Brunobkr/llama.cpp_AlgMor24_github

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")12class DotsOCRVisionModel(MmprojModel):13    def __init__(self, *args, **kwargs):14        super().__init__(*args, **kwargs)15        assert self.hparams_vision is not None16        self.hparams_vision["image_size"] = 0 # dynamic resolution17 18    def set_gguf_parameters(self):19        super().set_gguf_parameters()20        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.DOTSOCR)21        self.gguf_writer.add_vision_min_pixels(self.preprocessor_config["min_pixels"])22        self.gguf_writer.add_vision_max_pixels(self.preprocessor_config["max_pixels"])23        self.gguf_writer.add_vision_attention_layernorm_eps(self.find_vparam(["rms_norm_eps"]))24        self.gguf_writer.add_vision_projector_scale_factor(self.find_vparam(["spatial_merge_size"]))25        self.gguf_writer.add_vision_use_silu(True)26 27    @classmethod28    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:29        name, gen = item30 31        if not name.startswith("vision_tower."):32            return None33 34        if "vision_tower.blocks." in name and ".mlp." in name:35            # note: to avoid naming conflicts in tensor_mapping.py, we need to handle FFN renaming here36            # x = F.silu(self.fc1(x)) * self.fc3(x)37            # x = self.fc2(x)38            # fc1 -> gate, fc2 -> down, fc3 -> up39            # mapping original names to Qwen2.5 naming scheme40            name = name.replace("vision_tower.blocks.", "visual.blocks.")41            name = name.replace(".fc1", ".gate_proj")42            name = name.replace(".fc2", ".down_proj")43            name = name.replace(".fc3", ".up_proj")44 45        return super().filter_tensors((name, gen))46 47    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:48        yield from super().modify_tensors(data_torch, name, bid)49