*`i ddlmZddlZddlmZddgZ ddejjdedeed efd Z d ej d ej ded e ejjd ej f dZ dej dej d ej fdZ dS))OptionalN)TransformLocationget_transform_sizeapply_transform_weightmodulelocationhead_dimreturnc d}t|tjjr*|tjtjfvr|j}ns|j}nkt|tjj r*|tjtjfvr|j }n*|j }n"| tdt|d|4|0||zdkr't|d|dt|d||}|S)a Determine the size of a transform matrix given its location on the module :param module: module that matrix will be applied to :param location: location on module :param head_dim: size of head when transform is applied to mha :return: size of matrix NzTransforms on z# are not supported without head_dimrz must divide z for z at ) isinstancetorchnnLinearrINPUT WEIGHT_INPUT in_features out_features Embeddingnum_embeddings embedding_dimNotImplementedErrortype ValueError)rrr sizes }/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/compressed_tensors/transform/utils/matrix.pyrrs D&%(/**  )/1B1OP P P%DD&DD FEH. / / )/1B1OP P P(DD'DD  ! NT&\\ N N N     x1 4 4QQ$QQT&\\QQxQQ  Ktransform_weightvalue module_typec6|jd|jdksJt|rt||S|tjjkrK|tjkrt||jS|tj krt|j|SnU|tjj kr@|tjkrt||S|tj krt||Std|d|d)a; Using the transform location, apply the transform_weight to the given value wrt linear weights. For more info on input and output transforms, see `TransformLocation` The following explains how weights should be applied to values according to location let x be input activation W be weight, yh, xh, Wh be transformed output, input, weight note that y = (x W.T) // torch.nn.Linear Choose values for yh, xh, and Wh which incorporate matrix transforms let V, Vi be transform matrices on input side U, Ui be transform matrices on output side pick xh = (x V) Wh = (U.T W Vi.T) yh = (y U) The following shows that `yh = (xh) (Wh).T` for the chosen values of yh, xh, and Wh (xh) (Wh).T = (x V) (U.T W Vi.T).T = (x V) (Vi W.T U) // transpose matrix product identity = (x W.T) U = y U = yh :param transform_weight: transform weight to apply :param value: value to apply transform_weight to :param location: determines how weight should be applied :param model_type: result of type(module), passed in to determine application of weight transform :return: value after transform_weight has been applied rzApplying transforms to  z is not supported) shaper is_online_multihead_matmulr rrrT WEIGHT_OUTPUTrr)rrrrs rrrAs(Z  !! $(8(>q(A A A A A"",,..: (8999eho%% (5 5 5$U,<,>?? ? *8 8 8$%5%7?? ?9 * * * (5 5 5$%5u== = *8 8 8$U,<== = K+KKKKK  rABc|jd|jdkrN|jd}|jd|z}|d||f}||zddS|jd|jdkrN|jd}|jd|z}|d||f}||zddS||zS)a Performs A @ B for last two dims of two matrices A and B that possibly have different shapes, as is the case in multi-headed dimension. If shapes are different, this is equivalent to converting the last two dims of the smaller matrix into a block-diagonal matrix with the same shape as the last two dims of the larger matrix. E.g. if A is half the size of B, this function will perform [[A ] @ B [ A]] If B is a third of the size of A, this function will perform A @ [[B ] [ B ] [ B]] This function will error out if the shapes are not evenly divisble :param A: left-hand tensor :param B: right-hand tensor :return: result )r# unflattenflatten)r(r)r num_headss rr%r%s. wr{QWR[  72;GBK8+ KKY1 2 2Ar2&&& qwr{ " "72;GBK8+ KKY1 2 2Ar2&&&1u r)N)typingrr compressed_tensors.transformr__all__rModuleintrTensorrrr%rrr8s  :::::: !9 : #&& HO&&sm& &&&&RElE <E Eeho& E  \ EEEEP""%,"5<""""""r