m4/models/vbloom/configuration_vbloom.py

# coding=utf-8
# Copyright 2022 the Big Science Workshop and 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.
""" VBloom 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__)

BLOOM_PRETRAINED_CONFIG_ARCHIVE_MAP = {
    "bigscience/bloom": "https://huggingface.co/bigscience/bloom/resolve/main/config.json",
    "bigscience/bloom-560m": "https://huggingface.co/bigscience/bloom-560m/blob/main/config.json",
    "bigscience/bloom-1b1": "https://huggingface.co/bigscience/bloom-1b1/blob/main/config.json",
    "bigscience/bloom-1b7": "https://huggingface.co/bigscience/bloom-1b7/blob/main/config.json",
    "bigscience/bloom-3b": "https://huggingface.co/bigscience/bloom-3b/blob/main/config.json",
    "bigscience/bloom-7b1": "https://huggingface.co/bigscience/bloom-7b1/blob/main/config.json",
}


class VBloomConfig(PretrainedConfig):
    """
    This is the configuration class to store the configuration of a [`BloomModel`]. It is used to instantiate a Bloom
    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
    defaults will yield a similar configuration to the Bloom architecture
    [bigscience/bloom](https://huggingface.co/bigscience/bloom).

    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 50257):
            Vocabulary size of the Bloom model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`BloomModel`].
        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 embeddings and hidden states.
        n_layer (`int`, *optional*, defaults to 12):
            Number of hidden layers in the Transformer encoder.
        n_head (`int`, *optional*, defaults to 12):
            Number of attention heads for each attention layer in the Transformer encoder.
        attn_pdrop (`float`, *optional*, defaults to 0.1):
            The dropout ratio for the attention.
        layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):
            The epsilon to use in the layer normalization layers.
        initializer_range (`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.
        apply_residual_connection_post_layernorm (`bool`, *optional*, defaults to `False`):
            If enabled, use the layer norm of the hidden states as the residual in the transformer blocks
        skip_bias_add (`bool`, *optional*, defaults to `True`):
            If set to `True`, it will skip bias add for each linear layer in the transformer blocks
        skip_bias_add_qkv (`bool`, *optional*, defaults to `False`):
            If set to `True`, it will skip bias add for the first linear layer in the transformer blocks
        hidden_dropout (`float`, *optional*, defaults to 0.1):
            Dropout rate of the dropout function on the bias dropout.
        attention_dropout (`float`, *optional*, defaults to 0.1):
            Dropout rate applied to the attention probs
        use_cache (`bool`, *optional*, defaults to `True`):
            Whether or not the model should return the last key/values attentions (not used by all models).
        pretraining_tp (`int`, *optional*, defaults to `1`):
            Experimental feature. Tensor parallelism rank used during pretraining with Megatron. Please refer to [this
            document](https://huggingface.co/docs/transformers/parallelism) to understand more about it. This value is
            necessary to ensure exact reproducibility of the pretraining results. Please refer to [this
            issue](https://github.com/pytorch/pytorch/issues/76232). Note also that this is enabled only when
            `slow_but_exact=True`.
        slow_but_exact (`bool`, *optional*, defaults to `False`):
            Experimental feature. Whether to use slow but exact implementation of the attention mechanism. While
            merging the TP rank tensors, due to slicing operations the results may be slightly different between the
            model trained on Megatron and our model. Please refer to [this
            issue](https://github.com/pytorch/pytorch/issues/76232). A solution to obtain more accurate results is to
            enable this feature. Enabling this will hurt the computational time of the inference. Will be probably
            resolved in the future once the main model has been fine-tuned with TP_rank=1.

    Example:

    ```python
    >>> from transformers import BloomModel, BloomConfig

    >>> # Initializing a Bloom configuration
    >>> configuration = BloomConfig()

    >>> # Initializing a model from the configuration
    >>> model = BloomModel(configuration)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

    model_type = "vbloom"
    keys_to_ignore_at_inference = ["past_key_values"]
    attribute_map = {
        "num_hidden_layers": "n_layer",
        "num_attention_heads": "n_head",
    }

    def __init__(
        self,
        vocab_size=250880,
        additional_vocab_size=0,
        hidden_size=64,
        n_layer=2,
        n_head=8,
        layer_norm_epsilon=1e-5,
        initializer_range=0.02,
        alpha_initializer="ones",
        alphas_initializer_range=0.0,
        alpha_type="vector",
        use_cache=False,
        bos_token_id=1,
        eos_token_id=2,
        apply_residual_connection_post_layernorm=False,
        hidden_dropout=0.0,
        attention_dropout=0.0,
        pretraining_tp=1,  # TP rank used when training with megatron
        slow_but_exact=False,
        cross_layer_interval=1,
        tie_word_embeddings=False,
        freeze_text_layers=True,
        freeze_lm_head=False,
        freeze_vision_layers=True,
        vision_model_name="google/vit-base-patch16-224",
        vision_model_params="{}",
        vision_embed_dim=768,
        image_token_index=250880,
        use_resampler=False,
        resampler_n_latents=64,
        resampler_depth=6,
        resampler_n_heads=16,
        resampler_head_dim=96,
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.additional_vocab_size = additional_vocab_size
        # Backward compatibility with n_embed kwarg
        n_embed = kwargs.pop("n_embed", None)
        self.hidden_size = hidden_size if n_embed is None else n_embed
        self.n_layer = n_layer
        self.n_head = n_head
        self.layer_norm_epsilon = layer_norm_epsilon
        self.initializer_range = initializer_range
        self.alpha_initializer = alpha_initializer
        self.alphas_initializer_range = alphas_initializer_range
        self.alpha_type = alpha_type
        self.use_cache = use_cache
        self.pretraining_tp = pretraining_tp
        self.apply_residual_connection_post_layernorm = apply_residual_connection_post_layernorm
        self.hidden_dropout = hidden_dropout
        self.attention_dropout = attention_dropout

        self.bos_token_id = bos_token_id
        self.eos_token_id = eos_token_id
        self.slow_but_exact = slow_but_exact

        self.cross_layer_interval = cross_layer_interval
        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_lm_head = freeze_lm_head
        self.image_token_index = image_token_index

        self.vision_embed_dim = vision_embed_dim

        # 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.

        super().__init__(
            bos_token_id=bos_token_id, eos_token_id=eos_token_id, tie_word_embeddings=tie_word_embeddings, **kwargs
        )

    def check_compatibilities(self):
        if self.tie_word_embeddings and (self.freeze_text_layers != self.freeze_lm_head):
            raise ValueError(
                "if `tie_word_embeddings` is True, then `freeze_lm_head` and `freeze_text_layers` must be equal."
            )

        vision_model_params = eval(self.vision_model_params)
        config = AutoConfig.from_pretrained(self.vision_model_name, **vision_model_params)
        if hasattr(config, "vision_config"):
            vison_config = config.vision_config
        else:
            vison_config = config
        vision_embed_dim = vison_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})"
            )

    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
        outputs = super(VBloomConfig, 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