Why onnxruntime register custom opsets when custom operators are declared ?

[Johansmm]

Context: I am using ONNXRuntime extensions to declare a custom operation and compare it with its original version :

import onnx
import numpy as np
import onnxruntime
import onnxruntime.tools
import onnxruntime.tools.update_onnx_opset
from onnxruntime_extensions import get_library_path, onnx_op


@onnx_op(op_type="custom_domain::my_op")
def custom_op(x):
    return 5 * x


model1 = onnx.parser.parse_model("""
    < ir_version: 7, opset_import: ["" : 15, "custom_domain": 1] >
    graph (float[N, 10] X) => (float[N, 10] Y){
        Y = custom_domain.my_op(X)
    }""")

# Register onnxruntime library to recognize 'custom_op' in 'model'
so = onnxruntime.SessionOptions()
so.register_custom_ops_library(get_library_path())
sess1 = onnxruntime.InferenceSession(model1.SerializeToString(), so,
                                     providers=["CPUExecutionProvider"])

feed = {'X': np.random.rand(1, 10).astype('float32')}
y1 = sess1.run(None, feed)
print(y1)

model2 = onnx.parser.parse_model("""
    < ir_version: 7, opset_import: ["" : 15] >
    graph (float[N, 10] X) => (float[N, 10] Y){
        const = Constant<float=5.0>()
        Y = Mul(const, X)
    }""")

another_so = onnxruntime.SessionOptions()
another_so.optimized_model_filepath = 'opt_model_path.onnx'
another_so.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_BASIC
# Optimize model2
onnxruntime.InferenceSession(model2.SerializeToString(), another_so,
                             providers=["CPUExecutionProvider"])

opt_model = onnx.load(another_so.optimized_model_filepath)
opt_sess = onnxruntime.InferenceSession(opt_model.SerializeToString(),
                                        providers=["CPUExecutionProvider"])
y2 = opt_sess.run(None, feed)
print(y2)

However, I am surprised to see ONNXRuntime optimizer has added custom opsets in the second model that does not contain my custom operator:

print(opt_model.opset_import)
# [domain: ""
# version: 15
# , domain: "com.microsoft.nchwc"
# version: 1
# , domain: "ai.onnx.ml"
# version: 5
# , domain: "ai.onnx.training"
# version: 1
# , domain: "ai.onnx.preview.training"
# version: 1
# , domain: "com.microsoft"
# version: 1
# , domain: "com.microsoft.experimental"
# version: 1
# , domain: "org.pytorch.aten"
# version: 1
# , domain: "custom_domain" <- Added when I register the library in the first model
# version: 1000
# , domain: "ai.onnx.contrib"  <- Added when I register the library in the first model
# version: 1000
# , domain: "com.microsoft.extensions"  <- Added when I register the library in the first model
# version: 1000
# ]

Therefore, I have the following questions :

• Why has the onnxruntime optimizer added these opsets, seeing that opt_model has nothing to do with the custom operator definition ?
• Is there any way to restrict the opset version that the operator adds when optimizing? I mean, when declaring the custom op through @onnx_op, can I include the version of the opset I want (e.g. a value different from 1000)?