llama.cpp

ggml: use dynamic allocation for split graph inputs

#22789

Merged