🚀 DMax: Aggressive Parallel Decoding for dLLMs

DMax: Aggressive Parallel Decoding for dLLMs
Zigeng Chen, Gongfan Fang, Xinyin Ma, Ruonan Yu, Xinchao Wang
xML Lab, National University of Singapore

💪 Highlights

Aggressive Decoding Parallelism: Achieves 6.0 TPF on math and reasoning tasks and 6.6 TPF on code tasks while preserving accuracy.
Self-Revising dLLM: Extends a pretrained MDLM into a UDLM with an intrinsic ability to revise its own erroneous predictions during decoding.
Soft Parallel Decoding: Uses interpolation between mask and token embeddings to propagate confidence priors from previous steps.

Superior Parallelism-Accuracy Trade-off, Increased TPF with Maintained Accuracy.

💻 Model and Datasets

Model	Description	Source Model	Link
🤖 DMax-Math-16B	Highly parallel dLLM for math and reasoning.	LLaDA-2.0-mini	HF
🤖 DMax-Coder-16B	Highly parallel dLLM for code generation.	LLaDA-2.0-mini	HF

Dataset	Description	Link
📊 DMax-Math-Training-Data	math trajectories generated by LLaDA-2.0-mini	HF
📊 DMax-Code-Training-Data	code trajectories generated by LLaDA-2.0-mini	HF

🚀 Quick Start

import torch
from transformers import AutoModelForCausalLM
from transformers import AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "Zigeng/DMax-Coder-16B", trust_remote_code=True, device_map="cuda:0"
)
model = model.to(torch.bfloat16)
model.eval()
tokenizer = AutoTokenizer.from_pretrained("Zigeng/DMax-Coder-16B", trust_remote_code=True)

prompt = "Write a python function to find the first repeated character in a given string." + "\n\nPlease enclose your code within delimiters as follows:\n```python\n# YOUR CODE HERE\n```\n\n"

input_ids = tokenizer.apply_chat_template(
    [{"role": "user", "content": prompt}],
    add_generation_prompt=True,
    tokenize=True,
    return_tensors="pt",
)

nfe, generated_tokens = model.generate_spd(
    inputs=input_ids,
    gen_length=2048,
    block_length=32,
    threshold=0.65,
)

generated_answer = tokenizer.decode(
    generated_tokens[0],
    skip_special_tokens=True,
)

print(generated_answer)
print("nfe:",nfe,"token length",len(generated_tokens[0]))