The usual solution I’ve seen a few times is to define these intrinsics in terms of integer types – either i8 or larger, depending on the requirements of your target intrinsic. The bitcast from fp8 to i8 should be a noop.
There is a conversation that might be relevant to your problem in the MLIR subcategory.
Also I’d recommend creating a new post for specific problems in the MLIR category, for better visibility!