From Hours to Minutes: Lossless Acceleration of Ultra Long Sequence Generation

TokenSwift is a novel framework designed to substantially accelerate the generation process of ultra-long sequences, up to 100K tokens, while maintaining the target model's inherent quality.

Highlights	Description	Emoji
⚡ Speed	3× faster than vanilla Transformers	⏩
🎯 Lossless	Matches original model's output quality	✅
📈 Scalability	Linear time complexity for 100K+ sequences	📏
🛠️ Plug & Play	Works with most HuggingFace models	🤗

---

✨ News

[2025.2.28] Code Release.

[2025.2.27] Paper Release on Arxiv.

---

📦 Demo

tokenswift.mp4

---

📖 Table of contents

• Introduction
• Installation
  • Method 1: With pip
  • Method 2: From the source (recommended)
• Inference
  • Models Download
  • Getting Start
• Training Guide (Option)
  • Datasets Download
  • How to Train
• Citation
• Acknowledgment

---

Introduction

We propose TokenSwift, a novel framework that achieves lossless acceleration for ultra-long sequence generation (up to 100K tokens) while reducing computation time from hours to minutes.

Illustration of TOKENSWIFT Framework. First, target model (LLM) with partial KV cache and three linear layers outputs 4 logits in a single forward pass. Tree-based attention is then applied to construct candidate tokens. Secondly, top-k candidate 4-grams are retrieved accordingly. These candidates compose draft tokens, which are fed into the LLM with full KV cache to generate target tokens. The verification is performed by checking if draft tokens match exactly with target tokens. Finally, we randomly select one of the longest valid draft tokens, and update n-gram table and KV cache accordingly.

This repository contains:

• ✅ 100% reproducibility for all experiments
• 📊 Benchmark scripts for sequence lengths: 20K/40K/60K/80K/100K
• 🤖 Pre-trained model adapters for Any Structure

Visualization of our acceleration performance vs. baseline methods

---

Installation

Method 1: With pip

pip install tokenswift

Method 2: From the source (recommended)

git clone https://github.com/bigai-nlco/TokenSwift.git
cd TokenSwift
conda create -n tokenswift python=3.11
conda activate tokenswift
conda install nvidia::cuda-nvcc
pip install -r requirements.txt
pip install https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiFALSE-cp311-cp311-linux_x86_64.whl

---

Inference

Models Download

Model Name	Download Link
TokenSwift-Yarn-Llama-2-7b-128k	HuggingFace
TokenSwift-Llama-3.1-8B	HuggingFace
TokenSwift-Qwen2.5-1.5B	HuggingFace
TokenSwift-Qwen2.5-7B	HuggingFace
TokenSwift-Qwen2.5-14B	HuggingFace
TokenSwift-DeepSeek-R1-Distill-Qwen-32B	HuggingFace

Getting Start

Take LLaMA3.1-8B as an example:

torchrun  --master-port 1111 --nproc_per_node=1 main.py \
    --model_type llama3_1 \
    --ckpt_path your_checkpoint_path \
    --prefill_len 4096 \
    --retrival_max_budget 4096 \
    --gen_len 102400 \
    --gamma 4 \
    --min_p 0.1 \
    --temperature 1.0 \
    --tree_decoding \
    --ngram_topk 20 \
    --penalty 1.2 \
    --penalty_length 1024 \
    --prompt_id 0

  <NOTE: Modify the data and model path>

For other models, you can run the scripts in infer_scripts/ folder. For example：

bash infer_scripts/r1_qwen_32b.sh

---

Training Guide (Option)

Datasets Download

From the PG-19 training set, data larger than 8K are filtered out according to different tokenizer.

Or download processed training datasets from llama2-pg19, llama3.1-pg19, qwen2.5-pg19.

How to Train

Take LLaMA3.1-8B as an example:

torchrun --master-port 1111 --nproc_per_node=4 train/train_legacy.py \
    --model_name_or_path /your_model_path/Meta-Llama-3.1-8B \
    --llama_type llama3_1 \
    --data_path /your_data_path/llama3_1_pg19_8k_data \
    --output_dir /your_checkpoint_path/adapter_ckpts_llama3_1 \
    --max_steps 200 \
    --per_device_train_batch_size 3 \
    --gradient_accumulation_steps 10 \
    --save_steps 200 \
    --learning_rate 5e-3 \
    --weight_decay 0.1 \
    --warmup_steps 50 \
    --lr_scheduler_type cosine \
    --logging_steps 5 \
    --report_to tensorboard \
    --bf16 True \
    --medusa_heads 3 \
    --remove-unused-columns false
  
  <NOTE: Modify the data and model path>

For other models, you can run the scripts in train/scripts/ folder. For example：

cd train
bash scripts/train_R1_qwen2_5_32b.sh

---

Citation

@misc{wu2025hoursminuteslosslessacceleration,
      title={From Hours to Minutes: Lossless Acceleration of Ultra Long Sequence Generation up to 100K Tokens}, 
      author={Tong Wu and Junzhe Shen and Zixia Jia and Yuxuan Wang and Zilong Zheng},
      year={2025},
      eprint={2502.18890},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.18890}, 
}

Acknowledgment

This codebase is influenced by remarkable projects from the LLM community, including Medusa and TriForce.