DFlash 2: Keep Drafting Parallel

DFlash 2, an advancement in speculative decoding technology, increases inference throughput by 16-25% by enabling parallel token prediction across entire blocks instead of one token at a time, adding minimal latency while maintaining accuracy. The technology, developed by Inco AI and already integrated into major inference engines like SGLang and vLLM, achieves 2.7-3.4× faster throughput compared to autoregressive decoding, addressing the critical bottleneck of token consumption in AI agent applications.