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("InternVisionModel")12@ModelBase.example("OpenGVLab/InternVL3-2B", "OpenGVLab/InternVL2_5-1B")13class InternVisionModel(MmprojModel):14 15    min_dynamic_tiles: int = 016    max_dynamic_tiles: int = 017 18    def __init__(self, *args, **kwargs):19        super().__init__(*args, **kwargs)20        assert self.hparams_vision is not None21        self.min_dynamic_tiles = self.global_config.get("min_dynamic_patch", 0)22        self.max_dynamic_tiles = self.global_config.get("max_dynamic_patch", 0)23 24    def set_gguf_parameters(self):25        assert self.hparams_vision is not None26        if isinstance(self.hparams_vision['image_size'], list):27            self.hparams_vision['image_size'] = self.hparams_vision['image_size'][0]28        if isinstance(self.hparams_vision['patch_size'], list):29            self.hparams_vision['patch_size'] = self.hparams_vision['patch_size'][0]30        super().set_gguf_parameters()31 32        hparams = self.hparams33        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.INTERNVL)34        self.gguf_writer.add_vision_attention_layernorm_eps(hparams["layer_norm_eps"])35        # hidden_act36        if hparams["hidden_act"] == "silu":37            self.gguf_writer.add_vision_use_silu(True)38        elif hparams["hidden_act"] == "gelu":39            self.gguf_writer.add_vision_use_gelu(True)40        else:41            raise ValueError(f"Unsupported hidden_act: {hparams['hidden_act']}")42        # downsample_ratio43        downsample_ratio = self.global_config.get("downsample_ratio")44        assert downsample_ratio is not None45        self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / downsample_ratio))46        # older models may not have min/max_dynamic_patch in config47        if self.min_dynamic_tiles > 0:48            self.gguf_writer.add_vision_preproc_min_tiles(self.min_dynamic_tiles)49        if self.max_dynamic_tiles > 0:50            self.gguf_writer.add_vision_preproc_max_tiles(self.max_dynamic_tiles)51 52    def tensor_force_quant(self, name, new_name, bid, n_dims):53        if ".position_embd." in new_name:54            return gguf.GGMLQuantizationType.F3255        return super().tensor_force_quant(name, new_name, bid, n_dims)56 57    @classmethod58    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:59        name, gen = item60 61        vision_prefix = ['vision_model', 'mlp', 'model.vision_tower', 'model.multi_modal_projector']62        if not any([name.startswith(prefix) for prefix in vision_prefix]):63            return None64        # deal with intern-s1 special case65        names_map = {66            "model.multi_modal_projector.layer_norm.bias": "mlp1.0.bias",67            "model.multi_modal_projector.layer_norm.weight": "mlp1.0.weight",68            "model.multi_modal_projector.linear_1.bias": "mlp1.1.bias",69            "model.multi_modal_projector.linear_1.weight": "mlp1.1.weight",70            "model.multi_modal_projector.linear_2.bias": "mlp1.3.bias",71            "model.multi_modal_projector.linear_2.weight": "mlp1.3.weight",72        }73        if name in names_map:74            name = names_map[name]75        # correct name76        if name.startswith("vision_model"):77            name = "vision_tower." + name78        if (".ls" in name or ".lambda_" in name or "position_embedding" in name) and not name.endswith(".weight"):79            name += ".weight"80 81        return super().filter_tensors((name, gen))82 83    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:84        # split QKV tensors if needed85        if ".qkv." in name:86            if data_torch.ndim == 2: # weight87                c3, _ = data_torch.shape88            else: # bias89                c3 = data_torch.shape[0]90            assert c3 % 3 == 091            c = c3 // 392            wq = data_torch[:c]93            wk = data_torch[c: c * 2]94            wv = data_torch[c * 2:]95            yield from super().modify_tensors(wq, name.replace("attn.qkv", "self_attn.q_proj"), bid)96            yield from super().modify_tensors(wk, name.replace("attn.qkv", "self_attn.k_proj"), bid)97            yield from super().modify_tensors(wv, name.replace("attn.qkv", "self_attn.v_proj"), bid)98        else:99            yield from super().modify_tensors(data_torch, name, bid)100