Background of Introducing the GE

GE Definition

Currently, mainstream deep learning frameworks provide the eager execution mode and graph mode. In eager execution mode, each computational operation is executed immediately after it is delivered. In graph mode, all computational operations are constructed into a graph and then delivered and executed as a graph. Compared with the mode of delivering single computational operations in sequence, the graph mode has a global view of the graph and can effectively simplify and optimize computational graph operations more effectively, thereby achieving better execution performance.

The Graph Engine (GE) serves as the control center for build and executing computational graphs on the Ascend platform. Its key functionalities include graph build, graph build optimization, and graph execution control. With the GE capability, algorithm models of mainstream AI frameworks such as PyTorch, TensorFlow, MindSpore, and PaddlePaddle can be converted into computational graphs (Ascend graphs) represented by Ascend intermediate representation (Ascend IR). In addition, the GE graph build acceleration technologies significantly improve the execution efficiency of computational graphs on Ascend hardware. Additionally, GE provides unified graph development APIs and supports user-defined graph structures, helping users quickly deploy neural network services on Ascend hardware.

Figure 1 GE logical architecture

Open Capabilities of GE

For upper-layer AI framework interconnection and service deployment scenarios: GE provides unified graph development APIs to connect to upper-layer graph-based open frameworks. Currently, mainstream AI frameworks such as PyTorch (TorchAir graph mode), TensorFlow, MindSpore, and PaddlePaddle are supported. In addition, GE supports user-defined graph structures, helping users efficiently deploy neural network services on Ascend hardware.
For AI model graph build and execution optimization scenarios: GE opens graph build optimization and graph execution capabilities, and supports functions such as user-defined graph fusion, enabling you to customize high-performance graph solutions.