Internal APIs

Internally monarch is implemented using a Rust library for actors called hyperactor.

This book provides more details about its design.

This page provides access to the Rust API documentation for Monarch.

The Monarch project consists of several Rust crates, each with specialized functionality:

Core Framework

• hyperactor - Core actor framework for distributed computing

• hyperactor_macros - Procedural macros for the hyperactor framework

• hyperactor_mesh - Mesh networking for hyperactor clusters

• hyperactor_mesh_macros - Macros for hyperactor mesh functionality

• hyperactor_config - Configuration framework for hyperactor

• hyperactor_telemetry - Telemetry and monitoring for hyperactor

CUDA and GPU Computing

• nccl-sys - NCCL (NVIDIA Collective Communications Library) bindings

• torch-sys2 - Simplified PyTorch Python API bindings for Rust

• torch-sys-cuda - CUDA-specific PyTorch FFI bindings

• monarch_tensor_worker - High-performance tensor processing worker

RDMA and High-Performance Networking

• monarch_rdma - Remote Direct Memory Access (RDMA) support for high-speed networking

• rdmaxcel-sys - Low-level RDMA acceleration bindings

Monarch Python Integration

• monarch_hyperactor - Python bindings bridging hyperactor to Monarch’s Python API

• monarch_extension - Python extension module for Monarch functionality

• monarch_messages - Message types for Monarch actor communication

System and Utilities

• hyper - Mesh admin CLI and HTTP utilities

• ndslice - N-dimensional array slicing and manipulation

• typeuri - Type URI system for message serialization

• wirevalue - Wire-level value serialization for actor messages

• serde_multipart - Zero-copy multipart serialization

Architecture Overview

The Rust implementation provides a comprehensive framework for distributed computing with GPU acceleration:

• Actor Model: Built on the hyperactor framework for concurrent, distributed processing

• GPU Integration: Native CUDA support for high-performance computing workloads

• Mesh Networking: Efficient communication between distributed nodes

• Tensor Operations: Optimized tensor processing with PyTorch integration

• Multi-dimensional Arrays: Advanced slicing and manipulation of n-dimensional data

For complete technical details, API references, and usage examples, explore the individual crate documentation above.