I found two confusing annotation in relax/frontend/onnx/onnx_frontend.py。
in public function from_onnx()，the model’s type should be onnx.onnx_ml_pb2.ModelProto，as we will take its graph member, and in class ONNXGraphImporter.from_onnx()，the type of graph should be onnx.onnx_ml_pb2.GraphProto。
but in onnx_frontend.py file, the two “from_onnx()” annotations are totally reversed! which may confusing readers, hope the tvm official will modify it later. 
tvm/python/tvm/relax/frontend/onnx/onnx_frontend.py at unity · apache/tvm (github.com)
class ONNXGraphImporter:
...
def from_onnx(self, graph: onnx.onnx_ml_pb2.ModelProto, opset: int) -> IRModule:
...
def from_onnx(
model: onnx.onnx_ml_pb2.GraphProto,
shape_dict: Optional[Dict[str, List]] = None,
dtype_dict: Optional[Union[str, Dict[str, str]]] = "float32",
opset: int = None,
keep_params_in_input: bool = False,
sanitize_input_names: bool = True,
) -> Tuple[IRModule, Dict]:
"""Convert a ONNX model into an equivalent Relax Function.
ONNX graphs are represented as Python Protobuf objects.
The current implementation assumes that the input model is after ONNX v1.1.0.
Parameters
----------
model : protobuf object
ONNX ModelProto after ONNX v1.1.0
...