variable-length-verification

SGLang now integrates DSpark, a confidence-driven speculative decoding method that verifies only confident draft tokens, improving inference efficiency under load.