Hailo-10H / DFC v5.3.0: a16_w16 on a transformer encoder is not a blanket allocator wall — it's a 3-stage cascade (attention crash → a16-conv nan exponent → needs super-defuse). What is the intended 16-bit path?

Hi ! I actually worked on trying to compile a LLM on hailo10h: I compiled an LLM for the hailo-10h let met know if I can help you !