The invention relates to the field of
language model pruning, and particularly discloses a task-adaptive visual language
large model collaborative
pruning method, which comprises the following steps of: S1, performing task
perception, acquiring a training
data set containing image text pairs, shielding a corresponding
data set of a single mode, and calculating loss value fixed mode dependency so as to adaptively bias and finely adjust visual and
language module shares; s2, calculating importance scores, dividing model weights, hierarchically calculating gradient norm approximate values, and multiplying the gradient norm approximate values with
modal shares to obtain scores of all
layers; s3, sparseness is distributed, summarizing scores are normalized, the sparseness of each layer is calculated in combination with a total parameter retention proportion, and a threshold value is set; and S4, structured sparseness is carried out, a
pruning strategy and token screening are combined, a fusion
score of
modal correlation degrees is calculated, effective tokens are reserved to calculate channel scores, and key channels are reserved according to groups. According to the technical scheme, the problems that traditional pruning is poor in
adaptation, task
modal dependence is not considered, and key parameters are easily
cut by mistake are solved, compression and reasoning efficiency is improved, cross-modal
information loss is avoided, and performance loss is reduced.