[TF:XLA] Don't register kernels with infeasible type constraints.
tensorflow::KernelAttrsMatch returns an error if it sees a KernelDef with an empty type constraint, and it is perfectly reasonable for an op (here TopK with a bfloat16 type) to only be supported on a subset of XLA backends. PiperOrigin-RevId: 202144682
Loading
Please sign in to comment