Welcome! Type "help" for available commands.
$
Machine Learning Model
Summary
This repository hosts the DFlash drafter model, a novel speculative decoding method that utilizes a lightweight block diffusion model to enhance inference speed. It is designed to be used in conjunction with the Qwen/Qwen3.6-27B target model to enable efficient, high-quality parallel drafting.
Highlights
- Introduces DFlash, a speculative decoding method using block diffusion for drafting.
- Serves as a specialized drafter component for the Qwen/Qwen3.6-27B model.
- Provides specific installation instructions for vLLM and SGLang to support architectural changes like causal SWA layers.
- Aims to push the limits of inference speed through efficient parallel drafting.
auto-generated
Context
Audience
Machine Learning Engineers and AI Researchers
DomainMachine Learning
Formatmodel repository
Accessopen source
Related
Speculative DecodingQwenvLLMSGLangBlock DiffusionLarge Language Models