[Question] why not using FP8 rope in sparse FP8 decoding?

[lyppg]

Hello,
I am reading this blog and the Sparse FP8 decoding kernel implementation, it seems you dequantize the fp8 KV cache into bf16 in the kernel because the rope is bf16. But why not using fp8 rope so no need to dequantize the fp8 KV? I see flashinfer also uses fp8 rope, and it seems no precision issue.
Thanks.