+ "ݳjT^RIt^RIHt.R OtRtRt!RR]4tR RRlltR#) N)UniformQuantizationObserverBasePerChannelNormObservercjVR8:dVR8gQR4h\P!\P!W,W,3^V, VR74p\P!WR^R7p\P!WR^R7pVP V4P 4P 4#)?gz*sparsity should be a value between 0 and 1)dtypedim)torch bernoullifullrepeat_interleaveto contiguouscuda)MN blocksizesparsityrAs&&&&& W/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torchao/sparsity/utils.pycreate_block_sparse_tensorrs s?x34 .  ANAN3Q\O A !4A !4A 44; ! ! # ( ( **c\P!^^W,^,34P4p\PPP V^R7P W4P4P\P4pV\P!W4P4V,,pVPV4#)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. ) num_classes) r randintrnn functionalone_hotreshaperrint32rand)rcrchoice_indicesmask sparse_weights&&& rcreate_semi_structured_tensorr's]]1a!%1*7<<>N ##N#B   EKK  EJJq,113d:;M   E ""rcNaa]tRt^3toRtV3RlV3RlltRtRtRtVt V;t #)rzX A custom observer that computes the L2 norm of each channel and stores it in a buffer. c<V^8dQhRR/#)returnN)format __classdict__s"r __annotate__#PerChannelNormObserver.__annotate__8s 7 7D 7rc *<\SV`!R R\PR\PRRRRRRR\P !\P 4P/VBRVnVPR \P!.44R#) rqscheme reduce_rangeF quant_minN quant_maxepsrnormr,) super__init__r quint8per_channel_affinefinfofloat32r6averaging_constantregister_buffertensor)selfkwargs __class__s&,rr9PerChannelNormObserver.__init__8s  ,, ,,       EMM*..   #& VU\\"%56rctVP4^8XdV#VP4p\VP44Uu.uFq3NK ppVR,V^,uV^&VR&VP V4p\ P !V^R7p\ PPV^R7^,pVPP4^8XdCVPPVP4VPPV4V#V;PV, un V#uupi)r) start_dimr) numeldetachranger permuter flattenlinalg vector_normr7resize_shapecopy_)rAx_origxi new_axis_listyr7s&& rforwardPerChannelNormObserver.forwardIs <<>Q M MMO%*!%%'N3NqN 3.;B.?qAQ+ a-+ IIm $ MM!q )||''q'1Q6 99??  ! II  djj ) IIOOD !  II I 4s D5c\R4h)z>PerChannelNormObserver is designed to store activations only. )NotImplementedError)rAs&rcalculate_qparams(PerChannelNormObserver.calculate_qparams_s! L  r)r>) __name__ __module__ __qualname____firstlineno____doc__r9rWr[__static_attributes____classdictcell__ __classcell__)rCr.s@@rrr3s# 7 7",  rcpV^8dQhR\PR\R\R\P/#)r*r@rrr+)r Tensorint)r-s"rr/r/es8 LL   \\ rcRpVP4V,^8wd\RVP RV R24hVP4V,pVP4P 4P WB4p\ P!V^R7RR\W!, 413,p\ P!VPVPR7pVP^V^R7P VP4pV#) 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 NzTensor of size z can't be evenly divided into z groupsr:NNN)device)r indexvalue) rH ValueErrorrPrIabsrr argsortrgonesriscatter_)r@rrr% num_groups tensor_temprjw_bs&&& r mask_creatorrtes D ||~Qfll^+I!G T  1$J--/%%'// >K MM+1 -a3qu:o >E **[&&{/A/A BC <rxs6  J  +#,/ </ dr