h SSKrSSKJr /SQrSrSr"SS\5rS S\RS \S \S \R4S jjr g)N)UniformQuantizationObserverBase)create_block_sparse_tensorcreate_semi_structured_tensorPerChannelNormObserver mask_creatorc<US::aUS:dS5e[R"[R"X-X-4SU- US95n[R"XRSS9n[R"XRSS9nUR U5R 5R 5$)N?gz*sparsity should be a value between 0 and 1)dtyperdim)torch bernoullifullrepeat_interleaveto contiguouscuda)MN blocksizesparsityr As P/home/james-whalen/.local/lib/python3.13/site-packages/torchao/sparsity/utils.pyrrs s?x34 .  ANAN3Q\O A !4A !4A 44; ! ! # ( ( **c[R"SSX-S-45R5n[RRR USS9R X5R5R[R5nU[R"X5R5U--nURU5$)z This function returns a 1:2 sparse matrix of size (r, c). Note that this means this matrix will also be 2:4 and 4:8 sparse as well. r) num_classes) rrandintrnn functionalone_hotreshaperrint32rand)rcr choice_indicesmask sparse_weights rrrs]]1a!%1*7<<>N ##N#B   EKK  EJJq,113d:;M   E ""rc<^\rSrSrSrSU4SjjrSrSrSrU=r $)r3zX A custom observer that computes the L2 norm of each channel and stores it in a buffer. c >[TU]"S[R[RSSS[R "[R 5RS.UD6 SUlURS[R"/55 g)NF)r qscheme reduce_range quant_min quant_maxepsr norm) super__init__rquint8per_channel_affinefinfofloat32r2averaging_constantregister_buffertensor)selfkwargs __class__s rr6PerChannelNormObserver.__init__8sq  ,,,, EMM*..   #& VU\\"%56rcFUR5S:XaU$UR5n[UR55Vs/sHo3PM nnUSUSsUS'US'UR U5n[ R "USS9n[ RRUSS9S-nURR5S:XaBURRUR5 URRU5 U$U=RU- sl U$s snf)Nrr ) start_dimr r) numeldetachranger permuterflattenlinalg vector_normr3resize_shapecopy_)r>x_origxi new_axis_listyr3s rforwardPerChannelNormObserver.forwardIs <<>Q M MMO%*!%%'N3NqN 3.;B.?qAQ+ a-+ IIm $ MM!q )||''q'1Q6 99??  ! II  djj ) IIOOD !  II I 4s Dc[S5e)Nz>PerChannelNormObserver is designed to store activations only. )NotImplementedError)r>s rcalculate_qparams(PerChannelNormObserver.calculate_qparams_s! L  r)r;)returnN) __name__ __module__ __qualname____firstlineno____doc__r6rTrX__static_attributes__ __classcell__)r@s@rrr3s 7",  rrr=rrrZcSnUR5U-S:wa[SURSUS35eUR5U-nUR5R 5R XB5n[ R"USS9SS2S[X!- 524n[ R"URURS9nURSUSS 9R UR5nU$) a Class for creating N:M sparsity masks. Masks will be created using the N:M ratio, where for every block of M weights, N will be pruned based on ranked weight value. Each mask will correspond to the given tensor. :param tensor: The input tensor to create a mask for :param N: The number of weights in a group to keep :param M: The size of a weight group :return: A mask tensor with the same shape as the input tensor NrzTensor of size z can't be evenly divided into z groupsr r )device)r indexvalue) rE ValueErrorrMrFabsr#rargsortintonesrcscatter_)r=rrr) num_groups tensor_temprdw_bs rrres D ||~Qfll^+I!G T  1$J--/%%'// >K MM+1 -a3qu:o >E **[&&{/A/A BC <rsse  J  +#,/ </ h  LL   \\ r