Multiple P-token locations and anchor tokens

[dsosnoski]

The reference https://huggingface.co/docs/peft/v0.11.0/package_reference/p_tuning says "prompt tokens can be added anywhere in the input sequence, and p-tuning also introduces anchor tokens for improving performance", but I don't see any options in the configuration to control this. How can I best handle a multipart input with P-tuning?

It's also confusing that the reference page above shows a PromptEncoder class that uses the PromptEncoderConfig, but the task example https://huggingface.co/docs/peft/main/en/task_guides/ptuning-seq-classification just passes the PromptEncoderConfig directly to get_peft_model. Is there a way of using PromptEncoder for more-flexible handling?

[BenjaminBossan]

but I don't see any options in the configuration to control this.

Yes, you're right, there is no way to configure this. I agree that the documentation is confusing in this regard and should be updated.

How can I best handle a multipart input with P-tuning?

Could you clarify what you mean by that?

It's also confusing that the reference page above shows a PromptEncoder class that uses the PromptEncoderConfig

What do you mean by that?

just passes the PromptEncoderConfig directly to get_peft_model

This is the intended usage. As a user, you should not have to create PromptEncoder manually.