Salesforce/codegen2-7B_P

1# coding=utf-82# Copyright 2022 Salesforce authors, The EleutherAI, and HuggingFace Teams. 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""" CodeGen model configuration"""16from collections import OrderedDict17from typing import Any, List, Mapping, Optional18 19from transformers import PreTrainedTokenizer, TensorType, is_torch_available20from transformers.configuration_utils import PretrainedConfig21from transformers.onnx import OnnxConfigWithPast, PatchingSpec22from transformers.utils import logging23 24 25logger = logging.get_logger(__name__)26 27 28CODEGEN_PRETRAINED_CONFIG_ARCHIVE_MAP = {29    "Salesforce/codegen-350M-nl": "https://huggingface.co/Salesforce/codegen-350M-nl/resolve/main/config.json",30    "Salesforce/codegen-350M-multi": "https://huggingface.co/Salesforce/codegen-350M-multi/resolve/main/config.json",31    "Salesforce/codegen-350M-mono": "https://huggingface.co/Salesforce/codegen-350M-mono/resolve/main/config.json",32    "Salesforce/codegen-2B-nl": "https://huggingface.co/Salesforce/codegen-2B-nl/resolve/main/config.json",33    "Salesforce/codegen-2B-multi": "https://huggingface.co/Salesforce/codegen-2B-multi/resolve/main/config.json",34    "Salesforce/codegen-2B-mono": "https://huggingface.co/Salesforce/codegen-2B-mono/resolve/main/config.json",35    "Salesforce/codegen-6B-nl": "https://huggingface.co/Salesforce/codegen-6B-nl/resolve/main/config.json",36    "Salesforce/codegen-6B-multi": "https://huggingface.co/Salesforce/codegen-6B-multi/resolve/main/config.json",37    "Salesforce/codegen-6B-mono": "https://huggingface.co/Salesforce/codegen-6B-mono/resolve/main/config.json",38    "Salesforce/codegen-16B-nl": "https://huggingface.co/Salesforce/codegen-16B-nl/resolve/main/config.json",39    "Salesforce/codegen-16B-multi": "https://huggingface.co/Salesforce/codegen-16B-multi/resolve/main/config.json",40    "Salesforce/codegen-16B-mono": "https://huggingface.co/Salesforce/codegen-16B-mono/resolve/main/config.json",41}42 43 44class CodeGenConfig(PretrainedConfig):45    r"""46    This is the configuration class to store the configuration of a [`CodeGenModel`]. It is used to instantiate a47    CodeGen model according to the specified arguments, defining the model architecture. Instantiating a configuration48    with the defaults will yield a similar configuration to that of the CodeGen49    [Salesforce/codegen-2B-mono](https://huggingface.co/Salesforce/codegen-2B-mono) architecture. Configuration objects50    inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the documentation from51    [`PretrainedConfig`] for more information.52 53    Args:54        vocab_size (`int`, *optional*, defaults to 50400):55            Vocabulary size of the CodeGen model. Defines the number of different tokens that can be represented by the56            `inputs_ids` passed when calling [`CodeGenModel`].57        n_positions (`int`, *optional*, defaults to 2048):58            The maximum sequence length that this model might ever be used with. Typically set this to something large59            just in case (e.g., 512 or 1024 or 2048).60        n_embd (`int`, *optional*, defaults to 4096):61            Dimensionality of the embeddings and hidden states.62        n_layer (`int`, *optional*, defaults to 28):63            Number of hidden layers in the Transformer encoder.64        n_head (`int`, *optional*, defaults to 16):65            Number of attention heads for each attention layer in the Transformer encoder.66        rotary_dim (`int`, *optional*, defaults to 64):67            Number of dimensions in the embedding that Rotary Position Embedding is applied to.68        n_inner (`int`, *optional*, defaults to None):69            Dimensionality of the inner feed-forward layers. `None` will set it to 4 times n_embd70        activation_function (`str`, *optional*, defaults to `"gelu_new"`):71            Activation function, to be selected in the list `["relu", "silu", "gelu", "tanh", "gelu_new"]`.72        resid_pdrop (`float`, *optional*, defaults to 0.1):73            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.74        embd_pdrop (`int`, *optional*, defaults to 0.1):75            The dropout ratio for the embeddings.76        attn_pdrop (`float`, *optional*, defaults to 0.1):77            The dropout ratio for the attention.78        layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):79            The epsilon to use in the layer normalization layers.80        initializer_range (`float`, *optional*, defaults to 0.02):81            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.82        scale_attn_weights (`bool`, *optional*, defaults to `True`):83            Scale attention weights by dividing by sqrt(hidden_size).84        use_cache (`bool`, *optional*, defaults to `True`):85            Whether or not the model should return the last key/values attentions (not used by all models).86 87    Example:88 89    ```python90    >>> from transformers import CodeGenModel, CodeGenConfig91 92    >>> # Initializing a CodeGen 6B configuration93    >>> configuration = CodeGenConfig()94 95    >>> # Initializing a model from the configuration96    >>> model = CodeGenModel(configuration)97 98    >>> # Accessing the model configuration99    >>> configuration = model.config100    ```"""101    model_type = "codegen"102    attribute_map = {103        "max_position_embeddings": "n_positions",104        "hidden_size": "n_embd",105        "num_attention_heads": "n_head",106        "num_hidden_layers": "n_layer",107    }108 109    def __init__(110        self,111        vocab_size=50400,112        n_positions=2048,113        n_ctx=2048,114        n_embd=4096,115        n_layer=28,116        n_head=16,117        rotary_dim=64,118        n_inner=None,119        activation_function="gelu_new",120        resid_pdrop=0.0,121        embd_pdrop=0.0,122        attn_pdrop=0.0,123        layer_norm_epsilon=1e-5,124        initializer_range=0.02,125        scale_attn_weights=True,126        use_cache=True,127        bos_token_id=50256,128        eos_token_id=50256,129        tie_word_embeddings=False,130        **kwargs131    ):132        self.vocab_size = vocab_size133        self.n_ctx = n_ctx134        self.n_positions = n_positions135        self.n_embd = n_embd136        self.n_layer = n_layer137        self.n_head = n_head138        self.n_inner = n_inner139        self.rotary_dim = rotary_dim140        self.activation_function = activation_function141        self.resid_pdrop = resid_pdrop142        self.embd_pdrop = embd_pdrop143        self.attn_pdrop = attn_pdrop144        self.layer_norm_epsilon = layer_norm_epsilon145        self.initializer_range = initializer_range146        self.scale_attn_weights = scale_attn_weights147        self.use_cache = use_cache148 149        self.bos_token_id = bos_token_id150        self.eos_token_id = eos_token_id151 152        super().__init__(153            bos_token_id=bos_token_id, eos_token_id=eos_token_id, tie_word_embeddings=tie_word_embeddings, **kwargs154        )155 156 157# Copied from transformers.models.gpt2.configuration_gpt2.GPT2OnnxConfig158class CodeGenOnnxConfig(OnnxConfigWithPast):159    def __init__(160        self,161        config: PretrainedConfig,162        task: str = "default",163        patching_specs: List[PatchingSpec] = None,164        use_past: bool = False,165    ):166        super().__init__(config, task=task, patching_specs=patching_specs, use_past=use_past)167        if not getattr(self._config, "pad_token_id", None):168            # TODO: how to do that better?169            self._config.pad_token_id = 0170 171    @property172    def inputs(self) -> Mapping[str, Mapping[int, str]]:173        common_inputs = OrderedDict({"input_ids": {0: "batch", 1: "sequence"}})174        if self.use_past:175            self.fill_with_past_key_values_(common_inputs, direction="inputs")176            common_inputs["attention_mask"] = {0: "batch", 1: "past_sequence + sequence"}177        else:178            common_inputs["attention_mask"] = {0: "batch", 1: "sequence"}179 180        return common_inputs181 182    @property183    def num_layers(self) -> int:184        return self._config.n_layer185 186    @property187    def num_attention_heads(self) -> int:188        return self._config.n_head189 190    def generate_dummy_inputs(191        self,192        tokenizer: PreTrainedTokenizer,193        batch_size: int = -1,194        seq_length: int = -1,195        is_pair: bool = False,196        framework: Optional[TensorType] = None,197    ) -> Mapping[str, Any]:198        common_inputs = super(OnnxConfigWithPast, self).generate_dummy_inputs(199            tokenizer, batch_size=batch_size, seq_length=seq_length, is_pair=is_pair, framework=framework200        )201 202        # We need to order the input in the way they appears in the forward()203        ordered_inputs = OrderedDict({"input_ids": common_inputs["input_ids"]})204 205        # Need to add the past_keys206        if self.use_past:207            if not is_torch_available():208                raise ValueError("Cannot generate dummy past_keys inputs without PyTorch installed.")209            else:210                import torch211 212                batch, seqlen = common_inputs["input_ids"].shape213                # Not using the same length for past_key_values214                past_key_values_length = seqlen + 2215                past_shape = (216                    batch,217                    self.num_attention_heads,218                    past_key_values_length,219                    self._config.hidden_size // self.num_attention_heads,220                )221                ordered_inputs["past_key_values"] = [222                    (torch.zeros(past_shape), torch.zeros(past_shape)) for _ in range(self.num_layers)223                ]224 225        ordered_inputs["attention_mask"] = common_inputs["attention_mask"]226        if self.use_past:227            mask_dtype = ordered_inputs["attention_mask"].dtype228            ordered_inputs["attention_mask"] = torch.cat(229                [ordered_inputs["attention_mask"], torch.ones(batch, past_key_values_length, dtype=mask_dtype)], dim=1230            )231 232        return ordered_inputs233 234    @property235    def default_onnx_opset(self) -> int:236        return 13