llama.cpp

Allocate all temporary ggml_tensor_extra_gpu from a fixed-size buffer

#2220

Merged