TorchForge — агентское обучение с подкреплением на PyTorch

★ 7.7 · ai-tooling

torchforge is a Claude Code skill that provides guidance for working with torchforge, Meta's PyTorch-native library for agentic reinforcement learning that separates infrastructure concerns from algorithm concerns. The skill covers implementing RL algorithms — GRPO, DAPO, CISPO, GSPO, and SAPO — in roughly 100 lines of code, scaling training from a single GPU to thousands via the Monarch actor system, and integrating TorchTitan for model parallelism with vLLM handling inference. Required dependencies are torch>=2.9.0, torchtitan>=0.2.0, vllm, and monarch. It targets RL researchers who want clean PyTorch-native abstractions without a Ray dependency and need to rapidly experiment with new loss functions or train reasoning models on benchmarks like GSM8K and MATH.