pytorch/Phi-4-mini-instruct-parq-3w-4e-shared

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 model configuration"""17 18from transformers.configuration_utils import PretrainedConfig19from transformers.utils import logging20 21 22logger = logging.get_logger(__name__)23 24 25class Phi3Config(PretrainedConfig):26    r"""27    This is the configuration class to store the configuration of a [`Phi3Model`]. It is used to instantiate a Phi-328    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the29    defaults will yield a similar configuration to that of the30    [microsoft/Phi-3-mini-4k-instruct](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct).31 32    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the33    documentation from [`PretrainedConfig`] for more information.34 35    Args:36        vocab_size (`int`, *optional*, defaults to 32064):37            Vocabulary size of the Phi-3 model. Defines the number of different tokens that can be represented by the38            `inputs_ids` passed when calling [`Phi3Model`].39        hidden_size (`int`, *optional*, defaults to 3072):40            Dimension of the hidden representations.41        intermediate_size (`int`, *optional*, defaults to 8192):42            Dimension of the MLP representations.43        num_hidden_layers (`int`, *optional*, defaults to 32):44            Number of hidden layers in the Transformer decoder.45        num_attention_heads (`int`, *optional*, defaults to 32):46            Number of attention heads for each attention layer in the Transformer decoder.47        num_key_value_heads (`int`, *optional*):48            This is the number of key_value heads that should be used to implement Grouped Query Attention. If49            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if50            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When51            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed52            by meanpooling all the original heads within that group. For more details checkout [this53            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to54            `num_attention_heads`.55        resid_pdrop (`float`, *optional*, defaults to 0.0):56            Dropout probability for mlp outputs.57        embd_pdrop (`int`, *optional*, defaults to 0.0):58            The dropout ratio for the embeddings.59        attention_dropout (`float`, *optional*, defaults to 0.0):60            The dropout ratio after computing the attention scores.61        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):62            The non-linear activation function (function or string) in the decoder.63        max_position_embeddings (`int`, *optional*, defaults to 4096):64            The maximum sequence length that this model might ever be used with.65        original_max_position_embeddings (`int`, *optional*, defaults to 4096):66            The maximum sequence length that this model was trained with. This is used to determine the size of the67            original RoPE embeddings when using long scaling.68        initializer_range (`float`, *optional*, defaults to 0.02):69            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.70        rms_norm_eps (`float`, *optional*, defaults to 1e-05):71            The epsilon value used for the RMSNorm.72        use_cache (`bool`, *optional*, defaults to `True`):73            Whether or not the model should return the last key/values attentions (not used by all models). Only74            relevant if `config.is_decoder=True`. Whether to tie weight embeddings or not.75        tie_word_embeddings (`bool`, *optional*, defaults to `False`):76            Whether to tie weight embeddings77        rope_theta (`float`, *optional*, defaults to 10000.0):78            The base period of the RoPE embeddings.79        rope_scaling (`dict`, *optional*):80            The scaling strategy for the RoPE embeddings. If `None`, no scaling is applied. If a dictionary, it must81            contain the following keys: `type`, `short_factor` and `long_factor`. The `type` must be `longrope` and82            the `short_factor` and `long_factor` must be lists of numbers with the same length as the hidden size83            divided by the number of attention heads divided by 2.84        partial_rotary_factor (`float`, *optional*, defaults to 1.0):85            Percentage of the query and keys which will have rotary embedding. Must be between 0.0 and 1.0.86        bos_token_id (`int`, *optional*, defaults to 1):87            The id of the "beginning-of-sequence" token.88        eos_token_id (`int`, *optional*, defaults to 32000):89            The id of the "end-of-sequence" token.90        pad_token_id (`int`, *optional*, defaults to 32000):91            The id of the padding token.92        sliding_window (`int`, *optional*):93            Sliding window attention window size. If `None`, no sliding window is applied.94 95    Example:96 97    ```python98    >>> from transformers import Phi3Model, Phi3Config99 100    >>> # Initializing a Phi-3 style configuration101    >>> configuration = Phi3Config.from_pretrained("microsoft/Phi-3-mini-4k-instruct")102 