l used arcgis python api to perform deeplearning inference tasks. When one task successed, proceed to another task, the job of RasterAnalysisTools failed. Job messages prompted GPU memory is occupied.
l found instance of RasterProcessingGPU do not release GPU memory.

l tried some ways but not optimal.
1. Adjust RasterprocessingGPU service recycling interval to short time
2 Restarted RasterprocessingGPU service after each execution of the script.
But my question is instance of service can release GPU memory automatically ?