Tag
DSpark is a speculative decoding framework that combines semi-autoregressive draft generation with confidence-scheduled verification to accelerate LLM inference and improve throughput in high-concurrency settings.