[EMNLP 2024 & AAAI 2026] A powerful toolkit for compressing large models including LLMs, VLMs, and video generative models.

📚 Collection of token-level model compression resources.

A paper list about Token Merge, Reduce, Resample, Drop for MLLMs.

[NeurIPS 2025] HoliTom: Holistic Token Merging for Fast Video Large Language Models

[CVPR 2025] PACT: Pruning and Clustering-Based Token Reduction for Faster Visual Language Models

Official implementation of CVPR 2024 paper "vid-TLDR: Training Free Token merging for Light-weight Video Transformer".

[CVPR'25] MergeVQ: A Unified Framework for Visual Generation and Representation with Token Merging and Quantization

The official implementation of "Learning Compact Vision Tokens for Efficient Large Multimodal Models"

[ICLR 2026] MergeMix: A Unified Augmentation Paradigm for Visual and Multi-Modal Understanding

😎 Awesome papers on token redundancy reduction

DRIP: Dynamic Patch Pooling for Efficient Vision Transformers

[ICLR 2026] Official code of PPE: Positional Preservation Embedding for Token Compression in Multimodal Large Language Models.

Graph-Guided Token Merging (G2TM) is a lightweight one-shot module designed to eliminate redundant tokens early in the ViT architecture. It performs a single merging step after a shallow attention block, enabling all subsequent layers to operate on a compact token set. It leverages graph theory to identify groups of semantically redundant patches.

Implementation of Vision Transformers (ViT) with a token merging mechanism

Efficient vision-language pre-training toolkit: Token Merging, LoRA/QLoRA fine-tuning, Knowledge Distillation

Compress context data to optimize memory and performance in C++ large language model applications within the llm-cpp toolkit.

🚀 Collect and manage tokens effortlessly with this simple, efficient framework for building and maintaining your own token collections.