Open-Sourced DSpark Speculator Boosts Inkling Throughput by 1.89x

Developers have open-sourced a new DSpark speculator for the Inkling NVFP4 model, built using SpecForge on an online SGLang engine. This speculative decoding implementation successfully boosts decoding throughput by 1.89 times.

2026-07-23 ~ 2026-07-23 · 3 related posts

• Open-Sourced DSpark Speculator: Boosts Inkling Decode Throughput by 1.89X — BanghuaZ · 2026-07-23
• SGLang DSpark reaches 1.89× decode throughput on 8×B200 — BanghuaZ · 2026-07-23
• DSpark speculator trained on live SGLang lifts decode throughput 1.89× on B200s — ying11231 · 2026-07-23