Scaling reinforcement learning (RL) for large language models (LLMs) presents significant challenges, particularly when trying to implement it across hundreds of GPUs. The traditional bottlenecks include issues with distributed coordination, stability, and reproducibility. To address these hurdles, Meta's PyTorch team has released torchforge, a PyTorch-native RL library designed to simplify large-scale post-training. They tested it on a 512-GPU cluster, achieving streamlined setups and efficient training processes that were previously unachievable with existing tools. Torchforge enables researchers to focus on RL algorithms instead of infrastructure. It offers a suite of features such as pseudocode-like APIs, flexible synchronicity options, and components like Monarch for coordination and TorchStore for efficient weight synchronization. These elements work together to create a robust RL stack that allows for faster iteration on reward design and policy updates without the overhead of traditional distributed systems. The article also highlights Weaver, a weak verifier system used as a reward function in RL pipelines. Weaver evaluates model outputs and provides scalar rewards based on their correctness probability, which is crucial for training without relying on costly human data. Instead of using a single, strong verifier, Weaver aggregates multiple weak ones to enhance the reliability of feedback on complex reasoning tasks. This approach not only reduces the need for continuous human annotation but also supports high-throughput RL experiments by verifying thousands of generations per second.