gpu-kernel-partial-restart

To restart a stalled GPU kernel optimization session without losing progress, use a partial-restart workflow that masks half of the existing iteration memory and launches a fresh subagent. This breaks stale patterns while preserving valid historical data.

Your agent gets stuck repeating failed strategies during GPU kernel profiling because it falls into local minima. Masking half of the previous iteration memory forces the agent to explore new, productive directions and prevents repeating failed strategies.

Before running a partial-restart for kernel tuning, you need a kernel_opt workspace containing existing memory JSON files and a valid kernel.py starting point. Python is also required as a dependency to execute the workflow.

You should use a subagent to refresh context for GPU kernel tuning when the target performance has not been met and the agent has exhausted all obvious optimization paths. It re-evaluates the current kernel state to force an evidence-based restart.

The best way to break out of local minima during agent-based GPU kernel optimization is to mask half of the iteration memory and orchestrate a fresh subagent. This approach breaks stale patterns while retaining partial progress.