FriendliAI/Phi-tiny-MoE-instruct

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""" PyTorch Phi-MoE model."""17 18 19from transformers.configuration_utils import PretrainedConfig20from transformers.utils import logging21 22 23logger = logging.get_logger(__name__)24 25 26class PhiMoEConfig(PretrainedConfig):27    r"""28    This is the configuration class to store the configuration of a [`PhiMoEModel`]. It is used to instantiate a Phi-MoE29    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the30    defaults will yield a similar configuration to that of the31    [microsoft/Phi-3.5-MoE-instruct](https://huggingface.co/microsoft/Phi-3.5-MoE-instruct).32 33    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the34    documentation from [`PretrainedConfig`] for more information.35 36 37    Args:38        vocab_size (`int`, *optional*, defaults to 32064):39            Vocabulary size of the PhiMoE model. Defines the number of different tokens that can be represented by the40            `inputs_ids` passed when calling [`PhiMoEModel`]41        hidden_size (`int`, *optional*, defaults to 4096):42            Dimension of the hidden representations.43        intermediate_size (`int`, *optional*, defaults to 6400):44            Dimension of the MLP representations.45        num_hidden_layers (`int`, *optional*, defaults to 32):46            Number of hidden layers in the Transformer encoder.47        num_attention_heads (`int`, *optional*, defaults to 32):48            Number of attention heads for each attention layer in the Transformer encoder.49        num_key_value_heads (`int`, *optional*, defaults to 8):50            This is the number of key_value heads that should be used to implement Grouped Query Attention. If51            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if52            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When53            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed54            by meanpooling all the original heads within that group. For more details checkout [this55            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `8`.56        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):57            The non-linear activation function (function or string) in the decoder.58        max_position_embeddings (`int`, *optional*, defaults to `4096*32`):59            The maximum sequence length that this model might ever be used with. Mixtral's sliding window attention60            allows sequence of up to 4096*32 tokens.61        initializer_range (`float`, *optional*, defaults to 0.02):62            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.63        rms_norm_eps (`float`, *optional*, defaults to 1e-05):64            The epsilon used by the rms normalization layers.65        use_cache (`bool`, *optional*, defaults to `True`):66            Whether or not the model should return the last key/values attentions (not used by all models). Only67            relevant if `config.is_decoder=True`.68        pad_token_id (`int`, *optional*):69            The id of the padding token.70        bos_token_id (`int`, *optional*, defaults to 1):71            The id of the "beginning-of-sequence" token.72        eos_token_id (`int`, *optional*, defaults to 2):73            The id of the "end-of-sequence" token.74        tie_word_embeddings (`bool`, *optional*, defaults to `False`):75            Whether the model's input and output word embeddings should be tied.76        rope_theta (`float`, *optional*, defaults to 10000.0):77            The base period of the RoPE embeddings.78        rope_scaling (`dict`, *optional*):79            The scaling strategy for the RoPE embeddings. If `None`, no scaling is applied. If a dictionary, it must80            contain the following keys: `type`, `short_factor`, `long_factor`, `short_mscale`, `long_mscale` and81            `original_max_position_embeddings`. The `type` must be `longrope`, the `short_mscale` and `long_scale` must82            be numbers, the `short_factor` and `long_factor` must be lists of numbers with the same length as half of83            the attention head size and the `original_max_position_embeddings` must be an integer.84        sliding_window (`int`, *optional*):85            Sliding window attention window size. If not specified, will default to `262144`.86        attention_dropout (`float`, *optional*, defaults to 0.0):87            The dropout ratio for the attention probabilities.88        num_experts_per_tok (`int`, *optional*, defaults to 2):89            The number of experts to root per-token, can be also interpreted as the `top-p` routing90            parameter91        num_local_experts (`int`, *optional*, defaults to 16):92            Number of experts per Sparse MLP layer.93        output_router_logits (`bool`, *optional*, defaults to `False`):94            Whether or not the router logits should be returned by the model. Enabeling this will also95            allow the model to output the auxiliary loss. See [here]() for more details96        router_aux_loss_coef (`float`, *optional*, defaults to 0.0):97            The aux loss factor for the total loss.98        router_jitter_noise (`float`, *optional*, defaults to 0.01):99            Amount of noise to add to the router.100 101    ```python102    >>> from transformers import PhiMoEModel, PhiMoEConfig103 104    >>> # Initializing a Phi-3 style configuration105    >>> configuration = PhiMoEConfig.from_pretrained("microsoft/Phi-3.5-MoE-instruct")106 107    >>> # Initializing a model from the configuration108    >>> model = PhiMoEModel(configuration)109 110    >>> # Accessing the model configuration111    >>> configuration = model.config112    ```"""113    114    model_type = "phimoe"115    keys_to_ignore_at_inference = ["past_key_values"]116 