--- title: README emoji: 🐇 colorFrom: pink colorTo: indigo sdk: static pinned: false --- # Visual Informatics Group @ University of Texas at Austin ([VITA-Group](https://vita-group.github.io/)) At VITA group, we have unusually broad, and forever-evolving research interests spanning from the theory to the application aspects of machine learning (ML). Our current "research keywords" include, but are not limited to: sparsity (from classical optimization to modern neural networks); efficient training, inference or transfer (especially, of large foundation models); robustness and trustworthiness; learning to optimize (L2O); generative AI; graph learning, and more. ## Compressed LLM Model Zone **NOTE: All compressed LLMs are moved to a new repo at [compressed-llm](https://huggingface.co/compressed-llm).** The models are prepared by [Visual Informatics Group @ University of Texas at Austin (VITA-group)](https://vita-group.github.io/). Credits to Ajay Jaiswal, Zhenyu Zhang, Zhangheng Li, Lu Yin, Shiwei Liu and Junyuan Hong. License: [MIT License](https://opensource.org/license/mit/) Setup environment ```shell pip install torch==2.0.0+cu117 torchvision==0.15.1+cu117 torchaudio==2.0.1 --index-url https://download.pytorch.org/whl/cu117 pip install transformers==4.31.0 pip install accelerate pip install auto-gptq # for gptq ``` How to use pruned models ```python import torch from transformers import AutoModelForCausalLM, AutoTokenizer base_model = 'llama-2-7b' comp_method = 'magnitude_unstructured' comp_degree = 0.2 model_path = f'vita-group/{base_model}_{comp_method}' model = AutoModelForCausalLM.from_pretrained( model_path, revision=f's{comp_degree}', torch_dtype=torch.float16, low_cpu_mem_usage=True, device_map="auto" ) tokenizer = AutoTokenizer.from_pretrained('meta-llama/Llama-2-7b-hf') input_ids = tokenizer('Hello! I am a VITA-compressed-LLM chatbot!', return_tensors='pt').input_ids.cuda() outputs = model.generate(input_ids, max_new_tokens=128) print(tokenizer.decode(outputs[0])) ``` How to use wanda+gptq models ```python from transformers import AutoTokenizer from auto_gptq import AutoGPTQForCausalLM model_path = 'vita-group/llama-2-7b_wanda_2_4_gptq_4bit_128g' tokenizer_path = 'meta-llama/Llama-2-7b-hf' model = AutoGPTQForCausalLM.from_quantized( model_path, # inject_fused_attention=False, # or disable_exllama=True, device_map='auto', ) tokenizer = AutoTokenizer.from_pretrained(tokenizer_path, trust_remote_code=True) input_ids = tokenizer('Hello! I am a VITA-compressed-LLM chatbot!', return_tensors='pt').input_ids.to('cuda') outputs = model.generate(input_ids=input_ids, max_length=128) tokenizer.decode(outputs[0]) ``` How to use gptq models ```python from transformers import AutoTokenizer from auto_gptq import AutoGPTQForCausalLM # model_path = 'vita-group/llama-2-7b_wanda_2_4_gptq_4bit_128g' # tokenizer_path = 'meta-llama/Llama-2-7b-hf' model_path = 'vita-group/vicuna-7b-v1.3_gptq' tokenizer_path = 'lmsys/vicuna-7b-v1.3' model = AutoGPTQForCausalLM.from_quantized( model_path, # inject_fused_attention=False, # or disable_exllama=True, device_map='auto', revision='2bit_128g', ) from transformers import AutoTokenizer tokenizer = AutoTokenizer.from_pretrained(tokenizer_path, trust_remote_code=True) input_ids = tokenizer('Hello! I am a VITA-compressed-LLM chatbot!', return_tensors='pt').input_ids.to('cuda') outputs = model.generate(input_ids=input_ids, max_length=128) tokenizer.decode(outputs[0]) ``` | | Base Model | Model Size | Compression Method | Compression Degree | |---:|:-------------|:-------------|:------------------------------------------------------------------------------------------------|:--------------------------------------------------------------------------------------------------| | 0 | Llama-2 | 13b | [magnitude_semistruct](https://huggingface.co/vita-group/llama-2-13b_magnitude_semistruct) | [0.5_2to4](https://huggingface.co/vita-group/llama-2-13b_magnitude_semistruct/tree/0.5_2to4) | | 1 | Llama-2 | 13b | [sparsegpt_semistruct](https://huggingface.co/vita-group/llama-2-13b_sparsegpt_semistruct) | [0.5_2to4](https://huggingface.co/vita-group/llama-2-13b_sparsegpt_semistruct/tree/0.5_2to4) | | 2 | Llama-2 | 7b | [magnitude_unstructured](https://huggingface.co/vita-group/llama-2-7b_magnitude_unstructured) | [s0.1](https://huggingface.co/vita-group/llama-2-7b_magnitude_unstructured/tree/s0.1) | | 3 | Llama-2 | 7b | [magnitude_unstructured](https://huggingface.co/vita-group/llama-2-7b_magnitude_unstructured) | [s0.2](https://huggingface.co/vita-group/llama-2-7b_magnitude_unstructured/tree/s0.2) | | 4 | Llama-2 | 7b | [magnitude_unstructured](https://huggingface.co/vita-group/llama-2-7b_magnitude_unstructured) | [s0.3](https://huggingface.co/vita-group/llama-2-7b_magnitude_unstructured/tree/s0.3) | | 5 | Llama-2 | 7b | [magnitude_unstructured](https://huggingface.co/vita-group/llama-2-7b_magnitude_unstructured) | [s0.5](https://huggingface.co/vita-group/llama-2-7b_magnitude_unstructured/tree/s0.5) | | 6 | Llama-2 | 7b | [magnitude_unstructured](https://huggingface.co/vita-group/llama-2-7b_magnitude_unstructured) | [s0.6](https://huggingface.co/vita-group/llama-2-7b_magnitude_unstructured/tree/s0.6) | | 7 | Llama-2 | 7b | [sparsegpt_unstructured](https://huggingface.co/vita-group/llama-2-7b_sparsegpt_unstructured) | [s0.1](https://huggingface.co/vita-group/llama-2-7b_sparsegpt_unstructured/tree/s0.1) | | 8 | Llama-2 | 7b | [sparsegpt_unstructured](https://huggingface.co/vita-group/llama-2-7b_sparsegpt_unstructured) | [s0.2](https://huggingface.co/vita-group/llama-2-7b_sparsegpt_unstructured/tree/s0.2) | | 9 | Llama-2 | 7b | [sparsegpt_unstructured](https://huggingface.co/vita-group/llama-2-7b_sparsegpt_unstructured) | [s0.3](https://huggingface.co/vita-group/llama-2-7b_sparsegpt_unstructured/tree/s0.3) | | 10 | Llama-2 | 7b | [sparsegpt_unstructured](https://huggingface.co/vita-group/llama-2-7b_sparsegpt_unstructured) | [s0.5](https://huggingface.co/vita-group/llama-2-7b_sparsegpt_unstructured/tree/s0.5) | | 11 | Llama-2 | 7b | [sparsegpt_unstructured](https://huggingface.co/vita-group/llama-2-7b_sparsegpt_unstructured) | [s0.6](https://huggingface.co/vita-group/llama-2-7b_sparsegpt_unstructured/tree/s0.6) | | 12 | Llama-2 | 7b | [wanda_gptq](https://huggingface.co/vita-group/llama-2-7b_wanda_2_4_gptq_4bit_128g) | 4bit_128g | | 13 | Llama-2 | 7b | [wanda_unstructured](https://huggingface.co/vita-group/llama-2-7b_wanda_unstructured) | [s0.1](https://huggingface.co/vita-group/llama-2-7b_wanda_unstructured/tree/s0.1) | | 14 | Llama-2 | 7b | [wanda_unstructured](https://huggingface.co/vita-group/llama-2-7b_wanda_unstructured) | [s0.2](https://huggingface.co/vita-group/llama-2-7b_wanda_unstructured/tree/s0.2) | | 15 | Llama-2 | 7b | [wanda_unstructured](https://huggingface.co/vita-group/llama-2-7b_wanda_unstructured) | [s0.3](https://huggingface.co/vita-group/llama-2-7b_wanda_unstructured/tree/s0.3) | | 16 | Llama-2 | 7b | [wanda_unstructured](https://huggingface.co/vita-group/llama-2-7b_wanda_unstructured) | [s0.5](https://huggingface.co/vita-group/llama-2-7b_wanda_unstructured/tree/s0.5) | | 17 | Llama-2 | 7b | [wanda_unstructured](https://huggingface.co/vita-group/llama-2-7b_wanda_unstructured) | [s0.6](https://huggingface.co/vita-group/llama-2-7b_wanda_unstructured/tree/s0.6) | | 18 | Llama-2-chat | 13b | [magnitude_semistruct](https://huggingface.co/vita-group/llama-2-13b-chat_magnitude_semistruct) | [0.5_2to4](https://huggingface.co/vita-group/llama-2-13b-chat_magnitude_semistruct/tree/0.5_2to4) | | 