AutoRound ===========================

Advanced Quantization Algorithm for LLMs

[![python](https://img.shields.io/badge/python-3.9%2B-blue)](https://github.com/intel/auto-round) [![version](https://img.shields.io/badge/release-0.4.5-green)](https://github.com/intel/auto-round) [![license](https://img.shields.io/badge/license-Apache%202-9C27B0)](https://github.com/intel/auto-round/blob/main/LICENSE) Model Checkpoints ---
AutoRound is an advanced quantization algorithm for low-bits LLM/VLM inference. It's tailored for a wide range of models. AutoRound adopts sign gradient descent to fine-tune rounding values and minmax values of weights in just 200 steps, which competes impressively against recent methods without introducing any additional inference overhead and keeping low tuning cost. The below image presents an overview of AutoRound. Check out our paper on [arxiv](https://arxiv.org/pdf/2309.05516) for more details and quantized models in several Hugging Face Spaces, e.g. [OPEA](https://huggingface.co/OPEA), [Kaitchup](https://huggingface.co/kaitchup) and [fbaldassarri](https://huggingface.co/fbaldassarri).
![](docs/imgs/autoround_overview.png)
## What's New * [2024/01] We provide experimental support for GGUF q4_0 and q4_1 formats. * [2024/11] We provide experimental support for VLM quantization, please check out the [README](./auto_round/mllm/README.md) * [2024/11] We provide some tips and tricks for LLM&VLM quantization, please check out [this blog](https://medium.com/@NeuralCompressor/10-tips-for-quantizing-llms-and-vlms-with-autoround-923e733879a7) ## Installation ### Install from pypi ```bash # GPU pip install auto-round[gpu] # CPU pip install auto-round[cpu] # HPU pip install auto-round-lib ```
Build from Source ```bash # GPU pip install .[gpu] # CPU pip install .[cpu] # HPU python setup.py install lib ```
## Model Quantization ### Basic Usage (Gaudi2/CPU/GPU) A user guide detailing the full list of supported arguments is provided by calling ```auto-round -h``` on the terminal. Set the format you want in `format` and multiple formats exporting has been supported. Please check out [step-by-step-instruction](./docs/step_by_step.md) for more details about calibration dataset or evaluation. ```bash auto-round \ --model facebook/opt-125m \ --bits 4 \ --group_size 128 \ --format "auto_gptq,auto_awq,auto_round" \ --disable_eval \ --output_dir ./tmp_autoround ``` We provide two recipes for best accuracy and fast running speed with low memory. Details as below.
Other Recipes ```bash ## best accuracy, 3X slower, low_gpu_mem_usage could save ~20G but ~30% slower auto-round-best \ --model facebook/opt-125m \ --bits 4 \ --group_size 128 \ --low_gpu_mem_usage \ --disable_eval ``` ```bash ## fast and low memory, 2-3X speedup, slight accuracy drop at W4G128 auto-round-fast \ --model facebook/opt-125m \ --bits 4 \ --group_size 128 \ --disable_eval ```
