I agree, we're too far down a chain of hypotheticals motivated by "but ONNX must be bad compared to $MODELX.cpp?" Wouldn't make sense to deploy 4-bit quantizat

I agree, we're too far down a chain of hypotheticals motivated by "but ONNX must be bad compared to $MODELX.cpp?" Wouldn't make sense to deploy 4-bit quantization as a product either.