Brunobkr/llama.cpp_AlgMor24_github

1from __future__ import annotations2 3import json4 5from typing import Iterable, TYPE_CHECKING6 7if TYPE_CHECKING:8    from torch import Tensor9 10from .base import ModelBase, TextModel, gguf, logger11 12 13@ModelBase.register("PanguEmbeddedForCausalLM")14class PanguEmbeddedModel(TextModel):15    model_arch = gguf.MODEL_ARCH.PANGU_EMBED16 17    def set_vocab(self):18        self._set_vocab_sentencepiece()19 20        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'21        if tokenizer_config_file.is_file():22            with open(tokenizer_config_file, "r", encoding="utf-8") as f:23                tokenizer_config_json = json.load(f)24                if "add_prefix_space" in tokenizer_config_json:25                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])26 27    def set_gguf_parameters(self):28        super().set_gguf_parameters()29        hparams = self.hparams30        self.gguf_writer.add_vocab_size(hparams["vocab_size"])31 32        # PanguEmbedded's hparam loaded from config.json without head_dim33        if (rope_dim := hparams.get("head_dim")) is None:34            rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]35        self.gguf_writer.add_rope_dimension_count(rope_dim)36 37        if hparams.get("head_dim") is None:38            self.gguf_writer.add_key_length(rope_dim)39            self.gguf_writer.add_value_length(rope_dim)40 41    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:42        if name == "lm_head.weight":43            if self.hparams.get("tie_word_embeddings", False):44                logger.info("Skipping tied output layer 'lm_head.weight'")45                return46        yield from super().modify_tensors(data_torch, name, bid)47