ByteDance/Sa2VA-InternVL3-8B

1# Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.2#3# Licensed under the Apache License, Version 2.0 (the "License");4# you may not use this file except in compliance with the License.5# You may obtain a copy of the License atd6#7#     http://www.apache.org/licenses/LICENSE-2.08#9# Unless required by applicable law or agreed to in writing, software10# distributed under the License is distributed on an "AS IS" BASIS,11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.12# See the License for the specific language governing permissions and13# limitations under the License.14 15""" Phi-3 model configuration"""16 17 18from transformers.configuration_utils import PretrainedConfig19from transformers.utils import logging20 21logger = logging.get_logger(__name__)22 23PHI3_PRETRAINED_CONFIG_ARCHIVE_MAP = {24    'microsoft/Phi-3-mini-4k-instruct': 'https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/resolve/main/config.json',25    'microsoft/Phi-3-mini-128k-instruct': 'https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/resolve/main/config.json',26}27 28 29class Phi3Config(PretrainedConfig):30    r"""31    This is the configuration class to store the configuration of a [`Phi3Model`]. It is used to instantiate a Phi-332    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the33    defaults will yield a similar configuration to that of the34    [microsoft/Phi-3-mini-4k-instruct](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct).35 36    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the37    documentation from [`PretrainedConfig`] for more information.38 39    Args:40        vocab_size (`int`, *optional*, defaults to 32064):41            Vocabulary size of the Phi-3 model. Defines the number of different tokens that can be represented by the42            `inputs_ids` passed when calling [`Phi3Model`].43        hidden_size (`int`, *optional*, defaults to 3072):44            Dimension of the hidden representations.45        intermediate_size (`int`, *optional*, defaults to 8192):46            Dimension of the MLP representations.47        num_hidden_layers (`int`, *optional*, defaults to 32):48            Number of hidden layers in the Transformer decoder.49        num_attention_heads (`int`, *optional*, defaults to 32):50            Number of attention heads for each attention layer in the Transformer decoder.51        num_key_value_heads (`int`, *optional*):52            This is the number of key_value heads that should be used to implement Grouped Query Attention. If53            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if54            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When55            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed56            by meanpooling all the original heads within that group. For more details checkout [this57            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to58            `num_attention_heads`.59        resid_pdrop (`float`, *optional*, defaults to 0.0):60            Dropout probability for mlp outputs.61        embd_pdrop (`int`, *optional*, defaults to 0.0):62            The dropout ratio for the embeddings.63        attention_dropout (`float`, *optional*, defaults to 0.0):64            The dropout ratio after computing the attention scores.65        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):66            The non-linear activation function (function or string) in the decoder.67        max_position_embeddings (`int`, *optional*, defaults to 4096):68            The maximum sequence length that this model might ever be used with.69        original_max_position_embeddings (`int`, *optional*, defaults to 4096):70            The maximum sequence length that this model was trained with. This is used to determine the size of the71            original RoPE embeddings when using long scaling.72        initializer_range (`float`, *optional*, defaults to 0.02):73            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.74        rms_norm_eps (`float`, *optional*, defaults to 1e-05):75            The epsilon value used for the RMSNorm.76        use_cache (`bool`, *optional*, defaults to `True`):77            Whether or not the model should return the last key/values attentions (not used by all models). Only78            relevant if `config.is_decoder=True`. Whether to tie weight embeddings or not.79        tie_word_embeddings (`bool`, *optional*, defaults to `False`):80            Whether to tie weight embeddings81        rope_theta (`float`, *optional*, defaults to 10000.0):82            The base period of the RoPE embeddings.83        rope_scaling (`dict`, *optional*):84            The scaling strategy for the RoPE embeddings. If `None`, no scaling is applied. If a dictionary, it must85            contain the following keys: `type`, `short_factor` and `long_factor`. The `type` must be either `su` or `yarn` and86            the `short_factor` and `long_factor` must be lists of numbers with the same length as the hidden size87            divided by the number of attention heads divided by 2.88        