关键词: 模型压缩, 深度学习编译器, FP8量化, MLIR, Yolov5s模型

Abstract: With the development of object detection models and language models,network models are becoming increasingly large.In order to better deploy the model on the end-to-end hardware,model quantization technology is usually used to compress the model.The existing model quantization strategies are mainly implemented based on FP16,BF16,INT8,and other types.Among them,the 8-bit data type is the most significant in reducing inference memory usage and deployment costs,but the INT8 type relies on specific calibration algorithms and fails to handle models with large dynamic ranges and multiple outliers well.The FP8 type can better fit the data distribution in neural networks,and has multiple formats that can be flexibly adjusted in terms of expression range and accuracy.However,the current MLIR lacks support for quantifying the FP8 type.To this end,a FP8 quantization simulation strategy based on MLIR is proposed,which includes two formats:FP8E4M3 and FP8E5M2.By quantifying and simulating the operators in the network,the impact of the two formats on the inference accuracy of the model is evaluated.A memory reuse strategy based on define use chain is proposed to address the issue of memory allocation redundancy in inference engines,further reducing the peak memory usage during the model inference process.Typical Yolov5s and Resnet50 models are selected for testing and verification,and the results show that,compared to the existing INT8 quantization strategy,the FP8 quantization strategy can maintain better model accuracy,and does not rely on specific calibration algorithms,making deployment more convenient.In terms of model accuracy,the test cases achieve an accuracy of 55.5% and 77.8%,respectively.After memory reuse optimization,the peak memory usage is reduced by about 15%~20%.

Key words: Model compression, Deep learning compiler, FP8 quantification, MLIR, Yolov5s model