* Add unpack_inputs to remaining models * remove stray use of inputs in the templates; fix tf.debugging of attn masks
tokenizer_config.json