llama.cpp

convert : use self.block_count everywhere instead of reading hparams

#17359

Merged