GPU Goldmines: Tunable WebGPU Kernels for Edge LLM Inference
GPU Goldmines is a framework that allows users to specify, execute, and tune WebGPU kernels for LLM inference by creating a centralized performance database. It makes kernels tunable by parameterizing them, enabling systematic exploration across diverse hardware. This approach derives data-driven policies that accelerate tuning on new devices and predict high-performing configurations for unseen input sizes.