Introduction
What Is Rec SDK
Product Background
As artificial intelligence (AI) technology evolves, industries such as e-commerce, long- and short-form video, and social media increasingly demand effective search, recommendation, and advertising (SRA) systems. The rapid growth of the Internet has led to a massive increase in user data, product information, and video content, underscoring the importance of effective SRA systems. The growing demand for such systems inevitably increases the need for computing power. A primary focus for practitioners is how to deploy more computing power and fully utilize its potential.
Definition
Rec SDK offers an SDK for search, recommendation, and advertising services in the Internet market. It offers a framework for these services based on the Ascend platform to meet related model training requirements, thus supporting large-scale SRA scenarios and facilitating efficient training of models for such scenarios.
Product Value
Table 1 Product value
| Product Feature | Product Value |
|---|---|
| Ease of use | Build algorithm models quickly using minimalist APIs. |
| Accuracy | Achieve less than 0.01% accuracy error in standard model verification. |
| Performance | Maximize performance with efficient multi-level pipeline acceleration, high-speed collective communication, and extreme optimization. |
Component Overview
Rec SDK provides sparse recommendation frameworks based on TensorFlow and PyTorch. Based on how computing tasks are distributed between the CPU and NPU, the frameworks support two modes: Non-fully-offloaded and Fully-offloaded.
- Non-fully-offloaded: a hybrid mode in which some computing tasks run on the NPU and others run on the CPU
- Fully-offloaded: a mode in which all computing tasks are offloaded to the NPU
The components are as follows:
| Component | Adaptation Status | Description |
|---|---|---|
| tf_rec_v1 | Non-fully-offloaded | Non-fully-offloaded sparse recommendation framework based on TensorFlow, adapted for NPU devices |
| tf_rec_v2 | Fully-offloaded | Fully-offloaded sparse recommendation framework based on TensorFlow, adapted for NPU devices |
| torch_rec_v1 | Non-fully-offloaded | Non-fully-offloaded sparse recommendation framework based on the open-source PyTorch and TorchRec software, adapted for NPU devices |
| torch_rec_v2 | Fully-offloaded | Fully-offloaded sparse recommendation framework based on the open-source PyTorch and TorchRec software, adapted for NPU devices |