[E:onnxruntime:Default, provider_bridge_ort.cc:1862 TryGetProviderInfo_CUDA] /onnxruntime_src/onnxruntime/core/session/provider_bridge_ort.cc:1539 onnxruntime::Provider& onnxruntime::ProviderLibrary::Get() [ONNXRuntimeError] : 1 : FAIL : Failed to load library libonnxruntime_providers_cuda.so with error: libcudnn_adv.so.9: cannot open shared object file: No such file or directory

find ./venv -type f -name 'libcudnn_adv.so.9'
./venv/lib/python3.11/site-packages/nvidia/cudnn/lib/libcudnn_adv.so.9

pip install torch==2.6.0+cu124
pip install onnxruntime-gpu==1.20.1

import onnx
import onnx.helper as helper
import torch # this used to help ort find the libs but now this trick stop working
import onnxruntime as ort
import numpy as np

def test_onnxruntime_cuda_provider():
    # 1. Create a simple ONNX model (addition of two inputs)
    # Define input shapes and types
    input_shape = [2, 2]
    input_type = np.float32

    # Create ONNX model
    input1 = helper.make_tensor_value_info('input1', onnx.TensorProto.FLOAT, input_shape)
    input2 = helper.make_tensor_value_info('input2', onnx.TensorProto.FLOAT, input_shape)
    output = helper.make_tensor_value_info('output', onnx.TensorProto.FLOAT, input_shape)

    # Define a simple addition node
    add_node = helper.make_node('Add', inputs=['input1', 'input2'], outputs=['output'])

    # Create the graph
    graph = helper.make_graph([add_node], 'add_graph', [input1, input2], [output])

    # Create the model
    model = helper.make_model(
        graph,
        opset_imports=[helper.make_opsetid("", 21)]
    )


    # 2. Run inference using ONNX Runtime with CUDA provider
    # Check available providers
    available_providers = ort.get_available_providers()
    assert 'CUDAExecutionProvider' in available_providers, "CUDAExecutionProvider is not available"

    # Create an inference session with the CUDA provider
    session = ort.InferenceSession(model.SerializeToString(), providers=['CUDAExecutionProvider'])

    # Prepare input data
    input_data1 = np.array([[1.0, 2.0], [3.0, 4.0]], dtype=input_type)
    input_data2 = np.array([[5.0, 6.0], [7.0, 8.0]], dtype=input_type)

    # Run inference
    outputs = session.run(None, {'input1': input_data1, 'input2': input_data2})

    # Verify the output
    expected_output = np.array([[6.0, 8.0], [10.0, 12.0]], dtype=input_type)
    np.testing.assert_allclose(outputs[0], expected_output, rtol=1e-5, atol=1e-5)

    print("Test passed! Output matches expected result.")

if __name__ == "__main__":
    test_onnxruntime_cuda_provider()