llama.cpp

07b0e7a5 - convert : use self.block_count everywhere instead of reading hparams (#17359)