configuration_diffusionvl_qwen2_5_vl.py

# coding=utf-8
# Copyright 2025 The HustVL Team and The HuggingFace Inc. team. All rights reserved.
#
# This code is based on Qwen2.5-VL, which is derived from EleutherAI's GPT-NeoX library
# and the GPT-NeoX and OPT implementations. It has been modified to create DiffusionVL.
#
# 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.

"""DiffusionVL (Qwen2.5-VL based) model configuration."""

from typing import List, Optional, Union

from transformers.configuration_utils import PretrainedConfig


class DiffusionVL_Qwen2_5_VL_VisionConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`DiffusionVL_Qwen2_5_VL_VisionModel`].
    It is used to instantiate the vision encoder according to the specified arguments.

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

    Args:
        depth (`int`, *optional*, defaults to 32):
            Number of vision transformer layers.
        hidden_size (`int`, *optional*, defaults to 1280):
            Dimensionality of the encoder layers and the pooler layer.
        hidden_act (`str`, *optional*, defaults to `"silu"`):
            The non-linear activation function in the encoder.
        intermediate_size (`int`, *optional*, defaults to 3420):
            Dimensionality of the "intermediate" (i.e., feed-forward) layer.
        num_heads (`int`, *optional*, defaults to 16):
            Number of attention heads for each attention layer.
        in_channels (`int`, *optional*, defaults to 3):
            Number of input channels.
        patch_size (`int`, *optional*, defaults to 14):
            The size of each image patch.
        spatial_merge_size (`int`, *optional*, defaults to 2):
            The spatial merge size for patch merging.
        temporal_patch_size (`int`, *optional*, defaults to 2):
            The temporal patch size for video processing.
        tokens_per_second (`int`, *optional*, defaults to 4):
            Number of tokens per second for video processing.
        window_size (`int`, *optional*, defaults to 112):
            Window size for windowed attention.
        out_hidden_size (`int`, *optional*, defaults to 3584):
            Output hidden size after the vision encoder.
        fullatt_block_indexes (`List[int]`, *optional*):
            Indices of blocks that use full attention instead of windowed attention.
        initializer_range (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing weight matrices.

    Example:

    ```python
    >>> from configuration_diffusionvl_qwen2_5_vl import DiffusionVL_Qwen2_5_VL_VisionConfig

    >>> # Initializing a DiffusionVL vision configuration
    >>> configuration = DiffusionVL_Qwen2_5_VL_VisionConfig()
    ```
    """

    model_type = "diffusionvl_qwen2_5_vl_vision"
    base_config_key = "vision_config"

    def __init__(
        self,
        depth: int = 32,
        hidden_size: int = 1280,
        hidden_act: str = "silu",
        intermediate_size: int = 3420,
        num_heads: int = 16,
        in_channels: int = 3,
        patch_size: int = 14,
        spatial_merge_size: int = 2,
        temporal_patch_size: int = 2,
        tokens_per_second: int = 4,
        window_size: int = 112,
        out_hidden_size: int = 3584,
        fullatt_block_indexes: Optional[List[int]] = None,
        initializer_range: float = 0.02,
        **kwargs,
    ):
        super().__init__(**kwargs)

        self.depth = depth
        self.hidden_size = hidden_size
        self.hidden_act = hidden_act
        self.intermediate_size = intermediate_size
        self.num_heads = num_heads
        self.in_channels = in_channels
        self.patch_size = patch_size
        self.spatial_merge_size = spatial_merge_size
        self.temporal_patch_size = temporal_patch_size
        self.tokens_per_second = tokens_per_second
        self.window_size = window_size
        self.out_hidden_size = out_hidden_size
        self.fullatt_block_indexes = fullatt_block_indexes or [7, 15, 23, 31]
        self.initializer_range = initializer_range


class DiffusionVL_Qwen2_5_VL_Config(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`DiffusionVL_Qwen2_5_VL_ForConditionalGeneration`].
    It is used to instantiate a DiffusionVL model according to the specified arguments.

    DiffusionVL extends Qwen2.5-VL architecture with BD3LM (Block Diffusion Language Model)
    for diffusion-based text generation instead of autoregressive decoding.

