Hey guys,

When running model generation on Jetson AGX Xavier, we get the following warning:

[TensorRT] INFO:   Calibrated batch 41 in 2.60759 seconds.
Warning: pre_top_k need to be reduced for devices with arch 7.2, got pre_top_k=1000
Warning: pre_top_k need to be reduced for devices with arch 7.2, got pre_top_k=1000

As I understand from the jetson documentation, this only applies to TX2 and Nano but we get the warning also on AGX. Is this intentional? FYI, the model generation appears to work fine on AGX - even when leavning pre_top_k unchanged or reducing it to 1000, as I did in the above example.

Do we reduce pre_top_k due to a cuda Compute Capability limitation or a GPU memory limitation?

I seem to remember an issue describing why pre_top_k must be reduced on some Jetson platforms - but I cannot find it.