Bug-Fix sets an upper VRAM limit for cached ggml_cuda graphs to prevent VRAM memory leaks#21673

kmorennv wants to merge 25 commits into

ggml-org:masterggml-org/llama.cpp:masterfrom

kmorennv:kmoren/Bug-Fix-19639-CUDA-memory-leak-with-MTMD-Gemma3kmorennv/llama.cpp:kmoren/Bug-Fix-19639-CUDA-memory-leak-with-MTMD-Gemma3Copy head branch name to clipboard