Felipe97/llama-cpp-compiled

1from __future__ import annotations2 3import json4 5from typing import Iterable, TYPE_CHECKING6 7if TYPE_CHECKING:8    from torch import Tensor9 10from .base import MmprojModel, ModelBase, gguf, logger11 12from .llama import LlamaModel13 14 15@ModelBase.register(16    "LlavaForConditionalGeneration", # pixtral17    "Mistral3ForConditionalGeneration", # mistral small 3.118)19@ModelBase.example("mistral-community/pixtral-12b", "mistralai/Mistral-Small-3.1-24B-Instruct-2503")20class LlavaVisionModel(MmprojModel):21    img_break_tok_id = -122    use_break_tok = True23 24    def __init__(self, *args, **kwargs):25        super().__init__(*args, **kwargs)26        if self.hparams.get("model_type") == "pixtral":27            # layer_norm_eps is not in config.json, it is hard-coded in modeling_pixtral.py28            self.hparams["layer_norm_eps"] = self.hparams.get("layer_norm_eps", 1e-5)29            if self.use_break_tok:30                self.img_break_tok_id = self.get_token_id("[IMG_BREAK]")31        elif self.is_mistral_format:32            # hparams is already vision config here so norm_eps is only defined in global_config.33            self.hparams["norm_eps"] = self.global_config.get("norm_eps", None)34            assert self.hparams["norm_eps"] is not None, "norm_eps not found in params.json"35            if self.use_break_tok:36                self.img_break_tok_id = self.find_vparam(["image_break_token_id"])37 38                # params.json may ship -1 placeholders (Mistral Medium 3.5)39                # resolve the real id from the bundled tokenizer in that case40                if self.img_break_tok_id < 0:41                    self.img_break_tok_id = self.get_mistral_token_id("[IMG_BREAK]")42        else:43            raise ValueError(f"Unsupported model type: {self.hparams['model_type']}")44        logger.info(f"Image break token id: {self.img_break_tok_id}")45 46    def get_token_id(self, token: str) -> int:47        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'48        with open(tokenizer_config_file, "r", encoding="utf-8") as f:49            added_tokens_decoder = json.load(f).get('added_tokens_decoder') or {}50            for id_, token_data in added_tokens_decoder.items():51                if token_data.get("content") == token:52                    return int(id_)53            # fallthrough to tokenizer.json54        with open(self.dir_model / "tokenizer.json", "r", encoding="utf-8") as f:55            tokenizer_json = json.load(f)56            for token_data in tokenizer_json["added_tokens"]:57                if token_data["content"] == token:58                    return int(token_data["id"])59        raise ValueError(f"Token '{token}' not found in tokenizer config.")60 61    def get_mistral_token_id(self, token: str) -> int:62        # mistral native format ships tekken.json or a versioned spm tokenizer63        tekken_file = self.dir_model / "tekken.json"64        if tekken_file.is_file():65            with open(tekken_file, "r", encoding="utf-8") as f:66                data = json.load(f)67            for entry in data.get("special_tokens", []):68                if entry.get("token_str") == token:69                    return int(entry["rank"])70        tokenizer_json_file = self.dir_model / "tokenizer.json"71        if tokenizer_json_file.is_file():72            with open(tokenizer_json_file, "r", encoding="utf-8") as f:73                data = json.load(f)74            for entry in data.get("added_tokens", []):75                if entry.get("content") == token:76                    return int(entry["id"])77        raise ValueError(f"Token '{token}' not found in mistral tokenizer files.")78 79    def set_gguf_parameters(self):80        super().set_gguf_parameters()81        hparams = self.hparams82        if hparams.get("model_type") == "pixtral":83            self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.PIXTRAL)84            self.gguf_writer.add_vision_attention_layernorm_eps(hparams["layer_norm_eps"])85 86            # hidden_act87            if hparams["hidden_act"] == "silu":88                self.gguf_writer.add_vision_use_silu(True)89            elif hparams["hidden_act"] == "gelu":90                self.gguf_writer.add_vision_use_gelu(True)91            else:92                raise ValueError(f"Unsupported hidden_act: {hparams['hidden_act']}")93 94            # spatial_merge_size95            if "spatial_merge_size" in self.global_config:96                self.gguf_writer.add_vision_spatial_merge_size(self.global_config["spatial_merge_size"])97 98    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:99        n_head = (100            self.hparams["num_attention_heads"] if not self.is_mistral_format else self.find_vparam(["num_attention_heads"])101        )102        n_kv_head = n_head103 104        valid_prefixes = (105            "multi_modal_projector.",106            "vision_tower.",107            "vision_encoder.",108            "vision_language_adapter.",109            "patch_merger.",110            "pre_mm_projector_norm",111        )112 113        if any(name.startswith(prefix) for prefix in valid_prefixes):114            # process vision tensors115            if name.endswith(("q_proj.weight", "q_proj.bias")) and not self.is_mistral_format:116                data_torch = LlamaModel.permute(data_torch, n_head, n_head)117            if name.endswith(("k_proj.weight", "k_proj.bias")) and not self.is_mistral_format:118                data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)119            yield from super().modify_tensors(data_torch, name, bid)120            return121 122        embed_key = "embed_tokens.weight" if not self.is_mistral_format else "tok_embeddings.weight"123        if self.img_break_tok_id > 0 and embed_key in name:124            logger.info(f"Extracting [IMG_BREAK] token embedding from {name}")125            # for pixtral model, we need to extract the [IMG_BREAK] token embedding126            img_break_embd = data_torch[self.img_break_tok_id]127            name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK]128            yield from super().modify_tensors(img_break_embd, name, bid)129 130        return # skip other tensors131