SGLang Launches New Plugin System for NVIDIA and AMD AI Kernels

SGLang has launched a new plugin system using two extensions, sgl-nvidia-mega and sgl-amd-mono, to allow hardware vendors to deliver low-latency kernels as external libraries. The NVIDIA plugin is currently limited to DeepSeek-V4.1-Flash running on 4x GB300 GPUs. The system is already enabling third-party optimizations, such as a Zyphra kernel that increased single-user decode speeds at 1M context from 43.5 to 61.1 tok/s for DeepSeek-V4. SGLang is developing additional kernels for GLM-5.3, K3, and MiniMax M3 models for both NVIDIA and AMD hardware.