U Kc)@s6UddlZddlmZmZmZmZmZddlZddlZddlm Z ej j Z iZ eej jefed<ej jZejjZddZiZ ddZdd d Zeejje e d d dZeejje ee e fd ddZe e eeedddZeeje e eee e ee ee eeeee ee ee fd ddZeedddZeeje e ee ee ee ee ee eeeeee ee ee fd ddZeejjjjddeejjj jeejjj!jeejjj"jeejjj#jeejjjjeejjjjeejjjjeejjj$jdS) N)CallableDictListOptionalTuple)Tensordecomposition_table_for_jvpcsfdd}|S)Ncs,zt|WStk r&|YSXdSN)register_decomposition Exception)fopH/tmp/pip-unpacked-wheel-gikjz4vx/torch/_decomp/decompositions_for_jvp.py decorator&sz/maybe_register_decomposition..decoratorr)rrrr rmaybe_register_decomposition%s rcCs t|tdS)N)registry)r r)fnrrrregister_decomposition_for_jvp5srFcCs|tkrt}n|tkrt}ntd|||}|rntj|}t|}dd}||}tj|j j }ntj |j }tj ||dS)Nz!could not find decomposition for cSsFdd|jD}dd|jD}dd|dd|dS)NcSsg|] }|qSrr.0Z param_strrrr KszQ_register_jit_decomposition_for_jvp..get_function_def..cSsg|] }|qSrrrrrrrLszdef wrapped_decomp(z, z): return decomp_fn(z) ) parametersvalueskeysjoin)sigZ param_defZ param_userrrget_function_defJsz=_register_jit_decomposition_for_jvp..get_function_def) rdecomposition_table RuntimeErrortorchZjitignoreinspect signatureZCompilationUnitZwrapped_decompgraphscriptZ_register_decomposition)decomp use_pythonZdecomposition_table_usedZ decomp_fnrrZf_strr%rrr#_register_jit_decomposition_for_jvp9s  r))selfreturncCstt|Sr )r!sumZdiag)r*rrrtraceZsr-cCsLt|d|}tt| }|jr6|d}n|}|t||fS)Nr)r)r!Zminimum new_zerosexpabsZis_cudalog1p)r*minzbufferrrrlog_sigmoid_forward_s  r5)inputrstdinner_dim_indiceskeepdimcCsXtj|||d}tj||d|d}td|d|}|}dt||}||fS)N)dimr9F)r:Zunbiasedr9)r!meanvarpowdetachsqrt)r6r7r8r9r=r>epsrrrrecompute_mean_varjs rC) grad_outr6normalized_shaper=r7weightbias output_maskr+cCs|j}|} | t|} || d} |d| } tt| | } ttd| }d}| D] }||9}qVd}| D] }||9}ql|dks|dkr||||| d||| dfSt||| dd\}}|||}|dk r||}n|}||}t|| d}t ||}t|| d}t ||}|||}|drL|||}n t |}|dr|dk rt|dkrt|||d}n||}n |dk rt |}n t d}|dr|dk rt|dkrt||d}n| }n |dk rt |}n t d}|||fS)Nrr;Tr9Frr<) shaper:lenlistranger.rCr!r,mul zeros_likezerosclone)rDr6rEr=r7rFrGrH input_shapeZ input_ndimaxisZ inner_dimsZ outer_dimsr8Zouter_dim_indicesNiMZmean_Zrstd_Zx_hatZ grad_x_hatabZc1c2c3innerZd_inputZd_weightZd_biasrrrnative_layer_norm_backwardxsf                     r\)xcCsd}|D] }||9}q|S)Nr;r)r]rrUrrrprods r_) rDr6rF running_mean running_var save_mean save_invstdtrainrBrHr+c Cs|j} |} | dkstdd} t| | | } |}|}|r|dk rN|dk sVtddgttd|}|dk sztt|||dd\}}n&|dk r|dk st|}t||}|dk r|dk stdg| }| | || <g}t| D]}|| kr| |qt ||}d| }t ||}t ||||}t |||}t t |||||}|dkrt ||d}nt |||}|r|||}||||}n||}| dr||}n |dk rt |}n td }| dr|}n t |}|||fS) Nr<z$rank of the input must be at least 2r;z7when train=True, save_mean and save_invstd are requiredrFrIg?r)rJr:AssertionErrorr_rLrMrCr!ZrsqrtappendZreshaper,rNrOrP)rDr6rFr`rarbrcrdrBrHrRZ input_rankrSZ num_featuresr=ZinvstdZreduciton_dimsZbroadcast_maskZreduction_axesrUZnormZgrad_output_sumZdot_pZ grad_meanZ proj_scaleZ grad_scaleZprojZ grad_inputZ grad_weightZ grad_biasrrrnative_batch_norm_backwardsj               rgT)r()F)%r#typingrrrrrr!Z torch._decompr_decomprrZ_opsZ OpOverload__annotations__r opsZatenrrr)r-defaultr5intboolrCr\r_rgfloatZnll_loss_backwardZnll_loss2d_backwardZ_log_softmax_backward_dataZ_softmax_backward_dataZcudnn_batch_norm_backwardrrrrsp   !    LQ