AdaFlash
The official implementation of a speculative-decoding method that pairs a large language model with an adaptively trained diffusion drafter model to speed up inference.
AdaFlash is the official implementation of a speculative-decoding framework that uses an on-policy distilled diffusion model as the draft model alongside a candidate-length head that adapts how many tokens the drafter proposes at a time. It supports asynchronous on-policy inference and training with hot-swappable draft checkpoints, and provides reproducible benchmarks against several other decoding methods, including EAGLE-3, DFlash, and MTP.
The repository includes configurations and training recipes for the Qwen3, Qwen3-Coder, and Qwen3.5 model families, and its training framework is adapted from SpecForge. It serves models using a patched build of SGLang and includes scripts to prepare benchmark datasets and to regenerate training data from a target model and an initial drafter.
The paper behind this code was accepted as a COLM 2026 spotlight.