o wZОhMу@s╝UddlZddlmZddlmZddlZddlmZddlm Z ddl mZGddДde ГZGd d Дd e ГZ GddДde ГZGd dДde ГZdaeeeejed<dejfddДZdS)щN)┌chain)┌Optionalй┌_get_device_index)┌Function)┌commc@є$eZdZeddДГZeddДГZdS)┌ Broadcastcs╕tddД|DГГs JdГВddД|DГ}||_t|ГdkrdSt|Г|_|dаб|_tа||jб}g}t|j ddЕГD]\Й}|sO|а Зfd dД|DГбq=|j|Оtt а|бГS) NcsєБ|] }|jjdkVqdSй┌cpuNй┌device┌typeй┌.0┌iйr·K/var/www/auris/lib/python3.10/site-packages/torch/nn/parallel/_functions.py┌ єА z$Broadcast.forward..z2Broadcast function not implemented for CPU tensorscSєg|]}t|dГСqSйTrйr┌xrrr┌ єz%Broadcast.forward..rrщc3sБ|]}|ИVqdSйNr)r┌outputй┌idxrrrєА)┌all┌target_gpus┌len┌ num_inputs┌ get_device┌input_devicerZbroadcast_coalesced┌ enumerateZneeds_input_grad┌extendZmark_non_differentiable┌tupler┌ from_iterable)┌ctxr$┌inputs┌outputsZnon_differentiablesZinput_requires_gradrr r┌forwards& ■ А zBroadcast.forwardcGsdtj|j|jg|вRОS)Nr)┌ReduceAddCoalesced┌applyr(r&йr-Zgrad_outputsrrr┌backwards zBroadcast.backwardNй┌__name__┌ __module__┌__qualname__┌staticmethodr0r4rrrrr s r c@r)r1csLЗfddДtdtИГИГDГ|_ЗЗfddДtdtИГИГDГ}tа||бS)Ncsg|]}И|абСqSrйr'r)┌gradsrrr)s z.ReduceAddCoalesced.forward..rcsg|] }И||ИЕСqSrrrйr;r&rrr-s)┌ranger%r$rZreduce_add_coalesced)r-┌destinationr&r;Zgrads_rr<rr0's zReduceAddCoalesced.forwardcGsdtj|jg|вRОS)NйNN)r r2r$r3rrrr40s¤zReduceAddCoalesced.backwardNr5rrrrr1&s r1c@r)┌Gathercs┬tddД|DГГs JdГВ|dkrdИ_nt|dГ}|И_|И_tddД|DГГИ_tddД|DГГrI|dkrItd dД|DГГ}tаd бdИ_ndИ_tЗfddД|DГГИ_ t а|ИjИjбS) Ncsr rr rrrrr;rz!Gather.forward..z/Gather function not implemented for CPU tensorsrTcssБ|]}|абVqdSrr:rrrrrDr"cssБ|] }|абdkVqdSйrN)┌dimйr┌trrrrEєАrcssБ|]}|аdбVqdS)rN)┌viewrCrrrrFsАzvWas asked to gather along dimension 0, but all input tensors were scalars; will instead unsqueeze and return a vector.Fc3sБ|] }|аИjбVqdSr)┌sizerBrйr-rrrOrE)r#┌ target_devicerrBr+┌ input_gpus┌warnings┌warn┌unsqueezed_scalar┌input_sizesr┌gather)r-rIrBr.rrHrr09s( ■ zGather.forwardcCs6tа|j|j|j|б}|jrtddД|DГГ}d|S)NcssБ|]}|dVqdSrAr)r┌grrrrXr"z"Gather.backward..r?)┌Scatterr2rJrNrBrMr+)r-┌grad_outputZscattered_gradsrrrr4Rs zGather.backwardNr5rrrrr@8s r@c@r)rQc Cs╪ddД|DГ}||_|jjdkr|абnd|_d}tjабr*|jdkr*ddД|DГ}tа ||||j|б}|durjt |ГD]-\}}tjа||бПtjаб} | а||б|а | бWdГn1sdwYq<|S)NcSrrrrrrrr_rz#Scatter.forward..rщ cSsg|] }ttаd|бГСqS)┌cuda)┌_get_stream┌torchr)rrrrrres )rBrrr'r(rVrTZis_availablerZscatterr)Zcurrent_streamZwait_streamZ record_stream) r-r$Zchunk_sizesrB┌input┌streamsr/rrZmain_streamrrrr0]s$ ¤АzScatter.forwardcGs dddtj|j|jg|вRОfSr)r@r2r(rB)r-rRrrrr4rs zScatter.backwardNr5rrrrrQ\s rQ┌_streamsrcCsh|jdkrdStt|jdГ}|durdStdurdg|абat|jdur/|а|jбt|j<t|jS)zBGet a background stream for copying between CPU and target device.rN)r┌getattrrVrYZdevice_count┌index┌Stream)rZ device_modrrrrU{s rU)rK┌ itertoolsr┌typingrrVZtorch._utilsrZtorch.autogradrZtorch.nn.parallelrr r1r@rQrY┌listr\┌__annotations__rrUrrrr┌s $