DC-Leap slashes dLLM inference by 105x with training-free decoding trick
New method blasts past JPDE bottleneck, delivering 53x speedup on long sequences—no training required.
DC-Leap, introduced by Yanhua Jiao and eight co-authors, solves a key inefficiency in Diffusion Large Language Models (dLLMs): overly conservative confidence thresholds that cause redundant denoising steps. This redundancy stems from the Joint Probability Dependence Error (JPDE), which forces existing parallel decoding strategies to play it safe. DC-Leap’s core innovation is a Dynamic Contiguous Verification mechanism that integrates strictly-ordered causal constraints into parallel token generation, progressively validating dependencies to neutralize JPDE without retraining the model.
On top of that, DC-Leap deploys a draft-guided decoding strategy: a lightweight draft model proposes multiple tokens at once, allowing the main model to “leap” forward in the sequence while retaining bidirectional attention benefits. Tests on standard benchmarks show dramatic speedups: up to 53.19x on MBPP for long-sequence code generation and up to 105.02x when paired with KV-Cache, all with comparable generation quality. The team has released the code, making it easy for engineers to integrate into existing dLLM pipelines.
- DC-Leap eliminates the JPDE bottleneck using Dynamic Contiguous Verification, enabling faster parallel decoding without training.
- Achieves up to 53.19x speedup on MBPP (long-sequence code generation) and up to 105.02x with KV-Cache.
- Draft-guided leaping lets the model skip multiple tokens while preserving bidirectional attention context.
Why It Matters
Makes dLLMs practical for real-time applications by slashing inference latency without sacrificing quality or requiring retraining.