| # Shortened LLM Model Card | |
| Shortened LLM is a depth-pruned version of large language models for efficient text generation. | |
| - **Developed by:** [Nota AI](https://www.nota.ai/) | |
| - **License:** Non-commercial license | |
| - **Repository:** https://github.com/Nota-NetsPresso/shortened-llm | |
| - **Paper:** https://arxiv.org/abs/2402.02834 | |
| ## Compression Method | |
| * After identifying unimportant Transformer blocks, we perform **one-shot pruning**. | |
| * In retraining pruned models for quality recovery, we leverage **continued pretraining (CPT)**, which involves updating all parameters, on a large-scale pretraining corpus. | |
| * Once CPT is completed, the model in this card is further finetuned with **low-rank adaptation (LoRA)** on an instruction tuning dataset. | |
| ## Models from Aggressive Pruning & CPT Retraining (arXiv-v2): | |
| | Source<br>Model | Pruning<br>Ratio | Pruning<br>Criterion | Retraining<br>Method | HF Models<br>Link | | |
| |:---:|:---:|:---:|:---:| :---:| | |
| | Vicuna-v1.3-7B | 20% | PPL | CPT | [nota-ai/cpt_st-vicuna-v1.3-5.5b-ppl](https://huggingface.co/nota-ai/cpt_st-vicuna-v1.3-5.5b-ppl) | | |
| | Vicuna-v1.3-7B | 45% | PPL | CPT | [nota-ai/cpt_st-vicuna-v1.3-3.7b-ppl](https://huggingface.co/nota-ai/cpt_st-vicuna-v1.3-3.7b-ppl) | | |
| | Vicuna-v1.3-7B | 60% | PPL | CPT | [nota-ai/cpt_st-vicuna-v1.3-2.7b-ppl](https://huggingface.co/nota-ai/cpt_st-vicuna-v1.3-2.7b-ppl) | | |
| | Vicuna-v1.3-7B | 80% | PPL | CPT | [nota-ai/cpt_st-vicuna-v1.3-1.5b-ppl](https://huggingface.co/nota-ai/cpt_st-vicuna-v1.3-1.5b-ppl) | | |
| | Vicuna-v1.3-7B | 20% | PPL | CPT⇒LoRA | [nota-ai/cpt-lora_st-vicuna-v1.3-5.5b-ppl](https://huggingface.co/nota-ai/cpt-lora_st-vicuna-v1.3-5.5b-ppl) | | |
| | Vicuna-v1.3-7B | 45% | PPL | CPT⇒LoRA | [nota-ai/cpt-lora_st-vicuna-v1.3-3.7b-ppl](https://huggingface.co/nota-ai/cpt-lora_st-vicuna-v1.3-3.7b-ppl) | | |
| | Vicuna-v1.3-7B | 60% | PPL | CPT⇒LoRA | [nota-ai/cpt-lora_st-vicuna-v1.3-2.7b-ppl](https://huggingface.co/nota-ai/cpt-lora_st-vicuna-v1.3-2.7b-ppl) | | |
| | Vicuna-v1.3-7B | 80% | PPL | CPT⇒LoRA | [nota-ai/cpt-lora_st-vicuna-v1.3-1.5b-ppl](https://huggingface.co/nota-ai/cpt-lora_st-vicuna-v1.3-1.5b-ppl) | | |
| <details> | |
| <summary> | |
| Click to see the results: | |
| </summary> | |
| - EleutherAI/lm-evaluation-harness version [3326c54](https://github.com/EleutherAI/lm-evaluation-harness/tree/3326c547a733d598b4377e54be96e194861b964c) | |
| <img alt="results" img src="https://netspresso-research-code-release.s3.us-east-2.amazonaws.com/compressed-llm/st_llm-cpt_results.png" width="100%"> | |
| </details> | |
| #### Experimental Setup for CPT of Pruned Vicuna-7B | |
| * Dataset: [SlimPajama-627B](https://huggingface.co/datasets/cerebras/SlimPajama-627B) | |
| * Training using 8 NVIDIA H100 GPUs. | |
| * 5.5B parameters: 37B training tokens (for 6 days) | |
| * 3.7B parameters: 74B tokens (for 8 days) | |
| * 2.7B parameters: 150B tokens (for 12 days) | |
| * 1.5B parameters: 271B tokens (for 11 days) | |
| * AdamW optimizer with (β1, β2)=(0.9, 0.95); a learning rate of 0.0001; a weight decay of 0.1. | |
| * Global batch size: 512 (micro-batch size of 2 × 32 gradient accumulation steps × 8 GPUs). | |
| <details> | |
| <summary> | |
| Click to see the learning curve: | |
| </summary> | |
| **Zero-shot performance over the course of training for models from Vicuna-7B-v1.3 at different pruning ratios.** For each model size, the CPT duration was limited to a two-week period, but additional training could further improve the quality. | |
| <img alt="results" img src="https://netspresso-research-code-release.s3.us-east-2.amazonaws.com/compressed-llm/st_llm-cpt_learning-curve.png" width="100%"> | |
| </details> | |
| #### Experimental Setup for LoRA Instruction Tuning | |
