Will this bring any improvment to a TensorRT model ?

[joihn]

I guess the speedup is only aplicable to pytorch inference, right ?
If if then convert it to onnx then to tensorRT, there won't be any speedup compared to convert a model containing classic cat + conv layers ?