Developed end-to-end 3D pooling support for the FlagOpen/FlagGems repository, enabling robust 3D convolutional neural network workloads. Implemented max_pool3d and avg_pool3d operators using Triton kernels in Python, covering both forward and backward paths. Integrated these operators with the existing dispatch system to ensure parity with 2D pooling patterns, and reorganized tests and benchmarks into dedicated files to improve CI reliability and maintainability. Comprehensive accuracy tests and performance benchmarks were added across multiple configurations and data types. This work established a solid foundation for future optimizations and expanded the repository’s capabilities in deep learning and GPU programming.