Does Prefix Tuning support T5 later?

[fade-color]

I saw T5's source code doens't support adding past_key_values ​​to the encoder in their code base. in

OpenPrompt/openprompt/prompts/prefix_tuning_template.py

Lines 18 to 26 in b8d9230

	class PrefixTuningTemplate(Template):
	r"""This template different from most template in that this emplate doesn't need to
	wrap the input sentences with the template. The new tokens are always prepended to
	the language model. A mapping is used to map the new_tokens embeddings in to the
	past_key_value, and then input into the language model. The mask token of this
	template is automatically the last token. Currently, our implementation of
	prefix_tuning doens't support adding past_key_values to the encoder side of an
	encoder_decoder architecture such as T5 without modifying the T5 source code.
	(T5's source code doens't support adding past_key_values to the encoder in their code base. )

, Does this show that the prefix tuning is not compatible with T5? Will T5 will be supported later? thanks.

[ningding97]

I saw T5's source code doens't support adding past_key_values ​​to the encoder in their code base. in

OpenPrompt/openprompt/prompts/prefix_tuning_template.py

Lines 18 to 26 in b8d9230

	class PrefixTuningTemplate(Template):
	r"""This template different from most template in that this emplate doesn't need to
	wrap the input sentences with the template. The new tokens are always prepended to
	the language model. A mapping is used to map the new_tokens embeddings in to the
	past_key_value, and then input into the language model. The mask token of this
	template is automatically the last token. Currently, our implementation of
	prefix_tuning doens't support adding past_key_values to the encoder side of an
	encoder_decoder architecture such as T5 without modifying the T5 source code.
	(T5's source code doens't support adding past_key_values to the encoder in their code base. )

, Does this show that the prefix tuning is not compatible with T5? Will T5 will be supported later? thanks.

Hi, yes, we are working on it.

[ShengdingHu]

It has been fixed, see #27