RLM Serving

Official code for ICLR 2026 paper "Reasoning Language Model Inference Serving Unveiled: An Empirical Study"

Project Status

• Core benchmark pipeline is available.
• Paper links and citation are available.
• Clean and release pilot study scripts for RLLM serving.
• Add complete reproduction guide with exact commands and expected outputs.

Paper

• Title: Reasoning Language Model Inference Serving Unveiled: An Empirical Study
• Authors: Qi Li, Junpan Wu, Xiang Liu, Yuxin Wang, Zeyu Li, Zhenheng Tang, Yuhan Chen, Shaohuai Shi, Xiaowen Chu
• Venue: ICLR 2026
• Project Homepage: RLM Serving Homepage
• Code: GitHub Repository
• arXiv: arXiv:2510.18672
• OpenReview: OpenReview Discussion

Environment

This project uses Conda for the base Python environment and pip for most Python packages.

1. Install Conda

Install Miniconda or Anaconda first.

2. Create the Conda environment

conda env create -f environment.yml

If the environment already exists:

conda env update -f environment.yml

3. Activate the environment

conda activate rllm

4. Install Python packages with pip

pip install -r req.txt

environment.yml only contains Conda packages. req.txt contains the Python packages installed by pip.

Pilot Study Scripts (RLLM + vLLM)

The cleaned pilot scripts are in pilot-study/ and are intended for public release.

Script responsibilities

• pilot-study/start-vllm-service.sh: starts and stops vllm serve, then orchestrates dataset-level benchmark runs.
• pilot-study/benchmark-rllm.sh: runs the actual benchmark by calling main.py over datasets and token budgets.
• pilot-study/7b-rllm.sh: preset entrypoint for DeepSeek-R1-Distill-Qwen-7B.
• pilot-study/14b-rllm.sh: preset entrypoint for DeepSeek-R1-Distill-Qwen-14B.
• pilot-study/32b-rllm.sh: preset entrypoint for DeepSeek-R1-Distill-Qwen-32B.

How to run

From repository root:

bash pilot-study/7b-rllm.sh
bash pilot-study/14b-rllm.sh
bash pilot-study/32b-rllm.sh

Environment behavior for pilot scripts

• Default mode uses conda run -n <your_env> ... inside scripts.
• You do not need to manually run conda activate in default mode.
• Set environment variable RLLM_CONDA_ENV or pass --conda_env <your_env>.
• Set environment variable RLLM_API_KEY or pass --api_key <your_key>.
• You can disable conda wrapping with --use_conda_run false and then run in your currently active environment.

Outputs

• Server logs: pilot-study/logs/server/
• Benchmark raw outputs: pilot-study/outputs/
• Benchmark CSV logs: pilot-study/logs/

Citation (BibTeX)

  @inproceedings{
              rllm-serving,
              title={Reasoning Language Model Inference Serving Unveiled: An Empirical Study},
              author={Li, Qi and Wu, Junpan and Liu, Xiang and Wang, Yuxin and Li, Zeyu and Tang, Zhenheng and Chen, Yuhan and Shi, Shaohuai and Chu, Xiaowen},
              booktitle={The Fourteenth International Conference on Learning Representations},
              year={2026},
              url={https://openreview.net/forum?id=6CGjZYp6ft}
}