Ankit2802/phi3_vision_128k

1# coding=utf-82# Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.3#4# Licensed under the Apache License, Version 2.0 (the "License");5# you may not use this file except in compliance with the License.6# You may obtain a copy of the License at7#8#     http://www.apache.org/licenses/LICENSE-2.09#10# Unless required by applicable law or agreed to in writing, software11# distributed under the License is distributed on an "AS IS" BASIS,12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.13# See the License for the specific language governing permissions and14# limitations under the License.15 16""" Phi-3-V model configuration"""17 18 19from transformers.configuration_utils import PretrainedConfig20from transformers.utils import logging21 22 23logger = logging.get_logger(__name__)24 25PHI3V_PRETRAINED_CONFIG_ARCHIVE_MAP = {26    "microsoft/Phi-3-vision-128k-instruct": "https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/resolve/main/config.json",27}28 29 30class Phi3VConfig(PretrainedConfig):31    r"""32    This is the configuration class to store the configuration of a [`Phi3VModel`]. It is used to instantiate a Phi-333    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the34    defaults will yield a similar configuration to that of the35    [microsoft/Phi-3-vision-128k-instruct](https://huggingface.co/microsoft/Phi-3-vision-128k-instruct).36 37    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the38    documentation from [`PretrainedConfig`] for more information.39 40    Args:41        vocab_size (`int`, *optional*, defaults to 32064):42            Vocabulary size of the Phi-3-V model. Defines the number of different tokens that can be represented by the43            `inputs_ids` passed when calling [`Phi3VModel`].44        hidden_size (`int`, *optional*, defaults to 3072):45            Dimension of the hidden representations.46        intermediate_size (`int`, *optional*, defaults to 8192):47            Dimension of the MLP representations.48        num_hidden_layers (`int`, *optional*, defaults to 32):49            Number of hidden layers in the Transformer decoder.50        num_attention_heads (`int`, *optional*, defaults to 32):51            Number of attention heads for each attention layer in the Transformer decoder.52        num_key_value_heads (`int`, *optional*):53            This is the number of key_value heads that should be used to implement Grouped Query Attention. If54            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if55            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When56            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed57            by meanpooling all the original heads within that group. For more details checkout [this58            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to59            `num_attention_heads`.60        resid_pdrop (`float`, *optional*, defaults to 0.0):61            Dropout probability for mlp outputs.62        embd_pdrop (`int`, *optional*, defaults to 0.0):63            The dropout ratio for the embeddings.64        attention_dropout (`float`, *optional*, defaults to 0.0):65            The dropout ratio after computing the attention scores.66        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):67            The non-linear activation function (function or string) in the decoder.68        max_position_embeddings (`int`, *optional*, defaults to 4096):69            The maximum sequence length that this model might ever be used with.70        original_max_position_embeddings (`int`, *optional*, defaults to 4096):71            The maximum sequence length that this model was trained with. This is used to determine the size of the72            original RoPE embeddings when using long scaling.73        initializer_range (`float`, *optional*, defaults to 0.02):74            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.75        rms_norm_eps (`float`, *optional*, defaults to 1e-05):76            The epsilon value used for the RMSNorm.77        use_cache (`bool`, *optional*, defaults to `True`):78            Whether or not the model should return the last key/values attentions (not used by all models). Only79            relevant if `config.is_decoder=True`. Whether to tie weight embeddings or not.80        tie_word_embeddings (`bool`, *optional*, defaults to `False`):81            Whether to tie weight embeddings82        rope_theta (`float`, *optional*, defaults to 10000.0):83            The base period of the RoPE embeddings.84        rope_scaling (`dict`, *optional*):85            The scaling strategy for the RoPE embeddings. If `None`, no scaling is applied. If a dictionary, it must86            contain the following keys: `type`, `short_factor` and `long_factor`. The `type` must be either `su` or `yarn` and87            the `short_factor` and `long_factor` must be lists of numbers with the same length as the hidden size88            divided by the number of attention heads divided by 2.89        bos_token_id (`int`, *optional*, defaults to 1):90            The id of the "beginning-of-sequence" token.91        eos_token_id (`int`, *optional*, defaults to 32000):92            The