Some questions about half-precision

I have got a tensorrt model by:
python tools/trt.py -f exps/example/yolox_voc/yolox_myvoc_m.py -c best_ckpt.pth
but I get :
RuntimeError: Input type (torch.cuda.FloatTensor) and weight type (torch.cuda.HalfTensor) should be the same
when I used the tensorrt model to inference by:
python tools/demo.py image -f exps/example/yolox_voc/yolox_myvoc_m.py --path dog.jpg --trt --save_result --fp16

this is because of the code in the tools/demo.py:
`

    if trt_file is not None:
        from torch2trt import TRTModule

        model_trt = TRTModule()
        model_trt.load_state_dict(torch.load(trt_file))

        x = torch.ones(1, 3, exp.test_size[0], exp.test_size[1]).cuda()
        self.model(x)
        self.model = model_trt

`

the dtype of x is torch.float32 but the dtype of the models weights is torch.float16 at this time.

However,when i add a line of code (x = x.half()) befor the self.model(x)，like this:
`

     if trt_file is not None:
         from torch2trt import TRTModule

         model_trt = TRTModule()
         model_trt.load_state_dict(torch.load(trt_file))

         x = torch.ones(1, 3, exp.test_size[0], exp.test_size[1]).cuda()

         x = x.half()

         self.model(x)
         self.model = model_trt

`

and to run the demo.py by:
python tools/demo.py image -f exps/example/yolox_voc/yolox_myvoc_m.py --path dog.jpg --trt --save_result --fp16

I got the following Error finnaly:
RuntimeError: CUDA error: an illegal memory access was encountered

So,how could I overcome this problem？

Start with tools/demo.py and the TensorRT setup path invoked by tools/trt.py, then reproduce the reported command using --trt and --fp16. Compare the input dtype used during TensorRT initialization with the loaded model weights and investigate the illegal CUDA memory access. Done means the demo inference runs successfully with the reported commands.