1. News
2. Highlights
3. How to contribute
4. Requirements
5. Installation and quick guide
6. LongAlpaca Data
7. Models
8. Training
9. Evaluation
10. Demo
11. Streaming Inference
12. Data Generation via Pdf2Text
13. Examples
14. Citation
15. Acknowledgement
16. License

• [2024.1.17] LongLoRA has been accepted by ICLR 2024 as an Oral presentation.
• [2023.11.19] We release a new version of LongAlpaca models, LongAlpaca-7B-16k, LongAlpaca-7B-16k, and LongAlpaca-7B-16k. These models are fine-tuned on a subset LongAlpaca-12k dataset with LongLoRA in SFT, LongAlpaca-16k-length. We evaluate the LongAlpaca-7B-16k model on LongBench and L-Eval benchmarks and results can be found here.
• [2023.11.2] We have updated our LongAlpaca models from alpaca prompting to llama2 prompting, which is consistent to their pre-trained models. Please refer to the inference code with the llama2 prompting.
• [2023.10.23] We support the combination of QLoRA and LongLoRA in the supervised fine-tuning, for further reduction of the GPU memory cost. We release the LoRA weights of a 7B model at LongAlpaca-7B-qlora-weights.
• [2023.10.18] We support StreamingLLM inference on our LongAlpaca models. This increases the context-length of the multi-round dialogue in StreamingLLM.
• [2023.10.8] We release the long instruction-following dataset, LongAlpaca-12k and the corresponding models, LongAlpaca-7B, LongAlpaca-13B, and LongAlpaca-70B.
• (The previous sft models, Llama-2-13b-chat-longlora-32k-sft and Llama-2-70b-chat-longlora-32k-sft, have been deprecated.)
• [2023.10.3] We add support GPTNeoX models. Please refer to this PR for usage. Thanks for @naubull2 for this contribution.
• [2023.9.22] We release all our fine-tuned models, including 70B-32k models, LLaMA2-LongLoRA-70B-32k, LLaMA2-LongLoRA-7B-100k. Welcome to check them out!
• [2023.9.22] We release Paper and this GitHub repo, including training and evaluation code.

LongLoRA: Efficient Fine-tuning of Long-Context Large Language Models [Paper]
Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhijian Liu, Song Han, Jiaya Jia

1. In LongLoRA approach, The proposed shifted short attention is easy to implement, compatible with Flash-Attention, and is not required during inference.
2. We released all our models, including models from 7B to 70B, context length from 8k to 100k, including LLaMA2-LongLoRA-7B-100k, LLaMA2-LongLoRA-13B-64k, and LLaMA2-LongLoRA-70B-32k.
3. We built up a long-context instruction-following dataset, LongAlpaca-12k. We released the corresponding LongAlpaca-7B, LongAlpaca-13B and LongAlpaca-70B models. To our best knowledge, this is the first open-sourced long-context 70B model.

• Make sure to have git installed.
• Create your own fork of the project.
• Clone the repository on your local machine, using git clone and pasting the url of this project.
• Read both the Requirements and Installation and Quick Guide sections below.
• Commit and push your changes.
• Make a pull request when finished modifying the project.

To download and use the pre-trained weights you will need:

1. Hugging Face (HF) account with valid email. Note, the email used for HF must alse be used for the license agreement.
2. Accept the Meta license and acceptable use policy

To install and run the application:

1. Fork this repo on github
2. Clone the repository on your local machine, using git clone and pasting the url of this project.
3. Run the following code:

pip install -r requirements.txt
pip install flash-attn --no-build-isolation

4. Use either a Released model or Fine tune a model to fit your preferences.
5. Test your model by chat.
6. Deploy your own demo.

LongAlpaca-12k contains 9k long QA data that we collected and 3k short QA sampled from the original Alpaca data. This is to avoid the case that the model might degrade at short instruction following. The data we collect contains various types and amounts as the following figure.

Following the original Alpaca format, our Long QA data uses the following prompts for fine-tuning:

• instruction: str, describes the task the model should perform. For example, to answer a question after reading a book section or paper. We vary the contents and questions to make instructions diverse.
• output: str, the answer to the instruction.

We did not use the input format in the Alpaca format for simplicity.

We use LLaMA2 models as the pre-trained weights and fine-tune them to long context window sizes. Download based on your choices.

