AI Accelerators Overview: Training & Inference Architecture - Platform For AI

PAI provides AI accelerators for training and inference acceleration using dataset optimization and computing acceleration methods to improve speed and stability.

Features

AI accelerators support the following technical methods and features:

Technical method	Feature
EPL (distributed training framework)	Supports data parallelism, operator splitting, and pipeline parallelism. Supports automatic parallelism for optimal distributed training performance.
Rapidformer (Transformer training acceleration)	Optimizes PyTorch Transformer model training and incorporates multiple optimization technologies for seamless integration with Transformer model libraries.
PAI-Blade (inference optimization)	Supports TensorFlow, PyTorch, and mainstream acceleration devices such as GPU, CPU, and end devices. Supports graph optimization, vendor-optimized libraries, AI compiler optimization, high-performance operator libraries, mixed precision, and automatic compression. Provides standard Python SDK for easy optimization implementation.

Use AI accelerators

Refer to the following documents to get started with AI accelerators:

EPL (distributed training framework)

EPL is an efficient framework for distributed model training that enables high-performance training at low cost. For more information, see Use EPL to accelerate AI model training.
Rapidformer (Transformer training acceleration)

Rapidformer is a training optimization tool for PyTorch Transformer models that combines optimization technologies to improve training speed and efficiency. For more information, see Rapidformer overview.
PAI-Blade (inference optimization)

PAI-Blade is an inference optimization tool that integrates various optimization technologies to optimize trained model inference performance for optimal results. For more information, see Blade overview.