Felipe97/llama-cpp-compiled

1from __future__ import annotations2 3from typing import Callable, TYPE_CHECKING4 5if TYPE_CHECKING:6    from torch import Tensor7 8from .base import ModelBase, TextModel, gguf, logger9 10 11@ModelBase.register("WavTokenizerDec")12@ModelBase.example("novateur/WavTokenizer-large-speech-75token")13class WavTokenizerDecModel(TextModel):14    model_arch = gguf.MODEL_ARCH.WAVTOKENIZER_DEC15 16    @classmethod17    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:18        name, gen = item19 20        if \21                name.endswith("codebook.cluster_size") or \22                name.endswith("codebook.embed_avg") or \23                name.endswith("codebook.inited"):24            logger.debug(f"Skipping {name!r}")25            return None26 27        return super().filter_tensors(item)28 29    def set_vocab(self):30        self._set_vocab_none()31 32    def set_gguf_parameters(self):33        super().set_gguf_parameters()34        self.gguf_writer.add_vocab_size         (self.hparams["vocab_size"])35        self.gguf_writer.add_features_length    (self.hparams["n_embd_features"])36        self.gguf_writer.add_feed_forward_length(self.hparams["n_ff"])37        self.gguf_writer.add_group_norm_eps     (self.hparams["group_norm_epsilon"])38        self.gguf_writer.add_group_norm_groups  (self.hparams["group_norm_groups"])39 40        self.gguf_writer.add_posnet_embedding_length(self.hparams["posnet"]["n_embd"])41        self.gguf_writer.add_posnet_block_count     (self.hparams["posnet"]["n_layer"])42 43        self.gguf_writer.add_convnext_embedding_length(self.hparams["convnext"]["n_embd"])44        self.gguf_writer.add_convnext_block_count     (self.hparams["convnext"]["n_layer"])45 46        self.gguf_writer.add_causal_attention(False)47