The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
PSRL
PSRL is a post-training framework for LLMs that supports both synchronous and asynchronous training paradigms through staleness control, streaming rollout and parameter server architecture.
Quick Start
Installation
Requirements:
- GCC >= 9
- Python 3.9+
- PyTorch 2.7.1
- CUDA 12.8
# Use conda to manage the environment
conda create -n psrl python=3.10
conda activate psrl
# Install all dependencies (including Megatron)
# If you have an existing **editable** vLLM or veRL installation,
# you can pass in VLLM_PATH and VERL_PATH to the script.
USE_MEGATRON=1 bash scripts/install_env.sh
# Install PSRL
pip install -e .
Training
You can find training examples in the examples directory. We provide examples for different training paradigms, including:
- Algorithms: PPO, GRPO
- Training Backends: FSDP, Megatron-LM
- Rollout Modes: Batch, Streaming
- Synchronous/Asynchronous Training: controlled by
stalenessparameter (0 for synchronous, >0 for asynchronous) - Rollout Strategies: Heterogeneous, Homogeneous TP/PP
As an example, you can run the following command to start training:
# For PPO with FSDP backend, batch rollout, homogeneous TP/PP and synchronous training
bash examples/ppo_trainer/fsdp/train.sh
Acknowledgements
PSRL is built upon the foundations of verl, an open-source RLHF framework from ByteDance Seed.
We also appreciate all the pioneering and inspirable projects from the community, including but not limited to vLLM, OpenRLHF, AReaL and NeMo-RL.
- Downloads last month
- 58