Trilu op still not work with INT32 input about onnxruntime

[1826133674]

Also, I'v tried different version of Onnxruntime from 1.15.0 to 1.19.0-dev-nightly。But they all failed！The Error I get is all the same。like this：

sess.initialize_session(providers, provider_options, disabled_optimizers)
onnxruntime.capi.onnxruntime_pybind11_state.NotImplemented: [ONNXRuntimeError] : 9 : NOT_IMPLEMENTED : Could not find an implementation for Trilu(14) node with name 'Trilu'

[snnn]

Now we support the following types for the first input:

1. tensor(bool)
2. tensor(double)
3. tensor(float)
4. tensor(int64)

[mindest]

As mentioned above, Trilu for CPUExecutionProvider only supports four types, but for CUDAExecutionProvider all main types are supported. You can try CUDA EP before CPU EP support is added, or you can change the input dtype to int64 if it also meets your need.

And there are some issues with your script, you may need modify it accordingly:

script

import onnx
from onnx import helper, TensorProto
import numpy as np
import onnxruntime

input_1 = helper.make_tensor_value_info('input_1', TensorProto.INT32, [1,1,1024,1024])
- input_2 = helper.make_tensor_value_info('input_2', TensorProto.INT64, [1])
+ input_2 = helper.make_tensor_value_info('input_2', TensorProto.INT64, [])
output = helper.make_tensor_value_info('output', TensorProto.INT32,[1,1,1024,1024])

def build_model():

    node_def = helper.make_node(
        'Trilu',  
        ['input_1', 'input_2'], 
        ['output'],
        name="Trilu" 
    )

    graph_def = helper.make_graph(
        nodes=[node_def], 
        name='SingleOperatorGraph',  
        inputs=[input_1, input_2],  
        outputs=[output], 
        initializer=[]
    )
   
    model_def = helper.make_model(
        graph_def,
        producer_name='MyModel',
        opset_imports=[helper.make_opsetid("", 14)],
        
    
    )
    return model_def

def run_model(model_path):
-   input_data_1 = np.random.randint(10, size=(1,1,1024, 1024)).astype(np.int64)
+   input_data_1 = np.random.randint(10, size=(1,1,1024, 1024)).astype(np.int32)
    input_data_2 = np.array(-1).astype(np.int64)
    np.save("trilu_i1",input_data_1)
    np.save("trilu_i2",input_data_2)
    session = onnxruntime.InferenceSession(model_path,providers=['CPUExecutionProvider'])
    
    input_names = [input.name for input in session.get_inputs()]
    inputs = {input_names[0]: input_data_1, input_names[1]: input_data_2}
    output = session.run(None, inputs)
    print(output)

model_def = build_model()

onnx.checker.check_model(model_def)

save_path = "single_operator_model.onnx"
onnx.save_model(model_def, "single_operator_model.onnx")
print(f"{onnxruntime.__version__=}")
run_model(save_path)