sglang v0.5.16 introduces DSpark speculative decoding and Inkling support
▼ WHAT HAPPENED
The latest sglang release includes a significant update with the introduction of DSpark, a confidence-driven speculative decoding algorithm that enhances performance by dynamically sizing verify windows based on draft confidence rather than fixed lengths. This feature reaches impressive speeds of 383.7 tok/s at an accept length of ~5 on DeepSeek-V4-Pro and TP8 on B300 (bs=1). Additionally, the release adds support for Inkling, a large-scale model that expands sglang's capabilities further. Operators should consider updating to leverage these performance improvements and new model support.
▼ OPERATOR ANGLE
Update to sglang v0.5.16 to take advantage of DSpark speculative decoding and Inkling support, enhancing both efficiency and model diversity in local AI operations.