o 0ёhMу@s╝UddlZddlmZddlmZddlZddlmZddlm Z ddl mZGddДde ГZGd d Дd e ГZ GddДde ГZGd dДde ГZdaeeeejed<dejfddДZdS)щN)┌chain)┌Optionalй┌_get_device_index)┌Function)┌commc@є$eZdZeddДГZeddДГZdS)┌ Broadcastcs╕tddД|DГГs JdГВddД|DГ}||_t|ГdkrdSt|Г|_|dаб|_tа||jб}g}t|j ddЕГD]\Й}|sO|а Зfd dД|DГбq=|j|Оtt а|бГS) NcsєБ|] }|jjdkVqdSй┌cpuNй┌device┌typeй┌.0┌iйr·P/var/www/vscode/kcb/lib/python3.10/site-packages/torch/nn/parallel/_functions.py┌ єА z$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_devicer┌broadcast_coalesced┌ enumerate┌needs_input_grad┌extend┌mark_non_differentiable┌tupler┌ from_iterable)┌ctxr$┌inputs┌outputs┌non_differentiables┌input_requires_gradrr r┌forwards& ■ А zBroadcast.forwardcGsdtj|j|jg|вRОS)Nr)┌ReduceAddCoalesced┌applyr(r&йr0┌grad_outputsrrr┌backwards zBroadcast.backwardNй┌__name__┌ __module__┌__qualname__┌staticmethodr5r:rrrrr s r c@r)r6csLЗfddДtdtИГИГDГ|_ЗЗfddДtdtИГИГDГ}tа||бS)Ncsg|]}И|абСqSrйr'r)┌gradsrrr)s z.ReduceAddCoalesced.forward..rcsg|] }И||ИЕСqSrrrйrAr&rrr-s)┌ranger%r$r┌reduce_add_coalesced)r0┌destinationr&rA┌grads_rrBrr5's zReduceAddCoalesced.forwardcGsdtj|jg|вRОS)NйNN)r r7r$r8rrrr:0s¤zReduceAddCoalesced.backwardNr;rrrrr6&s r6c@r)┌Gathercs┬tddД|DГГs JdГВ|dkrdИ_nt|dГ}|И_|И_tddД|DГГИ_tddД|DГГrI|dkrItd dД|DГГ}tаd бdИ_ndИ_tЗfddД|DГГИ_ t а|ИjИjбS) Ncsr rr rrrrr;rz!Gather.forward..z/Gather function not implemented for CPU tensorsrTcssБ|]}|абVqdSrr@rrrrrDr"cssБ|] }|абdkVqdSйrN)┌dimйr┌trrrrEєАrcssБ|]}|аdбVqdS)rN)┌viewrKrrrrFsАzvWas asked to gather along dimension 0, but all input tensors were scalars; will instead unsqueeze and return a vector.Fc3sБ|] }|аИjбVqdSr)┌sizerJrйr0rrrOrM)r#┌ target_devicerrJr.┌ input_gpus┌warnings┌warn┌unsqueezed_scalar┌input_sizesr┌gather)r0rQrJr1rrPrr59s( ■ zGather.forwardcCs6tа|j|j|j|б}|jrtddД|DГГ}d|S)NcssБ|]}|dVqdSrIr)r┌grrrrXr"z"Gather.backward..rG)┌Scatterr7rRrVrJrUr.)r0┌grad_output┌scattered_gradsrrrr:Rs zGather.backwardNr;rrrrrH8s rHc@r)rYc Cs╪ddД|DГ}||_|jjdkr|абnd|_d}tjабr*|jdkr*ddД|DГ}tа ||||j|б}|durjt |ГD]-\}}tjа||бПtjаб} | а||б|а | бWdГn1sdwYq<|S)NcSrrrrrrrr_rz#Scatter.forward..rщ cSsg|] }ttаd|бГСqS)┌cuda)┌_get_stream┌torchr)rrrrrres )rJrrr'r(r_r]┌is_availabler┌scatterr*┌current_stream┌wait_stream┌ record_stream) r0r$┌chunk_sizesrJ┌input┌streamsr2rr┌main_streamrrrr5]s$ ¤АzScatter.forwardcGs dddtj|j|jg|вRОfSr)rHr7r(rJ)r0rZrrrr:rs zScatter.backwardNr;rrrrrY\s rY┌_streamsrcCsh|jdkrdStt|jdГ}|durdStdurdg|абat|jdur/|а|jбt|j<t|jS)zBGet a background stream for copying between CPU and target device.rN)r┌getattrr_ri┌device_count┌index┌Stream)r┌ device_modrrrr^{s r^)rS┌ itertoolsr┌typingrr_┌torch._utilsr┌torch.autogradr┌torch.nn.parallelrr r6rHrYri┌listrm┌__annotations__rr^rrrr┌s $