m4/models/vopt/configuration_vopt.py

# coding=utf-8
# Copyright 2022 The Metaseq Authors and The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" OPT model configuration"""
import os
from typing import Tuple, Union

from transformers import AutoConfig
from transformers.configuration_utils import PretrainedConfig
from transformers.utils import logging


logger = logging.get_logger(__name__)

OPT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
    "facebook/opt-125m": "https://huggingface.co/facebook/opt-125m/blob/main/config.json",
    "facebook/opt-350m": "https://huggingface.co/facebook/opt-350m/blob/main/config.json",
    "facebook/opt-1.3b": "https://huggingface.co/facebook/opt-1.3b/blob/main/config.json",
    "facebook/opt-2.7b": "https://huggingface.co/facebook/opt-2.7b/blob/main/config.json",
    "facebook/opt-6.7b": "https://huggingface.co/facebook/opt-6.7b/blob/main/config.json",
    "facebook/opt-13b": "https://huggingface.co/facebook/opt-13b/blob/main/config.json",
}


class VOPTConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`OPTModel`]. It is used to instantiate a OPT model
    according to the specified arguments, defining the model architecture. Instantiating a configuration with the
    defaults will yield a similar configuration to that of the OPT
    [facebook/opt-350m](https://huggingface.co/facebook/opt-350m) architecture.

    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.

    TODO: this doc is completely out of sync with the actual args

    Args:
        vocab_size (`int`, *optional*, defaults to 50272):
            Vocabulary size of the OPT model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`OPTModel`]
        additional_vocab_size (`int`, *optional`, defaults to 0):
            Additional vocabulary size of the model, typically for the special "<img>" token. Additional vocab tokens
            are always trainable whereas regular vocab tokens can be frozen or not.
        hidden_size (`int`, *optional*, defaults to 768):
            Dimensionality of the layers and the pooler layer.
        num_hidden_layers (`int`, *optional*, defaults to 12):
            Number of decoder layers.
        ffn_dim (`int`, *optional*, defaults to 3072):
            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
        num_attention_heads (`int`, *optional*, defaults to 12):
            Number of attention heads for each attention layer in the Transformer decoder.
        activation_function (`str` or `function`, *optional*, defaults to `"relu"`):
            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
            `"relu"`, `"silu"` and `"gelu_new"` are supported.
        max_position_embeddings (`int`, *optional*, defaults to 2048):
            The maximum sequence length that this model might ever be used with. Typically set this to something large
            just in case (e.g., 512 or 1024 or 2048).
        do_layer_norm_before (`bool`, *optional*, defaults to `True`):
            Whether to perform layer normalization before the attention block.
        word_embed_proj_dim (`int`, *optional*):
            `word_embed_proj_dim` can be set to down-project word embeddings, *e.g.* `opt-350m`. Defaults to
            `hidden_size`.
        dropout (`float`, *optional*, defaults to 0.1):
            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.
        layerdrop: (`float`, *optional*, defaults to 0.0):
            The LayerDrop probability. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556) for more
            details.
        init_std (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        alpha_initializer (`str`, *optional*, defaults to `"ones"`):
            Initialization type for the alphas.
        alphas_initializer_range (`float`, *optional*, defaults to 0.0):
            The standard deviation of the truncated_normal_initializer for initializing the alphas in the Gated Cross Attention.
        alpha_type (`str`, *optional*, defaults to `"vector"`):
            Whether the gating alphas should be vectors or single floats.
        use_cache (`bool`, *optional*, defaults to `True`):
            Whether or not the model should return the last key/values attentions (not used by all models).
        cross_layer_interval (`int`, *optional*, default to 1)
            Interval for cross attention (from text to image) layers.
    Example:

    ```python
    >>> from transformers import OPTModel, OPTConfig

    >>> # Initializing a OPT facebook/opt-large style configuration
    >>> configuration = OPTConfig()

    >>> # Initializing a model from the facebook/opt-large style configuration
    >>> model = OPTModel(configuration)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""
    model_type = "vopt"
    keys_to_ignore_at_inference = ["past_key_values"]

