Roboflow's SDK doesn't have explicit documentation for device management, but it likely leverages the underlying machine learning frameworks (like PyTorch or TensorFlow) for device allocation. Here's how you can approach the three questions based on typical behavior in ML libraries:

1. Is the model loaded to GPU by default?

   • Not necessarily. If you don't explicitly specify the device, most ML libraries like PyTorch or TensorFlow will default to the CPU unless a GPU is available and set as the default device.

2. What happens when multiple GPUs are available?

   • By default, if multiple GPUs are available, the first GPU (cuda:0) is usually selected. However, this depends on the configuration of the underlying library (PyTorch, TensorFlow, etc.). Roboflow SDK would likely follow the default behavior of the framework it's using.

3. How do I define in which GPU I want the model allocated?

   • You can typically set the desired GPU using environment variables or explicit device allocation commands. For example, if Roboflow uses PyTorch, you can set the device like this:

     device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
     model.to(device)
     

   You'd need to ensure that Roboflow provides access to the underlying model for manual device allocation if it's not set automatically.

To confirm exact behavior, you may need to refer to Roboflow's support or inspect the framework behind the SDK for device handling.