U Йd8у@s┤UddlZddlZddlmZddlmZddlmZddlm Z m Z GddДdeГZGd d Дd eГZGddДdeГZ Gd dДdeГZdae e e ejjed<edЬddДZdS)щNщ)┌comm)┌Functionй┌_get_device_index)┌List┌Optionalc@s$eZdZeddДГZeddДГZdS)┌ BroadcastcGs╛tddД|DГГstdГВddД|DГ}||_t|Гdkr@tГSt|Г|_|dаб|_tа ||jб}g}t |jddЕГD]$\}}|s||D]}|а||бqМq||j |ОtddД|DГГS) Ncss|]}|jjdkVqdSй┌cpuNй┌device┌typeй┌.0┌iйr·@/tmp/pip-unpacked-wheel-ua33x9lu/torch/nn/parallel/_functions.py┌ sz$Broadcast.forward..z2Broadcast function not implemented for CPU tensorscSsg|]}t|dГСqSйTrйr┌xrrr┌ sz%Broadcast.forward..rrcSsg|]}|D]}|СqqSrr)rZtensors┌trrrrs)┌all┌AssertionError┌target_gpus┌len┌tuple┌ num_inputs┌ get_device┌input_devicerZbroadcast_coalesced┌ enumerateZneeds_input_grad┌appendZmark_non_differentiable)┌ctxr┌inputs┌outputsZnon_differentiables┌idxZinput_requires_grad┌outputrrr┌forwards" zBroadcast.forwardcGsdtj|j|jf|ЮОS)NйN)┌ReduceAddCoalesced┌applyr!rйr$Zgrad_outputsrrr┌backward szBroadcast.backwardNй┌__name__┌ __module__┌__qualname__┌staticmethodr)r.rrrrr s r c@s$eZdZeddДГZeddДГZdS)r+csLЗfddДtdtИГИГDГ|_ЗЗfddДtdtИГИГDГ}tа||бS)Ncsg|]}И|абСqSrйr r)┌gradsrrr)sz.ReduceAddCoalesced.forward..rcsg|]}И||ИЕСqSrrrйr5rrrr+s )┌rangerrrZreduce_add_coalesced)r$Zdestinationrr5Zgrads_rr6rr)'s zReduceAddCoalesced.forwardcGsdtj|jf|ЮОS)NйNN)r r,rr-rrrr./szReduceAddCoalesced.backwardNr/rrrrr+%s r+c@s$eZdZeddДГZeddДГZdS)┌Gathercs┬tddД|DГГstdГВ|dkr*dИ_nt|dГ}|И_|И_tddД|DГГИ_tddД|DГГrТ|dkrТtd dД|DГГ}tаd бdИ_ ndИ_ tЗfddД|DГГИ_ tа|ИjИjбS) Ncss|]}|jjdkVqdSr rrrrrr8sz!Gather.forward..z/Gather function not implemented for CPU tensorsrTcss|]}|абVqdSr*r4rrrrrAscss|]}|абdkVqdSйrN)┌dimйrrrrrrBsrcss|]}|аdбVqdS)rN)┌viewr<rrrrCszvWas asked to gather along dimension 0, but all input tensors were scalars; will instead unsqueeze and return a vector.Fc3s|]}|аИjбVqdSr*)┌sizer;rйr$rrrJs) rr┌ target_devicerr;r┌ input_gpus┌warnings┌warn┌unsqueezed_scalar┌input_sizesrZgather)r$r@r;r%rr?rr)6s zGather.forwardcCs6tа|j|j|j|б}|jr.tddД|DГГ}d|S)Ncss|]}|dVqdSr:r)r┌grrrrQsz"Gather.backward..r8)┌Scatterr,rArEr;rDr)r$┌grad_outputZscattered_gradsrrrr.MszGather.backwardNr/rrrrr94s r9c@s$eZdZeddДГZeddДГZdS)rGc Cs─ddД|DГ}||_|jjdkr(|абnd|_d}tjабrT|jdkrTddД|DГ}tа ||||j|б}|dk r└t |ГD]F\}}tjа||бП(tjаб} | а||б|а | бW5QRXqx|S)NcSsg|]}t|dГСqSrrrrrrrYsz#Scatter.forward..rщ cSsg|]}t|ГСqSr)┌_get_stream)rr rrrr_s)r;r rr r!┌torch┌cudaZis_availablerZscatterr"Zcurrent_streamZwait_streamZ record_stream) r$rZchunk_sizesr;┌inputZstreamsr&rr(Zmain_streamrrrr)Ws zScatter.forwardcGsdddtj|j|jf|ЮОfSr*)r9r,r!r;)r$rHrrrr.jszScatter.backwardNr/rrrrrGUs rG┌_streamsйr cCsH|dkrdStdkr$dgtjабat|dkr@tjа|бt|<t|S)z8Gets a background stream for copying between CPU and GPUrIN)rNrKrLZdevice_count┌StreamrOrrrrJssrJ)rBrK┌rZtorch.autogradrZtorch._utilsr┌typingrrr r+r9rGrNrLrP┌__annotations__┌intrJrrrr┌s!