s r4r{CppBenchmarkRequest.__str__s#$""$%%r3)rrrrr rr}) r.r/r0r1rrwrr r{r2r~rs@r4rrnst ; ;? ;@ ; " ;  ;  ;; * ;G  6!&&r3rc,[RU5$)zG Do benchmarking in a subprocess and return the perf number (latency). ) tuning_poolrW)rs r4benchmark_in_sub_processrs   ))r3)rArGr)_ __future__r contextlibr dataclassesrrrr=rwrrcollections.abcrrconcurrent.futuresrrrr r typingr r r rrr.torch._inductor.async_compilertorch._dynamo.device_interfacertorch._dynamo.testingrtorch._inductorrtorch._inductor.codecacherrrrrtorch._inductor.utilsrrtorch._loggingrtorch.utils._ordered_setrmultiprocessing.processrmultiprocessing.queuesrtypesr torch._inductor.select_algorithmr codegen.commonr"rr#r$runtime.benchmarkingr% virtualizedr&r'rr.rL getLoggerrr*r7rOr:contextmanagerrC dataclassrErrrrLayoutOrBufferrrVr#r+rArFrrrrrr,r3r4rs"  .122@@ $!C.7,/3, E,--."8\: !       77( yyyx ~~~B ! ryy"))+, 3 3 3 l ]]]@+"  F--uU-uUp 79O  79O tM24DtMn<&13C<&~* '*&*r3