AWS Unveils Kiro and TorchNeuron to Streamline SAP Modernization and AI Development
TECHWORLD ·
✦ AI Summary
AWS has unveiled Kiro(Kiro) for SAP system modernization and TorchNeuron(TorchNeuron) for AI model development. Kiro automates SAP ABAP code analysis, conversion, and unit testing to support the move to S/4HANA and standard API refactoring. TorchNeuron reduces the burden of moving to AI chips by enabling existing PyTorch code to run on Trainium without rewriting it.
AWS has unveiled software tools designed to reduce the work needed to modernize enterprise SAP systems and develop AI models. The releases are the AI engineering agent Kiro and TorchNeuron.
Kiro is a tool that automates SAP code conversion. A sample Kiro-based agent analyzes SAP ABAP (Advanced Business Application Programming) code and middleware interfaces from a single prompt, converts SAP ABAP (Advanced Business Application Programming) code and middleware interfaces, and performs unit tests. Its focus is on automating the conversion of legacy code in long-running SAP environments for use in modern systems.
TorchNeuron supports running existing PyTorch code on AWS's own AI chip, Trainium. AWS also unveiled the tool as part of an effort to reduce the work required for AI model development.
Companies have accumulated large volumes of custom ABAP code within SAP environments used for core operations such as finance, order management, and supply chains. The approaching end of standard support for SAP BW, SAP PI/PO, and on-premises SAP ECC has also increased the need to move existing assets to new environments.
Accordingly, the sample agent supports converting ABAP code accumulated in SAP ECC for use with SAP S/4HANA and also supports refactoring to use standard SAP API. It also supports migration paths from SAP PI/PO interfaces to the SAP BTP Integration Suite and from SAP BW to Datasphere in SAP Business Data Cloud.
The sample agent includes functions for analyzing existing code and configurations, generating functional and technical specifications, and creating unit tests. In AWS internal benchmarks, it cut the manpower needed for modernization by up to 87% by automating ABAP code analysis, conversion, and unit testing.
The related capabilities are being tested or applied in real analysis and migration work by global companies and partners. Toyota Chile analyzed and documented hundreds of custom ABAP objects and used unused-code detection. NTT DATA Global Solutions applied its own migration services and said it reduced ABAP conversion effort by up to 95%.
Capgemini used the Kiro sample agent and developed a "Smart Integration AI Agent" that can reduce the analysis-to-build period for integration work by 40% to 60%. AWS SAP competency partners including Accenture, Deloitte, DXC, and Kyndryl also applied related functions to their own migration tools.
At the same time, safeguards are used during automation to preserve existing business logic. Steering files define the agent's operating rules, with more than 80 rules set for coding, security, and documentation. Unit tests are also run before and after code conversion, dependent ABAP code and libraries are analyzed together, and work history and decision logs are preserved.
AWS provides TorchNeuron not only for enterprise software but also to lower barriers to using its own infrastructure in the AI model development process. TorchNeuron supports migration to Trainium while preserving the existing PyTorch development environment.
TorchNeuron integrates with PyTorch's built-in hardware registration system and enables execution on AWS AI chip Trainium without rewriting existing PyTorch code. Developers can run existing PyTorch code on Trainium simply by changing the settings that specify which chip to use in the code. TorchNeuron is offered as open source under the Apache 2.0 license and also supports Trainium2 and Trainium3.
The purpose of this feature is to reduce the software transition burden when changing AI hardware. When running existing models on a new chip, code modifications and revalidation are often required, and TorchNeuron is designed to reduce that burden.
The technology supports large-scale distributed training and works with PyTorch's distributed training framework TorchTitan. It also enables model training to be distributed across multiple Trainium instances using fully sharded data parallelism (FSDP).
AWS is making the vLLM inference engine open source and is also providing performance-optimized kernels for Trainium as open source. In addition, AWS plans to contribute TorchNeuron-related functions to the original PyTorch code and is avoiding a policy of confining TorchNeuron functions to a proprietary chip. This reflects a goal of reducing code tied to specific hardware and expanding PyTorch's use of various AI accelerators.
Against this backdrop, applications are expanding into voice AI, drug development, robotics, imaging, autonomous driving, LLMs, and finance. Cartesia is running inference pipelines on Trainium with minimal changes to its existing PyTorch code, while the RIVER Brain Development Institute is reviewing the technology's potential for accelerating molecular dynamics simulations. Hugging Face is also working to apply native Trainium support to its AI development tools.
Kiro is focused on automating the conversion of accumulated legacy software assets in enterprises, while TorchNeuron is focused on easing the burden on AI model developers when changing hardware and modifying code. AWS is using the two tools to push ahead with broader enterprise system modernization and to expand the use of its own AI infrastructure.
Source: TECHWORLD · Kim Seung-gi
Original: https://www.epnc.co.kr/news/articleView.html?idxno=408921
References
This article was produced with the help of an automated content generation algorithm.
Source: TECHWORLD
View originalThis article was summarized and organized by BizCrush based on the original article from TECHWORLD. For exact quotations and full details, please refer to the original article.