Current Diffusion Language Models (DLMs) have been studied at a smaller scale compared to their autoregressive (AR) counterparts and lack fair comparison on language modeling benchmarks. Additionally, training diffusion models from scratch at scale remains challenging. We propose adapting existing AR models to build text diffusion models. We demonstrate connections between AR and diffusion modeling objectives and introduce a simple continual pre-training approach for training diffusion models.
Through systematic evaluation on language modeling, reasoning, and commonsense benchmarks, we show that we can convert AR models ranging from 127M to 7B parameters (GPT2 and LLaMA) into diffusion models DiffuGPT and DiffuLLaMA, using less than 200B tokens for training. Here we open-source adaptation code, efficient fine-tuning scripts, and evaluation toolkits.
The overview of adaptation process.
- 22 Nov 2024. Update evaluation scripts.
Basic environments:
python 3.11
transformers 4.44.2
torch 2.1.1+cu121For quick start:
pip install -r LLaMA-Factory/requirements.txt
pip install flash-attn==2.6.3 --no-build-isolationFlash-attention is not required, but we suggest to use for fast training/inference.
For training and finetuning, we provide our customed LLaMA-Factory based on 0.8.4.dev0.
cd LLaMA-Factory
pip install -e ".[torch,metrics]"📌 Note: We suggest you to set the HuggingFace cache_dir:
export HF_HOME=/path-to-huggingface/cache/. For efficient inference, please install flash-attention-2.
Explanation for folders
├── DiffuLLaMA-training/ # our code to adapt LLaMA2, implemented using transformers, considering that LLaMA-Factory is less efficient for pre-training
├── evaluation/ # evaluation scripts
├── example_output/ # some generation examples
├── LLaMA-Factory/ # our code to adapt GPT, and LoRA finetuning
│ ├── src/llamafactory # integrate diffusion training in LLaMA-Factory
│ │ ├── train/ddm # main folder for diffusion training
│ │ ├── model/loader # loading models
│ │ ├── data/loader # loading data
│ │ ├── attention_patch.py # we modify the way to pass the attention_mask variable
│ ├── examples # config for diffusion training
├── attention_patch.py # we modify the way to pass the attention_mask variable
├── model.py # defined model for quick inference
├── inf_diffugpt.py # quick start inference
├── inf_diffullama.py # quick start inferenceWe provide the example code for inference and you can directly do some unconditional and conditional generation.
### DiffuGPT-small inference example
python inf_diffugpt.py --model_name diffusionfamily/diffugpt-s --base_model_name gpt2
### DiffuGPT-medium inference example
python inf_diffugpt.py --model_name diffusionfamily/diffugpt-m --base_model_name gpt2-medium
### DiffuLLaMA inference example
python inf_diffullama.py --model_name diffusionfamily/diffullama --flash_attn flash_attention_2📌 Note: For
base_model_namehere, we will not download the original model but only use their config. For DiffuLLaMA, we support three types of attention:eager,sdpa,flash_attention_2. The output log can be seen at./example_output/.
We use pre-training corpus from FineWeb, TinyLLaMA.
For DiffuGPT, the example of data processing is in LLaMA-Factory/data/data_prepare.py. We only use 30B tokens from this subset.
python LLaMA-Factory/data/data_prepare.py
cat clean_00000.jsonl clean_00001.jsonl clean_00002.jsonl > fineweb-cc-24-0.jsonlAdd the following field to LLaMA-Factory/data/dataset_info.json.
"fineweb": {
"file_name": "fineweb-cc-24-0.jsonl",
"columns": {
"prompt": "text"
}
}We suggest to pre-tokenize the data, considering the training speed. Example: LLaMA-Factory/examples/train_full/gpt2_preprocess.yaml.
Considering the training efficiency, we choose not to use LLaMA-Factory for 7B pre-training. For DiffuLLaMA, please refer to DiffuLLaMA-training.
Use llamafactory-cli to start training, all configs are listed in LLaMA-Factory/examples/. If you want to use ZeRO, please install deepspeed==0.15.0.
cd LLaMA-Factory
export HF_HOME=yourpath
export WANDB_API_KEY=yourkey
# maybe export cuda path here
# FORCE_TORCHRUN=0 llamafactory-cli train examples/train_full/gpt2_preprocess.yaml
FORCE_TORCHRUN=1 llamafactory-cli train examples/train_full/gpt2_full_ddm.yamlYou can do unconditional generation for adapted diffusion language models.
FORCE_TORCHRUN=1 llamafactory-cli train examples/inference/gpt2_full_ddm-inf.yamlWe can set sampling parameters using logits_temp (top-k temperature) and diffusion_steps (decoding steps).
For DiffuGPT and DiffuLLaMA, we can finetune them to achieve better downstream tasks. For finetuning setting, please use stage: ddm-sft instead of stage: ddm for pre-training. For ddm-sft, we use the way of DiffuSeq to do the diffusion finetuning.
An example of DiffuGPT training config is in examples/train_full/gpt2_full_ddm-sft.yaml. We explain some keywords:
model_name_or_path: gpt2-medium # path or model name to AR base model
checkpoint_dir: output/gpt2-ddm/checkpoint-160000/ # local path to diffugpt
dataset: gsm # the customed data should be registered in `LLaMA-Factory/data/dataset_info.json`
cutoff_len: 256 # No need to set it as the pre-training length
anneal_steps: 1 # Do not use attention anneal for finetune, because the base model is already using bi-directional attention.
shift: trueAn example of DiffuLLaMA config is in examples/train_lora/llama2_lora_ddm-sft.yaml.
model_name_or_path: output/llama2-ddm # local path to diffullama
finetuning_type: lora
lora_target: all # lora params
additional_target: embed_tokens # finetune word embedding
lora_rank: 16 # lora paramsAn example of DiffuLLaMA inference: examples/inference/llama2_full_ddm-gsm-inf.yaml.
adapter_name_or_path: output/llama2-ddm-gsm-v3/checkpoint-64000 # the path to saved lora adapter ckpt📌 Note: If you skip the adapation training and choose to use our released checkpoint to do the finetuning, please specify
checkpoint_dirto the huggingface cache directory to these checkpoints, or using huggingface cli to download checkpoints to a local directory:huggingface-cli download model-name --local-dir /a-local-path/.
We build the zero-shot evaluation, please call the function in main. For example:
eval_hellaswag(model, tokenizer, args)We also open-source the implementation of baseline models Plaid and SEDD.
We sincerely appreciate the following people (works) for DiffuLLaMA:
- This work is built upon the LLaMA2 and GPT2 as the pre-trained models. We also use pre-training corpus from FineWeb, TinyLLaMA.
- We implement training and LoRA tuning based on LLaMAFactory.
- We thank Tianxiao Shen for discussion in the early stage of this project.
- We thank Yu Zhang for engineering discussion on this work.
@misc{gong2024scalingdiffusionlm,
title={Scaling Diffusion Language Models via Adaptation from Autoregressive Models},
author={Shansan Gong and Shivam Agarwal and Yizhe Zhang and Jiacheng Ye and Lin Zheng and Mukai Li and Chenxin An and Peilin Zhao and Wei Bi and Jiawei Han and Hao Peng and Lingpeng Kong},
year={2024},
eprint={2410.17891},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2410.17891},
}
- DiffuLLaMA & DiffuGPT is licensed under the Apache License 2.0. This means that it requires the preservation of copyright and license notices.