Disable/Enable graph level optimizations

[banu-narayana-dnjp]

I have noticed that the several graph optimizations like operator fusions, etc are occured by default while inferencing in TensorRT.
Is there any way to disable them or does TRT have levels of graph ops just as in ONNXRuntime?
Because I want to compare the performaces between before opt and after opt, for ex.

[lix19937]

mark all with polygraphy tool.

[asfiyab-nvidia]

@banu-narayana-dnjp Unfortunately, TensorRT API doesn't support disabling fusions at the moment.
However, you can experiment with the builder optimization level to achieve different fusion patterns. Please refer to doc on optimization level for usage

[banu-narayana-dnjp]

Thank you @asfiyab-nvidia and @lix19937 for your clarifications. Just one question, does MARK_ALL with the polygraphy tool result in any change in the final score(considering it prevents the fusions)?

[lix19937]

Mainly to prevents the fusions.