problem in converting fp32 to fp16

Hi guys, I'm trying to convert onnx model fp32 to fp16. but I faced This messages in the function 'remove_unnecessary_cast_node'

"The downstream node of the second cast node should be graph output" (onnxconverter_common/float16.py:557)

I knew That message caused by None of Downstream node I am trying to use the FP16 model as much as possible. Is there any other method I can try? Would using TensorRT instead of ONNX improve the situation? thank you