WIDE: Boosting Adaptive LLM Inference via Token-level Dynamic Width Pruning | Digital Library | PAMCET | PAMCET