This project also supports GPTNeoX models as the base model architecture. Some candidate pre-trained weights may include GPT-NeoX-20B, Polyglot-ko-12.8B and other variants.

torchrun --nproc_per_node=8 fine-tune.py  \
        --model_name_or_path path_to/Llama-2-7b-hf \
        --bf16 True \
        --output_dir path_to_saving_checkpoints       \
        --cache_dir path_to_cache \
        --model_max_length 8192 \
        --use_flash_attn True \
        --low_rank_training False \
        --num_train_epochs 1  \
        --per_device_train_batch_size 1     \
        --per_device_eval_batch_size 2     \
        --gradient_accumulation_steps 8     \
        --evaluation_strategy "no"     \
        --save_strategy "steps"     \
        --save_steps 1000     \
        --save_total_limit 2     \
        --learning_rate 2e-5     \
        --weight_decay 0.0     \
        --warmup_steps 20     \
        --lr_scheduler_type "constant_with_warmup"     \
        --logging_steps 1     \
        --deepspeed "ds_configs/stage2.json" \
        --tf32 True \
        --max_steps 1000

• Please remember to change path_to/Llama-2-7b-hf, path_to_saving_checkpoints, path_to_cache to your own directory.
• Note that you can change model_max_length to other values.
• You could change ds_configs/stage2.json to ds_configs/stage3.json if you want.
• Please set use_flash_attn as False if you use V100 machines or do not install flash attention.
• You can set low_rank_training as False if you want to use fully fine-tuning. It will cost more GPU memory and slower, but the performance will be a bit better.
• When training is finished, to get the full model weight:

cd path_to_saving_checkpoints && python zero_to_fp32.py . pytorch_model.bin

Note that the path_to_saving_checkpoints might be the global_step directory, which depends on the deepspeed versions.

torchrun --nproc_per_node=8 supervised-fine-tune.py  \
        --model_name_or_path path_to_Llama2_chat_models \
        --bf16 True \
        --output_dir path_to_saving_checkpoints       \
        --model_max_length 16384 \
        --use_flash_attn True \
        --data_path LongAlpaca-16k-length.json \
        --low_rank_training True \
        --num_train_epochs 5  \
        --per_device_train_batch_size 1     \
        --per_device_eval_batch_size 2     \
        --gradient_accumulation_steps 8     \
        --evaluation_strategy "no"     \
        --save_strategy "steps"     \
        --save_steps 98     \
        --save_total_limit 2     \
        --learning_rate 2e-5     \
        --weight_decay 0.0     \
        --warmup_steps 20     \
        --lr_scheduler_type "constant_with_warmup"     \
        --logging_steps 1     \
        --deepspeed "ds_configs/stage2.json" \
        --tf32 True

• There is no need to make supervised fine-tuning upon the fine-tuned context extended models. It is all right to directly use base model as Llama2-chat models, as the amount of long instruction following data is enough for SFT.
• Our long instruction following data can be found in LongAlpaca-12k.json.
• Note that supervised-fine-tune.py can be replaced by supervised-fine-tune-qlora.py if you want to try 4-bit quantized fine-tuning for further GPU memory reduction. This follows QLoRA.
• If you meet issue for saving pytorch_model.bin after the qlora sft, please refer to this issue.

In low-rank training, we set embedding and normalization layers as trainable. Please use the following line to extract the trainable weights trainable_params.bin from pytorch_model.bin

python3 get_trainable_weights.py --checkpoint_path path_to_saving_checkpoints --trainable_params "embed,norm"

Merge the LoRA weights of pytorch_model.bin and trainable parameters trainable_params.bin, save the resulting model into your desired path in the Hugging Face format:

python3 merge_lora_weights_and_save_hf_model.py \
        --base_model path_to/Llama-2-7b-hf \
        --peft_model path_to_saving_checkpoints \
        --context_size 8192 \
        --save_path path_to_saving_merged_model

For example,

python3 merge_lora_weights_and_save_hf_model.py \
        --base_model /dataset/pretrained-models/Llama-2-7b-hf \
        --peft_model /dataset/yukangchen/hf_models/lora-models/Llama-2-7b-longlora-8k \
        --context_size 8192 \
        --save_path /dataset/yukangchen/models/Llama-2-7b-longlora-8k-merged

To evaluate a model that is trained in the low-rank setting, please set both base_model and peft_model. base_model is the pre-trained weight. peft_model is the path to the saved checkpoint, which should contain trainable_params.bin, adapter_model.bin and adapter_config.json. For example,

python3 eval.py --seq_len 8192 --context_size 8192 --batch_size 1 --base_model path_to/Llama-2-7b-hf --peft_model path_to_saving_checkpoints --data_path pg19/test.bin

Or evaluate with multiple GPUs as follows.

torchrun --nproc_per_node=auto eval_distributed.py --seq_len 8192 --context_size 8192 --batch_size 1 --base_model path_to/Llama-2-7b-hf --peft_model path_to_saving_checkpoints --data_path pg19/test.bin

To evaluate a model that is fully fine-tuned, you only need to set base_model as the path to the saved checkpoint, which should contain pytorch_model.bin and config.json. peft_model should be ignored.

python3 eval.py --seq_len 8192 --context_size 8192 --batch_size 1 --base_model path_to_saving_checkpoints --data_path pg19/test.bin

Or evaluate with multiple GPUs as follows.

torchrun --nproc_per_node=auto eval_distributed.py --seq_len 8192 --context_size 8192 --batch_size 1 --base_model path_to_saving_checkpoints --data_path pg19/test.bin

• Note that --seq_len is to set the sequence length for evaluation. --context_size is to set the context length of the model during fine-tuning. --seq_len should not be larger than --context_size.

