Most agentic MLSys work were focused on individual kernels because they’re easy to benchmark, validate, and iterate on. But in production, end-to-end performance is what matters ultimately.
Glad to see more work in this direction. I’ve been thinking about this space too, and