    def __init__(
        self,
        vocab_size=50272,
        additional_vocab_size=0,
        hidden_size=768,
        num_hidden_layers=12,
        ffn_dim=3072,
        max_position_embeddings=2048,
        do_layer_norm_before=True,
        _remove_final_layer_norm=False,
        word_embed_proj_dim=None,
        dropout=0.1,
        attention_dropout=0.0,
        num_attention_heads=12,
        activation_function="relu",
        layerdrop=0.0,
        init_std=0.02,
        alpha_initializer="ones",
        alphas_initializer_range=0.0,
        alpha_type="vector",
        use_cache=True,
        pad_token_id=1,
        bos_token_id=2,
        eos_token_id=2,
        cross_layer_interval=1,
        cross_layer_activation_function="swiglu",
        normformer_layer_norms=False,
        qk_layer_norms=False,
        rms_norm=False,
        qk_layer_norms_perceiver=False,
        tie_word_embeddings=False,
        freeze_text_layers=True,
        freeze_text_module_exceptions=[],
        freeze_lm_head=False,
        freeze_vision_layers=True,
        freeze_vision_module_exceptions=[],
        vision_model_name="google/vit-base-patch16-224",
        vision_model_params="{}",
        vision_embed_dim=768,
        vision_image_size=224,
        image_token_index=50257,  # TODO: change this to right value
        use_resampler=False,
        resampler_n_latents=64,
        resampler_depth=6,
        resampler_n_heads=16,
        resampler_head_dim=96,
        **kwargs,
    ):
        super().__init__(
            pad_token_id=pad_token_id,
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            tie_word_embeddings=tie_word_embeddings,
            **kwargs,
        )
        self.vocab_size = vocab_size
        self.additional_vocab_size = additional_vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.num_attention_heads = num_attention_heads
        self.word_embed_proj_dim = word_embed_proj_dim if word_embed_proj_dim is not None else hidden_size
        self.ffn_dim = ffn_dim
        self.hidden_size = hidden_size
        self.num_hidden_layers = num_hidden_layers
        self.dropout = dropout
        self.attention_dropout = attention_dropout
        self.activation_function = activation_function
        self.init_std = init_std
        self.alpha_initializer = alpha_initializer
        self.alphas_initializer_range = alphas_initializer_range
        self.alpha_type = alpha_type
        self.layerdrop = layerdrop
        self.use_cache = use_cache
        self.do_layer_norm_before = do_layer_norm_before

        # Note that the only purpose of `_remove_final_layer_norm` is to keep backward compatibility
        # with checkpoints that have been fine-tuned before transformers v4.20.1
        # see https://github.com/facebookresearch/metaseq/pull/164
        self._remove_final_layer_norm = _remove_final_layer_norm

        self.cross_layer_interval = cross_layer_interval
        self.cross_layer_activation_function = cross_layer_activation_function
        self.normformer_layer_norms = normformer_layer_norms
        self.qk_layer_norms = qk_layer_norms
        self.rms_norm = rms_norm
        self.qk_layer_norms_perceiver = qk_layer_norms_perceiver
        self.freeze_vision_layers = freeze_vision_layers
        self.vision_model_name = vision_model_name
        self.vision_model_params = vision_model_params

        self.tie_word_embeddings = tie_word_embeddings
        self.freeze_text_layers = freeze_text_layers
        self.freeze_text_module_exceptions = freeze_text_module_exceptions
        self.freeze_vision_module_exceptions = freeze_vision_module_exceptions
        self.freeze_lm_head = freeze_lm_head
        self.image_token_index = image_token_index

        self.vision_embed_dim = vision_embed_dim
        self.vision_image_size = vision_image_size

        # Resampler params
        self.use_resampler = use_resampler
        self.resampler_n_latents = resampler_n_latents
        self.resampler_depth = resampler_depth
        self.resampler_n_heads = resampler_n_heads
        self.resampler_head_dim = resampler_head_dim

        # IMPORTANT: Do not do any __init__ args-based checks in the constructor, since
        # PretrainedConfig.from_dict first instantiates the class with the config dict and only then
        # updates the config object with `kwargs` from from_pretrained, so during the instantiation
        # of this object many attributes have default values and haven't yet been overridden.
        # Do any required checks inside `from_pretrained` once the superclass' `from_pretrained` was run.

    def check_compatibilities(self):
        vision_model_params = eval(self.vision_model_params)
        config = AutoConfig.from_pretrained(self.vision_model_name, **vision_model_params)
        if hasattr(config, "vision_config"):
            vision_config = config.vision_config
        else:
            vision_config = config
        vision_embed_dim = vision_config.hidden_size
        if self.vision_embed_dim != vision_embed_dim:
            raise ValueError(
                f"vision_embed_dim ({self.vision_embed_dim}) must match the hidden size of the vision model"
                f" ({vision_embed_dim})"
            )
        vision_image_size = vision_config.image_size
        if self.vision_image_size != vision_image_size:
            raise ValueError(
                f"vision_image_size ({self.vision_image_size}) must match the hidden size of the vision model"
                f" ({vision_image_size})"
            )

    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
        outputs = super(VOPTConfig, cls).from_pretrained(pretrained_model_name_or_path, **kwargs)
        if isinstance(outputs, Tuple):
            # When called with return_unused_kwargs=True, the first item will be the config
            outputs[0].check_compatibilities()
        else:
            outputs.check_compatibilities()
        return outputs