• We have already tokenized the validation and test splits of PG19 and proof-pile dataset into pg19/validation.bin, pg19/test.bin, and proof-pile/test_sampled_data.bin, with the tokenizer of LLaMA. proof-pile/test_sampled_data.bin contains 128 documents that are randomly sampled from the total proof-pile test split. For each document, it has at least 32768 tokens. We also release the sampled ids in proof-pile/test_sampled_ids.bin. You can download them from the links below.

We provide a manner to test the passkey retrieval accuracy. For example,

python3 passkey_retrivial.py \
        --context_size 32768 \
        --base_model path_to/Llama-2-7b-longlora-32k \
        --max_tokens 32768 \
        --interval 1000

• Note that the context_size is the context length during fine-tuning.
• max_tokens is maximum length for the document in passkey retrieval evaluation.
• interval is the interval during the document length increasing. It is a rough number because the document increases by sentences.

To chat with LongAlpaca models,

python3 inference.py  \
        --base_model path_to_model \
        --question $question \
        --context_size $context_length \
        --max_gen_len $max_gen_len \
        --flash_attn True \
        --material $material_content

To ask a question related to a book:

python3 inference.py  \
        --base_model /data/models/LongAlpaca-13B \
        --question "Why doesn't Professor Snape seem to like Harry?" \
        --context_size 32768 \
        --max_gen_len 512 \
        --flash_attn True \
        --material "materials/Harry Potter and the Philosophers Stone_section2.txt"

To ask a question related to a paper:

python3 inference.py  \
        --base_model /data/models/LongAlpaca-13B \
        --question "What are the main contributions and novelties of this work?" \
        --context_size 32768 \
        --max_gen_len 512 \
        --flash_attn True \
        --material "materials/paper1.txt"

• Note that inference.py can be replaced by inference-qlora.py if you want to try 4-bit quantized fine-tuning for further GPU memory reduction. This follows QLoRA.

To deploy your own demo run

python3 demo.py  \
	--base_model path_to_model \
	--context_size $context_size \
	--max_gen_len $max_gen_len \
	--flash_attn True

Example

python3 demo.py  \
	--base_model /data/models/LongAlpaca-13B \
	--context_size 32768 \
	--max_gen_len 512 \
	--flash_attn True

• Note that flash_attn=True will make the generation slow but save much GPU memory.

We support the inference of LongAlpaca models with StreamingLLM. This increases the context-length of the multi-round dialogue in StreamingLLM. Here is an example,

python run_streaming_llama_longalpaca.py \
	----enable_streaming \
	--test_filepath outputs_stream.json \
	--use_flash_attn True \
	--recent_size 32768

• Note that please use a smaller recent_size if you meet OOM issues, for example 8192.
• test_filepath is the json file that contains prompts for inference. We provide an example file outputs_stream.json, which is a subset of LongAlpaca-12k. You can replace it to your own questions.

During our dataset collection, we convert paper and books from pdf to text. The conversion quality has a large influence on the final model quality. We think that this step is non-trivial. We release the tool for the pdf2txt conversion, in the folder pdf2txt. It is built upon pdf2image, easyocr, ditod and detectron2. Please refer to the README.md in pdf2txt for more details.

If you find this project useful in your research, please consider citing:

@inproceedings{longlora,
  author       = {Yukang Chen and Shengju Qian and Haotian Tang and Xin Lai and Zhijian Liu and Song Han and Jiaya Jia},
  title        = {LongLoRA: Efficient Fine-tuning of Long-Context Large Language Models},
  booktitle    = {The International Conference on Learning Representations (ICLR)},
  year         = {2024},
}

@misc{long-alpaca,
  author = {Yukang Chen and Shaozuo Yu and Shengju Qian and Haotian Tang and Xin Lai and Zhijian Liu and Song Han and Jiaya Jia},
  title = {Long Alpaca: Long-context Instruction-following models},
  year = {2023},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/dvlab-research/LongLoRA}},
}

• This work is built upon the LLaMA2 as the pre-trained models.
• This work can also be built upon the GPTNeoX-HF which is based upon EleutherAI/GPTNeoX as the pre-trained model architecture.
• This work is based on DeepSpeed, peft, and Flash-Attention2 for acceleration.
• Some evaluation code is modified upon Landmark Attention.
• We use LongChat for the retrieval evaluation.
• We follow StreamingLLM for streaming inference.
• We combine QLoRA with LongLoRA for supervised fine-tuning.

• LongLoRA is licensed under the Apache License 2.0. This means that it requires the preservation of copyright and license notices.
• Data and weights are under CC-BY-NC 4.0 License. They are licensed for research use only, and allowed only non-commercial. Models trained using the dataset should not be used outside of research purposes.