During February 2025, Ssy10011218 developed an optional ZeRO-3 weight gathering feature for policy models in the huggingface/trl repository, targeting improved sequence generation performance. By leveraging DeepSpeed ZeRO-3 and PyTorch, they enabled faster generation throughput for large policy models by gathering weights during inference, which facilitates more efficient experimentation and deployment. The implementation required careful consideration of VRAM constraints and compatibility, as the feature is not suitable for models exceeding single-GPU memory or for use with vLLM generation. This focused contribution demonstrated depth in model optimization and reinforcement learning, establishing a foundation for future policy-weight management enhancements in the codebase.