Zyphra's ZYFA cuts FP8 attention numerical error 62–78% versus AITER

Small attention weights can be incorrectly rounded to zero in FP8. ZYFA rescales blocks before rounding and normalizes using the rounded values.

At head dimension 128, this yields 62–78% less numerical error than AITER’s FP8 kernel on Mistral-7B and Qwen3-8B at 4K tokens.