Worked on enhancing memory management for CUDA devices across the ggml-org/ggml and ggml-org/llama.cpp repositories, focusing on Unified Memory Architecture (UMA) support for DGX Spark systems. Leveraged C++ and Linux system programming to integrate system memory information, enabling more efficient memory utilization by reading from /proc/meminfo on UMA platforms. Standardized UMA memory handling across both repositories to ensure consistent behavior and simplify maintenance. Addressed code quality by performing targeted cleanup and refining CUDA memory allocation based on pull request feedback. The work improved resource management and reliability for memory-intensive applications running on advanced Linux and CUDA environments.