iRUddlZddlmZddlZddlmZddlmZddlm Z GddeZ Gdd eZ Gd d eZ Gd d eZ daeej dzdzed<dej$fdZy)N)chain)_get_device_index)Function)commc,eZdZedZedZy) Broadcastctd|DsJd|Dcgc]}t|d}}||_t|dk(ryt||_|dj |_tj||j}g}t|jddD]"\}|r |jfd|D$|j|ttj|Scc}w)Nc3NK|]}|jjdk7ywcpuNdevicetype.0is V/var/www/html/engine/venv/lib/python3.12/site-packages/torch/nn/parallel/_functions.py z$Broadcast.forward.. :a188==E):#%z2Broadcast function not implemented for CPU tensorsTrc3(K|] }| ywNr)routputidxs rrz$Broadcast.forward..s*M66#;*Ms)allr target_gpuslen num_inputs get_device input_devicerbroadcast_coalesced enumerateneeds_input_gradextendmark_non_differentiabletupler from_iterable)ctxrinputsxoutputsnon_differentiablesinput_requires_gradrs @rforwardzBroadcast.forward s:6:: @ :c^dtj|j|jg|zS)Nr)ReduceAddCoalescedapplyr"r r* grad_outputss rbackwardzBroadcast.backwards4+11   cnn /;   N__name__ __module__ __qualname__ staticmethodr0r6rr7rrr s(33$  r7rc,eZdZedZedZy)r2ctdt||Dcgc]}||jc}|_tdt||Dcgc] }||||z }}t j ||Scc}wcc}w)Nr)rangerr!rrreduce_add_coalesced)r* destinationr gradsrgrads_s rr0zReduceAddCoalesced.forward&s,1CJ +K &'E!H   ! 6;1c%j*5UV%A N+VV((==  Ws A:A?cHdtj|jg|zS)NNN)rr3rr4s rr6zReduceAddCoalesced.backward/s(  OOCOO ;l ;< ><.:rrz/Gather function not implemented for CPU tensorsr Tc3<K|]}|jywr)r!rs rrz!Gather.forward..Cs>!q||~>sc3BK|]}|jdk(ywrN)dimrts rrz!Gather.forward..Ds,quuw!|,src3>K|]}|jdyw)rN)viewrNs rrz!Gather.forward..Es5166!95szvWas asked to gather along dimension 0, but all input tensors were scalars; will instead unsqueeze and return a vector.) stacklevelFc3TK|]}|jj!ywr)sizerM)rrr*s rrz!Gather.forward..Os@Asww@s%() r target_devicerrMr( input_gpuswarningswarnunsqueezed_scalar input_sizesrgather)r*rVrMr+s` rr0zGather.forward8s:6:: = : E ! %C -mTBM -C >v>> ,V, ,5f55F MM'   %)C !$)C !@@@{{6377C,=,=>>r7ctj|j|j|j|}|j rt d|D}d|zS)Nc3&K|] }|d ywrLr)rgs rrz"Gather.backward..Xs#BQAaD#BsrE)Scatterr3rWr[rMrZr()r* grad_outputscattered_gradss rr6zGather.backwardRsK!-- NNCOOSWWk   ##B/#BBOo--r7Nr8rr7rrGrG7s(??2..r7rGc,eZdZedZedZy)r`c|Dcgc]}t|d}}||_|jjdk7r|j nd|_d}t jjr:|j dk(r+|Dcgc] }tt j|"}}tj||||j|}|t|D]s\} } t jj|| 5t jj} | j|| | j!| dddu|Scc}wcc}w#1swYxYw)NTr )rrMrrr!r"torch accelerator is_available _get_streamrscatterr$ device_indexcurrent_stream wait_stream record_stream) r*r chunk_sizesrMinputr,streamsrr-rr main_streams rr0zScatter.forward]s?;FGa(D1G G161B1Be1K5++-QS    ) ) +0@0@B0FGRSV{5<<#78SGS,,uk;Q  &w/ 6 6&&33KNC6"'"3"3"B"B"DK++GAJ7((566 6 H T 66sE%EAEE& c`dddtj|j|jg|fSr)rGr3r"rM)r*ras rr6zScatter.backwardps+T4c.>.>!V+!VVVr7Nr8rr7rr`r`\s*$WWr7r`_streamsrc|jdk(stjjsytjj j|jk(sJt "dgtjj zat |j0tj|jt |j<t |jS)zBGet a background stream for copying between CPU and target device.r N) rrfrgrhcurrent_acceleratorrt device_countindexStream)rs rririys{{e5#4#4#A#A#C    0 0 2 7 76;; FF F6E--::<< %!&fll!; FLL !!r7)rX itertoolsrrf torch._utilsrtorch.autogradrtorch.nn.parallelrrr2rGr`rtlistry__annotations__rrirr7rrsy *#"  6<<$".X".JWhW4.2$u||d" #d *1 " "r7