🐛 [Bug] Torch-TRT generated engines are larger than ONNX-TRT's

[zewenli98]

Bug Description

Torch-TRT generated engines are slightly larger than ONNX-TRT's for many models.
BERT is extremely larger.

[zewenli98]

For BERT, when embedding layer is constant folded, the engine size will become larger and larger as inputs increase, while the latency is not affected though.

Currently, I only observed the issue on BERT but not on other models. More issue reports are welcome.

[cehongwang]

For WAN2.1, Torch-TRT's engine size is 40% larger than ONNX-TRT, depending on the batch size. When disabling torch.ops.aten.expand.default constant folding, the model size difference becomes a lot smaller