### API Usage (Gaudi2/CPU/GPU) ```python from transformers import AutoModelForCausalLM, AutoTokenizer model_name = "facebook/opt-125m" model = AutoModelForCausalLM.from_pretrained(model_name) tokenizer = AutoTokenizer.from_pretrained(model_name) from auto_round import AutoRound bits, group_size, sym = 4, 128, True autoround = AutoRound(model, tokenizer, bits=bits, group_size=group_size, sym=sym) ## the best accuracy, 3X slower, low_gpu_mem_usage could save ~20G but ~30% slower # autoround = AutoRound(model, tokenizer, nsamples=512, iters=1000, low_gpu_mem_usage=True, bits=bits, group_size=group_size, sym=sym) ## fast and low memory, 2-3X speedup, slight accuracy drop at W4G128 # autoround = AutoRound(model, tokenizer, nsamples=128, iters=200, seqlen=512, batch_size=4, bits=bits, group_size=group_size, sym=sym ) autoround.quantize() output_dir = "./tmp_autoround" ## format= 'auto_round'(default in version>0.3.0), 'auto_gptq', 'auto_awq' autoround.save_quantized(output_dir, format='auto_round', inplace=True) ```
Detailed Hyperparameters - `model`: The PyTorch model to be quantized. - `tokenizer`: An optional tokenizer for processing input data. If none, a dataset must be provided. - `bits (int)`: Number of bits for quantization (default is 4). - `group_size (int)`: Size of the quantization group (default is 128). - `sym (bool)`: Whether to use symmetric quantization (default is True). - `enable_quanted_input (bool)`: Whether to use the output of the previous quantized block as the input for the current block for tuning (default is True). - `enable_minmax_tuning (bool)`: Whether to enable weight min-max tuning (default is True). - `iters (int)`: Number of tuning iterations (default is 200). - `lr (float)`: The learning rate for rounding value (default is None, it will be set to 1.0/iters automatically). - `minmax_lr (float)`: The learning rate for min-max tuning (default is None, it will be set to lr automatically). - `nsamples (int)`: Number of samples for tuning (default is 128). - `seqlen (int)`: Data length of the sequence for tuning (default is 2048). - `batch_size (int)`: Batch size for training (default is 8). - `scale_dtype (str)`: The data type of quantization scale to be used (default is "float16"), different kernels have different choices. - `amp (bool)`: Whether to use automatic mixed precision (default is True). - `nblocks (int)`: Packing several blocks as one for tuning together (default is 1). - `gradient_accumulate_steps (int)`: Number of gradient accumulation steps (default is 1). - `low_gpu_mem_usage (bool)`: Whether to save GPU memory at the cost of ~20% more tuning time (default is False). - `dataset Union[str, list, tuple, torch.utils.data.DataLoader]`: The dataset name for tuning (default is " NeelNanda/pile-10k"). Local json file and combination of datasets have been supported, e.g. " ./tmp.json,NeelNanda/pile-10k:train, mbpp:train+validation+test" - `layer_config (dict)`: Configuration for weight quantization (default is None), mainly for mixed bits or mixed precision. - `device`: The device to be used for tuning. The default is set to 'auto', allowing for automatic detection.
