llama.cpp

vulkan: allow reducing graph submission batches to avoid device timeouts

#24872

Merged