19 | Llama-2-chat | 13b | [sparsegpt_semistruct](https://huggingface.co/vita-group/llama-2-13b-chat_sparsegpt_semistruct) | [0.5_2to4](https://huggingface.co/vita-group/llama-2-13b-chat_sparsegpt_semistruct/tree/0.5_2to4) | | 20 | vicuna | 13b | [magnitude_semistruct](https://huggingface.co/vita-group/vicuna-13b-v1.3_magnitude_semistruct) | [0.5_2to4](https://huggingface.co/vita-group/vicuna-13b-v1.3_magnitude_semistruct/tree/0.5_2to4) | | 21 | vicuna | 13b | [sparsegpt_semistruct](https://huggingface.co/vita-group/vicuna-13b-v1.3_sparsegpt_semistruct) | [0.5_2to4](https://huggingface.co/vita-group/vicuna-13b-v1.3_sparsegpt_semistruct/tree/0.5_2to4) | | 22 | vicuna-v1.3 | 13b | [gptq](https://huggingface.co/vita-group/vicuna-13b-v1.3_gptq) | [10bit_128g](https://huggingface.co/vita-group/vicuna-13b-v1.3_gptq/tree/10bit_128g) | | 23 | vicuna-v1.3 | 13b | [gptq](https://huggingface.co/vita-group/vicuna-13b-v1.3_gptq) | [12bit_128g](https://huggingface.co/vita-group/vicuna-13b-v1.3_gptq/tree/12bit_128g) | | 24 | vicuna-v1.3 | 13b | [gptq](https://huggingface.co/vita-group/vicuna-13b-v1.3_gptq) | [14bit_128g](https://huggingface.co/vita-group/vicuna-13b-v1.3_gptq/tree/14bit_128g) | | 25 | vicuna-v1.3 | 13b | [gptq](https://huggingface.co/vita-group/vicuna-13b-v1.3_gptq) | [2bit_128g](https://huggingface.co/vita-group/vicuna-13b-v1.3_gptq/tree/2bit_128g) | | 26 | vicuna-v1.3 | 13b | [gptq](https://huggingface.co/vita-group/vicuna-13b-v1.3_gptq) | [3bit_128g](https://huggingface.co/vita-group/vicuna-13b-v1.3_gptq/tree/3bit_128g) | | 27 | vicuna-v1.3 | 13b | [gptq](https://huggingface.co/vita-group/vicuna-13b-v1.3_gptq) | [4bit_128g](https://huggingface.co/vita-group/vicuna-13b-v1.3_gptq/tree/4bit_128g) | | 29 | vicuna-v1.3 | 13b | [gptq](https://huggingface.co/vita-group/vicuna-13b-v1.3_gptq) | [8bit_128g](https://huggingface.co/vita-group/vicuna-13b-v1.3_gptq/tree/8bit_128g) | | 30 | vicuna-v1.3 | 7b | [gptq](https://huggingface.co/vita-group/vicuna-7b-v1.3_gptq) | [10bit_128g](https://huggingface.co/vita-group/vicuna-7b-v1.3_gptq/tree/10bit_128g) | | 31 | vicuna-v1.3 | 7b | [gptq](https://huggingface.co/vita-group/vicuna-7b-v1.3_gptq) | [12bit_128g](https://huggingface.co/vita-group/vicuna-7b-v1.3_gptq/tree/12bit_128g) | | 32 | vicuna-v1.3 | 7b | [gptq](https://huggingface.co/vita-group/vicuna-7b-v1.3_gptq) | [14bit_128g](https://huggingface.co/vita-group/vicuna-7b-v1.3_gptq/tree/14bit_128g) | | 33 | vicuna-v1.3 | 7b | [gptq](https://huggingface.co/vita-group/vicuna-7b-v1.3_gptq) | [2bit_128g](https://huggingface.co/vita-group/vicuna-7b-v1.3_gptq/tree/2bit_128g) | | 34 | vicuna-v1.3 | 7b | [gptq](https://huggingface.co/vita-group/vicuna-7b-v1.3_gptq) | [3bit_128g](https://huggingface.co/vita-group/vicuna-7b-v1.3_gptq/tree/3bit_128g) | | 35 | vicuna-v1.3 | 7b | [gptq](https://huggingface.co/vita-group/vicuna-7b-v1.3_gptq) | [4bit_128g](https://huggingface.co/vita-group/vicuna-7b-v1.3_gptq/tree/4bit_128g) | | 37 | vicuna-v1.3 | 7b | [gptq](https://huggingface.co/vita-group/vicuna-7b-v1.3_gptq) | [8bit_128g](https://huggingface.co/vita-group/vicuna-7b-v1.3_gptq/tree/8bit_128g) | ## Citations If you are using models in this hub, please consider citing our papers. ```bibtex @article{jaiswal2023emergence, title={The Emergence of Essential Sparsity in Large Pre-trained Models: The Weights that Matter}, author={Jaiswal, Ajay and Liu, Shiwei and Chen, Tianlong and Wang, Zhangyang}, journal={arXiv}, year={2023} } @article{jaiswal2023compressing, title={Compressing LLMs: The Truth is Rarely Pure and Never Simple}, author={Ajay Jaiswal and Zhe Gan and Xianzhi Du and Bowen Zhang and Zhangyang Wang and Yinfei Yang}, year={2023}, journal={arXiv}, } ``` For any question, please contact [Junyuan Hong](mailto:jyhong@utexas.edu).