|
--- |
|
datasets: |
|
- allenai/c4 |
|
base_model: |
|
- meta-llama/Llama-3.1-8B |
|
library_name: transformers |
|
--- |
|
|
|
# MaskLLM: Learnable Semi-structured Sparsity for Large Language Models |
|
|
|
<div align="center"> |
|
<figure> |
|
<img src="https://github.com/NVlabs/MaskLLM/blob/main/assets/teaser.png?raw=true" style="width:70%; display:block; margin-left:auto; margin-right:auto;" |
|
</figure> |
|
</div> |
|
|
|
This work introduces [MaskLLM](https://github.com/NVlabs/MaskLLM), a **learnable** pruning method that establishes **Semi-structured (or ``N:M'') Sparsity** in LLMs, aimed at reducing computational overhead during inference. The proposed method is scalable and stands to benefit from larger training datasets. |
|
|
|
## Requirements |
|
We provide pre-computed masks for Huggingface Models such as Llama-2 7B and Llama-3 8B with the minimum requirements. It will not involve docker, Megatron or data preprocessing. |
|
```bash |
|
pip install transformers accelerate datasets SentencePiece |
|
``` |
|
|
|
## Pre-computed Masks |
|
|
|
The following masks were trained and provided by [@VainF](https://github.com/VainF). We use ``huggingface_hub`` to automatically download those masks and apply them to offcical LLMs for evaluation. Those mask files were compressed using [numpy.savez_compressed](tool_compress_mask.py). More results for baselines (SparseGPT, Wanda) can be found in the appendix. |
|
| Model | Pattern | Training Data | Training/Eval SeqLen | PPL (Dense) | PPL (SparseGPT) | **PPL (MaskLLM)** | Link | |
|
| :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | |
|
| LLaMA-2 7B | 2:4 | C4 (2B Tokens)| 4096 | 5.12 | 10.42 | **6.78** | [HuggingFace](https://huggingface.co/Vinnnf/LLaMA-2-7B-MaskLLM-C4) | |
|
| LLaMA-3 8B | 2:4 | C4 (2B Tokens) | 4096 | 5.75 | 17.64 | **8.49** | [HuggingFace](https://huggingface.co/Vinnnf/LLaMA-3-8B-MaskLLM-C4) | |
|
| LLaMA-3.1 8B | 2:4 | C4 (2B Tokens) | 4096 | 5.89 | 18.65 | **8.58** | [HuggingFace](https://huggingface.co/Vinnnf/LLaMA-3.1-8B-MaskLLM-C4) | |
|
|
|
## How to use it |
|
|
|
Please see [NVlabs/MaskLLM](https://github.com/NVlabs/MaskLLM?tab=readme-ov-file#1-pre-trained-masks-for-hugging-face-models-). |
