infrastructure
fact
bullish
Token-level dynamic width pruning enables fine-grained computation allocation by allowing each token to dynamically select attention-head groups and FFN-channel groups at neuron-block-level granularity
Machine Learning02 Aug 2026