Developed GPU-accelerated POOL2D support for MobileVLM and CLIP inference in the Vulkan backends of both Mintplex-Labs/whisper.cpp and rmusser01/llama.cpp. This work involved designing and integrating a dedicated POOL2D shader and updating the Vulkan pipeline using C++ and GLSL, with a focus on performance optimization and GPU programming. The implementation reduced inference latency from approximately 2.8 seconds on CPU to 0.7 seconds on GPU, enabling near real-time performance. Additionally, a parameter-ordering fix was introduced to ensure correct operation sequencing, resulting in improved throughput, reduced CPU load, and lower per-inference costs for large-scale deployments.