stabilityai/stable-code-3b

1# coding=utf-82# Copyright 2024 Stability AI 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""" StableLM model configuration """16 17from transformers.configuration_utils import PretrainedConfig18from transformers.utils import logging19 20 21logger = logging.get_logger(__name__)22 23STABLELM_PRETRAINED_CONFIG_ARCHIVE_MAP = {24    "stabilityai/stablelm-3b-4e1t": "https://huggingface.co/stabilityai/stablelm-3b-4e1t/resolve/main/config.json",25    # See all StableLM models at https://huggingface.co/models?filter=stablelm26}27 28 29class StableLmConfig(PretrainedConfig):30    r"""31    This is the configuration class to store the configuration of a [`~StableLmModel`].32    It is used to instantiate an StableLM model according to the specified arguments, defining the model33    architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of34    the StableLM [stabilityai/stablelm-3b-4e1t](https://huggingface.co/stabilityai/stablelm-3b-4e1t) architecture.35 36    Configuration objects inherit from  [`PretrainedConfig`] and can be used37    to control the model outputs. Read the documentation from  [`PretrainedConfig`]38    for more information.39 40 41    Args:42        vocab_size (`int`, *optional*, defaults to 50304):43            Vocabulary size of the StableLM model. Defines the number of different tokens that44            can be represented by the `inputs_ids` passed when calling [`StableLmModel`].45        intermediate_size (`int`, *optional*, defaults to 6912):46            Dimension of the MLP representations.47        hidden_size (`int`, *optional*, defaults to 2560):48            Number of hidden layers in the Transformer decoder.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 encoder.53        num_key_value_heads (`int`, *optional*, defaults to 32):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).63        max_position_embeddings (`int`, *optional*, defaults to 4096):64            The maximum sequence length that this model might ever be used with.65            Typically set this to something large just in case (e.g., 512 or 1024 or 2048).66        initializer_range (`float`, *optional*, defaults to 0.02):67            The standard deviation of the truncated_normal_initializer for initializing68             all weight matrices.69        layer_norm_eps (`float`, *optional*, defaults to 1e-05):70            The epsilon used by the normalization layers.71        use_cache (`bool`, *optional*, defaults to `True`):72            Whether or not the model should return the last key/values attentions73            (not used by all models). Only relevant if `config.is_decoder=True`.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            Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling80            strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is81            `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update82            `max_position_embeddings` to the expected new maximum. See the following thread for more information on how83            these scaling strategies behave:84            https://www.reddit.com/r/LocalLLaMA/comments/14mrgpr/dynamically_scaled_rope_further_increases/. This85            is an experimental feature, subject to breaking API changes in future versions.86        use_qkv_bias (`bool`, *optional*, defaults to `False`):87            Whether or not the model should use bias for qkv layers.88        hidden_dropout (`float`, *optional*, defaults to 0.0):89            The dropout ratio after applying the MLP to the hidden states.90        attention_dropout (`float`, *optional*, defaults to 0.0):91            The dropout ratio for the attention probabilities.92        partial_rotary_factor (`float`, *optional*, defaults to 0.25):93            Percentage of the query and keys which will have rotary embedding.94        bos_token_id (int, *optional*, defaults to 0):95            The id of the `BOS` token in the vocabulary.96        eos_token_id (int, *optional*, defaults to 0):97            The id of the `EOS` token in the vocabulary.98 99    Example:100 101    ```python102    >>> from transformers import StableLmModel, StableLmConfig103 104    >>> # Initializing a StableLM stablelm-3b style configuration105    >>> configuration = StableLmConfig()106    ```"""107 108    model_type = "stablelm"109    keys_to_ignore_at_inference = ["past_key_values"]110 111    def __init__(112        self,113        vocab_size=50304,114        intermediate_size=6912,115        hidden_size=2560,116        num_hidden_layers=32,117        num_attention_heads=32,118        num_key_value_heads=32,119        hidden_act="silu",120        max_position_embeddings=4096,121        initializer_range=0.02,122        layer_norm_eps=1.0e-5,123        use_cache=True,124        tie_word_embeddings=False,125        rope_theta=10_000,126        rope_scaling=None,127        use_qkv_bias=False,128        hidden_dropout=0.0,129        attention_dropout=0.0,130        partial_rotary_factor=0.25,131        bos_token_id=0,132        eos_token_id=0,133        **kwargs,134    ):135        self.vocab_size = vocab_size136        self.max_position_embeddings = max_position_embeddings137 138        self.hidden_size = hidden_size139        self.intermediate_size = intermediate_size140        self.num_hidden_layers = num_hidden_layers141        self.num_attention_heads = num_attention_heads142        self.num_key_value_heads = num_key_value_heads143        self.hidden_act = hidden_act144 145        self.initializer_range = initializer_range146        self.layer_norm_eps = layer_norm_eps147        self.use_cache = use_cache148        self.rope_theta = rope_theta149        self.rope_scaling = rope_scaling150        self.use_qkv_bias = use_qkv_bias151        self.hidden_dropout = hidden_dropout152        self.attention_dropout = attention_dropout153        self.partial_rotary_factor = partial_rotary_factor154        self._rope_scaling_validation()155 156        super().__init__(157            bos_token_id=bos_token_id,158            eos_token_id=eos_token_id,159            tie_word_embeddings=tie_word_embeddings,160            **kwargs,161        )162 163    # Copied from transformers.models.llama.configuration_llama.LlamaConfig._rope_scaling_validation164    def _rope_scaling_validation(self):165        """166        Validate the `rope_scaling` configuration.167        """168        if self.rope_scaling is None:169            return170 171        if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 2:172            raise ValueError(173                "`rope_scaling` must be a dictionary with with two fields, `type` and `factor`, "174                f"got {self.rope_scaling}"175            )176        rope_scaling_type = self.rope_scaling.get("type", None)177        rope_scaling_factor = self.rope_scaling.get("factor", None)178        if rope_scaling_type is None or rope_scaling_type not in ["linear", "dynamic"]:179            raise ValueError(180                f"`rope_scaling`'s type field must be one of ['linear', 'dynamic'], got {rope_scaling_type}"181            )182        if rope_scaling_factor is None or not isinstance(rope_scaling_factor, float) or rope_scaling_factor <= 1.0:183            raise ValueError(f"`rope_scaling`'s factor field must be a float > 1, got {rope_scaling_factor}")184