March 2026 monthly summary for sglang development across repositories sgl-project/sglang and ping1jing2/sglang. Focused on improving NPU attention efficiency and robustness, with a cache-based optimization for rotary embeddings and a bug fix in context prefill parallelism. Deliveries are backed by explicit commits for traceability and business value delivered through faster and more reliable inference.

Month 2025-12 — Consolidated performance and reliability gains for kvcache-ai/sglang. Delivered three primary enhancements: NPU Backend Performance Optimizations for Attention, Scheduling Enhancements for dp_attention, and CP Feature Enablement Validation. Key outcomes include measurable throughput improvements on NPU attention workloads, reduced prefill idle time via SchedulerEnhancer, and hardened server argument validation to prevent misconfiguration. These efforts reduce latency, increase throughput, and improve production reliability, leveraging quantization-based kvcache optimizations, multi-stream processing options, and environment-driven tuning. Tech focus: NPU optimization (parallel prefill, quantization, reshaping), scheduling engineering, assertion-based validation, and robust CI.