bos_token_id (`int`, *optional*, defaults to 1):89            The id of the "beginning-of-sequence" token.90        eos_token_id (`int`, *optional*, defaults to 32000):91            The id of the "end-of-sequence" token.92        pad_token_id (`int`, *optional*, defaults to 32000):93            The id of the padding token.94        sliding_window (`int`, *optional*):95            Sliding window attention window size. If `None`, no sliding window is applied.96 97    Example:98 99    ```python100    >>> from transformers import Phi3Model, Phi3Config101 102    >>> # Initializing a Phi-3 style configuration103    >>> configuration = Phi3Config.from_pretrained("microsoft/Phi-3-mini-4k-instruct")104 105    >>> # Initializing a model from the configuration106    >>> model = Phi3Model(configuration)107 108    >>> # Accessing the model configuration109    >>> configuration = model.config110    ```"""111 112    model_type = 'phi3'113    keys_to_ignore_at_inference = ['past_key_values']114 115    def __init__(116        self,117        vocab_size=32064,118        hidden_size=3072,119        intermediate_size=8192,120        num_hidden_layers=32,121        num_attention_heads=32,122        num_key_value_heads=None,123        resid_pdrop=0.0,124        embd_pdrop=0.0,125        attention_dropout=0.0,126        hidden_act='silu',127        max_position_embeddings=4096,128        original_max_position_embeddings=4096,129        initializer_range=0.02,130        rms_norm_eps=1e-5,131        use_cache=True,132        tie_word_embeddings=False,133        rope_theta=10000.0,134        rope_scaling=None,135        bos_token_id=1,136        eos_token_id=32000,137        pad_token_id=32000,138        sliding_window=None,139        **kwargs,140    ):141        self.vocab_size = vocab_size142        self.hidden_size = hidden_size143        self.intermediate_size = intermediate_size144        self.num_hidden_layers = num_hidden_layers145        self.num_attention_heads = num_attention_heads146 147        if num_key_value_heads is None:148            num_key_value_heads = num_attention_heads149 150        self.num_key_value_heads = num_key_value_heads151        self.resid_pdrop = resid_pdrop152        self.embd_pdrop = embd_pdrop153        self.attention_dropout = attention_dropout154        self.hidden_act = hidden_act155        self.max_position_embeddings = max_position_embeddings156        self.original_max_position_embeddings = original_max_position_embeddings157        self.initializer_range = initializer_range158        self.rms_norm_eps = rms_norm_eps159        self.use_cache = use_cache160        self.rope_theta = rope_theta161        self.rope_scaling = rope_scaling162        self._rope_scaling_validation()163        self.sliding_window = sliding_window164 165        super().__init__(166            bos_token_id=bos_token_id,167            eos_token_id=eos_token_id,168            pad_token_id=pad_token_id,169            tie_word_embeddings=tie_word_embeddings,170            **kwargs,171        )172 173    def _rope_scaling_validation(self):174        """175        Validate the `rope_scaling` configuration.176        """177        if self.rope_scaling is None:178            return179 180        if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 3:181            raise ValueError(182                '`rope_scaling` must be a dictionary with three fields, `type`, `short_factor` and `long_factor`, '183                f'got {self.rope_scaling}'184            )185        rope_scaling_type = self.rope_scaling.get('type', None)186        rope_scaling_short_factor = self.rope_scaling.get('short_factor', None)187        rope_scaling_long_factor = self.rope_scaling.get('long_factor', None)188        if rope_scaling_type is None or rope_scaling_type not in ['su', 'yarn']:189            raise ValueError(f"`rope_scaling`'s type field must be one of ['su', 'yarn'], got {rope_scaling_type}")190        if not (191            isinstance(rope_scaling_short_factor, list)192            and all(isinstance(x, (int, float)) for x in rope_scaling_short_factor)193        ):194            raise ValueError(195                f"`rope_scaling`'s short_factor field must be a list of numbers, got {rope_scaling_short_factor}"196            )197        if not len(rope_scaling_short_factor) == self.hidden_size // self.num_attention_heads // 2:198            raise ValueError(199                f"`rope_scaling`'s short_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_short_factor)}"200            )201        if not (202            isinstance(rope_scaling_long_factor, list)203            and all(isinstance(x, (int, float)) for x in rope_scaling_long_factor)204        ):205            raise ValueError(206                f"`rope_scaling`'s long_factor field must be a list of numbers, got {rope_scaling_long_factor}"207            )208        if not len(rope_scaling_long_factor) == self.hidden_size // self.num_attention_heads // 2:209            raise ValueError(210                f"`rope_scaling`'s long_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_long_factor)}"211            )212