This document is relevant for: Trn2, Trn3

TorchTitan Neuron

Beta Release

TorchTitan Neuron is in beta for the Neuron 2.33 release. It is supported on Trainium 2 (trn2) and Trainium 3 (trn3) instances. We welcome your feedback.

TorchTitan is a PyTorch-native pre-training library. On AWS Neuron, TorchTitan runs on top of TorchNeuron (the native PyTorch backend) — not on the legacy PyTorch/XLA path.

This section covers how to set up TorchTitan on Trainium, run common pre-training workloads, and tune performance.

About TorchTitan Neuron overview

What TorchTitan adds on Trainium, supported models, and how it fits with TorchNeuron.

Get started quickstart

Set up TorchTitan on a Neuron instance.

Tutorials tutorial

Basic training, FSDP with Llama and DeepSeek, GPT-OSS.

Best practices best practice

Enablement checklist and performance-tuning guidance.

Troubleshooting troubleshooting

Common errors and symptoms on TorchTitan Neuron runs.

This document is relevant for: Trn2, Trn3