AI
WaveFront Decoding: Parallelized Self-Speculative Decoding for Looped Language Models
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 25/09/2026, 20:00:00 UTC
Briefing created: 29/09/2026, 04:31:19 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 29/09/2026, 04:18:25 UTC.
Briefings
WaveFront Decoding (WFD) is a training-free self-speculative decoding framework designed for looped language models to mitigate latency caused by repeated weight-shared blocks. By organizing mixed-depth states into a diagonal wavefront, WFD concurrently batches drafting and verification within single recurrent calls, outperforming traditional draft-then-verify schedules. The method achieves 2.42x speedup on Ouro-2.6B and 3.54x on Huginn-3.5B across six Spec-Bench categories, with cross-recurrence KV sharing boosting performance to 4.81x on Huginn-3.5B.
Key points
- WaveFront Decoding (WFD) is a training-free self-speculative decoding framework that uses diagonal wavefronts to concurrently batch drafting and verification within single recurrent calls for looped language models.
- WFD achieves 2.42x speedup on Ouro-2.6B and 3.54x on Huginn-3.5B across six Spec-Bench categories, with cross-recurrence KV sharing increasing speedup to 4.81x on Huginn-3.5B.
Context and limitations
Results are limited to six Spec-Bench task categories; cross-recurrence KV sharing speedup is specific to Huginn-3.5B; code availability is restricted to the provided GitHub link.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 29/09/2026, 04:31:19 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeibaonsfnkli4xnkfbuja7p4e7hc2yzxazqqarxocvy476br57hj4i