How-to Guides¶

Task-oriented guides that help you accomplish specific goals with Flux.

Available Guides¶

Add a Custom Algorithm

Create and register your own RL algorithm

Guide
Implement Custom Rewards

Build reward functions for your specific task

Guide
Scale to Multiple Nodes

Distribute training across multiple machines

Guide
Debug Training Issues

Diagnose and fix common training problems

Guide
Optimize Performance

Get maximum throughput from your hardware

Guide
Monitor with Prometheus

Set up comprehensive training monitoring

Guide
Resume from Checkpoint

Save and restore training state

Guide

Quick Answers¶

How do I change the learning rate during training?¶

# Use a scheduler
from flux import FluxConfig

config = FluxConfig(
    learning_rate=1e-6,
    lr_scheduler="cosine",
    lr_scheduler_config={
        "warmup_steps": 100,
        "min_lr": 1e-7,
    }
)

How do I use multiple reward functions?¶

from flux.rewards import CompositeReward, LengthReward, KeywordReward

reward = CompositeReward([
    (LengthReward(target=200), 0.3),
    (KeywordReward(required=["answer"]), 0.7),
])

How do I save checkpoints more frequently?¶

checkpoint:
  save_steps: 100  # Save every 100 steps
  max_checkpoints: 10
  keep_best: 3

How do I use a different model for inference?¶

model_path: "Qwen/Qwen3-8B"  # Training model

sglang:
  model_path: "Qwen/Qwen3-8B-Instruct"  # Inference model (optional)
  base_url: http://localhost:8000

How do I adjust the async ratio manually?¶

adaptive_async:
  enabled: false  # Disable adaptive control
  fixed_async_ratio: 0.5  # Use fixed ratio

By Category¶

Training¶

Guide	Description
Custom Algorithm	Add your own RL algorithm
Custom Rewards	Build task-specific rewards
Debugging	Fix training issues
Checkpoints	Save and restore training

Scaling¶

Guide	Description
Multi-Node	Distribute across machines
Performance	Optimize throughput

Operations¶

Guide	Description
Monitoring	Set up Prometheus/Grafana

Contributing Guides¶

Have a guide to share? We welcome contributions!

Fork the repository
Add your guide to docs/how-to/
Update this index
Submit a pull request

See Contributing for details.