[{"data":1,"prerenderedAt":4},["ShallowReactive",2],{"readme:verl":3},"\u003Cdiv align=\"center\">\n 👋 Hi, everyone!\n    verl is a RL training library initiated by \u003Cb>ByteDance Seed team\u003C\u002Fb> and maintained by the verl community.\n    \u003Cbr \u002F>\n    \u003Cbr \u002F>\n\u003C\u002Fdiv>\u003Cdiv align=\"center\">\u003Cp>\u003Ca href=\"https:\u002F\u002Fdeepwiki.com\u002Fverl-project\u002Fverl\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fdevin.ai\u002Fassets\u002Fdeepwiki-badge.png\" alt=\"Ask DeepWiki.com\" \u002F>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fverl-project\u002Fverl\u002Fstargazers\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fverl-project\u002Fverl\" alt=\"GitHub Repo stars\" \u002F>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Ftwitter.com\u002Fverl_project\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Ftwitter\u002Ffollow\u002Fverl_project\" alt=\"Twitter\" \u002F>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fjoin.slack.com\u002Ft\u002Fverl-project\u002Fshared_invite\u002Fzt-3c6mc2khw-v0lo6NfDPuFP6OnkrZwfqw\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FSlack-verl-blueviolet?logo=slack&\" \u002F>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fpdf\u002F2409.19256\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fstatic\u002Fv1?label=EuroSys&message=Paper&color=red\" \u002F>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fverl.readthedocs.io\u002Fen\u002Flatest\u002F\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdocumentation-blue\" alt=\"Documentation\" \u002F>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fraw.githubusercontent.com\u002Feric-haibin-lin\u002Fverl-community\u002Frefs\u002Fheads\u002Fmain\u002FWeChat.JPG\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F微信-green?logo=wechat&\" \u002F>\u003C\u002Fa>\u003C\u002Fp>\n\u003C\u002Fdiv>\u003Cp>\u003Cimg src=\"https:\u002F\u002Fgithub.com\u002Fuser-attachments\u002Fassets\u002Fc42e675e-497c-4508-8bb9-093ad4d1f216\" alt=\"seed logo\" \u002F>\u003C\u002Fp>\n\u003Ch1>verl: Volcano Engine Reinforcement Learning for LLMs\u003C\u002Fh1>\u003Cp>verl is a flexible, efficient and production-ready RL training library for large language models (LLMs).\u003C\u002Fp>\n\u003Cp>verl is the open-source version of \u003Cstrong>\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2409.19256v2\" rel=\"nofollow ugc noopener\">HybridFlow: A Flexible and Efficient RLHF Framework\u003C\u002Fa>\u003C\u002Fstrong> paper.\u003C\u002Fp>\n\u003Cp>verl is flexible and easy to use with:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cp>\u003Cstrong>Easy extension of diverse RL algorithms\u003C\u002Fstrong>: The hybrid-controller programming model enables flexible representation and efficient execution of complex post-training dataflows. Build RL dataflows such as GRPO, PPO in a few lines of code.\u003C\u002Fp>\n\u003C\u002Fli>\n\u003Cli>\u003Cp>\u003Cstrong>Seamless integration of existing LLM infra with modular APIs\u003C\u002Fstrong>: Decouples computation and data dependencies, enabling seamless integration with existing LLM frameworks, such as FSDP, Megatron-LM, vLLM, SGLang, etc\u003C\u002Fp>\n\u003C\u002Fli>\n\u003Cli>\u003Cp>\u003Cstrong>Flexible device mapping\u003C\u002Fstrong>: Supports various placement of models onto different sets of GPUs for efficient resource utilization and scalability across different cluster sizes.\u003C\u002Fp>\n\u003C\u002Fli>\n\u003Cli>\u003Cp>Ready integration with popular HuggingFace models\u003C\u002Fp>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>verl is fast with:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cp>\u003Cstrong>State-of-the-art throughput\u003C\u002Fstrong>: SOTA LLM training and inference engine integrations and SOTA RL throughput.\u003C\u002Fp>\n\u003C\u002Fli>\n\u003Cli>\u003Cp>\u003Cstrong>Efficient actor model resharding with 3D-HybridEngine\u003C\u002Fstrong>: Eliminates memory redundancy and significantly reduces communication overhead during transitions between training and generation phases.\u003C\u002Fp>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cdiv align=\"center\">\n \u003Cimg src=\"https:\u002F\u002Fgithub.com\u002Fverl-project\u002Fverl-data\u002Fblob\u002Fmain\u002Fimages\u002Fverl-arch.png?raw=true\" width=\"400\" alt=\"verl-arch.png\" \u002F>\n\u003C\u002Fdiv>\u003Cp>\u003C\u002Fp>\u003Ch2>News\u003C\u002Fh2>\n\u003Cul>\n\u003Cli>[2026\u002F06] \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fverl-project\u002Fverl-SpeCo\" rel=\"nofollow ugc noopener\">verl-SpeCo\u003C\u002Fa> is pre-released: a co-training framework for speculative decoding across RL training and inference, keeping draft models aligned during training and reusable for accelerated serving, built on top of verl.\u003C\u002Fli>\n\u003Cli>[2026\u002F05] \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fverl-project\u002Funi-agent\" rel=\"nofollow ugc noopener\">uni-agent\u003C\u002Fa> is released: a unified agent framework to build, run, and train LLM agents at scale, built on top of verl.\u003C\u002Fli>\n\u003Cli>[2026\u002F05] \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fverl-project\u002Fverl-omni\" rel=\"nofollow ugc noopener\">VeRL-Omni\u003C\u002Fa> is pre-released: a unified RL stack for diffusion and omni-modal model post-training built on top of verl. Read the \u003Ca href=\"https:\u002F\u002Fvllm.ai\u002Fblog\u002F2026-05-14-verl-omni\" rel=\"nofollow ugc noopener\">blog post\u003C\u002Fa> for details.\u003C\u002Fli>\n\u003Cli>[2026\u002F05] verl's zero-mismatch HuggingFace rollout \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fverl-project\u002Fvexact\" rel=\"nofollow ugc noopener\">vexact\u003C\u002Fa> is released: with batch-invariant kernels, shared model definition with FSDP, and out-of-box examples compatible with VeOmni.\u003C\u002Fli>\n\u003Cli>[2026\u002F04] verl's Megatron backend LoRA and router replay support is showcased at [PyTorch Conference Europe 2026](\u003Ca href=\"https:\u002F\u002Fpytorchconferenceeu2026.sched.com\u002Fevent\u002F2Juce\u002Foptimizing-reinforcement-learning-at-trillion-parameter-scale-so\" rel=\"nofollow ugc noopener\">https:\u002F\u002Fpytorchconferenceeu2026.sched.com\u002Fevent\u002F2Juce\u002Foptimizing-reinforcement-learning-at-trillion-parameter-scale-so\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n",1784240408324]