dParallel: Learnable Parallel Decoding for dLLMs

TL;DR

The paper addresses the limited parallel decoding potential of open-source diffusion LLMs by identifying sequential certainty convergence as the core bottleneck. It introduces certainty-forcing distillation (CFD) to train models that preserve trajectory consistency while rapidly achieving high certainty on masked tokens in parallel. Empirical results show substantial decoding-step reductions and 8–10x speedups on GSM8K and MBPP with minimal accuracy loss, using LoRA-based fine-tuning. The work sets a new baseline for parallel decoding in dLLMs and suggests promising avenues for broader pretraining and data-scale improvements.

Abstract

Diffusion large language models (dLLMs) have recently drawn considerable attention within the research community as a promising alternative to autoregressive generation, offering parallel token prediction and lower inference latency. Yet, their parallel decoding potential remains largely underexplored, as existing open-source models still require nearly token-length decoding steps to ensure performance. To address this, we introduce dParallel, a simple and effective method that unlocks the inherent parallelism of dLLMs for fast sampling. We identify that the key bottleneck to parallel decoding arises from the sequential certainty convergence for masked tokens. Building on this insight, we introduce the core of our approach: certainty-forcing distillation, a novel training strategy that distills the model to follow its original sampling trajectories while enforcing it to achieve high certainty on masked tokens more rapidly and in parallel. Extensive experiments across various benchmarks demonstrate that our method can dramatically reduce the number of decoding steps while maintaining performance. When applied to the LLaDA-8B-Instruct model, dParallel reduces decoding steps from 256 to 30 on GSM8K, achieving an 8.5x speedup without performance degradation. On the MBPP benchmark, it cuts decoding steps from 256 to 24, resulting in a 10.5x speedup while maintaining accuracy. Our code is available at https://github.com/czg1225/dParallel

Figures (9)

• Figure 1: Our method achieves highly parallel decoding. Compared to the original LLaDA Model, dParallel decodes over 8 tokens per step on GSM8K while preserving the accuracy.
• Figure 2: Empirical Studies: (a) The average confidence score exhibits a positive correlation with generation accuracy. (b) Token confidence propagates sequentially during the decoding process. (c) Convergence trajectories of confidence for different tokens.
• Figure 3: Overview of proposed certainty-forcing distillation. The dLLM is self-distilled along its original generation trajectory, ensuring consistency with the trajectory throughout training while encouraging token certainty to converge faster in parallel rather than sequentially.
• Figure 4: Comparison of speed–accuracy trade-off curves between confidence-threshold decoding and our method. (a) and (b) show results on the LLaDA model for GSM8K and HumanEval, respectively. (c) and (d) present results on the Dream model for GSM8K and HumanEval benchmarks.
• Figure 5: Average token confidence at the 8th and 16th decoding steps for LLaDA-8B-Instruct Model on GSM8K. The proposed certainty-forcing strategy reshapes the original sequential certainty convergence into a faster and more parallel convergence process.