Added
- SLModel: supports quantization compression algorithm, reducing communication volume by 2-4 times.
- SLModel: supports pipeline strategy, which can accelerate model training by 2-4 times in most scenarios.
- SLModel: PyTorch backend supports GPU.
- SLModel: introduces two attack and defense algorithms, LIA and FIA, for testing model security in the research and development stage.
- SLModel: supports a mode where one party only provides labels without providing features.
- Component: GLM train and predict components.
- Support the usage of brpc link as a backend for cross-silo communication.
Changed
- Component: add more parameters for SGB components.