Runtimes for timm + TensorRT

[SimJeg]

Hello,

I created a script to compare inference runtimes with torch, torch.compile and torch_tensorrt.compile for any timm model, input shape and dtype and some runtimes are worse using TensorRT. I opened an issue here and would be interested to get your feedbacks if any,

Simon

[SimJeg]

I added a comparison using ONNX export + trtexec in my issue and also included the kwarg exportable=True. But I still get worse runtimes for the convnext model, and no big improvements for ViT models. @rwightman any hint ?