Awesome

$\text{LLM}\times\text{MapReduce}$: An Effective Divide-and-Conquer Framework for Long-Sequence Processing

📖 Introduction

Enlarging the context window of large language models (LLMs) has become a crucial research area, particularly for applications involving extremely long sequences. We introduce $\text{LLM}\times\text{MapReduce}$, a novel training-free framework for processing long sequences, utilizing a divide-and-conquer strategy to achieve comprehensive document understanding. The proposed $\text{LLM}\times\text{MapReduce}$ framework splits the entire document into several chunks for LLMs to read and then aggregates the intermediate answers to produce the final output. The main challenge for divide-and-conquer long-sequence processing frameworks lies in the risk of losing essential long-range information when splitting the document, which can lead the model to produce incomplete or incorrect answers based on the segmented texts. Disrupted long-range information can be classified into two categories: inter-chunk dependency and inter-chunk conflict. We design a structured information protocol to better cope with inter-chunk dependency and an in-context confidence calibration mechanism to resolve inter-chunk conflicts. Experimental results demonstrate that $\text{LLM}\times\text{MapReduce}$ can outperform representative open-source and commercial long-context LLMs, and is applicable to several different models.

🎉 News

• 20241012: Released our paper on arXiv. 🎇
• 20240912: Introducing the $\text{LLM}\times\text{MapReduce}$ framework, which delivers strong performance on long-sequence benchmarks and is compatible with various open-source LLMs. 🎊

✨ Features

1. Divide-and-Conquer Strategy: The entire document is divided into chunks, which are processed individually by LLMs.

2. Structured Information Protocol: a structured information protocol ensures that crucial information flows between the map and reduce stages, preventing information loss when documents are split into chunks and enabling coherent answers for complex questions.

3. In-Context Confidence Calibration Mechanism: a dynamic mechanism that resolves conflicts between outputs from different chunks, ensuring the final result is accurate, consistent, and contextually aligned across the entire document.

⚡️ Getting Started

To get started, ensure all dependencies listed in requirements.txt are installed. You can do this by running:

pip install -r requirements.txt

Starting the Parallel Processing Backend

To enable parallel processing, you need to start the parallel processing backend.

Run the following command:

bash URLs/start_gunicorn.sh --hf-model-name=your/model/path --per-proc-gpus 2 --quantization None --cuda-visible-devices 0,1,2,3,4,5,6,7 --port=5002

Where:

• --hf-model-name: Specifies the path to your Hugging Face model.
• --per-proc-gpus: Defines the number of GPUs required per worker to load the model.
• --quantization: Specifies the quantization method applied to the model, or None if no quantization is used.
• --cuda-visible-devices: Lists the GPUs to be utilized. Ensure the number of GPUs matches the formula per-proc-gpus * worker_num = len(cuda-visible-devices).
• --port: Specifies the port number on which the backend server will listen.
• --max-model-len: Model context length. If unspecified, will be automatically derived from the model config.

The worker_num is automatically calculated based on the formula len(cuda-visible-devices) / per-proc-gpus. While you don’t need to set it directly, you should ensure that worker_num is consistent with the max_work_count value set in your configuration when modifying the config later. A higher worker_num allows for more parallel processing, which can improve performance by enabling multiple tasks to be processed concurrently. However, ensure that you have sufficient GPU resources to support the number of workers.

We also provide example scripts located in URLs/scripts, which include the following models:

• Llama3-70b-instruct: You can modify the provided script URLs/scripts/start_Meta-Llama-3-70B-Instruct.sh.
• Qwen2-72B-Instruct: You can adjust the settings in URLs/scripts/start_Qwen2-72B-Instruct.sh.
• MiniCPM3-4B: Note that using MiniCPM3 requires setting up the environment. You can find the installation instructions in the MiniCPM GitHub repository. After setting up the environment, you can modify URLs/scripts/start_MiniCPM3-4B.sh to suit your setup.

You can modify these scripts according to your requirements to fit your specific setup.

Modify Config

The configuration file is located in the config/ directory. This file allows you to set various parameters for the model, including prompts for each stage of processing. Below is an example configuration:

llm: 
  name_or_path: your/model/path


url: http://localhost:5002/infer
max_work_count: 4

map_prompt: MAP_PROMPT

collapse_prompt: COLLAPSE_PROMPT

reduce_prompt: REDUCE_PROMPT

Key Fields

• llm.name_or_path: Specifies the path to the model, which should match the hf-model-name set in the backend.
• url: The endpoint for the inference service. The default port is 5002, which should align with the port specified in the backend.
• max_work_count: Specifies the maximum number of workers, which should match the worker_num set in the backend.
• map_prompt: The prompt template for the "map" stage.
• collapse_prompt: The prompt template for the "collapse" stage.
• reduce_prompt: The prompt template for the "reduce" stage.

You can modify these prompts and settings to suit your specific tasks. Be sure to adjust paths and parameters based on your environment and model setup.

📃 Evaluation

We provide scripts to evaluate our framework using InfiniteBench in the scripts/ directory. Follow the steps below to set up the evaluation:

1. Download the Dataset

Before running the evaluation, you need to download the InfiniteBench dataset. Refer to the InfiniteBench repository for detailed instructions on how to obtain the dataset. Once downloaded, note the directory where the dataset is stored. We recommend storing the dataset in the data/ directory, which is the default directory used in the provided scripts.

2. Modify the Evaluation Script

We provide evaluation scripts in the scripts/ directory. Here's an example script for evaluating the En.MC task:

output_dir='output/path'  #output path
task='longbook_choice_eng'
data_dir='your/data/dir'
mkdir ${output_dir}


export TOKENIZERS_PARALLELISM=false
python -u eval/infinitebench/eval_infinitebench_MR.py \
    --task=${task} \
    --output_dir=${output_dir} \
    --data_dir=${data_dir} \
    --config_file='config/qa.yaml' 

python -u eval/infinitebench/process_answer.py \
    --result_dir=${output_dir}

python eval/infinitebench/compute_scores.py \
    --task=${task} \
    --output_dir=${output_dir}/'processed' \

You can modify the following parameters as needed:

• task: Set the task you want to evaluate.
• data_dir: Specify the directory where the dataset is stored. Make sure this points to the correct path for the downloaded dataset.
• output_dir: Set the directory where the evaluation results will be saved.
• config: Define the path to your model configuration file. The prompts and settings in the config we provide are already properly aligned with the task, so no further changes should be necessary unless you have specific requirements.

Additionally, modify the 7th line of eval/infinitebench/eval_infinitebench_MR.py

sys.path.append('/path/to/the/project')

Replace /path/to/the/project with the root directory of your project.

3. Run the Evaluation

After modifying the script, run it to evaluate the performance of your framework. The results will be saved in the specified output_dir. Since the output is in a structured format, you can find the extracted answers in output_dir/processed after running the scripts.

📊 Experiment Results

Our experiments demonstrate the improved performance of various LLMs using the $\text{LLM}\times\text{MapReduce}$ framework on InfiniteBench tasks. Detailed results are provided below.

📝 Citation

@misc{zhou2024llmtimesmapreducesimplifiedlongsequenceprocessing,
      title={LLM$\times$MapReduce: Simplified Long-Sequence Processing using Large Language Models}, 
      author={Zihan Zhou and Chong Li and Xinyi Chen and Shuo Wang and Yu Chao and Zhili Li and Haoyu Wang and Rongqiao An and Qi Shi and Zhixing Tan and Xu Han and Xiaodong Shi and Zhiyuan Liu and Maosong Sun},
      year={2024},
      eprint={2410.09342},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2410.09342}, 
}