Felipe97/llama-cpp-compiled

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")14@ModelBase.example("FreedomIntelligence/openPangu-Embedded-7B-V1.1")15class PanguEmbeddedModel(TextModel):16    model_arch = gguf.MODEL_ARCH.PANGU_EMBED17 18    def set_vocab(self):19        self._set_vocab_sentencepiece()20 21        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'22        if tokenizer_config_file.is_file():23            with open(tokenizer_config_file, "r", encoding="utf-8") as f:24                tokenizer_config_json = json.load(f)25                if "add_prefix_space" in tokenizer_config_json:26                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])27 28    def set_gguf_parameters(self):29        super().set_gguf_parameters()30        hparams = self.hparams31        self.gguf_writer.add_vocab_size(hparams["vocab_size"])32 33        # PanguEmbedded's hparam loaded from config.json without head_dim34        if (rope_dim := hparams.get("head_dim")) is None:35            rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]36        self.gguf_writer.add_rope_dimension_count(rope_dim)37 38        if hparams.get("head_dim") is None:39            self.gguf_writer.add_key_length(rope_dim)40            self.gguf_writer.add_value_length(rope_dim)41 42    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:43        if name == "lm_head.weight":44            if self.hparams.get("tie_word_embeddings", False):45                logger.info("Skipping tied output layer 'lm_head.weight'")46                return47        yield from super().modify_tensors(data_torch, name, bid)48