LLaMA Efficient Tuning

👋 Join our WeChat.

Changelog

[23/07/05] Now we support training the Falcon-7B/40B models in this repo. Try --model_name_or_path tiiuae/falcon-7b and --lora_target query_key_value arguments to use the Falcon model.

[23/06/29] We provide a reproducible example of training a chat model using instruction-following datasets, see this HuggingFace Repo for details.

[23/06/22] Now we align the demo API with the OpenAI's format where you can insert the fine-tuned model in arbitrary ChatGPT-based applications.

[23/06/15] Now we support training the baichuan-7B model in this repo. Try --model_name_or_path baichuan-inc/baichuan-7B and --lora_target W_pack arguments to use the baichuan-7B model. If you want to train with RTX3090, use git checkout baichuan-7b-rtx3090 to switch to the baichuan-7b-rtx3090 branch and try the --baichuan_rtx_gpu true argument. (Other RTX series GPUs can also be tried)

[23/06/03] Now we support quantized training and inference (aka QLoRA). Try --quantization_bit 4/8 argument to work with quantized model. (experimental feature)

[23/05/31] Now we support training the BLOOM & BLOOMZ models in this repo. Try --model_name_or_path bigscience/bloomz-7b1-mt and --lora_target query_key_value arguments to use the BLOOMZ model.

Supported Models

LLaMA (7B/13B/33B/65B)
BLOOM & BLOOMZ (560M/1.1B/1.7B/3B/7.1B/176B)
Falcon (7B/40B)
baichuan (7B)

Supported Training Approaches

(Continually) pre-training
- Full-parameter tuning
- Partial-parameter tuning
- LoRA
- QLoRA
Supervised fine-tuning
- Full-parameter tuning
- Partial-parameter tuning
- LoRA
- QLoRA
RLHF
- LoRA
- QLoRA

Provided Datasets

Please refer to data/README.md for details.

Some datasets require confirmation before using them, so we recommend logging in with your HuggingFace account using these commands.

pip install --upgrade huggingface_hub
huggingface-cli login

Requirement

Python 3.8+ and PyTorch 1.13.1+
🤗Transformers, Datasets, Accelerate, PEFT and TRL
jieba, rouge-chinese and nltk (used at evaluation)
gradio and mdtex2html (used in web_demo.py)
uvicorn, fastapi and sse-starlette (used in api_demo.py)

And powerful GPUs!

If you want to enable quantized LoRA (QLoRA) on the Windows platform, you should install a pre-built version of bitsandbytes library, which supports CUDA 11.1 to 12.1.

pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/bitsandbytes-0.39.1-py3-none-win_amd64.whl

Getting Started

Data Preparation (optional)

Please refer to data/example_dataset for checking the details about the format of dataset files. You can either use a single .json file or a dataset loading script with multiple files to create a custom dataset.

Note: please update data/dataset_info.json to use your custom dataset. About the format of this file, please refer to data/README.md.

Dependence Installation (optional)

git clone https://github.com/hiyouga/LLaMA-Efficient-Tuning.git
conda create -n llama_etuning python=3.10
conda activate llama_etuning
cd LLaMA-Efficient-Tuning
pip install -r requirements.txt

LLaMA Weights Preparation (optional)

Download the weights of the LLaMA models.
Convert them to HF format using the following command.

python -m transformers.models.llama.convert_llama_weights_to_hf \
    --input_dir path_to_llama_weights --model_size 7B --output_dir path_to_llama_model

(Continually) Pre-Training

CUDA_VISIBLE_DEVICES=0 python src/train_pt.py \
    --model_name_or_path path_to_your_model \
    --do_train \
    --dataset wiki_demo \
    --finetuning_type lora \
    --output_dir path_to_pt_checkpoint \
    --overwrite_cache \
    --per_device_train_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 5e-5 \
    --num_train_epochs 3.0 \
    --plot_loss \
    --fp16

Supervised Fine-Tuning

CUDA_VISIBLE_DEVICES=0 python src/train_sft.py \
    --model_name_or_path path_to_your_model \
    --do_train \
    --dataset alpaca_gpt4_en \
    --finetuning_type lora \
    --output_dir path_to_sft_checkpoint \
    --overwrite_cache \
    --per_device_train_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 5e-5 \
    --num_train_epochs 3.0 \
    --plot_loss \
    --fp16

Reward Model Training

CUDA_VISIBLE_DEVICES=0 python src/train_rm.py \
    --model_name_or_path path_to_your_model \
    --do_train \
    --dataset comparison_gpt4_en \
    --finetuning_type lora \
    --output_dir path_to_rm_checkpoint \
    --per_device_train_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 1e-5 \
    --num_train_epochs 1.0 \
    --plot_loss \
    --fp16

PPO Training (RLHF)

CUDA_VISIBLE_DEVICES=0 python src/train_ppo.py \
    --model_name_or_path path_to_your_model \
    --do_train \
    --dataset alpaca_gpt4_en \
    --finetuning_type lora \
    --checkpoint_dir path_to_sft_checkpoint \
    --reward_model path_to_rm_checkpoint \
    --output_dir path_to_ppo_checkpoint \
    --per_device_train_batch_size 2 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 1e-5 \
    --num_train_epochs 1.0 \
    --resume_lora_training False \
    --plot_loss

Distributed Training

accelerate config # configure the environment
accelerate launch src/train_XX.py # arguments (same as above)

Example configuration for full-tuning with DeepSpeed ZeRO-2

compute_environment: LOCAL_MACHINE
deepspeed_config:
  gradient_accumulation_steps: 4
  gradient_clipping: 0.5
  offload_optimizer_device: none
  offload_param_device: none
  zero3_init_flag: false
  zero_stage: 2
distributed_type: DEEPSPEED
downcast_bf16: 'no'
machine_rank: 0
main_training_function: main
mixed_precision: fp16
num_machines: 1
num_processes: 4
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false

Evaluation (BLEU and ROUGE_CHINESE)

CUDA_VISIBLE_DEVICES=0 python src/train_sft.py \
    --model_name_or_path path_to_your_model \
    --do_eval \
    --dataset alpaca_gpt4_en \
    --checkpoint_dir path_to_checkpoint \
    --output_dir path_to_eval_result \
    --per_device_eval_batch_size 8 \
    --max_samples 50 \
    --predict_with_generate

We recommend using --per_device_eval_batch_size=1 and --max_target_length 128 at 4/8-bit evaluation.

API / CLI / Web Demo

python src/xxx_demo.py \
    --model_name_or_path path_to_your_model \
    --checkpoint_dir path_to_checkpoint

Export model

python src/export_model.py \
    --model_name_or_path path_to_your_model \
    --checkpoint_dir path_to_checkpoint \
    --output_dir path_to_export

License

This repository is licensed under the Apache-2.0 License.

Please follow the Model Card to use the LLaMA models.

Please follow the RAIL License to use the BLOOM & BLOOMZ models.

Please follow the Apache-2.0 License to use the Falcon models.

Please follow the baichuan-7B License to use the baichuan-7B model.

Citation

If this work is helpful, please cite as:

@Misc{llama-efficient-tuning,
  title = {LLaMA Efficient Tuning},
  author = {hiyouga},
  howpublished = {\url{https://github.com/hiyouga/LLaMA-Efficient-Tuning}},
  year = {2023}
}

Acknowledgement

This repo is a sibling of ChatGLM-Efficient-Tuning. They share a similar code structure of efficient tuning on large language models.

xiang-li-oss / llama-reasoning Goto Github PK

llama-reasoning's Introduction