Fast dLLM v2 7B EU AI Act Compliance Profile

Fast-dLLM v2 (7B) — Efficient Block-Diffusion LLM

📖 Introduction

Autoregressive (AR) large language models (LLMs) have achieved remarkable performance across a wide range of natural language tasks, yet their inherent sequential decoding limits inference efficiency.

We present Fast-dLLM v2 — a carefully designed block diffusion language model (dLLM) that efficiently adapts a pretrained AR model (Qwen2.5-7B-Instruct) into a diffusion-style decoder for parallel text generation.

✨ Key Innovations

• Block Diffusion Mechanism + Complementary Attention Mask
  Enables blockwise bidirectional context modeling without sacrificing AR objectives.
• Hierarchical Caching
  • Block-level cache: Stores historical context representations across blocks.
  • Sub-block cache: Parallel decoding within partially generated blocks.
• Token Shift Mechanism
  Retains autoregressive characteristics while supporting bidirectional context within blocks.
• Parallel Decoding Pipeline
  Achieves up to 2.5× speedup over standard AR decoding without compromising quality.

🚀 Fast-dLLM v2 uses only ~1B tokens for fine-tuning — a 500× reduction vs. full-attention diffusion LLMs (Dream: 580B tokens) — while matching or surpassing AR baselines in accuracy.

🛠 Model Overview

• Type: Block Diffusion Language Model (dLLM)
• Base Model: Qwen/Qwen2.5-7B-Instruct
• Architecture: Transformer w/ RoPE, SwiGLU activation, RMSNorm, Attention QKV bias
• Params: ~7B
• Layers: 28
• Attention Heads: 28 (Q), 4 (KV, GQA)
• Block Diffusion Size: 32 tokens
• Key Feature: Parallel block-wise decoding + hierarchical caching (block-level & sub-block)

📦 Installation

You will need transformers, torch, and our custom generation function:

pip install transformers torch numpy

🚀 Quick

Your obligations as a Fast dLLM v2 7B deployer

Track your progress

Sign up to track obligations, receive deadline reminders, and generate audit packages.

Tags

Languages

Base Model