# <font color=#0000F0>FSJD</font>: Accelerating Auto-regressive Text-to-Image Generation with Training-free <br><font color=#0000F0>S</font>peculative <font color=#0000F0>J</font>acobi <font color=#0000F0>D</font>ecoding


## 🚩 注意事项
如果你把workdir生成在了本地,不要把workdir的内容push到仓库!!!!

## 🚩 New Features/Updates

- ✅ Apr, 2025. 💥 **<font color=#0000F0>SJD</font>** has been integrated into [Lumina-mGPT2](https://github.com/Alpha-VLLM/Lumina-mGPT-2.0) and [SimpleAR](https://github.com/wdrink/SimpleAR).
- ✅ Jan, 2025. 💥 **<font color=#0000F0>SJD</font>** is accepted to ICLR 2025.
- ✅ Oct, 2024. Release **<font color=#0000F0>SJD</font>**'s code.

## 🚩 TODO List

- □ Integrating SJD into vLLM framework for further acceleration.

## Installing the dependencies

##### Environment: 

- Python 3.10
- CUDA 12.5
- Pytorch 2.5.1+cu124
- Transformers 4.47.1

##### Install from `yaml`:

```bash
conda env create -f environment.yaml
```

## 🚩 如果你想完成完整的数据集测试?

### target model: lumina-gpt

```
cd ./fsjd
isp='random' #init的填充方式
method='speculative_jacobi' #sjd
num_init_new_token=16 #填充数量
benchmark_way='order' #测试集的载入方式
prompt="MSCOCO2017Val" #测试集的名字
num_images=2 #测试集的图片数量
slice='0-1000' #测试集的区间
output_path=/home/leihaodong/AAAI25/exp/FSJD/${prompt}/${method}# 测试结果的输出文件夹

python /home/leihaodong/AAAI25/fsjd/main.py \
  --output_path=$output_path \
  --isp=$isp \
  --method=$method \
  --num_init_new_token=$num_init_new_token \
  --benchmark_way=$benchmark_way \
  --prompt=$prompt \
  --num_images=$num_images \
  --slice=$slice
```
### target model: lumina-gpt

benchmark 为 parti-prompt

```
cd /home/leihaodong/AAAI25/fsjd
gpu_ids=0,3
dataset_name=parti
dataset_anno_file=/home/leihaodong/AAAI25/fsjd/data/prompts/PartiPrompts.tsv
model_name=leloy/Anole-7b-v0.1-hf
max_num_new_tokens=16
multi_token_init_scheme=random
seed=42
target_size=1024
prefix_token_sampler_scheme=speculative_jacobi
temperature=1.0

output_path=/home/leihaodong/AAAI25/exp/SJD/${dataset_name}_anole_${temperature}/${prefix_token_sampler_scheme}
mkdir -p ${output_path}

nohup python eval_model.py \
 --gpu_ids $gpu_ids \
 --dataset_name $dataset_name \
 --dataset_anno_file $dataset_anno_file \
 --model_name $model_name \
 --max_num_new_tokens $max_num_new_tokens \
 --multi_token_init_scheme $multi_token_init_scheme \
 --seed $seed \
 --target_size $target_size \
 --prefix_token_sampler_scheme $prefix_token_sampler_scheme \
 --temperature $temperature \
 --output_dir $output_path \
 --return_accl True \
 --num_images 0 > ${output_path}.log 2>&1 &
```

#### Lumina-mGPT

```bash
CUDA_VISIBLE_DEVICES=0 python test_lumina_mgpt.py
```

#### Emu3

```bash
CUDA_VISIBLE_DEVICES=0 python test_emu3.py
```

#### LlamaGen

```bash
CUDA_VISIBLE_DEVICES=0 python test_llamagen.py
```

## Acknowledge

Our code is based on [Lumina-mGPT](https://github.com/Alpha-VLLM/Lumina-mGPT), [Emu3](https://github.com/Alpha-VLLM/Lumina-mGPT), [LlamaGen](https://github.com/FoundationVision/LlamaGen), [Anole](https://github.com/GAIR-NLP/anole), and [CLLM](https://github.com/hao-ai-lab/Consistency_LLM). We would like to express our gratitude to [Tianwei Xiong](https://github.com/SilentView) for his assistance.

## Citation

```bibtex
@article{teng2024accelerating,
  title={Accelerating auto-regressive text-to-image generation with training-free speculative jacobi decoding},
  author={Teng, Yao and Shi, Han and Liu, Xian and Ning, Xuefei and Dai, Guohao and Wang, Yu and Li, Zhenguo and Liu, Xihui},
  journal={arXiv preprint arXiv:2410.01699},
  year={2024}
}
```