MathLLMs/MathCoder-VL-2B

1# Copyright (c) The InternLM team and The HuggingFace Inc. team. All rights reserved.2#3# This code is based on transformers/src/transformers/models/llama/configuration_llama.py4#5# Licensed under the Apache License, Version 2.0 (the "License");6# you may not use this file except in compliance with the License.7# You may obtain a copy of the License at8#9#     http://www.apache.org/licenses/LICENSE-2.010#11# Unless required by applicable law or agreed to in writing, software12# distributed under the License is distributed on an "AS IS" BASIS,13# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.14# See the License for the specific language governing permissions and15# limitations under the License.16""" InternLM2 model configuration"""17 18from transformers.configuration_utils import PretrainedConfig19from transformers.utils import logging20 21logger = logging.get_logger(__name__)22 23INTERNLM2_PRETRAINED_CONFIG_ARCHIVE_MAP = {}24 25 26# Modified from transformers.model.llama.configuration_llama.LlamaConfig27class InternLM2Config(PretrainedConfig):28    r"""29    This is the configuration class to store the configuration of a [`InternLM2Model`]. It is used to instantiate30    an InternLM2 model according to the specified arguments, defining the model architecture. Instantiating a31    configuration with the defaults will yield a similar configuration to that of the InternLM2-7B.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 32000):39            Vocabulary size of the InternLM2 model. Defines the number of different tokens that can be represented by the40            `inputs_ids` passed when calling [`InternLM2Model`]41        hidden_size (`int`, *optional*, defaults to 4096):42            Dimension of the hidden representations.43        intermediate_size (`int`, *optional*, defaults to 11008):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*):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 to56            `num_attention_heads`.57        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):58            The non-linear activation function (function or string) in the decoder.59        max_position_embeddings (`int`, *optional*, defaults to 2048):60            The maximum sequence length that this model might ever be used with. Typically set this to something large61            just in case (e.g., 512 or 1024 or 2048).62        initializer_range (`float`, *optional*, defaults to 0.02):63            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.64        rms_norm_eps (`float`, *optional*, defaults to 1e-12):65            The epsilon used by the rms normalization layers.66        use_cache (`bool`, *optional*, defaults to `True`):67            Whether or not the model should return the last key/values attentions (not used by all models). Only68            relevant if `config.is_decoder=True`.69        tie_word_embeddings(`bool`, *optional*, defaults to `False`):70            Whether to tie weight embeddings71        Example:72 73    """74    model_type = 'internlm2'75    _auto_class = 'AutoConfig'76 77    def __init__(  # pylint: disable=W010278        self,79        vocab_size=103168,80        hidden_size=4096,81        intermediate_size=11008,82        num_hidden_layers=32,83        num_attention_heads=32,84        num_key_value_heads=None,85        hidden_act='silu',86        max_position_embeddings=2048,87        initializer_range=0.02,88        rms_norm_eps=1e-6,89        use_cache=True,90        pad_token_id=0,91        bos_token_id=1,92        eos_token_id=2,93        tie_word_embeddings=False,94        bias=True,95        rope_theta=10000,96        rope_scaling=None,97        attn_implementation='eager',98        **kwargs,99    ):100        self.vocab_size = vocab_size101        self.max_position_embeddings = max_position_embeddings102        self.hidden_size = hidden_size103        self.intermediate_size = intermediate_size104        self.num_hidden_layers = num_hidden_layers105        self.num_attention_heads = num_attention_heads106        self.bias = bias107 108        if num_key_value_heads is None:109            num_key_value_heads = num_attention_heads110        self.num_key_value_heads = num_key_value_heads111 112        self.hidden_act = hidden_act113        self.initializer_range = initializer_range114        self.rms_norm_eps = rms_norm_eps115        self.use_cache = use_cache116        self.rope_theta = rope_theta117        self.rope_scaling = rope_scaling118        self._rope_scaling_validation()119 120        self.attn_implementation = attn_implementation121        if self.attn_implementation is None:122            self.attn_implementation = 'eager'123        super().__init__(124            pad_token_id=pad_token_id,125            bos_token_id=bos_token_id,126            eos_token_id=eos_token_id,127            tie_word_embeddings=tie_word_embeddings,128            **kwargs,129        )130 131    def _rope_scaling_validation(self):132        """133        Validate the `rope_scaling` configuration.134        """135        if self.rope_scaling is None:136            return137 138        if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 2:139            raise ValueError(140                '`rope_scaling` must be a dictionary with with two fields, `type` and `factor`, '141                f'got {self.rope_scaling}'142            )143        rope_scaling_type = self.rope_scaling.get('type', None)144        rope_scaling_factor = self.rope_scaling.get('factor', None)145        if rope_scaling_type is None or rope_scaling_type not in ['linear', 'dynamic']:146            raise ValueError(147                f"`rope_scaling`'s type field must be one of ['linear', 'dynamic'], got {rope_scaling_type}"148            )149        if rope_scaling_factor is None or not isinstance(rope_scaling_factor, float) or rope_scaling_factor < 1.0:150            raise ValueError(f"`rope_scaling`'s factor field must be a float >= 1, got {rope_scaling_factor}")151