Hi DeepSpec team,
Thanks for open-sourcing DeepSpec. I wanted to ask whether you would be interested in supporting Domino in DeepSpec, or whether Domino-style support is already on your roadmap.
Domino is a speculative decoding method that keeps draft generation block-parallel while adding a lightweight causal correction head. The goal is to improve draft-token acceptance without paying the sequential overhead of an autoregressive drafter.
Pipeline:
Paper: https://arxiv.org/abs/2605.29707
Code / checkpoints: https://github.com/jianuo-huang/Domino
The paper reports up to 5.49x end-to-end speedup under Transformers and up to 5.8x throughput speedup under SGLang serving on Qwen3 models.
If the DeepSpec team is open to adding Domino support, I would be happy to help with the implementation, training/evaluation integration, or checkpoint conversion/evaluation. Is there a preferred abstraction or direction I should target first?
Hi DeepSpec team,
Thanks for open-sourcing DeepSpec. I wanted to ask whether you would be interested in supporting Domino in DeepSpec, or whether Domino-style support is already on your roadmap.
Domino is a speculative decoding method that keeps draft generation block-parallel while adding a lightweight causal correction head. The goal is to improve draft-token acceptance without paying the sequential overhead of an autoregressive drafter.
Pipeline:
Paper: https://arxiv.org/abs/2605.29707
Code / checkpoints: https://github.com/jianuo-huang/Domino
The paper reports up to 5.49x end-to-end speedup under Transformers and up to 5.8x throughput speedup under SGLang serving on Qwen3 models.
If the DeepSpec team is open to adding Domino support, I would be happy to help with the implementation, training/evaluation integration, or checkpoint conversion/evaluation. Is there a preferred abstraction or direction I should target first?