Try theta_2 cuda upload to use Triple sum

[whdlgp]

The following errors may occur when using the "Triple sum" function.

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

The problem is,

• theta_0 and theta_1 were uploaded to CUDA with the code below
• but theta_2 was not uploaded.
  code in "mergers.py", 418 line

        theta_0[key] = theta_0[key].to("cuda")
        theta_1[key] = theta_1[key].to("cuda")

To solve this, I add simple check code for theta_2

        theta_0[key] = theta_0[key].to("cuda")
        theta_1[key] = theta_1[key].to("cuda")
        try:
            theta_2[key] = theta_2[key].to("cuda")
        except NameError:
            None

This method increases GPU memory usage by putting theta_2 on the GPU, but it succeeds if the GPU memory capacity is sufficient.

In my testing, I can do Triplesum on an RTX3080Ti with 12GB of VRAM. But the memory usage is pretty close.

Please review my commit and I hope it will help you.

Added: Fix exception handling

Updated the exception handling from except NameError to the more general except Exception to handle 2 model merge

        except NameError:
            None

to

        except Exception as e:
            pass # Do nothing