openvino inference crachs after several successful inferences. my bug?

[grofitGV4M]

I convert torch model to scripted jit model, and than to openvino.
the model get 105 inputs and return 106 outputs.
the reasone for so many inputs and outputs is that this model is a stage on the way to stateful model.
however, I do:

        scripted_model = torch.jit.script(model.eval(), (dummy_inputs[0], prev_input, prev_ft, prev_mask, *buffers_list))
        scripted_model.eval()
        scripted_model_path = osp.join(osp.dirname(output_path), 'script_model.pt')
        scripted_model.save(scripted_model_path)
        torch._C._jit_pass_inline(scripted_model.graph)
        torch._C._jit_pass_dce(scripted_model.graph)  # Dead code elimination
        torch._C._jit_pass_constant_propagation(scripted_model.graph)
        torch._C._jit_pass_peephole(scripted_model.graph, True)
        scripted_model.eval()
        input_shapes = {'new_frame.1': ([1, 1, 160], ov.Type.f32), 'prev_input.1': ([1, 1, 160], ov.Type.f32), 'prev_frame_ft.1': ([1, 1, 321, 1, 2], ov.Type.f32),
                        'prev_mask.1': ([1, 1, 321, 1, 2], ov.Type.f32)}
        for i, buffer_shape in enumerate(buffer_manager.buffer_shapes):
            input_shapes[f'unet_buff_{i}.1'] = (buffer_shape, ov.Type.f32)
            print(f'input_shapes[unet_buff_{i}] = {buffer_shape}')
        example_input = [torch.randn([1, 1, 160]), torch.randn([1, 1, 160]), torch.randn([1, 1, 321, 1, 2]), torch.randn([1, 1, 321, 1, 2])]
        for shape in buffer_manager.buffer_shapes:
            example_input.append(torch.randn(shape))
        ov_model_32 = ov.convert_model(input_model=scripted_model, example_input=example_input, input=input_shapes, verbose=True, share_weights= False)
        ov.save_model(ov_model_32, osp.join(osp.dirname(output_path), 'ov_scripted_model.xml'), compress_to_fp16=False)
            ov_scripted_path = osp.join(osp.dirname(output_path), 'ov_scripted_model.xml')
            ovc_model = ov.runtime.compile_model(ov_scripted_path, config={'device': 'CPU'})
            infer_request1 = ovc_model.create_infer_request()
            ov_combined_buffer = combined_buffer.clone().zero_()
            manual_states = []
            prev_input = torch.zeros([1, 1, 160], dtype=torch.float32)
            prev_ft = torch.zeros([1, 1, 321, 1, 2], dtype=torch.float32)
            prev_mask = torch.zeros([1, 1, 321, 1, 2], dtype=torch.float32)
            manual_states.append(prev_input.numpy())
            manual_states.append(prev_ft.numpy())
            manual_states.append(prev_mask.numpy())
            ov_buffers = buffer_manager.get_buffers_list(ov_combined_buffer)
            for buff in ov_buffers:
                manual_states.append(buff.numpy())
            for i in range(num_frames):
                try:
                    inputs = {ovc_model.inputs[0].any_name: dummy_inputs[i].numpy()}
                    for k, state in enumerate(manual_states):
                        inputs[ovc_model.inputs[k+1].any_name] = state
                    infer_request1.infer(inputs=inputs)
                except Exception as e:
                    print(e)
                    print('Error at frame:', i)
                    break
                output_frame_ov = np.array(infer_request1.get_tensor(ovc_model.output(1)).data)
                frame_toll_rate = np.mean(np.abs(output_frame_ov - output_frame_th_list[i]) < 1e-4)
                print(f'frame {i} toll rate: {frame_toll_rate}')
                for k, state in enumerate(manual_states):
                    manual_states[k][:] = np.array(infer_request1.get_tensor(ovc_model.output(k+2)).data)

after some inferences that work correctly, with valid outputs:

Exception from src/inference/src/cpp/infer_request.cpp:223:
Check 'found_port.found()' failed at src/inference/src/dev/isync_infer_request.cpp:223:
Cannot find tensor for port opset1::Parameter new_frame.1[0]:f32[1,1,160]

Error at frame: 43

[Iffa-Intel]

@grofitGV4M it seems that the script that you share is incomplete for us to validate.

Could you help to clarify & provide:

1. The model files or the complete script to generate them
2. I believe this is a custom model?
3. That error persist when trying to convert from original format into OpenVINO format or during inferencing?
4. If during inferencing, could you share the complete inferencing code that you use? or are you referring to any of this OpenVINO sample?