You need to export the trained PaddlePaddle model to a Paddle inference model that TVM can compile to generate code which is suitable to run on a Cortex-M processor.
Corstone-300 combines an example subsystem, complementary System IP, and software and tools to streamline SoC development. Corstone-300 leverages Cortex-M55, Arm's most AI capable Cortex-M CPU, and allows for straightforward integration of the Ethos-U55 NPU.