Efficient-Large-Model/Fast_dLLM_v2_1.5B

Fast-dLLM v2 (1.5B) — 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 (Qwen2.5-1.5B-Instruct) into a diffusion-style decoder for parallel text generation.

Our approach introduces a novel decoding recipe incorporating a complementary attention mask and block diffusion mechanism, which together enable blockwise bidirectional context modeling while preserving the original AR training objectives and performance. To further enhance inference speed, we design a hierarchical caching mechanism: a block-level cache that stores historical context representations and a sub-block level cache that supports efficient parallel decoding within partially generated blocks.

✨ 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.

[image]

🛠 Model Overview

• —Type: Block Diffusion Language Model (dLLM)
• —Base Model: Qwen/Qwen2.5-1.5B-Instruct
• —Architecture: Transformer w/ RoPE, SwiGLU, RMSNorm, Attention QKV bias, tied embeddings
• —Params: 1.54B (non-embedding: 1.31B)
• —Layers: 28
• —Attention Heads: 12 (Q), 2 (KV, GQA)
• —Key Feature: Parallel block-wise decoding + hierarchical caching

📦 Installation

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

pip install transformers torch numpy

🚀 Quickstart

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Efficient-Large-Model/Fast_dLLM_1.5B"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True
)

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)

prompt = "Give me a short introduction to large language model."
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

# Fast-dLLM v2 parallel decoding
gen_ids = model.generate(
    inputs["input_ids"],
    tokenizer=tokenizer,
    max_new_tokens=512,
    small_block_size=8,
    threshold=0.9,
)

response = tokenizer.decode(
    gen_ids[0][inputs["input_ids"].shape[1]:], 
    skip_special_tokens=True
)
print(response)

📊 Performance & Benchmarks

▶ Real-time Throughput

Fast-dLLM v2 offers up to 2.54× higher throughput than Qwen2.5-7B-Instruct, without loss in quality.

[image]

🏆 Benchmark Results

We compare Fast-dLLM v2 against AR baselines and previous diffusion LLMs on diverse tasks: HumanEval, MBPP (code), GSM8K, Math (reasoning), IFEval (instruction), MMLU, GPQA (knowledge QA).

• —1B group: Fast-dLLM v2 (1.5B) achieves best average score: 45.0.
• —7B group: Fast-dLLM v2 (7B) achieves best average score: 60.3, surpassing LLaDA and Dream models.

[image]

📜 Citation

If you use Fast-dLLM v2 in your research or products, please cite:

@misc{wu2025fastdllmv2efficientblockdiffusion,
      title={Fast-dLLM v2: Efficient Block-Diffusion LLM}, 
      author={Chengyue Wu and Hao Zhang and Shuchen Xue and Shizhe Diao and Yonggan Fu and Zhijian Liu and Pavlo Molchanov and Ping Luo and Song Han and Enze Xie},
      year={2025},
      eprint={2509.26328},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2509.26328}, 
}

📄 License

Released under Apache 2.0, following the base Qwen2.5 license.

🔗 Resources

• —📄 Paper
• —💻 Code
• —🤗 HuggingFace Model