Having many implementations is beneficial, because each implementation is optimal in unique places. And modern mainstream languages don't have one implementation anyway. Think about Python: its main implementation is CPython, but thanks to JPython you can run Python on the JVM too; thanks to Stackless Python you can have massive concurrency thanks to microthreads; etc. Such implementations will be encompatible in some ways: JPython integrates seamlessly with Java, whilst CPython doesn't. Same for Ruby.
What you don't want is having many implementations which are incompatible to the bone. That's the case with Scheme, where you can't share libraries among implementations without rewriting a lot of code, because Schemers can't agree on how to import/export libraries. Common Lisp libraries, OTOH, because of standardization in core areas, are more likely to be portable, and facilities exist to conditionally write code handling each implementation's peculiarities. Actually, nowadays you may say that Common Lisp is defined by its implementations (think about the ASDF package installation library), just like mainstream languages.
