Model.to(device) method doesn't work on models instantiate as list

mps
1

Let me know please if this is expected behaviour.

    class ModelB(nn.Module):
        def __init__(self) -> None:
            super().__init__()
            self.layer = nn.Linear(1, 1)

        def forward(self, x):
            self.layer(x)
            return x

    class ModelA(nn.Module):
        def __init__(self) -> None:
            super().__init__()
            self.model = ModelB()
            self.models = [ModelB()]
        def forward(self, x):
            x = self.model(x)
            for model in self.models:
                x = model(x)
            return x

        
    X = torch.rand(10, 1).to('mps')
    model_a = ModelA().to('mps')
    out = model_a(X)
    print(out)

the error is on this line: x = model(x) while being almost the same as correct: x = self.model(x)

RuntimeError: Placeholder storage has not been allocated on MPS device!

when I try: self.models = nn.ModuleList([ModelB()]) it again works correctly.

2

This is expected and you would need to use nn.ModuleList. The same limitations apply to parameters, too (where you would need to use nn.ParameterList instead).