Felipe97/llama-cpp-compiled

1from __future__ import annotations2 3import json4import math5import re6 7from typing import Callable, Iterable, TYPE_CHECKING8 9import torch10 11if TYPE_CHECKING:12    from torch import Tensor13 14from .base import MmprojModel, ModelBase, TextModel, gguf15 16 17@ModelBase.register("Ernie4_5_ForCausalLM", "Ernie4_5ForCausalLM")18@ModelBase.example("baidu/ERNIE-4.5-0.3B-PT")19class Ernie4_5Model(TextModel):20    model_arch = gguf.MODEL_ARCH.ERNIE4_521 22    def set_vocab(self):23        self._set_vocab_sentencepiece()24 25        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'26        if tokenizer_config_file.is_file():27            with open(tokenizer_config_file, "r", encoding="utf-8") as f:28                tokenizer_config_json = json.load(f)29                if "add_prefix_space" in tokenizer_config_json:30                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])31 32    def set_gguf_parameters(self):33        super().set_gguf_parameters()34 35    @classmethod36    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:37        name, gen = item38 39        if "ernie." in name:40            name = name.replace("ernie.", "model.")41 42        return super().filter_tensors((name, gen))43 44    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:45        num_heads = self.hparams["num_attention_heads"]46        num_kv_heads = self.hparams["num_key_value_heads"]47        if (head_dim := self.hparams.get("head_dim")) is None:48            head_dim = self.hparams["hidden_size"] // num_heads49 50        # split the qkv weights51        # qkv_proj shape: [(num_heads + 2 * num_kv_heads) * head_dim, hidden_size]52        if "qkv_proj" in name:53            name_q = name.replace("qkv_proj.weight", "q_proj.weight")54            name_k = name.replace("qkv_proj.weight", "k_proj.weight")55            name_v = name.replace("qkv_proj.weight", "v_proj.weight")56            total_q_dim = num_heads * head_dim57            total_k_dim = num_kv_heads * head_dim58            total_v_dim = num_kv_heads * head_dim59            q_proj_weight, k_proj_weight, v_proj_weight = data_torch.split([total_q_dim, total_k_dim, total_v_dim], dim=0)60            yield from super().modify_tensors(q_proj_weight, name_q, bid)61            yield from super().modify_tensors(k_proj_weight, name_k, bid)62            yield from super().modify_tensors(v_proj_weight, name_v, bid)63        # split the up_gate_proj into gate and up64        # up_gate_proj shape: [2 * intermediate_size, hidden_size]65        elif "up_gate_proj" in name:66            name_up = name.replace("up_gate_proj.weight", "up_proj.weight")67            name_gate = name.replace("up_gate_proj.weight", "gate_proj.weight")68            dim_half = data_torch.shape[0] // 269            gate_proj_weight, up_proj_weight = data_torch.split(dim_half, dim=0)70            yield from super().modify_tensors(gate_proj_weight, name_gate, bid)71            yield from super().modify_tensors(up_proj_weight, name_up, bid)72        else:73            yield from super().modify_tensors(data_torch, name, bid)74 75 76@ModelBase.register("Ernie4_5_MoeForCausalLM")77@ModelBase.example("baidu/ERNIE-4.5-21B-A3B-PT")78class Ernie4_5MoeModel(Ernie4_5Model):79    model_arch = gguf.MODEL_ARCH.ERNIE4_5_MOE80    _experts: list[dict[str, Tensor]] | None = None81 82    def __init__(self, *args, **kwargs):83        super().__init__(*args, **kwargs)84        self._experts = [{} for _ in range(self.block_count)]85 86    def set_gguf_parameters(self):87        super().set_gguf_parameters()88        self.gguf_writer.add_expert_count(self.hparams["moe_num_experts"])89        self.gguf_writer.add_expert_used_count(self.hparams["moe_k"])90        self.gguf_writer.add_interleave_moe_layer_step(self.hparams["moe_layer_interval"])91        self.gguf_writer.add_leading_dense_block_count(self.hparams["moe_layer_start_index"])92        if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:93            self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)94        if (shared_expert_count := self.hparams.get('moe_num_shared_experts')) is not None:95            self.gguf_writer.add_expert_shared_count(shared_expert_count)96            if shared_expert_count > 0 and (shared_expert_intermediate_size := self.hparams.get('intermediate_size')) is not None and (num_key_value_heads := self.hparams.get('num_key_value_heads')) is not None:97                self.gguf_writer.add_expert_shared_feed_forward_length(shared_expert_intermediate_size // num_key_value_heads)98 99    @classmethod100    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:101        name, gen = item102 103        # skip Multi-Token Prediction (MTP) layers (again, same as DeepseekV2)104        match = re.match(r"model.mtp_block.