April 2026 performance summary across two sgLang repositories (bytedance-iaas/sglang and yhyang201/sglang). Delivered deployment-ready documentation and NPU-optimized tensor processing enhancements for large-model inference. Key outcomes include enabling low-latency deployment for Qwen3-30B-A3B via a detailed deployment guide and benchmarks, and delivering NPU ops for MiniMax attention along with fixes to hidden state capture in distributed attention modes, improving correctness, performance, and scalability. These efforts reduce time-to-value for customers, improve inference efficiency, and strengthen state management in multi-GPU scenarios.

March 2026 Monthly Summary – ping1jing2/sglang Focused on delivering high-impact features to boost model performance, efficiency, and real-time capability on NPU-backed backends. The month centered on advancing MoE decoding support and improving sequence handling in Ascend attention backend to raise throughput and accuracy in live inference scenarios. Overall, the team advanced core capabilities that enable faster, more reliable deployments of large-scale MoE models, with clear performance and resource utilization benefits for NPU-based workloads.