paper | huggingface

See also:

• DPKD: A simple improvement of MiniLLM using DPO.
• MiniPLM: Knowledge distillation for pre-training lanuage models.

pip3 install git+https://github.com/t1101675/transformers@minillm
pip3 install torch
pip3 install deepspeed
pip3 install numerize
pip3 install rouge-score
pip3 install torchtyping
pip3 install rich
pip3 install accelerate
pip3 install datasets
pip3 install peft

or

bash install.sh

Our data and pre-trained models are uploaded to our HuggingFace repo. We modified the transformers code base to support model (tensor) parallel and teacher-mixed sampling. The modified lines are wrapped with

# ### MiniLLM BEGIN ###
... SOME NEW CODES ...
# ### MiniLLM END ###

• The training/evaluation intruction-response data before processing can be downloaded from the following links: dolly, self-inst, vicuna, sinst, and uinst

  huggingface-cli download MiniLLM/dolly --repo-type dataset /PATH_TO/LMOps/minillm/data/dolly/
  huggingface-cli download MiniLLM/self-inst --repo-type dataset /PATH_TO/LMOps/minillm/data/self-inst/
  huggingface-cli download MiniLLM/Vicuna --repo-type dataset /PATH_TO/LMOps/minillm/data/vicuna/
  huggingface-cli download MiniLLM/sinst --repo-type dataset /PATH_TO/LMOps/minillm/data/sinst/
  huggingface-cli download MiniLLM/uinst --repo-type dataset /PATH_TO/LMOps/minillm/data/uinst/

• (Optional) The plain-text corpus $\mathcal{D}_\text{PT}$ can be download from the HugginFace datasets repository. For reproducibility, we recommend you to use the following preprocessed data.
• The processed data can be downloaded from the following links: dolly, openwebtext (Optional), roberta-corpus (Optional).

  huggingface-cli download MiniLLM/dolly-processed --repo-type dataset --local-dir /PATH_TO/LMOps/minillm/processed_data/dolly/
  huggingface-cli download MiniLLM/openwebtext-processed --repo-type dataset --local-dir /PATH_TO/LMOps/minillm/processed_data/openwebtext/gpt2/512/10M/ # Optional
  huggingface-cli download MiniLLM/roberta-corpus-processed --repo-type dataset --local-dir /PATH_TO/LMOps/minillm/processed_data/openwebtext/ # Optional

bash scripts/gpt2/tools/process_data_dolly.sh /PATH_TO/LMOps/minillm # Process Dolly Train / Validation Data
bash scripts/opt/tools/process_data_dolly.sh /PATH_TO/LMOps/minillm # Process Dolly Train / Validation Data
bash scripts/llama/tools/process_data_dolly.sh /PATH_TO/LMOps/minillm # Process Dolly Train / Validation Data

Get plain-text corpus $\mathcal{D}_\text{PT}$:

python3 tools/get_openwebtext.py

This script will replace the continuous \n in each document with a special token "<@x(x!>" and write each document in OpenWebText in a line, which is covenient for parallel processing. In data/openwebtext/data.txt, we give an example of the resulting format. You can follow this format to prepare other corpus beyond OpenWebText.

Tokenize the data and store them in binary files:

bash scripts/gpt2/tools/process_data_pretrain.sh /PATH_TO/LMOps/minillm # Process OpenWebText Train / Validation Data
bash scripts/opt/tools/process_data_pretrain.sh /PATH_TO/LMOps/minillm # Process RoBERTa Corpus Train / Validation Data
bash scripts/llama/tools/process_data_pretrain.sh /PATH_TO/LMOps/minillm # Process RoBERTa Corpus Train / Validation Data

• The pre-trained models (MiniLLM and the baselines) can be found in this collection.

To run fine-tuning or standard KD baselines, you need to download the model checkpoints from [Huggingface Model Hub] and put them in checkpoints/. For example, for gpt2-large, you can download the model from this link and put them in checkpoints/gpt2-large.

huggingface-cli download gpt2 --repo-type model /PATH_TO/LMOps/minillm/checkpoints/gpt2-base
huggingface-cli download gpt2-medium --repo-type model /PATH_TO/LMOps/minillm/checkpoints/gpt2-medium
huggingface-cli download gpt2-large --repo-type model /PATH_TO/LMOps/minillm/checkpoints/gpt2-large
huggingface-cli download gpt2-xl --repo-type model /PATH_TO/LMOps/minillm/checkpoints/gpt2-xlarge

Alternatively, you can also change the CKPT variable in each script to the corresponding model name to enable Transformers to download the base models automatically. For example, set CKPT="gpt2-large" in scripts/gpt2/sft/sft_large.sh causes download of the gpt2-large base model from the HugginFace model hub.

