Train Qwen3 on AWS Trainium with torchtitan

What happened

AWS provides a guide to training Qwen3 with torchtitan on a single trn2.48xlarge using Transformers v5, torchtitan’s native Qwen3 implementation, and NKI mixture-of-experts kernels. The configurations share a torchrun launch pattern and configuration structure, with examples for loading and saving Hugging Face checkpoints. The guide also demonstrates custom kernel registration and autograd integration, explicitly presenting that example as an integration demonstration rather than a performance improvement.

Read next

• More changes in today’s briefingMicrosoft Research and Collaborator Resolve the KLS Conjecture With Human-Directed AI

The story so far

AWS published a walkthrough for training Qwen3 on a single Trainium trn2.48xlarge instance with torchtitan, providing three training configurations and examples for loading and saving Hugging Face checkpoints. The configurations use Transformers v5, torchtitan’s native Qwen3 implementation, and NKI mixture-of-experts kernels, with a shared torchrun launch pattern and configuration structure. The walkthrough requires either the AWS Deep Learning AMI for Neuron or a Neuron PyTorch training container. Its Qwen3-30B-A3B example reduces the model to 4 layers and 64 experts so it fits on one trn2.48xlarge. AWS also demonstrates custom kernel registration and autograd integration, explicitly describing that example as an integration demonstration rather than a performance improvement.