Why Merged LoRA Barely Changes Inference Time

While my peer was benchmarking a sales conversion classifier fine-tuned on Qwen3-0.6B, a merged LoRA...