The GPU language already inside vLLM isn't a company yet. Keep it that way.

TileLang went from research DSL to hard dependency of the two biggest open inference engines. That's a great commons and a terrible startup.