This repository contains the original Google Research implementation of Prompt Tuning from the EMNLP 2021 paper. The technique enables efficient adaptation of large pre-trained language models (T5) by training soft prompt embeddings rather than modifying all model parameters. The implementation builds on T5X for model and training logic, Flaxformer for model computation, Flax for low-level layers, and JAX for execution. It provides code for training prompts, performing inference with trained prompts, and includes released model checkpoints and prompts.