Developed spatially-sharded decoding for the Wan VAE model in the vllm-project/vllm-omni repository, enabling memory-efficient and parallelizable high-resolution image generation. This feature shards feature maps along both height and width dimensions, allowing distributed processing and improved throughput for large-scale rendering tasks. The implementation leveraged deep learning techniques with PyTorch and incorporated distributed computing and parallel processing strategies to optimize performance. The work was delivered as a dedicated feature commit, with collaborative code review and contributions from multiple team members, and positions the repository for handling more demanding, high-resolution workloads in future development cycles. No bug fixes were recorded.