meta-pytorch/monarch

Health v2 · maintenance only

Adoption not part of the score

1073 stars · 172 forks observed · 2026-09-03

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

Monarch is a distributed programming framework for PyTorch built on scalable actor messaging, with actors grouped into meshes, supervision-tree fault tolerance, RDMA transfers, and distributed tensors. It exposes a simple Python API for spawning processes and actors across GPUs while being implemented in Rust.

Use cases

• orchestrate distributed PyTorch training across multiple GPUs
• build fault-tolerant distributed ML pipelines with actor supervision trees
• broadcast messages to collections of remote actors
• perform point-to-point RDMA transfers of GPU or CPU memory
• work with tensors sharded across processes
• spawn one trainer process per GPU from Python

When to choose

• you need fine-grained control over distributed PyTorch process and actor orchestration
• you want fault tolerance with supervision trees and fine-grained recovery
• you need cheap RDMA-based GPU/CPU memory transfers
• you prefer an imperative Python API over declarative launchers

When to avoid

• you only need single-node single-GPU training
• you want a mature turnkey distributed trainer like torchrun or DeepSpeed
• you need Windows support or non-Linux GPU environments

Facets

framework · maturity active

concurrency rpc machine-learning llm-training streaming machine-learning microservices deep-learning gpu-computing python rust pytorch actor-model distributed-training rdma supervision-trees actor-messaging distributed-tensors linux macos gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="meta-pytorch/monarch")