TheDavidTai(@TheDavidTai):Extending acceptance length by 27% is quite the achievement. Congrats to the DSpine team. I wish there was a way to train Spec Decoders for specific use cases in the future.

2026.11.10 18:50

Extending acceptance length by 27% is quite the achievement. Congrats to the DSpine team. I wish there was a way to train Spec Decoders for specific use cases in the future.

Benjamin Marie@bnjmn_marie

2026.11.10 02:22

DSpine: +27.8% acceptance length and +23.3% throughput over DFlash on Qwen3-8B, according to the paper. But the gain is much smaller with the released checkpoint. It was trained on Open-PerfectBlend, and DFlash2 and DSpark retrained on the same data are close behind: about 4% lower at temperature 0 and 5% lower at temperature 1. Still a real gain, for a round that costs about the same as DFlash2's. If you already serve Qwen3-8B or Qwen3.8-27B with SGLang, it's worth an A/B test. I would measure tokens/s on my own prompts rather than trust acceptance length on GSM8K...