Felipe97/llama-cpp-compiled

1from __future__ import annotations2 3from typing import Callable, Iterable, TYPE_CHECKING4 5if TYPE_CHECKING:6    from torch import Tensor7 8from .base import MmprojModel, ModelBase, gguf9 10 11@ModelBase.register("DotsOCRForCausalLM")12@ModelBase.example("rednote-hilab/dots.ocr")13class DotsOCRVisionModel(MmprojModel):14    def __init__(self, *args, **kwargs):15        super().__init__(*args, **kwargs)16        assert self.hparams_vision is not None17        self.hparams_vision["image_size"] = 0 # dynamic resolution18 19    def set_gguf_parameters(self):20        super().set_gguf_parameters()21        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.DOTSOCR)22        self.gguf_writer.add_vision_min_pixels(self.preprocessor_config["min_pixels"])23        self.gguf_writer.add_vision_max_pixels(self.preprocessor_config["max_pixels"])24        self.gguf_writer.add_vision_attention_layernorm_eps(self.find_vparam(["rms_norm_eps"]))25        self.gguf_writer.add_vision_projector_scale_factor(self.find_vparam(["spatial_merge_size"]))26        self.gguf_writer.add_vision_use_silu(True)27 28    @classmethod29    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:30        name, gen = item31 32        if not name.startswith("vision_tower."):33            return None34 35        if "vision_tower.blocks." in name and ".mlp." in name:36            # note: to avoid naming conflicts in tensor_mapping.py, we need to handle FFN renaming here37            # x = F.silu(self.fc1(x)) * self.fc3(x)38            # x = self.fc2(x)39            # fc1 -> gate, fc2 -> down, fc3 -> up40            # mapping original names to Qwen2.5 naming scheme41            name = name.replace("vision_tower.blocks.", "visual.blocks.")42            name = name.replace(".fc1", ".gate_proj")43            name = name.replace(".fc2", ".down_proj")44            name = name.replace(".fc3", ".up_proj")45 46        return super().filter_tensors((name, gen))47 48    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:49        yield from super().modify_tensors(data_torch, name, bid)50