U *Ç-eÝã@s´UddlZddlZddlmZddlmZddlmZddlm Z m Z Gdd„deƒZ Gd d „d eƒZ Gd d „d eƒZ Gd d„deƒZdae e e ejed<ejdœdd„ZdS)éNé)Úcomm)ÚFunction©Ú_get_device_index)ÚListÚOptionalc@s$eZdZedd„ƒZedd„ƒZdS)Ú BroadcastcGs¾tdd„|Dƒƒstdƒ‚dd„|Dƒ}||_t|ƒdkr@tƒSt|ƒ|_|d ¡|_t  ||j¡}g}t |j dd…ƒD]$\}}|s||D]}|  ||¡qŒq||j |Žtdd„|DƒƒS) Ncss|]}|jjdkVqdS©ÚcpuN©ÚdeviceÚtype©Ú.0Úi©rú]/var/www/html/Darija-Ai-Train/env/lib/python3.8/site-packages/torch/nn/parallel/_functions.pyÚ sz$Broadcast.forward..z2Broadcast function not implemented for CPU tensorscSsg|]}t|dƒ‘qS©Tr©rÚxrrrÚ sz%Broadcast.forward..rrcSsg|]}|D]}|‘q qSrr)rZtensorsÚtrrrrs)ÚallÚAssertionErrorÚ target_gpusÚlenÚtupleÚ num_inputsÚ get_deviceÚ input_devicerZbroadcast_coalescedÚ enumerateZneeds_input_gradÚappendZmark_non_differentiable)ÚctxrÚinputsÚoutputsZnon_differentiablesÚidxZinput_requires_gradÚoutputrrrÚforward s"ÿ   zBroadcast.forwardcGsdtj|j|jf|žŽS)N©N)ÚReduceAddCoalescedÚapplyr!r©r$Z grad_outputsrrrÚbackward szBroadcast.backwardN©Ú__name__Ú __module__Ú __qualname__Ú staticmethodr)r.rrrrr s r c@s$eZdZedd„ƒZedd„ƒZdS)r+csL‡fdd„tdtˆƒˆƒDƒ|_‡‡fdd„tdtˆƒˆƒDƒ}t ||¡S)Ncsg|]}ˆ| ¡‘qSr©r r)Úgradsrrr)sz.ReduceAddCoalesced.forward..rcsg|]}ˆ||ˆ…‘qSrrr©r5rrrr+sÿ)ÚrangerrrZreduce_add_coalesced)r$Ú destinationrr5Zgrads_rr6rr)'s   ÿzReduceAddCoalesced.forwardcGsdtj|jf|žŽS)N©NN)r r,rr-rrrr./szReduceAddCoalesced.backwardNr/rrrrr+%s r+c@s$eZdZedd„ƒZedd„ƒZdS)ÚGathercsÂtdd„|Dƒƒstdƒ‚|dkr*dˆ_nt|dƒ}|ˆ_|ˆ_tdd„|Dƒƒˆ_tdd„|Dƒƒr’|dkr’td d„|Dƒƒ}t d ¡dˆ_ nd ˆ_ t‡fd d„|Dƒƒˆ_ t   |ˆjˆj¡S) Ncss|]}|jjdkVqdSr r rrrrr8sz!Gather.forward..z/Gather function not implemented for CPU tensorsr Tcss|]}| ¡VqdSr*r4rrrrrAscss|]}| ¡dkVqdS©rN)Údim©rrrrrrBsrcss|]}| d¡VqdS)rN)Úviewr=rrrrCszvWas asked to gather along dimension 0, but all input tensors were scalars; will instead unsqueeze and return a vector.Fc3s|]}| ˆj¡VqdSr*)Úsizer<r©r$rrrJs) rrÚ target_devicerr<rÚ input_gpusÚwarningsÚwarnÚunsqueezed_scalarÚ input_sizesrÚgather)r$rAr<r%rr@rr)6s ÿ  zGather.forwardcCs6t |j|j|j|¡}|jr.tdd„|Dƒƒ}d|S)Ncss|]}|dVqdSr;r)rÚgrrrrQsz"Gather.backward..r9)ÚScatterr,rBrFr<rEr)r$Ú grad_outputZscattered_gradsrrrr.MszGather.backwardNr/rrrrr:4s r:c@s$eZdZedd„ƒZedd„ƒZdS)rIc CsÄdd„|Dƒ}||_|jjdkr(| ¡nd|_d}tj ¡rT|jdkrTdd„|Dƒ}t  ||||j|¡}|dk rÀt |ƒD]F\}}tj ||¡(tj  ¡} |   ||¡|  | ¡W5QRXqx|S)NcSsg|]}t|dƒ‘qSrrrrrrrYsz#Scatter.forward..r éÿÿÿÿcSsg|]}tt d|¡ƒ‘qS)Úcuda)Ú _get_streamÚtorchr )rr rrrr_s)r<r rr r!rNrLZ is_availablerZscatterr"Zcurrent_streamZ wait_streamZ record_stream) r$rZ chunk_sizesr<ÚinputÚstreamsr&rr(Z main_streamrrrr)Ws zScatter.forwardcGsdddtj|j|jf|žŽfSr*)r:r,r!r<)r$rJrrrr.jszScatter.backwardNr/rrrrrIUs rIÚ_streams)r cCsh|jdkrdStt|jdƒ}|dkr(dStdkr>dg| ¡at|jdkr^| |j¡t|j<t|jS)zBGets a background stream for copying between CPU and target devicer N)rÚgetattrrNrQZ device_countÚindexÚStream)r Z device_modrrrrMrs rM)rCrNÚrZtorch.autogradrZ torch._utilsrÚtypingrrr r+r:rIrQrTÚ__annotations__r rMrrrrÚs   !