Exploring Model Invariance with Discrete Search for Ultra-Low-Bit Quantization