GPU Runtime Idle Timeout Configuration and VM Recycling

Google Colab provides GPU runtimes with a maximum continuous usage limit of 12 hours. However, the platform also implements an idle timeout mechanism that can trigger VM recycling when no cell execution is detected for a period of time.

While documentation suggests a guideline of approximately 90 minutes for idle disconnections, the actual behavior appears inconsistent. There is uncertainty regarding whether this threshold is a static limit or a dynamic value that fluctuates based on current server load and resource availability.

For long-running computations that involve periods of inactivity between active processing stages, the lack of a configurable or guaranteed idle window creates a risk of unexpected state loss.

• Is the idle timeout threshold static across all GPU runtime tiers?
• Does the platform provide a mechanism to explicitly extend the idle window for a specific session?