flash_mla.flash_mla_with_kvcache have plans to support FP8 for the q data in the future or FP8 insufficient to support model accuracy?

[zhangfengwei]

Does the interface flash_mla.flash_mla_with_kvcache have plans to support FP8 for the q input in the future? Or from a model perspective, is quantizing q to FP8 insufficient to support model accuracy?