117    def __init__(118        self,119        vocab_size=32064,120        hidden_size=4096,121        intermediate_size=6400,122        num_hidden_layers=32,123        num_attention_heads=32,124        num_key_value_heads=8,125        head_dim=None, # added to control head dimension126        hidden_act="silu",127        max_position_embeddings=4096 * 32,128        initializer_range=0.02,129        rms_norm_eps=1e-5,130        use_cache=True,131        pad_token_id=None,132        bos_token_id=1,133        eos_token_id=2,134        tie_word_embeddings=False,135        rope_theta=1e6,136        rope_scaling=None,137        sliding_window=None,138        attention_dropout=0.0,139        num_experts_per_tok=2,140        num_local_experts=16,141        output_router_logits=False,142        router_aux_loss_coef=0.001,143        router_jitter_noise=0.01,144        input_jitter_noise=0.0,145        attention_bias = False,146        lm_head_bias = False,147        **kwargs,148    ):149        self.vocab_size = vocab_size150        self.max_position_embeddings = max_position_embeddings151        self.hidden_size = hidden_size152        self.intermediate_size = intermediate_size153        self.num_hidden_layers = num_hidden_layers154        self.num_attention_heads = num_attention_heads155        self.sliding_window = sliding_window156        self.attention_bias = attention_bias157        self.lm_head_bias = lm_head_bias158        # for backward compatibility159        if num_key_value_heads is None:160            num_key_value_heads = num_attention_heads161        if head_dim is None:162            head_dim = hidden_size // num_attention_heads163 164        self.head_dim = head_dim165        self.num_key_value_heads = num_key_value_heads166        self.hidden_act = hidden_act167        self.initializer_range = initializer_range168        self.rms_norm_eps = rms_norm_eps169        self.use_cache = use_cache170        self.rope_theta = rope_theta171        self.attention_dropout = attention_dropout172 173        self.num_experts_per_tok = num_experts_per_tok174        self.num_local_experts = num_local_experts175        self.output_router_logits = output_router_logits176        self.router_aux_loss_coef = router_aux_loss_coef177        self.router_jitter_noise = router_jitter_noise178        self.input_jitter_noise = input_jitter_noise179 180        self.rope_scaling = rope_scaling181        self._rope_scaling_validation()182 183        super().__init__(184            pad_token_id=pad_token_id,185            bos_token_id=bos_token_id,186            eos_token_id=eos_token_id,187            tie_word_embeddings=tie_word_embeddings,188            **kwargs,189        )190 191    def _rope_scaling_validation(self):192        """193        Validate the `rope_scaling` configuration.194        """195        if self.rope_scaling is None:196            return197 198        if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 6:199            raise ValueError(200                "`rope_scaling` must be a dictionary with three fields, `type`, `short_factor`, `long_factor`, "201                f"`short_mscale`, `long_mscale` and `original_max_position_embeddings`, got {self.rope_scaling}"202            )203        rope_scaling_type = self.rope_scaling.get("type", None)204        rope_scaling_short_factor = self.rope_scaling.get("short_factor", None)205        rope_scaling_long_factor = self.rope_scaling.get("long_factor", None)206        rope_scaling_short_mscale = self.rope_scaling.get("short_mscale", None)207        rope_scaling_long_mscale = self.rope_scaling.get("long_mscale", None)208        original_max_position_embeddings = self.rope_scaling.get("original_max_position_embeddings", None)209        if rope_scaling_type is None or rope_scaling_type not in ["longrope"]:210            raise ValueError(f"`rope_scaling`'s type field must be one of ['longrope'], got {rope_scaling_type}")211        if not (212            isinstance(rope_scaling_short_factor, list)213            and all(isinstance(x, (int, float)) for x in rope_scaling_short_factor)214        ):215            raise ValueError(216                f"`rope_scaling`'s short_factor field must be a list of numbers, got {rope_scaling_short_factor}"217            )218        if not len(rope_scaling_short_factor) == self.hidden_size // self.num_attention_heads // 2:219            raise ValueError(220                f"`rope_scaling`'s short_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_short_factor)}"221            )222        if not (223            isinstance(rope_scaling_long_factor, list)224            and all(isinstance(x, (int, float)) for x in rope_scaling_long_factor)225        ):226            raise ValueError(227                f"`rope_scaling`'s long_factor field must be a list of numbers, got {rope_scaling_long_factor}"228            )229        if not len(rope_scaling_long_factor) == self.hidden_size // self.num_attention_heads // 2:230            raise ValueError(231                f"`rope_scaling`'s long_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_long_factor)}"232            )233        if not isinstance(rope_scaling_short_mscale, (int, float)):234            raise ValueError(235                f"`rope_scaling`'s short_mscale field must be a number, got {rope_scaling_short_mscale}"236            )237        if not isinstance(rope_scaling_long_mscale, (int, float)):238            raise ValueError(239                f"`rope_scaling`'s long_mscale field must be a number, got {rope_scaling_long_mscale}"240            )241        if not isinstance(original_max_position_embeddings, int):242            raise ValueError(243                f"`rope_scaling`'s original_max_position_embeddings field must be an integer, got {original_max_position_embeddings}"244            )