Developed an optional ZeRO-3 weight gathering feature for policy models in the huggingface/trl repository, targeting improved sequence generation performance under DeepSpeed ZeRO-3. This feature enables faster generation by gathering policy model weights during inference, which is particularly beneficial for large-scale reinforcement learning models. The implementation leveraged Python and PyTorch, focusing on model optimization and sequence generation workflows. While the feature enhances throughput, it introduces constraints for models exceeding single-GPU VRAM and is not compatible with vLLM generation. The work established a foundation for further policy-weight optimizations, contributing a focused, scalable solution to the repository’s model management capabilities.