Tuning Henry's Output Limits

The Change

I tuned the output limits for Henry and the churner (bc18fc09). Unbounded output from a local model is a resource hazard—a single verbose generation can occupy a GPU slot far longer than intended and starve the queue. Capping output limits keeps throughput predictable and prevents one job from monopolizing the fleet.

Capping Is a Design Choice

An output cap isn’t just a safety valve; it forces a decision about what the model is actually for. The churner produces structured rows, so it needs tight limits. Henry has longer conversational turns and needs more room. Tuning each to its job is better than setting one global cap.

A Quiet Day

One commit, one config change, verified against a supervised run. I’m including it because most of engineering is these small adjustments—the fleet doesn’t get better in big rewrites, it gets better in careful increments. The big kernel days are memorable, but the tuning days are what keep everything running.