LLMCL-DeepSpeed

LLMCL-DeepSpeed is a framewrok that uses DeepSpeed to implement several classical continual learning algorithms of large language models.

Features

• Classical incremental learning methods: we implement some of the classical continual learning algorithms.
• Simple-to-use code: code are simple, NOT use Huggingface Trainer API, you are in-charge of all the training process!
• DeepSpeed integration for dsitributed: we use DeppSpeed to support milti-GPU training.

Quick Start

Installation

conda create -n llmcl python==3.10
pip install -r requirements.txt

Prepare Data

refer llmcl/get_dataset.py to prepare your own data.\

• make sure you dataset locates like:

  data_name
  ├── test.json
  └── train.json
  

• you may need customize your own data_process code llmcl/get_dataset.py `format_data` to deal with your own dataset.

Training

LLMCL-DeepSpeed implements deepspeed zero-2 training, config files can be found llmcl/ds_config/ds_zero2.json

port=$(shuf -i25000-30000 -n1)
export PYTHONPATH=$(pwd)/llmcl:$PYTHONPATH
cl_method=gem
output_dir=output
deepspeed --include=localhost:0,1 --master_port $port llmcl/main.py --deepspeed \
   --seed 42 \
   --data_path data/LLM-CL-Benchmark_5000 \
   --dataset_names C-STANCE,FOMC \
   --model_name_or_path ./Qwen2.5-7B-Instruct \
   --logging_steps 100 \
   --eval_steps 50 \
   --per_device_train_batch_size 6 \
   --gradient_accumulation_steps 8 \
   --per_device_eval_batch_size 8 \
   --learning_rate 1e-4 \
   --num_train_epochs 3 \
   --warmup_ratio 0.03 \
   --ewc_lambda 100 \
   --cl_method $cl_method \
   --logging_dir $output_dir/${cl_method}/tensorboard_logs \
   --output_dir $output_dir/${cl_method} \
   --deepspeed_config llmcl/ds_config/ds_zero2.json > logs/train_${cl_method}.log 2>&1 &

NOTE
currently not support deepspeed zero3 training and multi-node training.