from transformers import AutoTokenizer, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="cuda:3",
)

There are many GPUs on the server, but I can only use two of them. How should I configure device_map (or other parameters) so that the model runs on both GPUs?