MSALab/PerceptionDLM-Base

1# coding=utf-82# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.3#4# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX5# and OPT implementations in this library. It has been modified from its6# original forms to accommodate minor architectural differences compared7# to GPT-NeoX and OPT used by the Meta AI team that trained the model.8#9# Licensed under the Apache License, Version 2.0 (the "License");10# you may not use this file except in compliance with the License.11# You may obtain a copy of the License at12#13#     http://www.apache.org/licenses/LICENSE-2.014#15# Unless required by applicable law or agreed to in writing, software16# distributed under the License is distributed on an "AS IS" BASIS,17# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.18# See the License for the specific language governing permissions and19# limitations under the License.20""" LLaDA model configuration"""21 22from transformers.configuration_utils import PretrainedConfig23from transformers.utils import logging24 25 26logger = logging.get_logger(__name__)27 28LLaDA_PRETRAINED_CONFIG_ARCHIVE_MAP = {}29 30 31class LLaDAConfig(PretrainedConfig):32    r"""33    This is the configuration class to store the configuration of a [`LLaDAModel`]. It is used to instantiate an LLaDA34    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the35    defaults will yield a similar configuration to that of the LLaDA-8B.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 41    Args:42        vocab_size (`int`, *optional*, defaults to 32000):43            Vocabulary size of the LLaDA model. Defines the number of different tokens that can be represented by the44            `inputs_ids` passed when calling [`LLaDAModel`]45        hidden_size (`int`, *optional*, defaults to 4096):46            Dimension of the hidden representations.47        intermediate_size (`int`, *optional*, defaults to 11008):48            Dimension of the MLP representations.49        num_hidden_layers (`int`, *optional*, defaults to 32):50            Number of hidden layers in the Transformer decoder.51        num_attention_heads (`int`, *optional*, defaults to 32):52            Number of attention heads for each attention layer in the Transformer decoder.53        num_key_value_heads (`int`, *optional*):54            This is the number of key_value heads that should be used to implement Grouped Query Attention. If55            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if56            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When57            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed58            by meanpooling all the original heads within that group. For more details checkout [this59            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to60            `num_attention_heads`.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 2048):64            The maximum sequence length that this model might ever be used with. 65        initializer_range (`float`, *optional*, defaults to 0.02):66            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.67        rms_norm_eps (`float`, *optional*, defaults to 1e-06):68            The epsilon used by the rms normalization layers.69        use_cache (`bool`, *optional*, defaults to `True`):70            Whether or not the model should return the last key/values attentions (not used by all models). Only71            relevant if `config.is_decoder=True`.72        pad_token_id (`int`, *optional*):73            Padding token id.74        bos_token_id (`int`, *optional*, defaults to 1):75            Beginning of stream token id.76        eos_token_id (`int`, *optional*, defaults to 2):77            End of stream token id.78        pretraining_tp (`int`, *optional*, defaults to 1):79            Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this80            document](https://huggingface.co/docs/transformers/main/perf_train_gpu_many#tensor-parallelism) to understand more about it. This value is81            necessary to ensure exact reproducibility of the pretraining results. Please refer to [this82            issue](https://github.com/pytorch/pytorch/issues/76232).83        tie_word_embeddings (`bool`, *optional*, defaults to `False`):84            Whether to tie weight embeddings85        rope_theta (`float`, *optional*, defaults to 10000.0):86            The base period of the RoPE embeddings.87        rope_scaling (`Dict`, *optional*):88            Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling89            strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is90            `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update91            `max_position_embeddings` to the expected new maximum. 92        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):93            Whether to use a bias in the query, key, value and output projection layers during self-attention.94        attention_dropout (`float`, *optional*, defaults to 0.0):95            The dropout ratio for the attention probabilities.96    """97 98    model_type = "llada"99    keys_to_ignore_at_inference = ["past_key_values"]100 101    def __init__(102        self,103        vocab_size=32000,104        hidden_size=4096,105        intermediate_size=11008,106        num_hidden_layers=32,107        num_attention_heads=32,108        num_key_value_heads=None,109        hidden_act="silu",110        max_position_embeddings=2048,111        initializer_range=0.02,112        rms_norm_eps=1e-6,113        use_cache=True,114        pad_token_id=None,115        bos_token_id=1,116        eos_token_id=2,117        pretraining_tp=1,118        tie_word_embeddings=False,119        rope_theta=10000.0,120        rope_scaling=None,121        attention_bias=False,122        attention_dropout=0.0,123        **kwargs,124    ):125        self.vocab_size = vocab_size126        self.max_position_embeddings = max_position_embeddings127        self.hidden_size = hidden_size128        self.intermediate_size = intermediate_size129        self.num_hidden_layers = num_hidden_layers130        self.num_attention_heads = num_attention_heads131 132        # for backward compatibility133        if num_key_value_heads is None:134            num_key_value_heads = num_attention_heads135 136        self.num_key_value_heads = num_key_value_heads137        self.hidden_act = hidden_act138        self.initializer_range = initializer_range139        self.rms_norm_eps = rms_norm_eps140        self.pretraining_tp = pretraining_tp141        self.use_cache = use_cache142        self.rope_theta = rope_theta143        self.rope_scaling = rope_scaling144        self._rope_scaling_validation()145        self.attention_bias = attention_bias146        self.attention_dropout = attention_dropout147 148        super().__init__(149            pad_token_id=pad_token_id,150            bos_token_id=bos_token_id,151            eos_token_id=eos_token_id,152            tie_word_embeddings=tie_word_embeddings,153            **kwargs,154        )155 156    def _rope_scaling_validation(self):157        """158        Validate the `rope_scaling` configuration.159        """160        if self.rope_scaling is None:161            return162 163        if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 2:164            raise ValueError(165                "`rope_scaling` must be a dictionary with with two fields, `type` and `factor`, "166                f"got {self.rope_scaling}"167            )168        rope_scaling_type = self.rope_scaling.get("type", None)169        rope_scaling_factor = self.rope_scaling.get("factor", None)170        if rope_scaling_type is None or rope_scaling_type not in ["linear", "dynamic"]:171            raise ValueError(172                f"`rope_scaling`'s type field must be one of ['linear', 'dynamic'], got {rope_scaling_type}"173            )174        if rope_scaling_factor is None or not isinstance(rope_scaling_factor, float) or rope_scaling_factor <= 1.0:175            raise ValueError(f"`rope_scaling`'s factor field must be a float > 1, got {rope_scaling_factor}")176