Hao AI Lab 开源 JetSpec++ 投机解码，最高提速 7.45 倍
[4/5] Bigger targets leave more room. On Qwen3.8-27B, hiding the draft saves less, because verification dominates the round. But the same long window fits three
Hao AI Lab 开源 JetSpec++，通过因果并行草稿与隐藏状态预测器，将草稿生成与验证流水线化，在 Qwen3-8B 和 Qwen3.8-27B 上分别实现 1,361 TPS 和 367 TPS，较 DSpark 最高提速 1.54 倍、较自回归解码提速 7.45 倍且保持无损。在 Qwen3.8-27B 上，每轮接受的 token 数较 DSpark 提升 22%（8B 为 6%）；仅用 16 块训练即可外推到 32 块草稿，MATH-500 上每轮 14.55（8B）和 14.05（27B）个 token，DSpark 降至 8.66 和 8.49。结合 NVFP4 与 kernel 优化，两块 GB200 上 Qwen3-8B 超 4,000 TPS、Qwen3.8-27B 超 900 TPS。
正文
[4/5] Bigger targets leave more room. On Qwen3.8-27B, hiding the draft saves less, because verification dominates the round. But the same long window fits three refinement passes instead of one, so the gain moves to acceptance: +22% accepted tokens per round over DSpark on 27B, vs +6% on 8B (GSM8K, as served).
It also extrapolates. Trained on blocks of 16, JetSpec++ drafts blocks of 32 and gains: 14.55 tokens per round on MATH-500 on 8B and 14.05 on 27B, while DSpark drops to 8.66 and 8.49.
We would love to hear your feedback!