FAI-RL: Foundation AI - Reinforcement Learning Library

A production-ready framework for training, inference, evaluation using advanced reinforcement learning techniques. Built for researchers and practitioners who need a flexible, scalable solution for LLM fine-tuning.

Overview

FAI-RL provides a unified, extensible framework for fine-tuning language models with the state-of-the-art algorithms:

• 🎯 Supports Multiple RL Algorithms: DPO, PPO, GRPO, GSPO implementations as well as support for Supervised Fine-Tuning.
• 🚀 Production Ready: Validated on AWS p4d instances with 8x A100 GPUs
• 📦 Simple Configuration: YAML-based configs with CLI override support
• ⚡ Memory Efficient: Full support for LoRA, QLoRA, and DeepSpeed ZeRO-3
• 🔧 Highly Extensible: Custom reward functions, dataset templates, and API integrations

Table of Contents

• Installation
• Authentication & Setup
• Quick Start
  • Training
  • Inference
  • Evaluation
• Supported Methods
• Key Features
• Project Structure
• S3 Checkpoint Upload
• Memory Optimization
• System Requirements
• License

📦 Installation

Install the Package

For Linux/Windows with NVIDIA GPUs (CUDA):

pip install FAI-RL[cuda] --extra-index-url https://download.pytorch.org/whl/cu118

For macOS (Apple Silicon or Intel):

pip install FAI-RL==0.1.14

Clone the Repository for Configuration Recipes

git clone https://github.com/Roblox/FAI-RL.git
cd FAI-RL

Package: https://pypi.org/project/FAI-RL/
Note: The --extra-index-url flag ensures PyTorch is installed with CUDA 11.8 support (Linux/Windows only).

🔑 Authentication & Setup

Before training or using models, you'll need to authenticate with HuggingFace and optionally set up experiment tracking with Weights & Biases.

HuggingFace Authentication

Login to HuggingFace to access models and datasets:

huggingface-cli login

You'll be prompted to enter your HuggingFace access token. You can create a token at https://huggingface.co/settings/tokens.

What this enables:

• Access gated models (if you have permission)

Weights & Biases (Optional)

Login to Weights & Biases for experiment tracking and visualization:

wandb login

You'll be prompted to enter your W&B API key. Get your API key at https://wandb.ai/authorize.

Note: W&B integration is optional. If not logged in, training will proceed without experiment tracking.

🚀 Quick Start

Training

Train a model using any of the supported algorithms (DPO, PPO, GRPO, GSPO, SFT):

# Single GPU training with LoRA
fai-rl-train --recipe recipes/training/sft/llama3_3B_lora.yaml --num-gpus 1

# Multi-GPU training with DeepSpeed
fai-rl-train --recipe recipes/training/dpo/llama3_3B_lora.yaml --num-gpus 8

# Override parameters from CLI
fai-rl-train --recipe recipes/training/sft/llama3_3B_lora.yaml --num-gpus 4 \
  training.learning_rate=5e-5 \
  training.num_train_epochs=3

📖 Complete Training Guide →

Inference

Generate text completions from trained or base models:

# Run inference on a trained model
fai-rl-inference --recipe recipes/inference/llama3_3B.yaml

# Use debug mode for detailed logging
fai-rl-inference --recipe recipes/inference/llama3_3B.yaml --debug

📖 Complete Inference Guide →

Evaluation

Evaluate model performance on academic benchmarks (MMLU, GSM8K):

# Evaluate on MMLU benchmark
fai-rl-eval --recipe recipes/evaluation/mmlu/llama3_3B.yaml --debug

📖 Complete Evaluation Guide →

Supported Algorithms

FAI-RL implements five state-of-the-art reinforcement learning algorithms for language model fine-tuning:

Algorithm	Full Name	Description	Best For
SFT	Supervised Fine-Tuning	Direct supervised learning from labeled examples	Instruction fine-tuning and foundational model fine-tuning
DPO	Direct Preference Optimization	Alignment via preference learning without explicit reward models	Human preference alignment, chat model training
PPO	Proximal Policy Optimization	Policy gradient method with value function and reward model	Complex reward functions, multi-objective optimization
GRPO	Group Relative Policy Optimization	Efficient preference learning with group-based comparison	Reasoning tasks, competitive response generation
GSPO	Group Sequence Policy Optimization	Advanced sequence-level policy optimization	Complex multi-step reasoning, mathematical problem-solving

Training Configurations

All algorithms support three efficiency modes:

