Research review / opensource

Berth Adds Protected Training Leases to Shared GPU Scheduling

Berth 0.6.0, released November 14, adds scheduled GPU access, protected training runs and lease histories that survive daemon restarts. The project still labels itself alpha.

By Sandeep Belgavi · Published 2026-11-15

Berth 0.6.0 was released on November 14, 2026. The GPU resource manager's release notes add reservations, protected training leases and explanations of why a request waited or lost access.

Protected training changes admission

Training holds the requested cards until its command exits. While training holds or waits for a card, competing managed requests for that card are refused immediately; other cards can continue serving. The notes say protection survives a daemon restart, with competing requests refused while the training process reattaches.

Reservations can be made up to fourteen days ahead. Persistent events retain explanations across restarts. These are access-control and recovery features; this digest makes no claim that they accelerate model computation.

An alpha scheduler with explicit limits

The project README calls Berth alpha and describes ordinary leases as cooperation among managed engines. Unmanaged processes are reported and left running. A scheduler's refusal therefore needs to be interpreted within the set of workloads it manages.

The current README and license file identify Apache-2.0 licensing for the code, with the project name, mascot and artwork excluded by the README. This was checked on the current branch, not a tag-pinned license copy. No project artwork or implementation code is reproduced here.

A synthetic handoff test

Proposed experiment, not executed: use a tiny training workload with synthetic tensors and a managed inference stub that logs every admission. Start a protected training lease, submit an inference request, then restart the daemon while training remains active. Check for an immediate refusal, no inference admission during recovery, and an explanatory event history. Finally end training and confirm that inference can acquire the card.

The useful observation would be whether the ownership rule survives recovery. Keep that result separate from any later tokens-per-second measurement. This digest has not run the experiment or independently measured the project's reliability.

Sources accessed November 16, 2026. Views are my own and do not represent any employer, past or present.

Sources and evidence

GPU scheduling needs a recoverable ownership record alongside memory estimates. Test refusals and restart behavior before sharing a card between managed inference and training workloads.