LLaMA Efficient Tuning

👋 Join our WeChat.

 English | [中文](README_zh.md) 

Changelog

[23/08/18] Now we support resuming training, upgrade transformers to 4.31.0 to enjoy this feature.

[23/08/12] Now we support RoPE scaling to extend the context length of the LLaMA models. Try --rope_scaling linear argument in training and --rope_scaling dynamic argument at inference to extrapolate the position embeddings.

[23/08/11] Now we support DPO training for instruction-tuned models. See this example to train your models (experimental feature).

[23/08/03] Now we support training the Qwen-7B model in this repo. Try --model_name_or_path Qwen/Qwen-7B-Chat and --lora_target c_attn arguments to train the Qwen-7B model. Remember to use --template chatml argument when you are using the Qwen-7B-Chat model.

[23/07/31] Now we support dataset streaming. Try --streaming and --max_steps 10000 arguments to load your dataset in streaming mode.

[23/07/29] We release two instruction-tuned 13B models at Hugging Face. See these Hugging Face Repos (LLaMA-2 / Baichuan) for details.

[23/07/19] Now we support training the LLaMA-2 models in this repo. Try --model_name_or_path meta-llama/Llama-2-7b-hf argument to use the LLaMA-2 model. Remember to use --template llama2 argument when you are using the LLaMA-2-chat model.

[23/07/18] Now we develop an all-in-one Web UI for training, evaluation and inference. Try train_web.py to fine-tune models in your Web browser. Thank @KanadeSiina and @codemayq for their efforts in the development.

[23/07/11] Now we support training the Baichuan-13B model in this repo. Try --model_name_or_path baichuan-inc/Baichuan-13B-Base and --lora_target W_pack arguments to train the Baichuan-13B model. Remember to use --template baichuan argument when you are using the Baichuan-13B-Chat model.

[23/07/09] Now we release FastEdit ⚡🩹, an easy-to-use package for editing the factual knowledge of large language models efficiently. Please follow FastEdit if you are interested.

[23/07/07] Now we support training the InternLM-7B model in this repo. Try --model_name_or_path internlm/internlm-7b argument to use the InternLM model. Remember to use --template intern argument when you are using the InternLM-chat model.

[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 Baichuan-7B-sft 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.

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

[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

Model	Model size	Default module	Template
LLaMA	7B/13B/33B/65B	q_proj,v_proj	-
LLaMA-2	7B/13B/70B	q_proj,v_proj	llama2
BLOOM	560M/1.1B/1.7B/3B/7.1B/176B	query_key_value	-
BLOOMZ	560M/1.1B/1.7B/3B/7.1B/176B	query_key_value	-
Falcon	7B/40B	query_key_value	-
Baichuan	7B/13B	W_pack	baichuan
Baichuan2	7B/13B	W_pack	baichuan2
InternLM	7B	q_proj,v_proj	intern
Qwen	7B	c_attn	chatml
XVERSE	13B	q_proj,v_proj	xverse
ChatGLM2	6B	query_key_value	chatglm2

• Default module is used for the --lora_target argument, you can use --lora_target all to specify all the available modules.
• For the "base" models, the --template argument can be chosen from default, alpaca, vicuna etc. But make sure to use the corresponding template for the "chat" models.

Supported Training Approaches

Approach	Full-parameter	Partial-parameter	LoRA	QLoRA
Pre-Training	✅	✅	✅	✅
Supervised Fine-Tuning	✅	✅	✅	✅
Reward Modeling			✅	✅
PPO Training			✅	✅
DPO Training	✅		✅	✅

• Use --quantization_bit 4/8 argument to enable QLoRA.

Provided Datasets

• For pre-training:
  • Wiki Demo (en)
  • RefinedWeb (en)
  • StarCoder (en)
  • Wikipedia (en)
  • Wikipedia (zh)
• For supervised fine-tuning:
  • Stanford Alpaca (en)
  • Stanford Alpaca (zh)
  • GPT-4 Generated Data (en&zh)
  • Open Assistant (multilingual)
  • Self-cognition (zh)
  • ShareGPT (zh)
  • Guanaco Dataset (multilingual)
  • BELLE 2M (zh)
  • BELLE 1M (zh)
  • BELLE 0.5M (zh)
  • BELLE Dialogue 0.4M (zh)
  • BELLE School Math 0.25M (zh)
  • BELLE Multiturn Chat 0.8M (zh)
  • Firefly 1.1M (zh)
  • LIMA (en)
  • CodeAlpaca 20k (en)
  • Alpaca CoT (multilingual)
  • Web QA (zh)
  • UltraChat (en)
  • WebNovel (zh)
  • Ad Gen (zh)
• For reward modeling or DPO training:
  • HH-RLHF (en)
  • Open Assistant (multilingual)
  • GPT-4 Generated Data (en&zh)

Please refer to data/README.md for details.

Some datasets require confirmation before using them, so we recommend logging in with your Hugging Face 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
• sentencepiece and tiktoken
• jieba, rouge-chinese and nltk (used at evaluation)
• gradio and matplotlib (used in web_demo.py)
• uvicorn, fastapi and sse-starlette (used in api_demo.py)

And powerful GPUs!

