Move model to device after pytorch interpolation

[joeyballentine]

Fixes the bug mentioned by metin on discord

[RunDevelopment]

... Why does this solve the bug?

[joeyballentine]

Since the load model node moves the model to gpu, if you did an interp into another interp, where the second interp comes from a load model node (so model 1 and model 2 into interp 1, and interp 1 and model 3 into interp 2), then one model would be on the gpu and one would be on cpu. This fixes that and makes it consistent.

Alternatively, perhaps we could change the model input's enforce method to always move it to the specified device?

[RunDevelopment]

Okay, I just looked at the discord messages. This is not a fix.

The issue is that the node currently can't interpolate 2 models when they are on different devices. So you tried to make sure that this node puts the new model it returns on the same device as other nodes.

Why is this not a fix? You didn't solve the issue that this node cannot interpolate 2 models when they are on different devices. The cause of the issue still remains and there will be a bug report again in the future. For example, like this:

1. Add a Load model node and select a model.
2. Go to settings and change PyTorch from GPU to CPU (or vise versa).
3. Add another Load model node.
4. Setup the chain to interpolate both models are run it.
5. Same error.

Instead of assuming that models will always be on the same device, just move both models to the same device.

[RunDevelopment]

Alternatively, perhaps we could change the model input's enforce method to always move it to the specified device?

This would require inputs to have access to the current node execution context. Not ideal, since this would prevent optimizations in the future.