It seems to me that it would be best to put the documentation next to the implementation. The implementation is where you're most likely to make changes to the code, and when you change the code, you'll want to review the documentation to ensure it's still accurate, so you may as well have the two nearby.
Olaf mentions including a description at the interface as well, and you ask how the documentation-generator chooses which comments to use. An easy solution is to simply not use the documentation generator's special comment format in the interface. It should ignore those comments as being ordinary non-documentation comments, and so the implementation's comments will be the only ones the generator uses. Or you could put a short description at the interface and a long description in the implementation, and the generator can use whichever one is most appropriate for the context (e.g., a list of all methods with one-line descriptions versus in-depth coverage of a single method).
SrbShell suggests putting the documentation in the interface, and it's certainly convenient to have everything nearby, when not using a documentation generator. But that becomes much less important when you have a program that collects documentation from throughout your code and puts it all together. In that case, the expectation is that you'll be using the generated documentation, too. When you have prepared documentation, refer to that instead of having to find the documentation in the code. The IDE can even be configured to display descriptions in a pop-up window, so you don't have to go look at it in a separate place most of the time.
I've been talking about what I see as the "ideal" mode. If your particular tool makes that strategy no-so-ideal anymore, then adapt it to fit the tool, or find a different tool.