In Transformers, the quantized models are initialized first on the "meta" device before loading the weights.