Qwen3 can now be trained on AWS Trainium using torchtitan in three ways: via Hugging Face Transformers v5, torchtitan’s native Qwen3, or NKI mixture‑of‑experts kernels. The same torchrun launcher and config work unchanged across all paths, supporting loading and saving of Hugging Face safetensors. This simplifies large‑model training on AWS hardware.
Opening Kapyn…