RF-DETR has 52% higer latency than D-FINE

Hello Roboflow team, I am currently measuring the latency of detection models for a personal project. I found out that RF-DETR-L is slower than DFINE-L, YOLO(11 and 26)-L, and RT-DETR-50. I run this benchmark using an NVIDIA RTX A5500 and your `single_artifact_benchmarking` repo. The results are: ```shell Model Runtime FP16 mAP50 mAP50-95 AP75 Latency Throttled ----------------------------------------------------------------------------------------- rtdetr_r50_coco.onnx TRT yes 73.3 55.0 59.8 3.51 no rf-detr-medium.onnx TRT yes 73.5 54.7 59.1 1.95 no rf-detr-large.onnx TRT yes 77.3 59.0 64.3 5.88 no dfine_l_obj2coco_e25.opset17.onnx TRT yes 74.9 57.2 62.2 3.86 no yolo11l_nms_conf_0.01.onnx TRT yes 64.9 50.0 54.5 4.71 no yolo26l.onnx TRT yes 71.1 54.1 59.3 2.88 no ``` My question is: why is RF-DETR-L slow? It's even 3times slower than RF-DETR-M