Smolcluster adds Grove for zero-config training

This is less about inventing a new training algorithm and more about sanding down the rough edges that keep distributed learning annoying for small clusters. For people trying to learn FSDP, SyncPS, or other parallelism modes from first principles, that UX improvement is the difference between a demo and something you actually keep using.

• –Zero-config discovery is the real win here: mDNS on Mac and a TCP fallback on Linux/Jetson removes a bunch of network plumbing from the setup path
• –Grove’s live TUI turns distributed training into something you can inspect in real time instead of staring at logs after the fact
• –Smolcluster’s raw-socket implementations make the communication and synchronization behavior explicit, which is useful for education and debugging
• –The pitch is strongest for homelabs and local clusters, not large production training fleets
• –Automatic metrics forwarding into Grove makes the stack feel integrated rather than stitched together