#llama-cpp

Two CUDA PRs into llama.cpp, and what they taught me about a 100K-star codebase

A 95-line pooling kernel, a 9-line fix for a silent CPU fallback, and how ggml decides where your ops run