GB/T 45079-2024 in English

1. Introduction

With the rapid development of artificial intelligence technology, the demand for deep learning frameworks to adapt to multiple hardware platforms is increasing. GB/T45079-2024 "Technical Specifications for Adaptation of Artificial Intelligence Deep Learning Frameworks to Multiple Hardware Platforms" came into being, providing unified technical requirements and testing methods for the adaptation of deep learning frameworks to multiple hardware platforms.

2. Background of Standard Formulation

Artificial intelligence technology is widely used in fields such as image recognition and natural language processing. As a core tool, deep learning frameworks need to be efficiently adapted on different hardware platforms. However, the computing power, architecture, and interfaces of different hardware platforms vary, resulting in a complex adaptation process and a lack of unified standards.

Technology Evolution Analysis

From the early single computing platform to the current support for multiple hardware platforms, deep learning frameworks have undergone the following evolution:

• Single-machine single-card adaptation
• Single-machine multi-card and distributed training adaptation
• Device-side and cloud-side reasoning adaptation
• Multi-hardware platform unified framework adaptation

3. Comparison of standard frameworks

4. Interpretation of standard content

4.1 Environmental requirements

The adaptation of deep learning framework and hardware platform needs to meet the following environmental requirements:

• Training framework should support single-machine multi-card and multi-machine multi-card model training methods.
• Inference framework should have basic model inference function and support cloud-side and device-side deployment.

4.2 Interface Adaptation

The deep learning framework interacts with the hardware platform through the following interfaces:

• Device Management Module: includes device identification, memory management, and execution flow control.
• Computation Execution Module: supports operator development or mapping, subgraph/whole graph access, and neural network compiler backend adaptation.
• Distributed Communication Module: supports the communication needs of multi-machine and multi-card distributed training and inference.

4.3 Functional Requirements

The deep learning framework should have the following functions:

5. Test methods

The test is divided into three stages: environment test, interface test and function test:

• Environment test: Verify whether the software and hardware environment meets the standard requirements.
• Interface test: Check whether the function of the adapter interface is normal.
• Function test: Evaluate the performance indicators of model training and inference.

6. Implementation suggestions

In order to ensure the efficient adaptation of the deep learning framework and the hardware platform, the following measures are recommended:

1. Implement the standard requirements in stages, giving priority to solving the core function adaptation.
2. Provide detailed hardware access guides and development documents.
3. Establish a continuous optimization mechanism and regularly update the compatibility of the framework and hardware.