I understand I need to allocate both the input tensor and the model parameters to the mps device in order for PyTorch to use my Mac M1 GPU for training. I did just that and it still gave me this error message:

*Placeholder storage has not been allocated on MPS device! * Here is a snippet of my code.

if torch.backends.mps.is_available():
    mps_device = torch.device("mps")
    x = torch.ones(1, device=mps_device)
    print(x)
else:
    print("MPS device not found")


model = LSTMModel(input_size=197,
                  hidden_size=HIDDEN_UNITS,
                  output_size=1,layer_size=2)

# Transfer the model to the GPU
model = model.to(mps_device)


# iterate over the training data
    for i in range(100000):
        # send the input/labels to the GPU
        
        next_batch = next(train_generate)
        inputs = torch.from_numpy(next_batch[0]).float().to(mps_device)
        labels = torch.from_numpy(next_batch[1][-1]).float().to(mps_device)

        
        with torch.set_grad_enabled(True):
            outputs = model(inputs)
            loss = loss_function(outputs, labels)

            # backward
            loss.backward()
            optimizer.step()

I followed documentation of how to use M1 GPU for PyTorch but it didn't work.