cuda-agent-team

To speed up CUDA kernel tuning when single-agent optimization plateaus, run a parallel multi-agent approach that assigns distinct optimization directions to multiple worker agents exploring concurrently. This coordinates concurrent exploration across multiple GPUs to find better kernels faster.

Parallel GPU pinning works by assigning each worker agent to its own isolated GPU via CUDA_VISIBLE_DEVICES, ensuring contention-safe execution. This keeps speedup measurements comparable while each agent runs an isolated experiment loop for concurrent kernel optimization.

Yes, you need at least two GPUs to run parallel CUDA kernel optimization. The orchestration requires per-agent working directory isolation and GPU assignment via CUDA_VISIBLE_DEVICES to ensure isolated, contention-safe execution for comparable speedup measurements.

The merge protocol for combining CUDA kernel optimizations tracks per-worker perf logs to decide when to adopt a winning baseline. It merges compatible wins through defined synchronization rounds with mandatory correctness checks and speedup validation for any merged kernel.

You should split CUDA kernel tuning into orthogonal hypotheses across multiple GPUs when your tuning progress has flattened for multiple iterations. This allows you to explore distinct directions like memory restructuring versus compute restructuring concurrently to overcome single-agent plateaus.