llama.cpp

7cdd30bf - cuda : allocate all temporary ggml_tensor_extra_gpu from a fixed-size buffer (#2220)