vLLM-Watermark

Tiny. Hackable. Lightning-fast watermarking for researchers built on vLLM

Getting started

Installation Guide

Performance evaluation

Watermarking algorithms

Supported Algorithms

Algorithm	Description	Paper
Gumbel/OpenAI	Gumbel-Max trick for deterministic sampling	Aaronson (2023)
Power Law Detection	Near-optimal detection for Gumbel watermarks	Lattimore (2026)
Randomized Gumbel	Gumbel with double randomization for diversity	Verma & Phan (2025)
KGW/Maryland	Context-dependent green-red list with logit bias	Kirchenbauer et al. (2023)
PF (Permute-and-Flip)	Prefix-free coding with token permutations	Lean et al. (2024)
Unigram	Context-independent fixed green-red list	Zhao et al. (2024)
SynthID	Multi-layer tournament watermarking (non-distortionary)	Dathathri et al. (2024)
DIP (DiPmark)	Permutation-based probability redistribution	Wu et al. (2023)
SWEET	Entropy-selective green-list biasing	Lee et al. (2023)
Black-Box	Best-of-m rejection sampling (zero distortion)	Bahri & Wieting (2026)
Alignment Resampling	Best-of-N with reward model (wraps any watermark)	Verma & Phan (2025)

Note

Each algorithm has different trade-offs between detectability, robustness, and text quality. See individual algorithm pages for detailed theory and examples.

Quick start

Install the package (see Installation Guide)
Choose an algorithm from Watermarking Algorithms
Run the example code to try it locally

For detailed API information, refer to the docstrings in the repository code.

Citation

If you use vLLM-Watermark in your research, please cite:

@software{vllm_watermark,
  title  = {vLLM-Watermark: A tiny, hackable research framework for
            LLM watermarking experiments},
  author = {Apurv Verma},
  year   = {2025},
  url    = {https://github.com/dapurv5/vLLM-Watermark}
}