GLM-5 Serving Parameter Tuning for OpenClaw: Single-Deployment MaaS Inference Optimization for Long-Context Agent Workloads

Collection of arXiv papers (Jul 7 2026) covering system and ML infrastructure advances: GLM-5 serving parameter tuning for OpenClaw identifies a best single-node configuration (chunked-prefill=3072, tp=4, pp=4, max-running-requests=24) yielding ~10% cost and latency improvements for very long-context, tool-augmented requests. PEEK introduces a queue-informed KV-cache admission/eviction and multi-lane scheduler that substantially increases cache hits, TTFT, E2E latency and throughput on SGLang and vLLM. StateFlow proposes a distributed MoE inference policy that pins KV state to reduce KV-reloc/