Gradual structured pruning is a neural network compression technique that iteratively removes entire structural components, such as attention heads, channels, neurons, or layers, across multiple stages of training or fine-tuning. Unlike unstructured pruning, which eliminates individual weight parameters and creates irregular sparsity patterns, structured pruning removes complete computational blocks to yield dense, smaller models that achieve practical latency and memory improvements on standard hardware. By eliminating these structures gradually over time rather than in a single one-shot pass, the model is given continuous opportunities to recover and adapt its remaining parameters through retraining, resulting in higher task performance and better retention of accuracy at high compression rates.
