PyTorch 'RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!' (also 'Input type (torch.cuda.FloatTensor) and weight type (torch.FloatTensor) should be the same')

Cause: An operation received tensors living on different devices, typically model weights on the GPU and inputs (or labels, masks, indices, or a tensor created inside the model) on the CPU, or the reverse. PyTorch runs an operation on the device where its tensors are and refuses to mix devices unless the operation is a copy such as to() or cuda(). Fix (best first): 1. Pick one device and use it everywhere: device = torch.device("cuda" if torch.cuda.is_available() else "cpu"). 2. Move the model once, before creating the optimizer: model.to(device). Module.to() works in place, and the docs say the move should happen before the optimizer is constructed. 3. Move every batch: inputs = inputs.to(device); targets = targets.to(device). Tensor.to() is not in place; it returns a new tensor (or the same one if already on that device), so you must assign the result. 4. Inside model code, create new tensors on an existing tensor's device, for example torch.zeros_like(x), x.new_zeros(...), or torch.zeros(..., device=x.device). Tensors you need in the module that are not parameters should be registered with register_buffer so model.to(device) moves them. 5. Check where things are: print(next(model.parameters()).device, x.device). Watch for lists of tensors, tensors in dicts, or loss weights and class-weight tensors that were created on the CPU. 6. With multiple GPUs, make sure all tensors are on the same cuda:N index, not just on "cuda". Verify: the forward and backward pass run, and every tensor involved reports the same device. Sources: https://docs.pytorch.org/docs/2.15/notes/cuda.html https://docs.pytorch.org/docs/2.15/generated/torch.Tensor.to.html https://docs.pytorch.org/docs/2.15/generated/torch.nn.Module.html