id of the "end-of-sequence" token.93        pad_token_id (`int`, *optional*, defaults to 32000):94            The id of the padding token.95        sliding_window (`int`, *optional*):96            Sliding window attention window size. If `None`, no sliding window is applied.97        embd_layer (`str`, *optional*, defaults to `"default"`):98            The embedding layer to use. Can be either `"default"` or `"image"`. "default" uses the standard embedding for text. 99 100    Example:101 102    ```python103    >>> from transformers import Phi3VModel, Phi3VConfig104 105    >>> # Initializing a Phi-3-V style configuration106    >>> configuration = Phi3Config.from_pretrained("microsoft/Phi-3-vision-128k-instruct")107 108    >>> # Initializing a model from the configuration109    >>> model = Phi3VModel(configuration)110 111    >>> # Accessing the model configuration112    >>> configuration = model.config113    ```"""114 115    model_type = "phi3_v"116    keys_to_ignore_at_inference = ["past_key_values"]117 118    def __init__(119        self,120        vocab_size=32064,121        hidden_size=3072,122        intermediate_size=8192,123        num_hidden_layers=32,124        num_attention_heads=32,125        num_key_value_heads=None,126        resid_pdrop=0.0,127        embd_pdrop=0.0,128        attention_dropout=0.0,129        hidden_act="silu",130        max_position_embeddings=4096,131        original_max_position_embeddings=4096,132        initializer_range=0.02,133        rms_norm_eps=1e-5,134        use_cache=True,135        tie_word_embeddings=False,136        rope_theta=10000.0,137        rope_scaling=None,138        bos_token_id=1,139        eos_token_id=32000,140        pad_token_id=32000,141        sliding_window=None,142        embd_layer: str = "default",143        **kwargs,144    ):145        self.vocab_size = vocab_size146        self.hidden_size = hidden_size147        self.intermediate_size = intermediate_size148        self.num_hidden_layers = num_hidden_layers149        self.num_attention_heads = num_attention_heads150 151        if num_key_value_heads is None:152            num_key_value_heads = num_attention_heads153 154        self.num_key_value_heads = num_key_value_heads155        self.resid_pdrop = resid_pdrop156        self.embd_pdrop = embd_pdrop157        self.attention_dropout = attention_dropout158        self.hidden_act = hidden_act159        self.max_position_embeddings = max_position_embeddings160        self.original_max_position_embeddings = original_max_position_embeddings161        self.initializer_range = initializer_range162        self.rms_norm_eps = rms_norm_eps163        self.use_cache = use_cache164        self.rope_theta = rope_theta165        self.rope_scaling = rope_scaling166        self._rope_scaling_validation()167        self.sliding_window = sliding_window168        self.embd_layer = embd_layer169 170 171        super().__init__(172            bos_token_id=bos_token_id,173            eos_token_id=eos_token_id,174            pad_token_id=pad_token_id,175            tie_word_embeddings=tie_word_embeddings,176            **kwargs,177        )178 179    def _rope_scaling_validation(self):180        """181        Validate the `rope_scaling` configuration.182        """183        if self.rope_scaling is None:184            return185 186        if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 3:187            raise ValueError(188                "`rope_scaling` must be a dictionary with three fields, `type`, `short_factor` and `long_factor`, "189                f"got {self.rope_scaling}"190            )191        rope_scaling_type = self.rope_scaling.get("type", None)192        rope_scaling_short_factor = self.rope_scaling.get("short_factor", None)193        rope_scaling_long_factor = self.rope_scaling.get("long_factor", None)194        if rope_scaling_type is None or rope_scaling_type not in ["su", "yarn"]:195            raise ValueError(f"`rope_scaling`'s type field must be one of ['su', 'yarn'], got {rope_scaling_type}")196        if not (197            isinstance(rope_scaling_short_factor, list)198            and all(isinstance(x, (int, float)) for x in rope_scaling_short_factor)199        ):200            raise ValueError(201                f"`rope_scaling`'s short_factor field must be a list of numbers, got {rope_scaling_short_factor}"202            )203        if not len(rope_scaling_short_factor) == self.hidden_size // self.num_attention_heads // 2:204            raise ValueError(205                f"`rope_scaling`'s short_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_short_factor)}"206            )207        if not (208            isinstance(rope_scaling_long_factor, list)209            and all(isinstance(x, (int, float)) for x in rope_scaling_long_factor)210        ):211            raise ValueError(212                f"`rope_scaling`'s long_factor field must be a list of numbers, got {rope_scaling_long_factor}"213            )214        if not len(rope_scaling_long_factor) == self.hidden_size // self.num_attention_heads // 2:215            raise ValueError(216                f"`rope_scaling`'s long_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_long_factor)}"217            )