Domino: Decoupling Causal Modeling from Autoregressive Drafting in Speculative Decoding

Domino reports up to 5.8x throughput speedup on Qwen3 by decoupling causal modeling from autoregressive drafting in speculative decoding. The Reddit snippet links the arXiv paper, GitHub code, and Hugging Face models, but does not disclose hardware, baseline settings, dataset, or acceptance-rate details.

Why it matters: HKR-H/K/R all pass: 5.8x throughput is a concrete hook with open artifacts. Missing hardware, baseline config, and task set keep it in the good featured band, not same-day must-write.