Training Qwen3 on AWS Trainium: Three Flexible Approaches with torchtitan

AWS Trainium now supports multiple ways to train Qwen3 models using torchtitan on a single trn2.48xlarge instance—whether you prefer Hugging Face Transformers v5, native torchtitan, or NKI mixture-of-experts kernels. The best part? Your torchrun launcher and config stay identical across all three paths, making it super convenient to experiment and switch between approaches without rewriting your setup.