AI BriefingTencent Introduces DSpine: Parallel Speculative Decoding with Causal Injection
·2026.11.08 00:48
Key point
The method achieves a 27.8% increase in mean acceptance length over DFlash on Qwen3-8B at temperature zero.
Details
- DSpine is a new speculative decoding drafter that injects causal conditioning at every layer of the backbone to maintain consistent continuations during parallel drafting.
- On Qwen3-8B at temperature zero, DSpine raises the seven-benchmark mean acceptance length from 3.77 to 4.82, a 27.8% improvement over DFlash.
- In SGLang serving tests, DSpine delivers 23.3% higher average throughput than DFlash.
- The method uses a unified transfer space from the target model's output embeddings and layer-wise supervision to promote predicted feature formation in shallow layers.
- Implementation includes fused kernels and a transition cache for efficient parallel execution within SGLang.
- Code and model weights are available on GitHub, Hugging Face, and the project website.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
DeepSeek Open-Sources DeepInfer Framework for V4.1 Flash InferenceDecagon's AutoResearch Agent Automates Inference Stack Optimization, Achieving Up to 2.5x Throughput Gains