[MetaSchedule] TensorRT BYOC

[sunggg]

Enable TensorRT with meta schedule builder and runner.
Extended BuilderInput and RunnerInput to support graph-level execution by passing params.
Include test case for the sanity check.

[junrushao]

Suggested change

	params: dict
	params: Optional[Dict[str, NDArray]]

[junrushao]

Suggested change

	def __init__(self, mod: IRModule, target: Target, params: dict) -> None:
	def __init__(self, mod: IRModule, target: Target, params: Optional[Dict[str, NDArray]] = None) -> None:

[junrushao]

ditto

[junrushao]

I'm not sure but looks like we need to check nullability?

def _serialize_params(params: Optional[Dict[str, NDArray]]) -> Optional[bytearray]:
    if params is None:
        return None
    return save_param_dict(params) 

def _deserialize_params(params: Optional[bytearray]) -> Optional[Dict[str, NDArray]]:
    if params is None:
        return None
    return load_param_dict(params) 

[junrushao]

Suggested change

	params: dict,
	params: Optional[bytearray],

[junrushao]

The reason that we cannot use kwargs is that register_func accidentally converts a python function to TVM's PackedFunc (you may inspect the type using print(type(_build_module_no_factory))), which doesn't support kwargs. I will submit a workaround for this

[sunggg]

Gotcha. Will remove this comment and submit the updated PR.

[junrushao]

Thanks @sunggg!