[QUESTION] Limitation of TensorRT I/O format when using a Cutlass convolution

Hello,

This is not so much an issue with a particuliar TensorRT version as much a general compatibility issue between two NVIDIA software.
According to the documentation of TensorRT the only supported I/O format for INT8 data types are either kLinear or kCHWx. As I/O also applies to passed passed to and from plugins, this means an INT8 plugin should have this data format to avoid reformatting.
Now, I mainly want to leverage Cutlass in my plugin, and they have confirmed to me that their ImplicitGemm convolutions only support channels as the leading dimension. Plus, its still unclear that I can use bias when using an interleaved format, so my only real option right now is to use NHWC.
But this is unfortunate, as both these products are NVIDIA-made.

Is there any way to pass INT8+kHWC data to a plugin in TensorRT? I'm open to any tricks you might have. Thanks.