AI Reports · Amazon · 2026-10-21 · Blog / commentary
Train Qwen3 on AWS Trainium with torchtitan
This post trains Qwen3 three ways on a single trn2.48xlarge with torchtitan: the Hugging Face Transformers v5 modeling backend, torchtitan's native Qwen3, and NKI mixture-of-experts kernels. The same torchrun launcher and config structure work unchanged across all three paths, including loading and saving published Hugging Face safetensors. TorchNeuron registers Trainium as a native PyTorch device, letting torchtitan use it through standard PyTorch interfaces.