LW-DETR Latency

[probicheaux]

How were you able to measure ~16ms for LW-DETR xlarge when the original paper measures ~20ms (and I can reproduce that number on a T4)

[Peterande]

You can try using the latest TensorRT and ensure that you have exclusive access to the entire server bandwidth. We often encounter issues with unstable latency on certain server platforms, like Colab, but everything works fine locally.