Welcome to MindSpore AKG SIG. Let's build an industry-leading graph kernel fusion compiler!

Developers who have experience in deep learning projects may have similar requirements. Take the current popular large language model (LLM) as an example. We may use the MindSpore AI framework to train the Llama2 neural networks on the Ascend AI processor to solve natural language processing tasks. Here, we can use the functions in the Compute Architecture for Neural Networks (CANN) software stack to deploy and accelerate network training on the Ascend AI processor.

In most cases, however, model training and inference may not be performed on the same platform. The model may end up being deployed on CPUs or even mobile phones, which requires support from the platform architectures and other hardware. Developers usually train models using various deep learning top-level frameworks, such as MindSpore, and then deploy the trained models on various device backends. In addition to the Ascend AI processor, GPUs, CPUs, FPGAs, and other new AI accelerators are also included.

Moreover, learning and time costs have skyrocketed since developers need to write operators and ensure their performance under conditions such as different hardware features, different operators in the existing operator library, insufficient support of new accelerator operator library, and uncommon layers existing in unconventional neural networks. So the graph kernel fusion compilation technology is very helpful for developers.

AKG stands for Auto Kernel Generator. It accelerates the graph kernel fusion compilation in the MindSpore framework. Based on the polyhedral compilation technology, AKG can automatically generate schedules that meet parallelism and data locality requirements. Currently, AKG supports hardware such as NPUs, GPUs, and CPUs.

MindSpore keeps evolving since it was open-sourced in 2020. We hope that more people can try AKG to solve their business problems and involve enterprise users in software iteration. We also hope to expose AKG to more developers to implement the latest AI compilation technologies. What's more, AKG can perform its own technology iteration based on community practices.

Against this backdrop, MindSpore AKG SIG (special interest group) was established and started to recruit like-minded partners from the open source community.

1

Introduction to MindSpore AKG SIG

By making full use of MindSpore AKG's technical accumulation in the graph kernel fusion compilation field, our group continuously improves software functions, expands the community ecosystem, and provides scientific research personnel, teachers, and students with efficient and easy-to-use graph kernel fusion compilers. We also provide a platform for people with strong influence and great interest in this field to communicate and cooperate with each other.

2

MindSpore AKG SIG Mission

We aim to continuously introduce industry-leading operator compilation technologies, and iterate MindSpore AKG functions to build an operator compiler ahead of the industry.

Our group's current work priorities are as follows:

01. Keep learning advanced operator compilation technologies in the industry.

Learn, understand, and internalize the latest operator compilation technologies through technical sharing and communication in the open source community.

02. Iterate MindSpore AKG functions.

Iterate MindSpore AKG functions based on industry-leading technologies to meet internal and external service requirements.

3

MindSpore AKG SIG Work Plan

Focus on academic exchanges among group members and provide reference for MindSpore AKG evolution and function improvement.

Activity organization

Each year, the group will organize a large-scale activity and several small-scale activities, including a campus activity every quarter and a major summer school activity each year. Core experts and teachers in the group are invited to teach multiple topics for multiple days. The group teacher will lead members to conduct scientific and technological research, code repository function expansion, and bug fixing. Members can also use the MindSpore AKG software to conduct their own research and development.

Open source internships

The group will publish open source internship opportunities and crowd intelligence tasks in the community for students and teachers to participate. Currently, we have arranged the following tasks:

01. MindSpore AKG operator support

MindSpore AKG MLIR supports all important operators in the mainstream models. However, with the iteration of networks and improvement of algorithms, new operators keep coming out. Since MindSpore AKG supports multiple hardware backends, including NPUs, GPUs, and CPUs, we plan to support related operators based on the backend code generation capability. Our support includes:

· For new operators, providing expressions based on loops and mathematical expressions.

· For fused operators, providing expanded expression for the concatenation of existing operators.

02. MindSpore AKG backend iteration support

· MindSpore AKG supports multiple backends, including NPUs, GPUs, and CPUs. With the iteration of the MLIR community editions and hardware updates, we plan to keep updating the backend code generation capability.

· As more AI chips are emerging, we want to add more backend code generation capabilities.

4

MindSpore AKG SIG Composition

Team Leader

Zhao Jie, Professor at College of Computer Science and Electronic Engineering, Hunan University, and member of Huawei MindSpore Community Technical Expert Committee. His research focuses on deep learning-oriented compilation systems, code generation based on polyhedral models, and numerical program analysis and optimization. Professor Zhao Jie published multiple academic papers in top conferences and journals in the system software, architecture, and compiler fields as the first author and corresponding author, including ASE, CC, MICRO, MLSys, OSDI, PACT, PLDI, PPoPP, TACO and TOCS. The paper published at the MICRO conference in 2020 was nominated as the best paper, and the paper published on the TOCS in 2023 is his fourth independently-completed work in China.

Team Members

01. Zhang Renwei (@anyrenwei), SIG initiator, MindSpore technical expert

02. Ye Zichun (@zichun_ye), SIG lead and organizer, senior MindSpore engineer

03. Zhao Jie (@yaozhujia), SIG organizer and operations consultant, and senior expert in AI compilers

04. Leo (@zhanghanLeo), MindSpore engineer

05. Xinkai (@di-xinkai), MindSpore engineer

06. gent1e (@gent1ezzz) , MindSpore engineer

07. pudding, The Hong Kong University of Science and Technology, AI compiler expert

08. Dan, South China University of Technology, AI compiler developer

5

MindSpore AKG SIG Recruitment

Scan the QR code to join the MindSpore AKG SIG communication group.

Code repository

https://gitee.com/mindspore/akg

SIG repository

https://gitee.com/mindspore/community/tree/master/sigs/akg

About MindSpore SIG

The MindSpore Community welcomes industry experts and academic partners to set up special interest groups (SIGs) in the community. As the technical spokesperson of the community, the SIG builds a technical brand in the domain and builds a MindSpore open source ecosystem.

The purpose of establishing SIGs in the MindSpore community is to provide an open communication platform for experts, professors, and students in each field, promote technical exchange and win-win cooperation through activities such as sharing at meetings and project development, and improve the influence and technological capabilities of SIG members.

Want to establish your SIG?

Scan to add our assistant to know more.