AI Potluck
Model components / Fine-tuning code

verl

ByteDance Seed / Volcano Engine

verl (Volcano Engine Reinforcement Learning) is a post-training library for large language models, built on the HybridFlow hybrid-controller architecture from ByteDance Seed and the University of Hong Kong. It targets RLHF and reasoning RL on math and code tasks, implements PPO, GRPO, RLOO, GSPO, and DAPO, and runs on FSDP or Megatron-LM with vLLM and SGLang rollout. ByteDance Seed started the project; the verl community maintains it.

The recorded GitHub URL (volcengine/verl) 301-redirects to verl-project/verl following a January 2026 org migration noted in the README; repository content is unchanged. Verified 2026-08-09 via GitHub, the LICENSE body, the README, and PyPI.

Openness

5 high confidence
5.0
license
Apache-2.0(OSI)
source
public
governance
community(ByteDance-Seed origin)
core-gated
ungated

Apache-2.0 as unmodified boilerplate license text; the full training/rollout source ships in the public repo (recipe/ is a public verl-project/verl-recipe submodule); no paid or enterprise tier withholds any method from the published source.

  • https://raw.githubusercontent.com/volcengine/verl/main/LICENSE recorded 2026-08-09

    202-line, 11,358-byte plain-text Apache License, Version 2.0 (January 2004), unmodified through Section 9 and the appendix. The appendix still carries the unfilled boilerplate placeholder "Copyright [yyyy] [name of copyright owner]" followed by "Licensed under the Apache License, Version 2.0 (the "License")" -- no added use restriction or field-of-use clause.

  • https://github.com/volcengine/verl recorded 2026-08-09

    Rendered repo page for the URL on file. The request 301-redirected to https://github.com/verl-project/verl (final_url), and the embedded repo metadata reports "license":{"spdxId":"Apache-2.0","name":"Apache License 2.0"}. The redirect matches the README's January 2026 org migration note ("verl has been migrated to the verl-project"); repository content is otherwise the same.

  • https://raw.githubusercontent.com/volcengine/verl/main/README.md recorded 2026-08-09

    README states verl "is a RL training library initiated by ByteDance Seed team and maintained by the verl community" and calls it "the open-source version of" HybridFlow. Key Features lists FSDP/FSDP2/Megatron-LM for training and vLLM/SGLang/HF Transformers for rollout, full method coverage (PPO, GRPO, GSPO, ReMax, REINFORCE++, RLOO, PRIME, DAPO, DrGRPO, KL_Cov & Clip_Cov, SFT), and states verl "Scales up to 671B models and hundreds of GPUs with expert parallelism." The file names no paid, enterprise or hosted tier and gates no listed feature behind one.

  • https://raw.githubusercontent.com/volcengine/verl/main/pyproject.toml recorded 2026-08-09

    [project] declares name = "verl", license = {text = "Apache-2.0"}, description = "verl: Volcano Engine Reinforcement Learning for LLM", requires-python = ">=3.10", built via setuptools; dependencies/optional-dependencies/authors/urls are marked dynamic (read from files elsewhere in the repo) rather than split into a separate closed distribution.

  • https://pypi.org/project/verl/ recorded 2026-08-09

    Sidebar reads: Author "Bytedance - Seed - MLSys"; License "Apache-2.0"; Requires Python ">=3.10"; Provides Extra: test, prime, geo, gpu, math, vllm, sglang, trl, mcore, trtllm -- no enterprise, commercial or premium extra.

Adoption

2 high confidence
2.0

82,590 PyPI downloads in the trailing 30 days, which puts reach in the 10K-100K band. verl is a reinforcement-learning training framework, so its audience is model-training teams rather than a mass developer base.

Capability

5 high confidence
5.0

Frontier tier within this category: the broadest coverage of modern RL algorithms here -- PPO, GRPO, RLOO, ReMax, REINFORCE++, GSPO, PRIME, DAPO and DrGRPO alongside SFT and DPO -- built on the HybridFlow hybrid controller, with FSDP and Megatron backends and SGLang or vLLM rollout, and demonstrated use training production frontier reasoning models. It sits level with Megatron-LM, the pair that define the top of the scale for this category.

Verified 2026-08-09