Monarch for GPU Mode

A hands-on introduction to building distributed systems with Monarch

These are interactive marimo notebooks that accompany a GPU Mode presentation on February 7th, 2026 about Monarch — PyTorch's framework for distributed actors, fault tolerance, and async RL.

Want to run these yourself? Clone the repo and run with marimo:

git clone https://github.com/allenwang28/monarch-gpu-mode
cd monarch-gpu-mode && uv sync
uv run marimo edit notebooks/01_history_and_vision.py

Part 1 — Monarch Deep Dive

01 Monarch: History & Vision From tensor engine to general actor framework. Hyperactor, the World→Proc→Actor ontology, and why Monarch exists.
02 Interactive DevX: Monarch as Remote Torchrun SPMDJob, HostMesh, spawning actors, and running distributed PyTorch interactively.
03 Fault Tolerance in Monarch The canonical try/except pattern, supervision trees, and graceful failure handling.
04 Distributed Tensors in Monarch Mesh activation, broadcasting tensor compute, collectives, remote GPUs, and device mesh dimensions.

Part 2 — Async RL from Scratch

05 RL at Scale: The Systems Problem Why RL training is a distributed systems problem. Sync vs async, on-policy vs off-policy.
06 Services: Managing Worker Pools Round-robin routing, health tracking, failure recovery, and service discovery.
07 RDMA & Weight Synchronization The magic pointer pattern, CPU staging, circular buffers, and weight re-sharding.
07b RDMA Deep Dive ibverbs internals, memory registration benchmarks, and full implementations.
08 Closing the Loop: Async RL Training Full end-to-end async RL with trainers, generators, replay buffers, and weight sync.

What are marimo notebooks? marimo is a reactive Python notebook — like Jupyter, but cells re-run automatically when dependencies change, and notebooks are stored as plain .py files. The HTML exports above are read-only snapshots. To get the full interactive experience (sliders, live code execution), clone the repo and run with uv run marimo edit.