I am doing LlamaForCausalLM.from_pretrained, however, for fast development turnaround, is there a way to pass additional parameter to the from_pretrained method, to indicate that I just want to create a dummy model in memory, with same structure as the pretrained model (on disk), but just with empty or random weights, in order to increase speed for loading? I don’t care about weights being loaded from disk, just want to load it to memory as soon as possible.