configuration_mobilevlm.py

from typing import List, Optional
from transformers import PretrainedConfig

config = {
  "_name_or_path": "mtgv/MobileVLM-1.7B",
  "architectures": [
    "MobileLlamaForCausalLM"
  ],
  "bos_token_id": 1,
  "eos_token_id": 2,
  "freeze_mm_mlp_adapter": False,
  "hidden_act": "silu",
  "hidden_size": 2048,
  "image_aspect_ratio": "pad",
  "image_grid_pinpoints": None,
  "initializer_range": 0.02,
  "intermediate_size": 5632,
  "max_position_embeddings": 2048,
  "max_sequence_length": 2048,
  "mm_hidden_size": 1024,
  "mm_projector_type": "ldpnet",
  "mm_use_im_patch_token": False,
  "mm_use_im_start_end": False,
  "mm_vision_select_feature": "patch",
  "mm_vision_select_layer": -2,
  "mm_vision_tower": "openai/clip-vit-large-patch14-336",
  "model_type": "mobilevlm",
  "num_attention_heads": 16,
  "num_hidden_layers": 24,
  "num_key_value_heads": 16,
  "pad_token_id": 0,
  "pretraining_tp": 1,
  "rms_norm_eps": 1e-06,
  "rope_scaling": None,
  "rope_theta": 10000.0,
  "tie_word_embeddings": False,
  "torch_dtype": "bfloat16",
  "transformers_version": "4.33.1",
  "tune_mm_mlp_adapter": False,
  "use_cache": True,
  "use_mm_proj": True,
  "vision_tower_type": "clip",
  "vocab_size": 32000
}

class MobileVLMConfig(PretrainedConfig):
    model_type = "mobilevlm"

    def __init__(
        self,
        # _name_or_path: str = config["_name_or_path"],
        architectures: Optional[List[str]] = config["architectures"],
        bos_token_id: Optional[int] = config["bos_token_id"],
        eos_token_id: Optional[int] = config["eos_token_id"],
        freeze_mm_mlp_adapter: Optional[bool] = config["freeze_mm_mlp_adapter"],
        hidden_act: Optional[str] = config["hidden_act"],
        hidden_size: Optional[int] = config["hidden_size"],
        image_aspect_ratio: Optional[str] = config["image_aspect_ratio"],
        image_grid_pinpoints: Optional[bool] = config["image_grid_pinpoints"],
        initializer_range: Optional[float] = config["initializer_range"],
        intermediate_size: Optional[int] = config["intermediate_size"],
        max_position_embeddings: Optional[int] = config["max_position_embeddings"],
        max_sequence_length: Optional[int] = config["max_sequence_length"],
        mm_hidden_size: Optional[int] = config["mm_hidden_size"],
        mm_projector_type: Optional[str] = config["mm_projector_type"],
        mm_use_im_patch_token: Optional[bool] = config["mm_use_im_patch_token"],
        mm_use_im_start_end: Optional[bool] = config["mm_use_im_start_end"],
        mm_vision_select_feature: Optional[str] = config["mm_vision_select_feature"],
        mm_vision_select_layer: Optional[int] = config["mm_vision_select_layer"],
        mm_vision_tower: Optional[str] = config["mm_vision_tower"],
        # model_type: Optional[str] = config["model_type"],
        num_attention_heads: Optional[int] = config["num_attention_heads"],
        num_hidden_layers: Optional[int] = config["num_hidden_layers"],
        num_key_value_heads: Optional[int] = config["num_key_value_heads"],
        pad_token_id: Optional[int] = config["pad_token_id"],
        pretraining_tp: Optional[int] = config["pretraining_tp"],
        rms_norm_eps: Optional[float] = config["rms_norm_eps"],
        rope_scaling: Optional[float] = config["rope_scaling"],
        rope_theta: Optional[float] = config["rope_theta"],
        tie_word_embeddings: Optional[bool] = config["tie_word_embeddings"],
        torch_dtype: Optional[str] = config["torch_dtype"],
        transformers_version: Optional[str] = config["transformers_version"],
        tune_mm_mlp_adapter: Optional[bool] = config["tune_mm_mlp_adapter"],
        use_cache: Optional[bool] = config["use_cache"],
        use_mm_proj: Optional[bool] = config["use_mm_proj"],
        vision_tower_type: Optional[str] = config["vision_tower_type"],
        vocab_size: Optional[int] = config["vocab_size"],
        **kwargs,
    ):
        # self._name_or_path = _name_or_path
        self.architectures = architectures
        self.bos_token_id = bos_token_id
        self.eos_token_id = eos_token_id
        self.freeze_mm_mlp_adapter = freeze_mm_mlp_adapter
        self.hidden_act = hidden_act
        self.hidden_size = hidden_size
        self.image_aspect_ratio = image_aspect_ratio
        self.image_grid_pinpoints = image_grid_pinpoints
        self.initializer_range = initializer_range
        self.intermediate_size = intermediate_size
        self.max_position_embeddings = max_position_embeddings
        self.max_sequence_length = max_sequence_length
        self.mm_hidden_size = mm_hidden_size
        self.mm_projector_type = mm_projector_type
        self.mm_use_im_patch_token = mm_use_im_patch_token
        self.mm_use_im_start_end = mm_use_im_start_end
        self.mm_vision_select_feature = mm_vision_select_feature
        self.mm_vision_select_layer = mm_vision_select_layer
        self.mm_vision_tower = mm_vision_tower
        # self.model_type = model_type
        self.num_attention_heads = num_attention_heads
        self.num_hidden_layers = num_hidden_layers
        self.num_key_value_heads = num_key_value_heads
        self.pad_token_id = pad_token_id
        self.pretraining_tp = pretraining_tp
        self.rms_norm_eps = rms_norm_eps
        self.rope_scaling = rope_scaling
        self.rope_theta = rope_theta
        self.tie_word_embeddings = tie_word_embeddings
        self.torch_dtype = torch_dtype
        self.transformers_version = transformers_version
        self.tune_mm_mlp_adapter = tune_mm_mlp_adapter
        self.use_cache = use_cache
        self.use_mm_proj = use_mm_proj
        self.vision_tower_type = vision_tower_type
        self.vocab_size = vocab_size

        super().__init__(**kwargs)