Feature Request: P-Tuning v2 (Deep Prompt Tuning) Support

[hwilner]

Summary

I would like to propose adding support for P-Tuning v2 (also known as Deep Prompt Tuning) to the prompt-tuning library. This technique extends standard prompt tuning by adding trainable prompts at every layer of the transformer, not just the input layer.

Motivation

P-Tuning v2 has been shown to achieve performance comparable to full fine-tuning across various scales and tasks, while maintaining the parameter efficiency of prompt tuning. The key advantage is that it enables more fine-grained control over intermediate representations by inserting prompts at multiple layers.

Proposed Implementation

The implementation would include:

1. Layer-specific prompts: Trainable prompt parameters for each transformer layer
2. Optional shared prompts: Parameter-efficient variant where prompts are shared across layers
3. MLP-based prompt encoder: Generate layer-specific prompts from a shared representation
4. Compatibility: Work seamlessly with both encoder-decoder and decoder-only models

Key Features

• Prompts at every transformer layer (not just input)
• Configurable prompt length per layer
• Optional parameter sharing across layers
• Integration with existing T5X/Flaxformer infrastructure

Reference

Liu et al. (2022). "P-Tuning v2: Prompt Tuning Can Be Comparable to Fine-tuning Universally Across Scales and Tasks." arXiv:2110.07602

Additional Context

I have implemented a prototype of this technique that follows the library's design patterns and coding standards. The implementation is available in my fork at https://github.com/hwilner/prompt-tuning

Would the maintainers be interested in this enhancement? I'm happy to discuss the design and implementation details further.