The bottleneck associated with the key-value(KV) cache presents a significant challenge during the inference processes of large language models. While depth pruning accelerates inference, it requires extensive recovery training, which can take up to two weeks. On the other hand, width pruning retains much of the performance but offers slight speed gains. To tackle these challenges, we propose KVPruner to improve model efficiency while maintaining performance. Our method uses global perplexity-based analysis to determine the importance ratio for each block and provides multiple strategies to prune non-essential KV channels within blocks. Compared to the original model, KVPruner reduces runtime memory usage by 50% and boosts throughput by over 35%. Additionally, our method requires only two hours of LoRA fine-tuning techniques on small datasets to recover most of the performance.

Real-time systems involve tasks that may voluntarily suspend their execution as they await specific events or resources. Such self-suspension can introduce further delays and unpredictability in scheduling, making the analysis more challenging. Most current schedulability analysis methods of self-suspending tasks focus on fixed-priority scheduling or tasks with constrained deadlines. This paper proposes two schedulability analysis methods for self-suspending tasks with arbitrary deadlines under earliest-deadline-first-like (EDF-like) scheduling. Both methods are designed for preemptive uniprocessor systems. We first present a jitter-based response time analysis (JRTA) method. JRTA is designed based on a self-suspending response time analysis (SS-RTA) method under earliest-deadline-first (EDF) scheduling. We first convert self-suspensions to release jitters and then present a response time analysis (RTA) method of tasks with release jitters under EDF-like scheduling. To address the complexity of JRTA, we propose an improved schedulability analysis (ISA), a sufficiency blocking-based method. Finally, we provide many simulation experiments under some EDF-like scheduling algorithms. The results verify the effectiveness and efficiency of both proposed methods.