Online LLM serving is increasingly constrained by GPU memory capacity during decode. In length-skewed workloads, a small fraction of long requests keeps large key-value (KV) caches resident across many decode steps, exhausting GPU high-bandwidth memory (HBM), throttling admission, and reducing goodput for the short-req...
Dong Chen, Yu-Kun Zhou, Sheng Yao et al.· Proceedings of the 2026 ACM...· 0 citations
The booming LLM demand is driving a rapid increase in the scale and heterogeneity of modern AI infrastructure. To achieve high availability for sustained LLM workloads, infrastructure providers must contend with frequent and severe hardware errors. Without a holistic resilience understanding, providers may inadvertentl...
Xin-Kai Wang, Yun-Wei Li, Chao Li et al.· ACM Transactions on Architec...· 0 citations
Efficient serving of long-context reasoning-and-acting (ReAct) agents relies on KV cache reuse to reduce large language model (LLM) prefill latency and monetary cost. However, a semantic gap exists between agent harnesses and the underlying serving stack. Through context mutation, tool execution, and subagent coordinat...
Kai-Hua Fu, Yu-Kun Zhou, Chaokun Chang et al.· 0 citations
Model-as-a-service platforms call for continuous optimization as complex serving conditions expose inefficiencies missed before deployment. Detailed always-on profiling can incur substantial overhead, while lightweight collection omits information needed for diagnosis. We present Herschel, a continuous optimization sys...
Lu-Ping Wang, Wei-Gao Chen, Yi-Fei Wu et al.· 0 citations
Atrex-Bench is presented, a benchmark whose 30 operators and 440 shapes are sampled directly from full-cluster production inference traces of compute-limited, memory-rich GPUs, and a profile-driven kernel-optimization agent that combines iterative measure-revise search, optimization dropout for escaping stalled search...