AWS walkthrough: training Qwen3 on Trainium with torchtitan-neuron
Overview
AWS Machine Learning Blog publishes a walkthrough for training Qwen3 on a single trn2.48xlarge instance using torchtitan-neuron, which registers Trainium as a native PyTorch device.
The post covers three configurations: Hugging Face Transformers v5 as the modeling backend, torchtitan's native Qwen3, and Neuron Kernel Interface (NKI) MoE kernels. The tested stack is torch 2.14.0, torch_neuronx 2.14.3, transformers 5.16.0, and torchtitan-neuron release-2.33 (0.2.2+neuron.1), with 64 logical NeuronCores at LNC=2.
The blog says that to preserve the Neuron build of torch, torchtitan-neuron and torchtitan's remaining dependencies must be installed with --no-deps for these versions. It also warns that the Transformers v5 integration and Trainium-specific override modules can change between torchtitan releases.
Written by AI from the articles below · updated Oct 21, 4:18 PM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- AWS Machine Learning BlogOfficialTrain Qwen3 on AWS Trainium with torchtitan
AIAWS Machine Learning Blog shows how to train Qwen3 on a single trn2.48xlarge using torchtitan-neuron, with TorchNeuron registering Trainium as a native PyTorch device. The walkthrough covers three configurations: Hugging Face Transformers v5 as the modeling backend, torchtitan's native Qwen3, and Neuron Kernel Interface MoE kernels. The tested setup uses torch 2.14.0, torch_neuronx 2.14.3, transformers 5.16.0, and torchtitan-neuron release-2.33 (0.2.2+neuron.1), with 64 logical NeuronCores at LNC=2.
Heat Trend
Not enough continuous observations to show a trend yet.