The error is literally telling you the problem is not the input dtype but the input formatting:

"--custom_input_op_name_np_data_path is not specified, all input OPs must assume 4D tensor image data. INPUT Name: image_repr INPUT Shape: [1, 768] INPUT dtype: float32"

Your model is not using a 4D tensor image data as input currently, but one with the shape [1, 768]. I recommend testing the "--custom_input_op_name_np_data_path" argument first and see if that is enough to complete the quantization process. By the way, is your model using an image as input?