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

If you want to enable the quantized LoRA (QLoRA) on the Windows platform, you will be required to 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

All-in-one Web UI

CUDA_VISIBLE_DEVICES=0 python src/train_web.py

We strongly recommend using the all-in-one Web UI for newcomers since it can also generate training scripts automatically.

Currently the web UI only supports training on a single GPU.

Train on a single GPU

Pre-Training

CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
    --stage pt \
    --model_name_or_path path_to_llama_model \
    --do_train \
    --dataset wiki_demo \
    --finetuning_type lora \
    --lora_target q_proj,v_proj \
    --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_bash.py \
    --stage sft \
    --model_name_or_path path_to_llama_model \
    --do_train \
    --dataset alpaca_gpt4_en \
    --template default \
    --finetuning_type lora \
    --lora_target q_proj,v_proj \
    --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 Modeling

CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
    --stage rm \
    --model_name_or_path path_to_llama_model \
    --do_train \
    --dataset comparison_gpt4_en \
    --template default \
    --finetuning_type lora \
    --lora_target q_proj,v_proj \
    --resume_lora_training False \
    --checkpoint_dir path_to_sft_checkpoint \
    --output_dir path_to_rm_checkpoint \
    --per_device_train_batch_size 2 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 1e-6 \
    --num_train_epochs 1.0 \
    --plot_loss \
    --fp16

PPO Training

CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
    --stage ppo \
    --model_name_or_path path_to_llama_model \
    --do_train \
    --dataset alpaca_gpt4_en \
    --template default \
    --finetuning_type lora \
    --lora_target q_proj,v_proj \
    --resume_lora_training False \
    --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 \
    --plot_loss \
    --fp16

DPO Training

CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
    --stage dpo \
    --model_name_or_path path_to_llama_model \
    --do_train \
    --dataset comparison_gpt4_en \
    --template default \
    --finetuning_type lora \
    --lora_target q_proj,v_proj \
    --resume_lora_training False \
    --checkpoint_dir path_to_sft_checkpoint \
    --output_dir path_to_dpo_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 \
    --plot_loss \
    --fp16

Distributed Training

Use Huggingface Accelerate

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

Example config.yaml for training 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

Use DeepSpeed

deepspeed --num_gpus 8 --master_port=9901 src/train_bash.py \
    --deepspeed ds_config.json \
    ... # arguments (same as above)

Example ds_config.json for training with DeepSpeed ZeRO-2

{
  "train_micro_batch_size_per_gpu": "auto",
  "gradient_accumulation_steps": "auto",
  "gradient_clipping": "auto",
  "zero_allow_untested_optimizer": true,
  "fp16": {
    "enabled": "auto",
    "loss_scale": 0,
    "initial_scale_power": 16,
    "loss_scale_window": 1000,
    "hysteresis": 2,
    "min_loss_scale": 1
  },  
  "zero_optimization": {
    "stage": 2,
    "allgather_partitions": true,
    "allgather_bucket_size": 5e8,
    "reduce_scatter": true,
    "reduce_bucket_size": 5e8,
    "overlap_comm": false,
    "contiguous_gradients": true
  }
}

Export model

python src/export_model.py \
    --model_name_or_path path_to_llama_model \
    --template default \
    --finetuning_type lora \
    --checkpoint_dir path_to_checkpoint \
    --output_dir path_to_export

API Demo

python src/api_demo.py \
    --model_name_or_path path_to_llama_model \
    --template default \
    --finetuning_type lora \
    --checkpoint_dir path_to_checkpoint

Visit http://localhost:8000/docs for API documentation.

CLI Demo

python src/cli_demo.py \
    --model_name_or_path path_to_llama_model \
    --template default \
    --finetuning_type lora \
    --checkpoint_dir path_to_checkpoint

Web Demo

python src/web_demo.py \
    --model_name_or_path path_to_llama_model \
    --template default \
    --finetuning_type lora \
    --checkpoint_dir path_to_checkpoint

Evaluation (BLEU and ROUGE_CHINESE)

CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
    --stage sft \
    --model_name_or_path path_to_llama_model \
    --do_eval \
    --dataset alpaca_gpt4_en \
    --template default \
    --finetuning_type lora \
    --checkpoint_dir path_to_checkpoint \
    --output_dir path_to_eval_result \
    --per_device_eval_batch_size 8 \
    --max_samples 100 \
    --predict_with_generate

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

Predict

CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
    --stage sft \
    --model_name_or_path path_to_llama_model \
    --do_predict \
    --dataset alpaca_gpt4_en \
    --template default \
    --finetuning_type lora \
    --checkpoint_dir path_to_checkpoint \
    --output_dir path_to_predict_result \
    --per_device_eval_batch_size 8 \
    --max_samples 100 \
    --predict_with_generate

TODO

• Supporting flash attention (torch / xformers / flashattn).
• Implementing multi-query attention for faster inference.
• Supporting full-parameter RLHF training.

License

This repository is licensed under the Apache-2.0 License.

Please follow the model licenses to use the corresponding model weights:

• LLaMA
• LLaMA-2
• BLOOM
• Falcon
• Baichuan
• Baichuan2
• InternLM
• Qwen
• XVERSE
• ChatGLM2

Citation

If this work is helpful, please kindly 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.

Star History