A GPU endpoint is expensive but shows low GPU utilization and repeated small requests. What should the engineer evaluate before adding more GPUs?

1. Double the GPU count without any load test.
2. Remove required model weights to force faster execution.
3. Benchmark suitable instance sizing, model runtime settings, and supported request batching.
4. Use training accuracy as the only serving efficiency metric.