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("Llama4ForConditionalGeneration")12# [TAG_HF_EXAMPLE_GATED] meta-llama/Llama-4-Scout-17B-16E-Instruct is gated13@ModelBase.example("unsloth/Llama-4-Scout-17B-16E-Instruct")14class Llama4VisionModel(MmprojModel):15    def set_gguf_parameters(self):16        super().set_gguf_parameters()17        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.LLAMA4)18        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams["norm_eps"])19        self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / self.hparams["pixel_shuffle_ratio"]))20        assert self.hparams["hidden_act"] == "gelu"21        self.gguf_writer.add_vision_use_gelu(True)22 23    @classmethod24    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:25        name, gen = item26 27        if "multi_modal_projector" not in name and "vision_model" not in name:28            return None29 30        if "positional_embedding_vlm" in name and ".weight" not in name:31            name += ".weight"32 33        return super().filter_tensors((name, gen))34 35    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:36        if "multi_modal_projector.linear_1" in name:37            # despite the name with number postfix, this is a single fully connected layer38            yield (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_MMPROJ_FC] + '.weight', data_torch)39        else:40            yield from super().modify_tensors(data_torch, name, bid)41