Training jobs are supposed to be fire-and-forget. You pick hyperparameters, launch a GPU job, and wait hours for a learning curve that tells you what you should have changed on minute three.
What if you could change learning_rate, weight_decay, or dropout while the epoch loop is running — and the next batch already sees the new values?
That is what Kiponos.io is built for: a real-time config hub where connected SDKs hold the latest values in memory, updated over a permanent WebSocket. No polling. No restart. No redeploy.
This article shows the pattern for Python model training: keep Kiponos inside your inner training loop, read parameters as plain local variables, and let operators (or another algorithm) push delta-only changes from the dashboard.
The problem with "config files" in training loops
