o i4@s4ddlZgdZddZddZddZd d ZdS) N))sparse_semi_structured_from_dense_cutlass'sparse_semi_structured_to_dense_cutlass mask_creatorc Cs$tjd||ddddfd|}tjd||d|d}d}|jdkr'dnd}||||dd|dd ||d dd||dd }|ddk|ddk@tj}|ddk|ddk@tj} ||| 7}||| 8}d} || } || } | || || | d S) Nrdevice@ )torcharangerepeatitemsizetoint8view) m meta_ncols meta_dtyperdst_rowsdst_colsgroup_xgroup_ytopright bottomleft interleavecols_majcols_minr"h/home/ubuntu/.local/lib/python3.10/site-packages/compressed_tensors/utils/semi_structured_conversions.py*_calculate_meta_reordering_scatter_offsets-s,$       r$cCs|dkrtd|d|j\}}|j}tj}|jtjkr%tj}n|jtjtj tj tjfvr6tj }n td|jd|j dd}|dvrNtd |tjkrb|d d kratd |d n|dd krptd |d|d|d krtd|dd||jtj krd}| d|||}|d kd\}} } } nd}| d|||} | d kd\}} \} } |||} || @}|| @}|| @}|}|}||B| B}|| B}||tjd>B}||tjd>B}|jtj kr|d|d}|d|d}tj||fdd ||d}n| d|dd ||d}||d>B}| d| |f|}|dkrt|ddddd f|dddddfd>B|dddddfd>B|dddddfd>B}nk|dkr|ddddd f|dddddfd>B|dddddfd>B|dddddfd>B|dddddfd >B|dddddfd>B|dddddfd>B|dddddfd>B}||| f}t|| ||}|d || d|| || fS)Nr z)Expected 2-dimensional dense tensor, got -dimensional tensorInvalid datatype z of dense matrixr r )r r z6Invalid number of elements per meta element calculatedr rzNumber of rows of dense matrix z must be divisible by 16r z must be divisible by 32z"Number of columns of dense matrix z must be divisible by rrdim )r( RuntimeErrorshaperrrdtypeint32halfbfloat16floatint16rrunbindrint64gather unsqueezestack new_emptyr$scatter_)denserkrrquadbits_per_meta_elemksparsedense_4m0m1m2m3dense_2rexpr0expr1expr2bit0bit1bit2bit3idxs0idxs1sparse0sparse1sparsemeta_4meta_nmetameta_reordered meta_offsetsr"r"r#rNs          $          rcCs|dkrtd|d|j\}}|j}|dkr(td|d|j|kr9td|d|jd|j}|tjtjfvrLtd|d |jd d }|jtj kr[d nd}|j\}} ||krptd |d || ||d|krtd|d| ||ddt || ||} t | dd|  || } tj || d|f||d} |d kr)| d@| dddddf<| d?d@| dddddf<| d ?d@| dddddf<| d?d@| dddddf<| d ?d@| ddddd f<| d?d@| dddddf<| d?d@| dddddf<| d?d@| dddddf<n|d kr| d@| dddddf<| d?d@| dddddf<| d ?d@| dddddf<| d?d@| dddddf<| d ?d@| ddddd f<| d?d@| dddddf<| d?d@| dddddf<| d?d@| dddddf<| d?d@| ddddd f<| d?d@| dddddf<| d?d@| dddddf<| d ?d@| ddddd!f<| d"?d@| dddddf<| d#?d@| ddddd$f<| d%?d@| dddddf<| d&?d@| ddddd'f<| dtjdd||||d(d  dddd d} tj|d|f|j|d}|jtj kr\|d| |dn| tjd| | tj d| |d|S))Nr z*Expected 2-dimensional sparse tensor, got r%z(Expected 2-dimensional meta tensor, got zExpected meta matrix to be on z device, got matrix on z devicer&z of meta matrixr r zNumber of rows of meta matrix z4 must be equal to number of columns of spase matrix z#Number of columns of sparse matrix z different from the z<, expected according to the number of columns of meta matrixrr)r3rr)rr- r+r*r/r  r, r. r0r)r(r1r2rr3rr8r4rr7r$r;remptyrrzerosr?reshaper5)rUrYrrArrrBrC meta_nrowsrrZrXmeta_2 dense_offsetsr@r"r"r#rs          rcCsd}d}d}||dkrtd|jd|d||}|||}tj|dd dddt||f}tj |j|j d }|j d|dd |j}|S) aJ 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 N: The number of weights in a group to keep :param M: The size of a weight group r r NrzTensor of size z can't be evenly divided into z groupsrr'r)r(indexvalue) numel ValueErrorr2detachabsrgrargsortintonesrr?)tensorNMmask num_groups tensor_temprkw_br"r"r#r9s  &r)r__all__r$rrrr"r"r"r#s! a