Skip to content

Getting Started with Flux

Welcome to Flux! This guide will help you get up and running with the adaptive post-training framework for LLMs.

What is Flux?

Flux is an adaptive post-training framework that combines the best of synchronous and asynchronous RLHF training. Instead of forcing you to choose between stability and efficiency, Flux dynamically adjusts based on your training dynamics.

  • 5 min setup

    Get Flux installed and configured in under 5 minutes

    Installation

  • Quick Start

    Run your first training in 10 lines of code

    Quick Start

  • First Training

    Complete walkthrough of your first training run

    First Training

  • Configuration

    Learn the basics of Flux configuration

    Configuration

Prerequisites

Before you begin, ensure you have:

  • Python 3.10+
  • CUDA 12.0+ with compatible GPU (H100/A100 recommended)
  • 16GB+ GPU memory (for 8B models)
  • SGLang installed for inference

Installation Overview

pip install flux-rlhf

git clone https://github.com/flux-team/flux.git
cd flux
pip install -e ".[dev]"

docker pull fluxrlhf/flux:latest
docker run --gpus all -it fluxrlhf/flux

Full Installation Guide

Quick Example

from flux import FluxConfig, FluxTrainer

# Configure training
config = FluxConfig(
    model_path="Qwen/Qwen3-8B",
    num_steps=1000,
    batch_size=32,
    algorithm="grpo",
)

# Create trainer and run
trainer = FluxTrainer(config)
trainer.fit(prompts="data/prompts.jsonl")

Detailed Quick Start

Learning Path

New to Flux? Follow this recommended learning path:

graph LR
    A[Installation] --> B[Quick Start]
    B --> C[First Training]
    C --> D[Configuration]
    D --> E[Tutorials]
    E --> F[Advanced Topics]
  1. Installation - Set up your environment
  2. Quick Start - Run a minimal example
  3. First Training - Complete training walkthrough
  4. Configuration - Understand configuration options
  5. Tutorials - Deep-dive tutorials
  6. Concepts - Core concepts explained

Need Help?

  • Community

    Ask questions and get help from the community

    Discord

  • Issues

    Report bugs or request features

    GitHub Issues

  • FAQ

    Common questions and answers

    FAQ

FAQ

What GPUs are supported?

Flux supports NVIDIA GPUs with CUDA 12.0+. We recommend:

  • Development: RTX 4090, A100 (40GB)
  • Production: H100 (80GB), A100 (80GB)

Minimum 16GB VRAM for 8B models, 80GB for 70B models.

Can I use Flux with vLLM instead of SGLang?

Currently, Flux is optimized for SGLang integration. vLLM support is planned for future releases.

How does Flux compare to TRL?

TRL is excellent for simple RLHF setups. Flux is designed for:

  • Large-scale distributed training (64+ GPUs)
  • Maximum GPU utilization through adaptive async
  • Native Megatron integration for 3D parallelism

If you're training on 1-8 GPUs, TRL may be simpler. For scale, choose Flux.

Is Flux production-ready?

Flux is under active development. Core features are stable, but we recommend testing thoroughly before production deployment.

Next Steps

Ready to dive in? Start with the Installation Guide.