Class LoraModel

subprojects/llama.cpp/convert_lora_to_gguf.py:359–473

            sys.exit(1)

        class LoraModel(model_class):
            model_arch = model_class.model_arch

            lora_alpha: float

            def __init__(self, *args, dir_lora_model: Path, lora_alpha: float, **kwargs):

                super().__init__(*args, **kwargs)

                self.dir_model_card = dir_lora_model
                self.lora_alpha = float(lora_alpha)

            def set_vocab(self):
                pass

            def set_type(self):
                self.gguf_writer.add_type(gguf.GGUFType.ADAPTER)
                self.gguf_writer.add_string(gguf.Keys.Adapter.TYPE, "lora")

            def set_gguf_parameters(self):
                logger.debug("GGUF KV: %s = %d", gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
                self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
                alora_invocation_tokens = lparams.get("alora_invocation_tokens")
                invocation_string = lparams.get("invocation_string")
                if invocation_string and not alora_invocation_tokens:
                    logger.debug("Tokenizing invocation_string -> alora_invocation_tokens")
                    base_model_path_or_id = hparams.get("_name_or_path")
                    try:
                        tokenizer = AutoTokenizer.from_pretrained(base_model_path_or_id)
                    except ValueError:
                        logger.error("Unable to load tokenizer from %s", base_model_path_or_id)
                        raise
                    # NOTE: There's an off-by-one with the older aLoRAs where
                    # the invocation string includes the "<|start_of_turn|>"
                    # token, but the adapters themselves were trained to
                    # activate _after_ that first token, so we drop it here.
                    alora_invocation_tokens = tokenizer(invocation_string)["input_ids"][1:]
                if alora_invocation_tokens:
                    logger.debug("GGUF KV: %s = %s", gguf.Keys.Adapter.ALORA_INVOCATION_TOKENS, alora_invocation_tokens)
                    self.gguf_writer.add_key_value(
                        gguf.Keys.Adapter.ALORA_INVOCATION_TOKENS,
                        alora_invocation_tokens,
                        GGUFValueType.ARRAY,
                        GGUFValueType.UINT32,
                    )

            def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
                # Never add extra tensors (e.g. rope_freqs) for LoRA adapters
                return ()

            def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
                tensor_map: dict[str, PartialLoraTensor] = {}

                for name, tensor in lora_model.items():
                    if self.lazy:
                        tensor = LazyTorchTensor.from_eager(tensor)
                    base_name = get_base_tensor_name(name)
                    # note: mergekit-extract-lora also adds token embeddings to the adapter