AutoSmoothQuant

AutoSmoothQuant is an easy-to-use package for implementing smoothquant for LLMs. AutoSmoothQuant speeds up model inference under various workloads. AutoSmoothQuant was created and improved upon from the original work from MIT.

News or Update

[2024/03] We support model evaluation with lm-evaluation-harness
[2024/02] We support Mixtral and Baichuan model

Install

Prerequisites

Your GPU(s) must be of Compute Capability 8.0 or higher. Amphere and later architectures are supported.
Your CUDA version must be CUDA 11.4 or later.
Python 3.9+

Build from source

Currently this repo only support build form source. We will release package soon.

git clone https://github.com/AniZpZ/AutoSmoothQuant.git
cd AutoSmoothQuant
pip install -e .

Usage

quantize model

First add a config file named "quant_config.json" to model path. For currenttly supported models, config should be like:

{
  "qkv": "per-tensor",
  "out": "per-tensor",
  "fc1": "per-tensor",
  "fc2": "per-tensor"
}

"qkv" stands for QKV matmul of attention, "out" stands for out matmul of attention. "fc1" and "fc2" are the layers of the FFNs, which might be referred to as "gate_up" and "down" in Llama-like models. You can set the value to "per-tensor" or "per-token" to perform the quant granularity you want.

Once config is set, generate scales and do model quantization with following command:

cd autosmoothquant/examples
python3 smoothquant_model.py --model-path=/path/to/model --quantize-model=True --generate-scale=True --dataset-path=/path/to/dataset --smooth-strength=0.5

use following command for more information

python smoothquant_model.py -help

inference

inference with vLLM

Comming soon (this PR could be reference)

If you want to test quantized models with the PR mentioned above, only Llama is supported and quant config should be
```
{
  "qkv": "per-tensor",
  "out": "per-token",
  "fc1": "per-tensor",
  "fc2": "per-token"
}
```
inference in this repo

cd autosmoothquant/examples
python3 test_model.py --model-path=/path/to/model --tokenizer-path=/path/to/tokenizer --model-class=llama --prompt="something to say"

benchmark

inference speed

Comming soon (this PR could be reference)

model evaluation

Currently you need to install latest lm-evaluation-harness from source in the same path with AutoSmoothQuant repo.

git clone https://github.com/EleutherAI/lm-evaluation-harness.git
cd lm-evaluation-harness
pip install -e .
cd ../AutoSmoothQuant/autosmoothquant/example
python3 eval_model.py -model-path=/path/to/model --tokenizer-path=/path/to/tokenizer

Supported models

Model support list:

Models	Sizes
LLaMA-2	7B/13B/70B
LLaMA	7B/13B/30B/65B
Mixtral	8*7B
OPT	6.7B/13B/30B
Baichuan-2	7B/13B
Baichuan	7B/13B

Performance and inference efficency

Detailed data comming soon

Cases:

codellama-13b with A40. Tested with vLLM

llama-13b with A100. Tested with vLLM

Reference

If you find SmoothQuant useful or relevant to your research, please cite their paper:

@InProceedings{xiao2023smoothquant,
    title = {{S}mooth{Q}uant: Accurate and Efficient Post-Training Quantization for Large Language Models},
    author = {Xiao, Guangxuan and Lin, Ji and Seznec, Mickael and Wu, Hao and Demouth, Julien and Han, Song},
    booktitle = {Proceedings of the 40th International Conference on Machine Learning},
    year = {2023}
}