Mode	Memory Usage	Training Speed	Best For
Full Fine-tuning	High (baseline)	Fastest	Small models (<3B params), maximum performance
LoRA	Low (~10% of full)	Fast	Most use cases, balanced efficiency
QLoRA	Very Low (~3-4GB for 7B model)	Moderate	Large models on consumer GPUs

Additional features supported across all algorithms:

• ✅ Multi-GPU training with DeepSpeed ZeRO-3
• ✅ Gradient checkpointing for memory efficiency
• ✅ Custom reward functions and dataset templates
• ✅ Weights & Biases integration for experiment tracking
• ✅ Automatic S3 checkpoint upload (supports S3-compatible stores)

Key Features

🎯 Flexible Configuration System

• YAML-based recipes with comprehensive inline documentation for all parameters
• CLI overrides for runtime parameter changes without editing files
• Pre-configured templates for popular models (Llama 3, Qwen 3, etc.)
• Easy experimentation with hyperparameter tuning

🔧 Extensible Architecture

Custom Reward Functions:

• exact_match_reward_func - Accuracy-based rewards for verifiable tasks
• structured_xml_reward_func - Format-based rewards for structured outputs
• Easy to add your custom reward function

Dataset Templates:

• GSM8KTemplate - Math problem formatting with chain-of-thought
• OpenMathInstructTemplate - Mathematical instruction formatting

Pluggable Components:

• Extensible trainer base classes for new algorithms
• HuggingFace Transformers and TRL integration
• Custom dataset processing pipelines

🌐 Multi-Provider API Support

Native support for commercial LLM APIs with automatic provider detection for inference and evaluation:

Supported Providers:

• 🤖 OpenAI (GPT-5, GPT-4.5, GPT-4.1, etc.)
• 🧠 Google (Gemini Pro, Gemini Flash)
• 💬 Anthropic (Claude 4.5 Sonnet, Opus, etc.)
• 🏠 Hosted LLM (self-hosted or custom endpoints)

Configuration Example:

# OpenAI ChatGPT - provider detected from endpoint URL
inference:
  api_endpoint: "https://api.openai.com/v1/chat/completions"
  api_key: "sk-..."
  model: "gpt-4.1"  # Just the model name, no prefix needed!

# Google Gemini - provider detected from endpoint URL
inference:
  api_endpoint: "https://generativelanguage.googleapis.com/v1/models/gemini-pro:generateContent"
  api_key: "AIza..."
  model: "gemini-2.5-pro"

# Anthropic Claude - provider detected from endpoint URL
inference:
  api_endpoint: "https://api.anthropic.com/v1/messages"
  api_key: "sk-ant-..."
  model: "claude-sonnet-4-5-20250929"

# Hosted LLM - any custom or self-hosted model endpoint
inference:
  api_endpoint: "https://your-hosted-endpoint.com/v1/chat"
  api_key: "your-api-key"
  model: "your-model-name"

Customization for Custom APIs:

If your hosted LLM uses a non-OpenAI format, customize utils/hosted_llm_config.py:

• build_hosted_llm_request() - Modify request payload format
• parse_hosted_llm_response() - Customize response parsing
• build_hosted_llm_headers() - Adjust authentication headers

Each function includes detailed examples and inline documentation.

📁 Project Structure

FAI-RL/
├── core/                      # Core framework components
├── trainers/                  # Algorithm implementations
│   ├── rewards/               # Custom reward functions
│   │   ├── accuracy_rewards.py
│   │   └── format_rewards.py
│   └── templates/             # Dataset formatting templates
│       ├── gsm8k_template.py
│       └── openmathinstruct_template.py
├── inference/                 # Inference system
├── evaluations/               # Evaluation system
│   └── eval_datasets/         # Dataset-specific evaluation logic
│       ├── mmlu.py
│       └── gsm8k.py
├── recipes/                   # YAML configuration files
│   ├── training/              # Training recipes (sft/, dpo/, ppo/, grpo/, gspo/)
│   ├── inference/             # Inference recipes
│   └── evaluation/            # Evaluation recipes (mmlu/, gsm8k/)
├── configs/                   # DeepSpeed configurations
│   └── deepspeed/             # ZeRO-3 configs for 1/2/4/8 GPUs
├── utils/                     # Shared utilities
│   ├── s3_utils.py            # S3 checkpoint upload callback
│   └── hosted_llm_config.py   # Custom API endpoint configuration
└── [auto-generated]
    ├── models/                # Trained model checkpoints
    ├── outputs/               # Inference and evaluation results
    └── logs/                  # Training logs

☁️ S3 Checkpoint Upload

FAI-RL can automatically upload checkpoints and the final fine-tuned model to Amazon S3 (or any S3-compatible store such as MinIO). Uploads run in background threads so they never block training.

