The Hailo-8L can run models with a high resolution input.
Setting aside the error for now. I am not sure training a network on larger resolution is the right approach for what you are trying to achieve. I assume the object will only cover a small area of the image. This means the network will need to learn to detect the same objects in different parts of the image and parts of the network will do the same thing.
Have a look at the Tilling examples in our Tappas repository. They will achieve the same long distance detection capability while keeping the network itself smaller.
GitHub - Hailo Tappas - Tiling
Without thinking through all the details I suspect the performance for the tilling approach should be higher especially with the single PCIe lane and the Hailo-8L having fewer compute clusters and therefore the model likely requiring multiple contexts. You can then use the --batch-size parameter to infer the tiles more efficiently reducing the context switching overhead.
Regarding the error, it looks like it is related to the NMS. I would first try to convert the model without NMS and check whether my assumption about the network size vs performance is true. For performance comparison you can use the hailortcli run command without real data and NMS.