Abstract
A model quantization method includes dividing a target model into a plurality of parts based on a connection relationship between network structures in the target model, and a plurality of rounds of quantization processes are performed on the target model. The plurality of rounds of quantization processes are sequentially performed based on the network structures of the model. A part of the network structures in the target model are quantized in each round of quantization process. A quantity of parameters loaded onto a processing device at a time is reduced and the quantization process of the target model is performed.
Full Text
What is claimed is:
A model quantization method includes dividing a target model into a plurality of parts based on a connection relationship between network structures in the target model, and a plurality of rounds of quantization processes are performed on the target model. The plurality of rounds of quantization processes are sequentially performed based on the network structures of the model. A part of the network structures in the target model are quantized in each round of quantization process. A quantity of parameters loaded onto a processing device at a time is reduced and the quantization process of the target model is performed.
Timeline
Filed
05/19/2026Published
09/17/2026Granted
Not AvailableIPC Codes(1)
G06N 3/09:Supervised learning