NOTE:

1. LLaMA models require license and cannot be directly downloaded.
2. If you want to use model parallel for training, it is recommended to download the models to checkpoints because you need to run tools/convert_mp.py to change their model parallel sizes (see next section).

If you find the model is too large to fit in your GPUs, you can increase/decrease the tensor parallel sizes with

python3 tools/convert_mp.py \
    --input_path results/llama/train/minillm/7B-init-13B-sft \
    --source_mp_size 1 \
    --target_mp_size 4 \
    --model_type llama # choose from opt and llama

To use the model with Model Parallel, we provide two example scripts for training and evaluation.

NOTE: Model parallelism is not applied to gpt2 because these models are generally sufficiant small to fit in common GPUs.

bash scripts/gpt2/eval/run_eval.sh /PATH_TO/LMOps/minillm
bash scripts/opt/eval/run_eval.sh /PATH_TO/LMOps/minillm
bash scripts/llama/eval/run_eval.sh /PATH_TO/LMOps/minillm

We provide example commands for GPT-2 models. Similar scripts for model families can be found in scripts/opt and scripts/llama. All our experiments are conducted on 16 * 32V100, which can be reduced for small models. Some large models require tensor parallel size = 4, which is set in the scripts with --model-parallel and --model-parallel-size options.

The final checkpoints are selected by the Rouge-L scores.

bash scripts/gpt2/sft/sft_xlarge.sh /PATH_TO/LMOps/minillm

bash scripts/gpt2/sft/sft_base.sh /PATH_TO/LMOps/minillm
bash scripts/gpt2/sft/sft_medium.sh /PATH_TO/LMOps/minillm
bash scripts/gpt2/sft/sft_large.sh /PATH_TO/LMOps/minillm

bash scripts/gpt2/kd/kd_base.sh /PATH_TO/LMOps/minillm
bash scripts/gpt2/kd/kd_medium.sh /PATH_TO/LMOps/minillm
bash scripts/gpt2/kd/kd_large.sh /PATH_TO/LMOps/minillm

Generate and process responses with the teacher:

bash scripts/gpt2/tools/generate_data_seqkd.sh /PATH_TO/LMOps/minillm
bash scripts/gpt2/tools/process_pseudo_data_seqkd.sh /PATH_TO/LMOps/minillm

Fine-tune the model with SeqKD:

bash scripts/gpt2/seqkd/seqkd_base.sh /PATH_TO/LMOps/minillm
bash scripts/gpt2/seqkd/seqkd_medium.sh /PATH_TO/LMOps/minillm
bash scripts/gpt2/seqkd/seqkd_large.sh /PATH_TO/LMOps/minillm

The final checkpoints are selected by the validation loss.

bash scripts/gpt2/sft/sft_base.sh /PATH_TO/LMOps/minillm
bash scripts/gpt2/sft/sft_medium.sh /PATH_TO/LMOps/minillm
bash scripts/gpt2/sft/sft_large.sh /PATH_TO/LMOps/minillm

The final checkpoints are selected by the Rouge-L scores.

bash scripts/gpt2/minillm/train_base_xl.sh /PATH_TO/LMOps/minillm
bash scripts/gpt2/minillm/train_medium_xl.sh /PATH_TO/LMOps/minillm
bash scripts/gpt2/minillm/train_large_xl.sh /PATH_TO/LMOps/minillm

For the data we use:

• PROMPT_DATA_DIR is the SFT data ($\mathcal{D}$, Dolly), which is required.
• LM_DATA_DIR is the plain-text corpus ($\mathcal{D}_\text{PT}$), which is optional. See minillm/scripts/gpt2/minillm/train_base_xl_no_pt.sh for training without LM_DATA_DIR (by just commenting out the OPTS+=" --lm-data-dir ${LM_DATA_DIR}" line).

Multi-Node training is launched by deepspeed. We provide an example script in scripts/llama/sft/sft_7B_mn.sh for multi-node training. Compared to single-node scripts, some of the DISTRIBUTED_ARGS are changed, and you need to specify a hostfile like configs/hostfiles/node_0_1 to tell the script which nodes to use. For more information, please refer to HuggingFace's tutorial.

@inproceedings{minillm,
  title={MiniLLM: Knowledge Distillation of Large Language Models},
  author={Gu, Yuxian and Dong, Li and Wei, Furu and Huang, Minlie},
  booktitle={Proceedings of ICLR},
  year={2024}
}