ScalableMath/Lean-STaR-plus

1# coding=utf-82# Copyright (c) The InternLM team and The HuggingFace Inc. team. All rights reserved.3#4# This code is based on transformers/src/transformers/models/llama/configuration_llama.py5#6# Licensed under the Apache License, Version 2.0 (the "License");7# you may not use this file except in compliance with the License.8# You may obtain a copy of the License at9#10#     http://www.apache.org/licenses/LICENSE-2.011#12# Unless required by applicable law or agreed to in writing, software13# distributed under the License is distributed on an "AS IS" BASIS,14# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.15# See the License for the specific language governing permissions and16# limitations under the License.17""" InternLM2 model configuration"""18 19from transformers.configuration_utils import PretrainedConfig20from transformers.utils import logging21 22logger = logging.get_logger(__name__)23 24INTERNLM2_PRETRAINED_CONFIG_ARCHIVE_MAP = {}25 26 27# Modified from transformers.model.llama.configuration_llama.LlamaConfig28class InternLM2Config(PretrainedConfig):29    r"""30    This is the configuration class to store the configuration of a [`InternLM2Model`]. It is used to instantiate31    an InternLM2 model according to the specified arguments, defining the model architecture. Instantiating a32    configuration with the defaults will yield a similar configuration to that of the InternLM2-7B.33 34    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the35    documentation from [`PretrainedConfig`] for more information.36 37 38    Args:39        vocab_size (`int`, *optional*, defaults to 32000):40            Vocabulary size of the InternLM2 model. Defines the number of different tokens that can be represented by the41            `inputs_ids` passed when calling [`InternLM2Model`]42        hidden_size (`int`, *optional*, defaults to 4096):43            Dimension of the hidden representations.44        intermediate_size (`int`, *optional*, defaults to 11008):45            Dimension of the MLP representations.46        num_hidden_layers (`int`, *optional*, defaults to 32):47            Number of hidden layers in the Transformer encoder.48        num_attention_heads (`int`, *optional*, defaults to 32):49            Number of attention heads for each attention layer in the Transformer encoder.50        num_key_value_heads (`int`, *optional*):51            This is the number of key_value heads that should be used to implement Grouped Query Attention. If52            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if53            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When54            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed55            by meanpooling all the original heads within that group. For more details checkout [this56            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to57            `num_attention_heads`.58        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):59            The non-linear activation function (function or string) in the decoder.60        max_position_embeddings (`int`, *optional*, defaults to 2048):61            The maximum sequence length that this model might ever be used with. Typically set this to something large62            just in case (e.g., 512 or 1024 or 2048).63        initializer_range (`float`, *optional*, defaults to 0.02):64            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.65        rms_norm_eps (`float`, *optional*, defaults to 1e-12):66            The epsilon used by the rms normalization layers.67        use_cache (`bool`, *optional*, defaults to `True`):68            Whether or not the model should return the last key/values attentions (not used by all models). Only69            relevant if `config.is_decoder=True`.70        tie_word_embeddings(`bool`, *optional*, defaults to `False`):71            Whether to tie weight embeddings72        Example:73 74    """75    model_type = "internlm2"76    _auto_class = "AutoConfig"77 78    def __init__(  # pylint: disable=W010279        self,80        vocab_size=103168,81        hidden_size=4096,82        intermediate_size=11008,83        num_hidden_layers=32,84        num_attention_heads=32,85        num_key_value_heads=None,86        hidden_act="silu",87        max_position_embeddings=2048,88        initializer_range=0.02,89        rms_norm_eps=1e-6,90        use_cache=True,91        pad_token_id=0,92        bos_token_id=1,93        eos_token_id=2,94        tie_word_embeddings=False,95        bias=True,96        rope_theta=10000,97        rope_scaling=None,98        attn_implementation="eager",99        **kwargs,100    ):101        self.vocab_size = vocab_size102        self.max_position_embeddings = max_position_embeddings103        self.hidden_size = hidden_size104        self.intermediate_size = intermediate_size105        self.num_hidden_layers = num_hidden_layers106        self.num_attention_heads = num_attention_heads107        self.bias = bias108 109        if num_key_value_heads is None:110            num_key_value_heads = num_attention_heads111        self.num_key_value_heads = num_key_value_heads112 113        self.hidden_act = hidden_act114        self.initializer_range = initializer_range115        self.rms_norm_eps = rms_norm_eps116        self.use_cache = use_cache117        self.rope_theta = rope_theta118        self.rope_scaling = rope_scaling119        self._rope_scaling_validation()120 121        self.attn_implementation = attn_implementation122        if self.attn_implementation is None:123            self.attn_implementation = "eager"124        super().__init__(125            pad_token_id=pad_token_id,126            bos_token_id=bos_token_id,127            eos_token_id=eos_token_id,128            tie_word_embeddings=tie_word_embeddings,129            **kwargs,130        )131 132    def _rope_scaling_validation(self):133        """134        Validate the `rope_scaling` configuration.135        """136        if self.rope_scaling is None:137            return138 139        if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 2:140            raise ValueError(141                "`rope_scaling` must be a dictionary with with two fields, `type` and `factor`, "142                f"got {self.rope_scaling}"143            )144        rope_scaling_type = self.rope_scaling.get("type", None)145        rope_scaling_factor = self.rope_scaling.get("factor", None)146        if rope_scaling_type is None or rope_scaling_type not in ["linear", "dynamic"]:147            raise ValueError(148                f"`rope_scaling`'s type field must be one of ['linear', 'dynamic'], got {rope_scaling_type}"149            )150        if rope_scaling_factor is None or not isinstance(rope_scaling_factor, float) or rope_scaling_factor < 1.0:151            raise ValueError(f"`rope_scaling`'s factor field must be a float >= 1, got {rope_scaling_factor}")