OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Developer Tools & SDKs

Verl Omni

VeRL-Omni is a Python framework built on verl for reinforcement-learning post-training of diffusion and other multimodal generative models.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Verl Omni — OpenViST project profile in AI Developer Tools & SDKs.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use VeRL-Omni when building an RL post-training workflow for a multimodal generative model and the repository’s model and algorithm examples fit the experiment.

Who it is for

  • Researchers developing reinforcement-learning methods for generative models.
  • ML engineering teams planning multimodal model post-training workflows.

Documented uses

  • Train diffusion generators with reinforcement-learning methods.
  • Set up image, video, or video-and-audio generation experiments using model and algorithm combinations listed by the project.
  • Choose among documented rollout and training backend options for a multimodal training pipeline.
Before you choose

Requirements & limits

The public repository identifies an Apache-2.0 license and provides installation and quickstart documentation.

  • Check the model-and-algorithm support table for a match with the intended generator and training method.
  • The project is built on verl and describes vLLM-Omni rollout, selectable VeOmni or FSDP2 backends, and USP, TP, and DP parallelism.
  • The documentation provides installation and quickstart entry points, but the repository does not specify a universal deployment mode.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Install VeRL-Omni on a CUDA 13 NVIDIA environment

Clone VeRL-Omni and install its GPU training and rollout dependencies in a Python virtual environment.

Before you start

  • Python 3.11 or newer
  • CUDA 13.0 or newer and an NVIDIA driver 580 or newer; datacenter GPUs with drivers 535 or newer may use the documented CUDA forward-compatibility fallback
  • The uv package manager

Steps

  1. Clone the repository and change into its directory: the command below and cd verl-omni.

    Documented command
    git clone https://github.com/verl-project/verl-omni.git

  2. Create and activate a Python 3.12 virtual environment: the command below and source .venv/bin/activate.

    Documented command
    uv venv --python 3.12 --seed

  3. Install vLLM: the command below.

    Documented command
    uv pip install vllm==0.30.0 --torch-backend=auto

  4. Install VeRL-Omni's GPU extra: the command below.

    Documented command
    uv pip install -e ".[gpu]" --torch-backend=auto

  5. Run the documented import/version checks for torch, vLLM, vLLM-Omni, verl, and verl_omni to check the installation.

Expected outcome

The repository and GPU dependencies are installed, and the documented checks can confirm that the principal packages import.

Guide limitations

  • This recipe is for the documented NVIDIA CUDA environment, not the separate AMD GPU or Ascend NPU installation paths.
  • Use uv, not pip; the documentation warns that pip may fail to resolve this repository's dependency pins.
  • The checks verify package imports and versions; they do not establish that a model-training run succeeds.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗Guide documentation ↗
Source record details
Source availability
Active
Sources checked
13 October 2026
Original website title
verl-project/verl-omni · Official homepage title · retained in its source language
How we review sourcesSuggest a correction