Prerequisites

Configure AWS credentials using any standard method (environment variables, ~/.aws/credentials, IAM role, etc.):

# Option 1: Environment variables
export AWS_ACCESS_KEY_ID="..."
export AWS_SECRET_ACCESS_KEY="..."
export AWS_DEFAULT_REGION="us-east-1"

# Option 2: AWS CLI
aws configure

Configuration

Add an s3 section to your training recipe YAML:

s3:
  enabled: true                                          # Enable S3 upload
  bucket: "your-s3-bucket"                               # S3 bucket name
  prefix: "your-s3-prefix"                               # Key prefix (folder path inside bucket)
  region: null                                           # AWS region (null = use default)
  endpoint_url: null                                     # Custom S3-compatible endpoint (e.g. MinIO)
  upload_checkpoints: true                               # Upload intermediate checkpoints (at every save_steps)
  upload_final_model: true                               # Upload the final model at end of training
  delete_local_after_upload: false                       # Delete local files after successful upload

Parameter	Type	Default	Description
enabled	bool	false	Master switch for the S3 upload feature
bucket	string	""	Target S3 bucket name (required when enabled)
prefix	string	""	Key prefix under which all uploads are stored
region	string	null	AWS region; falls back to AWS_DEFAULT_REGION or boto3 default
endpoint_url	string	null	Custom endpoint for S3-compatible stores (e.g. http://minio:9000)
upload_checkpoints	bool	true	Upload each intermediate checkpoint saved at save_steps intervals
upload_final_model	bool	true	Upload the final model directory at the end of training
delete_local_after_upload	bool	false	Remove local checkpoint directory after a successful upload

How It Works

1. Intermediate checkpoints -- When the trainer saves a checkpoint (every training.save_steps steps), the S3 callback uploads the entire checkpoint directory to s3://<bucket>/<prefix>/checkpoint-<step>/ in a background thread.
2. Final model -- At the end of training, the output directory is uploaded to s3://<bucket>/<prefix>/final/.
3. Non-blocking -- All uploads happen on daemon threads. Training continues while files are being transferred. At the end of training, the callback waits for any remaining uploads to finish before the process exits.

S3 Upload Structure

Given the example config above, the resulting S3 layout would be:

s3://your-s3-bucket/
└── checkpoints/qwen3-4B-inst-dpo-lora-150k/
    ├── checkpoint-100/
    │   ├── adapter_config.json
    │   ├── adapter_model.safetensors
    │   └── ...
    ├── checkpoint-200/
    │   └── ...
    └── final/
        ├── adapter_config.json
        ├── adapter_model.safetensors
        └── ...

Memory Optimization

FAI-RL provides multiple techniques for efficient training of large models on limited hardware:

Optimization Techniques

Technique	Memory Savings	Speed Impact	Configuration
LoRA	~90% reduction	Minimal	use_lora: true + LoRA params
QLoRA	~95% reduction	Moderate	load_in_4bit: true + LoRA params
8-bit Quantization	~50% reduction	Minimal	load_in_8bit: true
Gradient Checkpointing	~30-50% reduction	20% slower	gradient_checkpointing: true
DeepSpeed ZeRO-3	Distributed across GPUs	Varies	Auto-enabled for multi-GPU

Optimization Strategy

1. Start with QLoRA if GPU memory is limited (<16GB)
2. Use LoRA for balanced efficiency on mid-range GPUs (16-40GB)
3. Full fine-tuning only for small models or high-end GPUs (80GB+)
4. Enable gradient checkpointing if still encountering OOM errors
5. Use DeepSpeed ZeRO-3 for multi-GPU setups to distribute memory load

🧪 System Requirements

Validated on Hardware

This framework has been validated on:

• Instance: AWS EC2 p4d.24xlarge
• GPUs: 8 x NVIDIA A100-SXM4-80GB (80GB VRAM each)
• CPU: 96 vCPUs
• Memory: 1152 GiB
• Storage: 8TB NVMe SSD
• Network: 400 Gbps

📄 License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

For Maintainers

Publishing a New Release

1. Update version in pyproject.toml:

[project]
name = "FAI-RL"
version = "X.Y.Z"  # Increment version

2. Build and publish:

# Install build tools
pip install --upgrade pip build twine

# Clean previous builds
rm -rf dist/ build/ *.egg-info

# Build the package
python -m build

# Upload to PyPI (requires credentials)
python -m twine upload dist/*

# Or upload to test PyPi (requires credentials)
python -m twine upload --repository testpypi dist/*