(\d+)", name)105        if match:106            return None107 108        # skip all other MTP tensors for now109        match = re.match(r"model.mtp_emb_norm.(\d+)", name)110        if match:111            return None112 113        match = re.match(r"model.mtp_hidden_norm.(\d+)", name)114        if match:115            return None116 117        match = re.match(r"model.mtp_linear_proj.(\d+)", name)118        if match:119            return None120 121        return super().filter_tensors(item)122 123    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:124        # process the experts separately125        if name.find("mlp.experts") != -1:126            n_experts = self.hparams["moe_num_experts"]127            assert bid is not None128 129            if self._experts is None:130                self._experts = [{} for _ in range(self.block_count)]131 132            self._experts[bid][name] = data_torch133 134            if len(self._experts[bid]) >= n_experts * 3:135                # merge the experts into a single 3d tensor136                for w_name in ["gate_proj", "up_proj", "down_proj"]:137                    datas: list[Tensor] = []138 139                    for xid in range(n_experts):140                        ename_to_retrieve = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"141                        datas.append(self._experts[bid][ename_to_retrieve])142                        del self._experts[bid][ename_to_retrieve]143 144                    data_torch = torch.stack(datas, dim=0)145                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"146                    yield from super().modify_tensors(data_torch, merged_name, bid)147        else:148            yield from ModelBase.modify_tensors(self, data_torch, name, bid)149 150    def prepare_tensors(self):151        super().prepare_tensors()152 153        if self._experts is not None:154            # flatten `list[dict[str, Tensor]]` into `list[str]`155            experts = [k for d in self._experts for k in d.keys()]156            if len(experts) > 0:157                raise ValueError(f"Unprocessed experts: {experts}")158 159 160@ModelBase.register("PaddleOCRVLForConditionalGeneration")161@ModelBase.example("PaddlePaddle/PaddleOCR-VL")162class PaddleOCRModel(Ernie4_5Model):163    model_arch = gguf.MODEL_ARCH.PADDLEOCR164 165 166@ModelBase.register("PaddleOCRVisionModel")167@ModelBase.example("PaddlePaddle/PaddleOCR-VL")168class PaddleOCRVisionModel(MmprojModel):169    # PaddleOCR-VL uses a modified version of Siglip170    min_pixels: int = 0171    max_pixels: int = 0172 173    def __init__(self, *args, **kwargs):174        super().__init__(*args, **kwargs)175        assert self.hparams_vision is not None176        self.min_pixels = self.preprocessor_config["min_pixels"]177        self.max_pixels = self.preprocessor_config["max_pixels"]178        self.hparams_vision["image_size"] = int(math.sqrt(self.max_pixels))179 180    def set_gguf_parameters(self):181        super().set_gguf_parameters()182        assert self.hparams_vision is not None183        hparams = self.hparams_vision184        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.PADDLEOCR)185        self.gguf_writer.add_vision_max_pixels(self.max_pixels)186        self.gguf_writer.add_vision_min_pixels(self.min_pixels)187        self.gguf_writer.add_vision_use_gelu(True)188        self.gguf_writer.add_vision_attention_layernorm_eps(hparams.get("rms_norm_eps", 1e-6))189 190    @classmethod191    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:192        name, gen = item193 194        if "vision_model" not in name and "mlp_AR" not in name:195            return None196        name = name.replace("visual.", "model.")197        if "packing_position_embedding" in name:198            # unused199            return None200        if "vision_model.head" in name:201            # we don't yet support image embeddings for this model202            return None203 204        return super().filter_tensors((name, gen))205