flash-llm-rl 0.5.0

flash llm rl

⚡ FlashRL ⚡

Fast RL training with Quantized Rollouts (Blog)

What is FlashRL? • Quick Start • Experiments • Citation

What is FlashRL?

FlashRL patches the inference package (vLLM) to enable: 1) accurate rollout logprob computation for RL training; and 2) online quantization to generate rollouts in INT8 & FP8.

⚡ Quick Start

1. Installation

pip install flashrl # need to be installed in all nodes in multi-node training

2. RL Logprob Patch Only

flashrl setup --fn bf16 -o $PATH_TO_PROFILE_PT_OUTPUT

export FLASHRL_CONFIG=$PATH_TO_PROFILE_PT_OUTPUT 
# alternatively, for submitting multi-node jobs via `ray submit`
# add `FLASHRL_CONFIG: $PATH_TO_CONFIG_YAML_OUTPUT` to runtime env
# as in TODO:PUT_AN_EXAMPLE
bash ... 

3. RL Rollout Quantization -- Simple Setup

Use our pre-set quantization profiles for simple setup.

# for Qwen2.5-0.5B-instruct
export FLASHRL_CONFIG=LiyuanLucasLiu/Qwen2-0.5B-Instruct-quantized.w8a8-RedHatAI/flashrl_config.yaml

# run Qwen2.5-0.5B experiments 
cd verl & bash TODO:UPLOAD_SCRIPT 

# for Qwen2.5-32B-instruct
export FLASHRL_CONFIG=LiyuanLucasLiu/Qwen2.5-32B-quantized.w8a8/flashrl_config.yaml

# run Qwen2.5-32B experiments 
cd verl & bash TODO:UPLOAD_SCRIPT

3. More Advanced

3.1 Profiling

flashrl profile -m $PATH_TO_MODEL -qm $PATH_TO_QUANTIZED_MODEL -o $PATH_TO_PROFILE_PT_OUTPUT --fn int8/fp8

3.2 Setup

flashrl setup --fn int8/fp8/bf16 -m $PATH_TO_MODEL -p $PATH_TO_PROFILE_PT_OUTPUT -o $PATH_TO_CONFIG_YAML_OUTPUT

3.3 RL Training

# for Qwen2.5-0.5B-instruct
export FLASHRL_CONFIG=$PATH_TO_CONFIG_YAML_OUTPUT

# run Qwen2.5-0.5B experiments 
cd verl & bash ... 

# for Qwen2.5-32B
export FLASHRL_CONFIG=$PATH_TO_CONFIG_YAML_OUTPUT
# or, alternatively, for submitting multi-node jobs via `ray submit`
# add `FLASHRL_CONFIG: $PATH_TO_CONFIG_YAML_OUTPUT` to runtime env
# as in TODO:PUT_AN_EXAMPLE

# run Qwen2.5-32B experiments 
cd verl & bash ... 

Example: Accelerating DAPO-Qwen2.5-32B with INT8

🚧 Roadmap & Future Improvements

We're working on several improvements to Flash-RL:

• Support of Other RL Toolkits: Currently Flash-RL only supports VeRL, we are working on rolloing out support for other packages like OpenRLHF
• Support of Other LLM Inference Toolkits: Currently Flash-RL only supports vLLM, we are working on rolloing out support for other tollkits like SgLang
• Further Throughput Optimization: We are working on implementing efficient GPU kernels to accelerate online quantization

📚 Citation

If you find our work useful, please cite us:

@misc{yao2025offpolicy,
  title = {Your Efficient RL Framework Secretly Brings You Off-Policy RL Training},
  url = {https://fengyao.notion.site/off-policy-rl},
  author = {Yao, Feng and Liu, Liyuan and Zhang, Dinghuai and Dong, Chengyu and Gao, Jianfeng},
  journal = {Feng Yao's Notion},
  year = {2025},
  month = aug,
}
@misc{yao2025flashrl,
  title = {Flash-RL: Fast RL training with Quantized Rollouts},
  url = {https://fengyao.notion.site/flash-rl,
  author = {Yao, Feng and Liu, Liyuan and Zhang, Dinghuai and Dong, Chengyu and Gao, Jianfeng},
  journal = {Feng Yao's Notion},
  year = {2025},
  month = aug,
}

Questions?

If you have any questions related to the code or the blog, feel free to reach out to us at Liyuan Liu