103    >>> # Initializing a model from the configuration104    >>> model = Phi3Model(configuration)105 106    >>> # Accessing the model configuration107    >>> configuration = model.config108    ```"""109 110    model_type = "phi3"111    keys_to_ignore_at_inference = ["past_key_values"]112 113    def __init__(114        self,115        vocab_size=32064,116        hidden_size=3072,117        intermediate_size=8192,118        num_hidden_layers=32,119        num_attention_heads=32,120        num_key_value_heads=None,121        resid_pdrop=0.0,122        embd_pdrop=0.0,123        attention_dropout=0.0,124        hidden_act="silu",125        max_position_embeddings=4096,126        original_max_position_embeddings=4096,127        initializer_range=0.02,128        rms_norm_eps=1e-5,129        use_cache=True,130        tie_word_embeddings=False,131        rope_theta=10000.0,132        rope_scaling=None,133        partial_rotary_factor=1.0,134        bos_token_id=1,135        eos_token_id=32000,136        pad_token_id=32000,137        sliding_window=None,138        **kwargs,139    ):140        self.vocab_size = vocab_size141        self.hidden_size = hidden_size142        self.intermediate_size = intermediate_size143        self.num_hidden_layers = num_hidden_layers144        self.num_attention_heads = num_attention_heads145 146        if num_key_value_heads is None:147            num_key_value_heads = num_attention_heads148 149        self.num_key_value_heads = num_key_value_heads150        self.resid_pdrop = resid_pdrop151        self.embd_pdrop = embd_pdrop152        self.attention_dropout = attention_dropout153        self.hidden_act = hidden_act154        self.max_position_embeddings = max_position_embeddings155        self.original_max_position_embeddings = original_max_position_embeddings156        self.initializer_range = initializer_range157        self.rms_norm_eps = rms_norm_eps158        self.use_cache = use_cache159        self.rope_theta = rope_theta160        self.rope_scaling = rope_scaling161        self.partial_rotary_factor = partial_rotary_factor162        self._rope_scaling_adjustment()163        self._rope_scaling_validation()164        self.sliding_window = sliding_window165 166        super().__init__(167            bos_token_id=bos_token_id,168            eos_token_id=eos_token_id,169            pad_token_id=pad_token_id,170            tie_word_embeddings=tie_word_embeddings,171            **kwargs,172        )173 174    def _rope_scaling_adjustment(self):175        """176        Adjust the `type` of the `rope_scaling` configuration for backward compatibility.177        """178        if self.rope_scaling is None:179            return180 181        rope_scaling_type = self.rope_scaling.get("type", None)182 183        # For backward compatibility if previous version used "su" or "yarn"184        if rope_scaling_type is not None and rope_scaling_type in ["su", "yarn"]:185            self.rope_scaling["type"] = "longrope"186 187    def _rope_scaling_validation(self):188        """189        Validate the `rope_scaling` configuration.190        """191        if self.rope_scaling is None:192            return193 194        if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 3:195            raise ValueError(196                "`rope_scaling` must be a dictionary with three fields, `type`, `short_factor` and `long_factor`, "197                f"got {self.rope_scaling}"198            )199        rope_scaling_type = self.rope_scaling.get("type", None)200        rope_scaling_short_factor = self.rope_scaling.get("short_factor", None)201        rope_scaling_long_factor = self.rope_scaling.get("long_factor", None)202        if rope_scaling_type is None or rope_scaling_type not in ["longrope"]:203            raise ValueError(f"`rope_scaling`'s type field must be one of ['longrope'], got {rope_scaling_type}")204        if not (205            isinstance(rope_scaling_short_factor, list)206            and all(isinstance(x, (int, float)) for x in rope_scaling_short_factor)207        ):208            raise ValueError(209                f"`rope_scaling`'s short_factor field must be a list of numbers, got {rope_scaling_short_factor}"210            )211        rotary_ndims = int(self.hidden_size // self.num_attention_heads * self.partial_rotary_factor)212        if not len(rope_scaling_short_factor) == rotary_ndims // 2:213            raise ValueError(214                f"`rope_scaling`'s short_factor field must have length {rotary_ndims // 2}, got {len(rope_scaling_short_factor)}"215            )216        if not (217            isinstance(rope_scaling_long_factor, list)218            and all(isinstance(x, (int, float)) for x in rope_scaling_long_factor)219        ):220            raise ValueError(221                f"`rope_scaling`'s long_factor field must be a list of numbers, got {rope_scaling_long_factor}"222            )223        if not len(rope_scaling_long_factor) == rotary_ndims // 2:224            raise ValueError(225                f"`rope_scaling`'s long_factor field must have length {rotary_ndims // 2}, got {len(rope_scaling_long_factor)}"226            )227