✨[Feature] Provide missing operators for YoloR

[rodja]

Is your feature request related to a problem? Please describe.

While compiling YOLOR, Torch-TensorRT reports the following missing opperators:

• aten::Int.Tensor(Tensor a) -> (int)
• aten::div.Tensor_mode(Tensor self, Tensor other, *, str? rounding_mode) -> (Tensor)
• prim::PythonOp(...) -> (...)

Describe the solution you'd like

I would like these converters to be available in Torch-TensorRT so I can run YoloR inferences.

Describe alternatives you've considered

I tried executing this operations with Torch instead of TensorRT by calling

ttrt_module = torch_tensorrt.compile(
    model,
    inputs=inputs,
    enabled_precisions=enabled_precisions,
    require_full_compilation=False,
    torch_executed_ops=['aten::Int', 'aten::div', 'prim::PythonOp'],
)

But the errors remained.

[rodja]

It looks like aten::Int.Tensor(Tensor a) is not so easy to implement: #513

[narendasan]

Yeah we are working on a solution for aten::Int. The suggestion here is to try to help torchscript here with explicit casting if possible for now. We can work on aten::div.Tensor_mode. prim::PythonOp I dont think we can support. Partial compilation would be the only option here

[github-actions]

This issue has not seen activity for 90 days, Remove stale label or comment or this will be closed in 10 days