The authors first analyzed DFlash, a parallel drafter that predicts a block of tokens in one backbone pass.
Depth-wise predecessor conditioning makes parallel speculative drafts more consistent
DSpine passes predicted features between adjacent draft positions throughout the network, increasing accepted draft length and serving throughput.
Chinese Tech
Haohui Zhang · Keyu Chen · Haocheng Sun · Weibo Gu · Ruizhi Qiao · Xing Sun · +1 more
Shanghai Jiao Tong University · Tencent YouTu Lab · Xiamen University
Research Digest··3 min read
Zhang et al.
Why this paper
From Tencent YouTu Lab and 2 others
In one line
DSpine injects causal conditioning at every layer, increasing acceptance length by 27.8% and throughput by 23.3% over DFlash.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks (2 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§