| * Dataset: [Refined Alpaca](https://huggingface.co/datasets/yahma/alpaca-cleaned) | |
| * Training using 1 NVIDIA A100 GPU. | |
| * The retraining costs are low, with the entire process being executed on a single GPU. | |
| * For example, LoRA retraining of a 20%-pruned model from 7B parameters requires about 2 hours and 22GB VRAM. | |
| * A LoRA rank of 8; AdamW optimizer with a learning rate of 0.0001. | |
| * A batch size of 64 over 2 epochs. | |
| ## Models from Moderate Pruning & LoRA Retraining (arXiv-v1): | |
| | Source<br>Model | Pruning<br>Ratio | Pruning<br>Criterion | HF Models<br>Link | | |
| |:---:|:---:|:---:|:---:| | |
| | LLaMA-1-7B | 20% | PPL | [nota-ai/st-llama-1-5.5b-ppl](https://huggingface.co/nota-ai/st-llama-1-5.5b-ppl) | | |
| | LLaMA-1-7B | 20% | Taylor+ | [nota-ai/st-llama-1-5.5b-taylor](https://huggingface.co/nota-ai/st-llama-1-5.5b-taylor) | | |
| | Vicuna-v1.3-7B | 20% | PPL | [nota-ai/st-vicuna-v1.3-5.5b-ppl](https://huggingface.co/nota-ai/st-vicuna-v1.3-5.5b-ppl) | | |
| | Vicuna-v1.3-7B | 20% | Taylor+ | [nota-ai/st-vicuna-v1.3-5.5b-taylor](https://huggingface.co/nota-ai/st-vicuna-v1.3-5.5b-taylor) | | |
| | Vicuna-v1.3-13B | 21% | PPL | [nota-ai/st-vicuna-v1.3-10.5b-ppl](https://huggingface.co/nota-ai/st-vicuna-v1.3-10.5b-ppl) | | |
| | Vicuna-v1.3-13B | 21% | Taylor+ | [nota-ai/st-vicuna-v1.3-10.5b-taylor](https://huggingface.co/nota-ai/st-vicuna-v1.3-10.5b-taylor) | | |
| <details> | |
| <summary> | |
| Click to see the results: | |
| </summary> | |
| - EleutherAI/lm-evaluation-harness version [3326c54](https://github.com/EleutherAI/lm-evaluation-harness/tree/3326c547a733d598b4377e54be96e194861b964c) | |
| <img alt="results" img src="https://netspresso-research-code-release.s3.us-east-2.amazonaws.com/compressed-llm/st-llama_zero-shot_scores.png" width="100%"> | |
| </details> | |
| ## License | |
| - All rights related to this repository and the compressed models are reserved by Nota Inc. | |
| - The intended use is strictly limited to research and non-commercial projects. | |
| ## Acknowledgments | |
| - [Microsoft for Startups Founders Hub](https://www.microsoft.com/en-us/startups) and [Gwangju AICA](http://www.aica-gj.kr/main.php) for generously providing GPU resources. | |
| - [LLM-Pruner](https://github.com/horseee/LLM-Pruner), which utilizes [LM Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness), [PEFT](https://github.com/huggingface/peft), and [Alpaca-LoRA](https://github.com/tloen/alpaca-lora). Thanks for the pioneering work on structured pruning of LLMs! | |
| - [LLaMA](https://github.com/facebookresearch/llama), [Vicuna](https://github.com/lm-sys/FastChat/blob/main/docs/vicuna_weights_version.md), [SlimPajama](https://huggingface.co/datasets/cerebras/SlimPajama-627B), and [Alpaca-Cleaned](https://huggingface.co/datasets/yahma/alpaca-cleaned). Thanks for the open-source LLMs and data! | |
| ## Citation | |
| ```bibtex | |
| @article{kim2024shortened, | |
| title={Shortened LLaMA: Depth Pruning for Large Language Models with Comparison of Retraining Methods}, | |
| author={Kim, Bo-Kyeong and Kim, Geonmin and Kim, Tae-Ho and Castells, Thibault and Choi, Shinkook and Shin, Junho and Song, Hyoung-Kyu}, | |
| journal={arXiv preprint arXiv:2402.02834}, | |
| year={2024}, | |
| url={https://arxiv.org/abs/2402.02834} | |
| } | |
| ``` | |
| ```bibtex | |
| @article{kim2024mefomo, | |
| title={Shortened LLaMA: A Simple Depth Pruning for Large Language Models}, | |
| author={Kim, Bo-Kyeong and Kim, Geonmin and Kim, Tae-Ho and Castells, Thibault and Choi, Shinkook and Shin, Junho and Song, Hyoung-Kyu}, | |
| journal={ICLR Workshop on Mathematical and Empirical Understanding of Foundation Models (ME-FoMo)}, | |
| year={2024}, | |
| url={https://openreview.net/forum?id=18VGxuOdpu} | |
| } | |
| ``` |