Pythorch What is the difference between these two codes?Please tell me who knows why only one of them has an error.

a=torch.tensor([10]]).to("cuda:0").half()
b=torch.tensor([2]).to("cuda:0").half()
print(type(a),a)
print(type(b),b)
print(a+b)
print("ok")

<class'torch.Tensor'>tensor([10.]], device='cuda:0', dtype=torch.float16)
<class'torch.Tensor'>tensor([2.], device='cuda:0', dtype=torch.float16)
tensor([12.]], device='cuda:0', dtype=torch.float16)
ok

 target=(gamma**multireward_steps) * targetQN.forward (memory.buffer [idx][0], "net_v")
rew=memory.buffer [idx][2].to("cuda:0")
print(type(targ), targ)
print(type(rew),rew)
targets[i] = rew+targ

<class'torch.Tensor'>tensor([0.0208]], device='cuda:0', dtype=torch.float16)
<class'torch.Tensor'>tensor([0.], device='cuda:0', dtype=torch.float16)

---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
<ipython-input-1-dff82183a33b>in<module>
    420 trin.pioritized_experience_replay(batch_size,gamma,step=episode,
    421 state_size=state_, action_size=athon,
-->422 multilireward_steps=multireward_steps)
    423 trin. Done (episode)
    424mainQN.Done()

<ipython-input-1-dff82183a33b>implemented_experience_replay(self, batch_size, gamma, step, state_size, action_size, multilireward_steps)
    289 print (type(targ), targ)
    290 print (type(rew), rew)
-->291 targets[i] = rew+targ
    292 
    293 priority = rank_sum (memory_TDerror.buffer [idx], self.alpha)

~\Anaconda3\envs\pyflan\lib\site-packages\torch\tensor.py in __array__(self, dtype)
    492 return self.numpy()
    493 else:
-->494 return self.numpy().astype(dtype, copy=False)
    495 
    496#Wrap Numpy array again in a sustainable sensor when done, to support e.g.

TypeError: can't convert cuda:0 device type sensor to numpy.Use Tensor.cpu() to copy the sensor to host memory first.