torchforge-rl-training

What problem does it solve?

This Skill simplifies Reinforcement Learning (RL) development and training in PyTorch by abstracting away complex infrastructure concerns, allowing researchers to focus on algorithms.

Core Features & Use Cases

• PyTorch-Native RL: Implement RL algorithms directly in PyTorch without external dependencies like Ray.
• Scalable Training: Supports training from single-GPU setups to large-scale distributed clusters using Monarch and TorchTitan.
• Algorithm Experimentation: Provides clean abstractions for rapid experimentation with RL algorithms like GRPO, DAPO, and SAPO.
• Use Case: Train a custom RL agent for a game or simulation by defining your reward function and using torchforge to handle distributed training, inference, and weight synchronization.

Quick Start

Launch GRPO training for Qwen3.1.7B using the provided configuration file.