Porting a 128-expert MoE (Gemma-4 26B-A4B) to AWS Inferentia2 — where every rank weighted the wrong experts

· Dev.to