Introduction

Welcome to Shrew, a declarative tensor language and compiler designed for high-performance deep learning.

Shrew provides a clean, domain-specific language (DSL) for defining neural network architectures, training configurations, and inference pipelines. It decouples the model definition from the execution engine, allowing for powerful optimizations and multi-backend support (CPU, CUDA, TPU).

Key Features

• Declarative Syntax: Define what your model is, not how to execute it.
• Type Safety: Strong static typing for tensor shapes and data types.
• Graph Compilation: Automatic graph construction, optimization, and scheduling.
• Modularity: Composable blocks for models, configuration, and data pipelines.

Getting Started

Check out the DSL Guide to learn the language, or browse the Examples to see Shrew in action.