Software Architecture

Figure 1 Software architecture

Built upon TorchRec, mainstream recommendation frameworks, CANN, and diverse hardware and network architectures, Rec SDK Torch addresses the specific requirements of search, recommendation, and advertising model training. It provides high-performance, streamlined APIs designed for ease of use, enabling Ascend AI Processors to achieve highly efficient training for search, recommendation, and advertising models.

**Table 1** Architecture modules
Rec SDK Torch Module	Description
Recommendation interface layer	Provides easy-to-use APIs to simplify user access and reduce migration costs. Supports large-scale user growth.
Recommendation function layer	Core function implementation layer, meeting user requirements.
Recommendation acceleration layer	Provides core components to build performance competitiveness and offer superb performance for the entire system.
Recommendation storage layer	Supports distributed storage of sparse tables.
Torchrec-npu	Ascend adaptation layer of the open source TorchRec.

Parent topic: Introduction