### API Usage for VLMs **This feature is experimental and may be subject to changes**, including potential bug fixes, API modifications, or adjustments to default hype-parameters By default, AutoRoundMLLM only quantizes the text module of VLMs and uses `NeelNanda/pile-10k` for calibration. To quantize the entire model, you can enable `quant_nontext_module` by setting it to True, though support for this feature is limited. For more information, please refer to the AutoRoundMLLM [readme](./auto_round/mllm/README.md). ```python from auto_round import AutoRoundMLLM from transformers import Qwen2VLForConditionalGeneration, AutoProcessor, AutoTokenizer ## load the model model_name = "Qwen/Qwen2-VL-2B-Instruct" model = Qwen2VLForConditionalGeneration.from_pretrained( model_name, trust_remote_code=True) tokenizer = AutoTokenizer.from_pretrained(model_name) processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True) ## quantize the model bits, group_size, sym = 4, 128, True autoround = AutoRoundMLLM(model, tokenizer, processor, bits=bits, group_size=group_size, sym=sym) autoround.quantize() # save the quantized model, set format='auto_gptq' or 'auto_awq' to use other formats output_dir = "./tmp_autoround" autoround.save_quantized(output_dir, format='auto_round', inplace=True) ``` #### Export Formats **AutoRound Format**: This format is well-suited for CPU, HPU devices, 2 bits, as well as mixed-precision inference. **[2,4] bits are supported**. However, it has not yet gained widespread community adoption. **AutoGPTQ Format**: This format is well-suited for symmetric quantization on CUDA devices and is widely adopted by the community, **[2,3,4,8] bits are supported**. However, **the asymmetric kernel has issues** that can cause considerable accuracy drops, particularly at 2-bit quantization and small models. **AutoAWQ Format**: This format is well-suited for asymmetric 4-bit quantization on CUDA devices and is widely adopted within the community, **only 4-bits quantization is supported**. **GGUF** Format: This format is well-suited for CPU devices and is widely adopted by the community, **only q4_0 and q4_1 (W4G32) is supported in our repo**. ### Quantization Costs Testing was conducted on the Nvidia A100 80G using the nightly version of PyTorch 2.6.0.dev20241029+cu124. Please note that data loading and packing costs have been excluded from the evaluation. **We enable torch.compile for Torch 2.6, but not for 2.5 due to encountered issues.** To optimize GPU memory usage, in addition to activating `low_gpu_mem_usage`, you can set `gradient_accumulate_steps=8` and a `batch_size=1`, though this may increase tuning time. The 3B and 14B models were evaluated on Qwen 2.5, the 8X7B model is Mixtral, while the remaining models utilized LLaMA 3.1. | Torch version/Config W4G128 | 3B | 8B | 14B | 70B | 8X7B | |---------------------------------------------------------------------------------------------|---------------|----------------|----------------|-----------------|----------------| | 2.6 with torch compile | 7min
10GB | 12min
18GB | 23min
22GB | 120min
42GB | 28min
46GB | | 2.6 with torch compile
low_gpu_mem_usage=True | 12min
6GB | 19min
10GB | 33min
11GB | 140min
25GB | 38min
36GB | | 2.6 with torch compile
low_gpu_mem_usage=True
gradient_accumulate_steps=8,bs=1 | 15min
3GB | 25min
6GB | 45min
7GB | 187min
19GB | 75min
36GB | | 2.5 w/o torch compile | 8min
10GB | 16min
20GB | 30min
25GB | 140min
49GB | 50min
