torchforge-rl-training

What problem does it solve?

This Skill simplifies Reinforcement Learning development by separating RL algorithms from complex distributed training infrastructure, enabling faster experimentation and scalable training.

Core Features & Use Cases

• PyTorch-Native: Avoids Ray dependencies, offering clean abstractions.
• Algorithm Experimentation: Easily implement and test RL algorithms like GRPO, DAPO, SAPO.
• Scalable Training: Leverages Monarch for distributed training across many GPUs.
• Use Case: A researcher wants to quickly test a new variant of GRPO on a large dataset. They can use this Skill to set up the distributed training environment and focus solely on implementing their novel loss function.

Quick Start

Use the torchforge skill to launch GRPO training for math reasoning using the provided configuration file.