Overall Architecture

This document describes the overall architecture of MindSpore.

The MindSpore framework consists of the Frontend Expression layer, Graph Engine layer, and Backend Runtime layer.

architecture

  • MindSpore Frontend Expression layer

    This layer contains Python APIs, MindSpore intermediate representation (IR), and graph high level optimization (GHLO).

    • Python APIs provide users with a unified API for model training, inference, and export, and a unified API for data processing and format transformation.

    • GHLO includes optimization irrelevant to hardware (such as dead code elimination), auto parallel, and auto differentiation.

    • MindSpore IR provides unified intermediate representations, based on which MindSpore performs pass optimization.

  • MindSpore Graph Engine layer

    This layer contains graph low level optimization (GLLO) and graph execution.

    • GLLO includes hardware-related optimization and in-depth optimization related to the combination of hardware and software, such as operator fusion and buffer fusion.

    • Graph execution provides communication APIs required for offline graph execution and distributed training.

  • MindSpore Backend Runtime layer

    This layer contains the efficient running environments on the cloud, edge and device.