Because they're internal hash functions, they have access to the original key being used, so they can tolerate a few collisions -- in the implementation of CHD,

Because they're internal hash functions, they have access to the original key being used, so they can tolerate a few collisions -- in the implementation of CHD, there's a -t parameter for number of collisions in a bin to provide a bit of flexibility. So they don't require a completely collision-free hash function, just one that on the input dataset is not going to produce more than a few.