RLinf

RLinf is an open-source framework for Reinforcement Learning training, focused on supporting the reinforcement learning training of large language models (LLMs) and vision-language models (VLAs). RLinf provides advanced features such as unified programming interfaces, flexible execution modes, automatic scheduling, and elastic communication, supporting various reinforcement learning algorithms like PPO and GRPO.

RLinf supports real-time experiment tracking, allowing you to stream loss curves, accuracy, GPU utilization, and any custom metrics to SwanLab.

You can use RLinf to quickly conduct reinforcement learning model training while using SwanLab for experiment tracking and visualization.

RLinf Official Documentation: https://rlinf.readthedocs.io/zh-cn/latest/

Quick Integration with SwanLab

RLinf supports enabling SwanLab as a logging backend in the YAML configuration file. Simply add "swanlab" to the runner.logger.logger_backends in the configuration file.

Add the following configuration to your YAML configuration file:

runner:
  task_type: math  # or "embodied", etc.
  logger:
    log_path: ${runner.output_dir}/${runner.experiment_name}
    project_name: rlinf
    experiment_name: ${runner.experiment_name}
    logger_backends: ["swanlab"]  
  experiment_name: grpo-1.5b
  output_dir: ./logs

After running the training, RLinf will create a subdirectory for the enabled SwanLab backend:

logs/grpo-1.5b/
├── checkpoints/
├── converted_ckpts/
├── log/
├── swanlab/            # SwanLab runtime directory

Setting Project Name and Experiment Name

You can set the project name and experiment name by configuring project_name and experiment_name in the YAML configuration file:

runner:
  project_name: rlinf
  experiment_name: grpo-1.5b

Setting Log File Save Location

You can set the log file save location by configuring log_path in the YAML configuration file:

runner:
  log_path: ./logs