49GB | ## Model Inference Please run the quantization code first ### AutoRound format **CPU**: pip install intel-extension-for-pytorch(much higher speed on Intel CPU) or pip install intel-extension-for-transformers, **HPU**: docker image with Gaudi Software Stack is recommended. More details can be found in [Gaudi Guide](https://docs.habana.ai/en/latest/). **CUDA**: no extra operations for sym quantization, for asym quantization, need to install auto-round from source #### CPU/HPU/CUDA ```python from transformers import AutoModelForCausalLM, AutoTokenizer from auto_round import AutoRoundConfig backend = "auto" ##cpu, hpu, cuda quantization_config = AutoRoundConfig( backend=backend ) quantized_model_path = "./tmp_autoround" model = AutoModelForCausalLM.from_pretrained(quantized_model_path, device_map=backend.split(':')[0], quantization_config=quantization_config) tokenizer = AutoTokenizer.from_pretrained(quantized_model_path) text = "There is a girl who likes adventure," inputs = tokenizer(text, return_tensors="pt").to(model.device) print(tokenizer.decode(model.generate(**inputs, max_new_tokens=50)[0])) ```
Evaluation ```bash auto-round --model saved_quantized_model \ --eval \ --task lambada_openai \ --eval_bs 1 ```
### AutoGPTQ/AutoAWQ format ```python from transformers import AutoModelForCausalLM, AutoTokenizer quantized_model_path = "./tmp_autoround" model = AutoModelForCausalLM.from_pretrained(quantized_model_path, device_map="auto") tokenizer = AutoTokenizer.from_pretrained(quantized_model_path) text = "There is a girl who likes adventure," inputs = tokenizer(text, return_tensors="pt").to(model.device) print(tokenizer.decode(model.generate(**inputs, max_new_tokens=50)[0])) ``` ## Support List AutoRound supports basically all the major large language models. Please note that an asterisk (*) indicates third-party quantized models, which may lack accuracy data and use a different recipe. We greatly appreciate their efforts and encourage more users to share their models, as we cannot release most of the models ourselves. Model | Supported | |-------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | nvidia/Llama-3.1-Nemotron-70B-Instruct-HF | [model-opea-int4-sym-autoround](https://huggingface.co/OPEA/Llama-3.1-Nemotron-70B-Instruct-HF-int4-sym-inc), [model-opea-int4-sym-autogptq](https://huggingface.co/OPEA/Llama-3.1-Nemotron-70B-Instruct-HF-int4-sym-inc), | | meta-llama/Llama-3.2-90B-Vision-Instruct | [model-opea-int4-sym-autoround](https://huggingface.co/OPEA/Llama-3.2-90B-Vision-Instruct-int4-sym-inc), [model-opea-int4-sym-autogptq](https://huggingface.co/OPEA/Llama-3.2-90B-Vision-Instruct-int4-sym-inc) | | Qwen/QwQ-32B-Preview | [model-opea-int4-sym-autoround-mixed](https://huggingface.co/OPEA/QwQ-32B-Preview-int4-sym-mixed-inc),[model-opea-int4-sym-autoawq-mixed](https://huggingface.co/OPEA/QwQ-32B-Preview-int4-sym-mixed-awq-inc) | | THUDM/cogvlm2-llama3-chat-19B | [model-opea-int4-sym-autoround](https://huggingface.co/OPEA/cogvlm2-llama3-chat-19B-int4-sym-inc) | | Qwen/Qwen2-VL-Instruct | [model-opea-int4-sym-autoround](https://huggingface.co/OPEA/Qwen2-VL-7B-Instruct-int4-sym-inc),[model-opea-int4-sym-autogptq](https://huggingface.co/OPEA/Qwen2-VL-7B-Instruct-int4-sym-inc) | | meta-llama/Llama-3.2-11B-Vision | [model-opea-int4-sym-autoround](https://huggingface.co/OPEA/Llama-3.2-11B-Vision-Instruct-int4-sym-inc), [model-opea-int4-sym-autogptq](https://huggingface.co/OPEA/Llama-3.2-11B-Vision-Instruct-int4-sym-inc) | | microsoft/Phi-3.5-vision-instruct | [model-opea-int4-sym-autoround](https://huggingface.co/OPEA/Phi-3.5-vision-instruct-int4-sym-inc), [model-opea-int4-sym-gptq](https://huggingface.co/OPEA/Phi-3.5-vision-instruct-int4-sym-inc) | | liuhaotian/llava-v1.5-7b | [model-opea-int4-sym-autoround](https://huggingface.co/OPEA/llava-v1.5-7b-int4-sym-inc),[model-opea-int4-sym-autogptq](https://huggingface.co/OPEA/llava-v1.5-7b-int4-sym-inc) | | Qwen/Qwen2.5-7B-Instruct | [model-opea-int4-sym-autoround](https://huggingface.co/OPEA/Qwen2.5-7B-Instruct-int4-sym-inc),[model-opea-int4-sym-autogptq](https://huggingface.co/OPEA/Qwen2.5-7B-Instruct-int4-sym-inc) [model-kaitchup-autogptq-int4*](https://huggingface.co/kaitchup/Qwen2.5-7B-Instruct-AutoRound-GPTQ-asym-4bit), [recipe](./docs/Qwen2.5-7B-Instruct-sym.md) | | Qwen/Qwen2.5-14B-Instruct | [model-opea-int4-sym-autoround](https://huggingface.co/OPEA/Qwen2.5-14B-Instruct-int4-sym-inc),[model-opea-int4-sym-autogptq](https://huggingface.co/OPEA/Qwen2.5-14B-Instruct-int4-sym-inc) | | Qwen/Qwen2.5-32B-Instruct | [model-opea-int4-sym-autoround](https://huggingface.co/OPEA/Qwen2.5-32B-Instruct-int4-sym-inc) | | Qwen/Qwen2.5-Coder-32B-Instruct | [model-kaitchup-autogptq-int4*](https://huggingface.co/kaitchup/Qwen2.5-Coder-32B-Instruct-AutoRound-GPTQ-4bit) | | Qwen/Qwen2.5-72B-Instruct | [model-opea-int4-sym-autoround](https://huggingface.co/OPEA/Qwen2.5-72B-Instruct-int4-sym-inc),[model-opea-int4-sym-autogptq](https://huggingface.co/OPEA/Qwen2.5-72B-Instruct-int4-sym-inc), [model-kaitchup-autogptq-int4*](https://huggingface.co/kaitchup/Qwen2.5-72B-Instruct-AutoRound-GPTQ-4bit), [model-kaitchup-autogptq-int2*](https://huggingface.co/kaitchup/Qwen2.5-72B-Instruct-AutoRound-GPTQ-2bit), [recipe](./docs/Qwen2.5-72B-Instruct-sym.md) | | meta-llama/Meta-Llama-3.1-70B-Instruct | [model-opea-int4-sym-autoround](https://huggingface.co/OPEA/Meta-Llama-3.1-70B-Instruct-int4-sym-inc), [model-opea-int4-sym-autogptq](https://huggingface.co/OPEA/Meta-Llama-3.1-70B-Instruct-int4-sym-inc),[model-opea-int4-asym-autoround](https://huggingface.co/OPEA/Meta-Llama-3.1-70B-Instruct-int4-asym-inc) | | meta-llama/Meta-Llama-3.1-8B-Instruct | [model-opea-int4-sym-autoround](https://huggingface.co/OPEA/Meta-Llama-3.1-8B-Instruct-int4-sym-inc),[model-opea-int4-sym-autogptq](https://huggingface.co/OPEA/Meta-Llama-3.1-8B-Instruct-int4-sym-inc),[model-kaitchup-autogptq-int4*](https://huggingface.co/kaitchup/Meta-Llama-3.1-8B-Instruct-autoround-gptq-4bit-asym), [model-kaitchup-autogptq-sym-int4*](https://huggingface.co/kaitchup/Meta-Llama-3.1-8B-Instruct-autoround-gptq-4bit-sym), [recipe](https://huggingface.co/Intel/Meta-Llama-3.1-8B-Instruct-int4-inc) | | meta-llama/Meta-Llama-3.1-8B | [model-kaitchup-autogptq-sym-int4*](https://huggingface.co/kaitchup/Meta-Llama-3.1-8B-autoround-gptq-4bit-sym) | | Qwen/Qwen2-7B | [model-autoround-sym-int4](https://huggingface.co/Intel/Qwen2-7B-int4-inc), [model-autogptq-sym-int4](https://huggingface.co/Intel/Qwen2-7B-int4-inc) | | THUDM/glm-4-9b-chat | [model-opea-int4-sym-autoround](https://huggingface.co/OPEA/glm-4-9b-chat-int4-sym-inc),[model-opea-int4-sym-autogptq](https://huggingface.co/OPEA/glm-4-9b-chat-int4-sym-inc) | | Qwen/Qwen2-57B-A14B-Instruct | [model-autoround-sym-int4](https://huggingface.co/Intel/Qwen2-57B-A14B-Instruct-int4-inc),[model-autogptq-sym-int4](https://huggingface.co/Intel/Qwen2-57B-A14B-Instruct-int4-inc) | | 01-ai/Yi-1.5-9B | [model-LnL-AI-autogptq-int4*](https://huggingface.co/LnL-AI/Yi-1.5-9B-4bit-gptq-autoround) | | 01-ai/Yi-1.5-9B-Chat | [model-LnL-AI-autogptq-int4*](https://huggingface.co/LnL-AI/Yi-1.5-9B-Chat-4bit-gptq-autoround) | | Intel/neural-chat-7b-v3-3 | [model-autogptq-int4](https://huggingface.co/Intel/neural-chat-7b-v3-3-int4-inc) | | Intel/neural-chat-7b-v3-1 | [model-autogptq-int4](https://huggingface.co/Intel/neural-chat-7b-v3-1-int4-inc) | | TinyLlama-1.1B-intermediate | [model-LnL-AI-autogptq-int4*](https://huggingface.co/LnL-AI/TinyLlama-1.1B-intermediate-step-1341k-3T-autoround-lm_head-symFalse) | | mistralai/Mistral-7B-v0.1 | [model-autogptq-lmhead-int4](https://huggingface.co/Intel/Mistral-7B-v0.1-int4-inc-lmhead), [model-autogptq-int4](https://huggingface.co/Intel/Mistral-7B-v0.1-int4-inc) | | google/gemma-2b | [model-autogptq-int4](https://huggingface.co/Intel/gemma-2b-int4-inc) | | tiiuae/falcon-7b | [model-autogptq-int4-G64](https://huggingface.co/Intel/falcon-7b-int4-inc) | | sapienzanlp/modello-italia-9b | [model-fbaldassarri-autogptq-int4*](https://huggingface.co/fbaldassarri/modello-italia-9b-autoround-w4g128-cpu) | | microsoft/phi-2 | [model-autoround-sym-int4](https://huggingface.co/Intel/phi-2-int4-inc) [model-autogptq-sym-int4](https://huggingface.co/Intel/phi-2-int4-inc) | | microsoft/Phi-3.5-mini-instruct | [model-kaitchup-autogptq-sym-int4*](https://huggingface.co/kaitchup/Phi-3.5-Mini-instruct-AutoRound-4bit) | | mistralai/Mistral-7B-Instruct-v0.2 | [outdated-recipe](./docs/Mistral-7B-Instruct-v0.2-asym-recipe.md) | | mistralai/Mixtral-8x7B-Instruct-v0.1 | [outdated-recipe](./docs/Mixtral-8x7B-Instruct-v0.1-asym-recipe.md) | | mistralai/Mixtral-8x7B-v0.1 | [outdated-recipe](./docs/Mixtral-8x7B-v0.1-asym-acc.md) | | meta-llama/Meta-Llama-3-8B-Instruct | [outdated-recipe](./docs/Meta-Llama-3-8B-Instruct-asym-recipe.md) | | google/gemma-7b | [outdated-recipe](./docs/gemma-7b-asym-recipe.md) | | meta-llama/Llama-2-7b-chat-hf | [outdated-recipe](./docs/Llama-2-7b-chat-hf-asym-recipe.md) | | baichuan-inc/Baichuan2-7B-Chat | [outdated-recipe](./docs/baichuan2-7b-cha-asym-recipe.md) | | 01-ai/Yi-6B-Chat | [outdated-recipe](./docs/Yi-6B-Chat-asym-recipe.md) | | facebook/opt-2.7b | [outdated-recipe](./docs/opt-2.7b-asym-recipe.md) | | bigscience/bloom-3b | [outdated-recipe](./docs/bloom-3B-asym-recipe.md) | | EleutherAI/gpt-j-6b | [outdated-recipe](./docs/gpt-j-6B-asym-recipe.md) | ## Integration AutoRound has been integrated into multiple repositories. [Intel Neural Compressor](https://github.com/intel/neural-compressor) [ModelCloud/GPTQModel](https://github.com/ModelCloud/GPTQModel) [pytorch/ao](https://github.com/pytorch/ao) ## Reference If you find AutoRound useful for your research, please cite our paper: ```bash @article{cheng2023optimize, title={Optimize Weight Rounding via Signed Gradient Descent for the Quantization of LLMs}, author={Cheng, Wenhua and Zhang, Weiwei and Shen, Haihao and Cai, Yiyang and He, Xin and Lv, Kaokao and Liu, Yi}, journal={arXiv preprint arXiv:2309.05516}, year={2023} } ```