Monarch: an API to your supercomputer

The PyTorch team has launched Monarch, a distributed programming framework designed to eliminate the pain of developing on massive GPU clusters. By exposing the entire supercomputer through a coherent Python API, Monarch makes a cluster of thousands of GPUs feel like a local development environment. Developers can define a complete, complex training system—including setups for distributed reinforcement learning—in a single Python program. The framework provides core primitives that enable higher-level capabilities like fault tolerance and orchestration to be built as reusable libraries, fundamentally changing how engineers interact with large-scale compute.

Monarch is specifically optimized for agentic development, providing the consistent infrastructure abstractions and SQL-based telemetry APIs that AI agents excel at using. Key technical features power this: an RDMA-powered remote filesystem distributes code and dependencies to every host in milliseconds, a distributed SQL engine collects live state and traces from all processes for easy agent-led debugging, and a Jobs API allows provisioning resources once to run multiple jobs. Since its October 2025 launch, major updates include first-class Kubernetes support with just-in-time pod provisioning, expanded RDMA backend support (including AWS EFA), and an external gateway for out-of-cluster access. This toolbox collectively slashes the iteration cycle for distributed training from hours to minutes.