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

    Args:
        vocab_size (`int`, *optional*, defaults to 152064):
            Vocabulary size of the DiffusionVL model.
        hidden_size (`int`, *optional*, defaults to 3584):
            Dimension of the hidden representations.
        intermediate_size (`int`, *optional*, defaults to 18944):
            Dimension of the MLP representations.
        num_hidden_layers (`int`, *optional*, defaults to 28):
            Number of hidden layers in the Transformer encoder.
        num_attention_heads (`int`, *optional*, defaults to 28):
            Number of attention heads for each attention layer.
        num_key_value_heads (`int`, *optional*, defaults to 4):
            Number of key-value heads for Grouped Query Attention (GQA).
        hidden_act (`str`, *optional*, defaults to `"silu"`):
            The non-linear activation function in the decoder.
        max_position_embeddings (`int`, *optional*, defaults to 128000):
            The maximum sequence length that this model might ever be used with.
        initializer_range (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing weight matrices.
        rms_norm_eps (`float`, *optional*, defaults to 1e-6):
            The epsilon used by the RMS normalization layers.
        use_cache (`bool`, *optional*, defaults to `True`):
            Whether to use the past key/values attentions.
        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
            Whether the model's input and output word embeddings should be tied.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.
        vision_config (`DiffusionVL_Qwen2_5_VL_VisionConfig`, *optional*):
            The configuration for the vision encoder.
        image_token_id (`int`, *optional*, defaults to 151655):
            The token index for image placeholder.
        video_token_id (`int`, *optional*, defaults to 151656):
            The token index for video placeholder.
        vision_start_token_id (`int`, *optional*, defaults to 151652):
            The token index denoting start of vision input.
        vision_end_token_id (`int`, *optional*, defaults to 151653):
            The token index denoting end of vision input.
        enable_bd3lm (`bool`, *optional*, defaults to `True`):
            Whether to enable BD3LM diffusion-based generation.
        bd3lm_block_size (`int`, *optional*, defaults to 8):
            Block size for BD3LM generation.
        bd3lm_cross_attn (`bool`, *optional*, defaults to `True`):
            Whether to use cross-attention in BD3LM.
        mask_token_id (`int`, *optional*, defaults to 151671):
            The token index for mask token used in diffusion.
        rope_theta (`float`, *optional*, defaults to 1000000.0):
            The base period of the RoPE embeddings.
        rope_scaling (`Dict`, *optional*):
            Dictionary containing the scaling configuration for RoPE embeddings.

    Example:

    ```python
    >>> from transformers import AutoModelForCausalLM
    >>> from configuration_diffusionvl_qwen2_5_vl import DiffusionVL_Qwen2_5_VL_Config

    >>> # Initializing a DiffusionVL configuration
    >>> configuration = DiffusionVL_Qwen2_5_VL_Config()

    >>> # Initializing a model from the configuration
    >>> model = AutoModelForCausalLM.from_pretrained(
    ...     "path/to/model", config=configuration, trust_remote_code=True
    ... )

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

    model_type = "diffusionvl_qwenvl"
    sub_configs = {"vision_config": DiffusionVL_Qwen2_5_VL_VisionConfig}
    keys_to_ignore_at_inference = ["past_key_values"]

    def __init__(
        self,
        vocab_size: int = 152064,
        hidden_size: int = 3584,
        intermediate_size: int = 18944,
        num_hidden_layers: int = 28,
        num_attention_heads: int = 28,
        num_key_value_heads: int = 4,
        hidden_act: str = "silu",
        max_position_embeddings: int = 128000,
        initializer_range: float = 0.02,
        rms_norm_eps: float = 1e-6,
        use_cache: bool = True,
        tie_word_embeddings: bool = False,
        attention_dropout: float = 0.0,
        # Vision configuration
        vision_config: Optional[Union[DiffusionVL_Qwen2_5_VL_VisionConfig, dict]] = None,
        # Multimodal token IDs
        image_token_id: int = 151655,
        video_token_id: int = 151656,
        vision_start_token_id: int = 151652,
        vision_end_token_id: int = 151653,
        # BD3LM diffusion parameters
        enable_bd3lm: bool = True,
        bd3lm_block_size: int = 8,
        bd3lm_cross_attn: bool = True,
        bd3lm_antithetic_sampling: bool = True,
        bd3lm_sampling_eps_min: float = 1e-3,
        bd3lm_sampling_eps_max: float = 1.0,
        mask_token_id: int = 151671,
        # RoPE parameters
        rope_theta: float = 1000000.0,
        rope_scaling: Optional[dict] = None,
        **kwargs,
    ):
        # Remove text_config from kwargs to avoid GenerationConfig issues
        # (text_config is only needed for train code, HF config uses flattened params)
        kwargs.pop("text_config", None)

        # Text model configuration
        self.vocab_size = vocab_size
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.num_key_value_heads = num_key_value_heads
        self.hidden_act = hidden_act
        self.max_position_embeddings = max_position_embeddings
        self.initializer_range = initializer_range
        self.rms_norm_eps = rms_norm_eps
        self.use_cache = use_cache
        self.attention_dropout = attention_dropout
        self.rope_theta = rope_theta
        self.rope_scaling = rope_scaling or {
            "mrope_section": [16, 24, 24],
            "rope_type": "default",
            "type": "default",
        }

        # Vision configuration
        if vision_config is None:
            self.vision_config = DiffusionVL_Qwen2_5_VL_VisionConfig()
        elif isinstance(vision_config, dict):
            self.vision_config = DiffusionVL_Qwen2_5_VL_VisionConfig(**vision_config)
        elif isinstance(vision_config, DiffusionVL_Qwen2_5_VL_VisionConfig):
            self.vision_config = vision_config
        else:
            self.vision_config = DiffusionVL_Qwen2_5_VL_VisionConfig()

        # Multimodal token IDs
        self.image_token_id = image_token_id
        self.video_token_id = video_token_id
        self.vision_start_token_id = vision_start_token_id
        self.vision_end_token_id = vision_end_token_id

        # BD3LM diffusion configuration
        self.enable_bd3lm = enable_bd3lm
        self.bd3lm_block_size = bd3lm_block_size
        self.bd3lm_cross_attn = bd3lm_cross_attn
        self.bd3lm_antithetic_sampling = bd3lm_antithetic_sampling
        self.bd3lm_sampling_eps_min = bd3lm_sampling_eps_min
        self.bd3lm_sampling_eps_max = bd3lm_sampling_eps_max
        self.mask_token_id = mask_token_id

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


__all__ = ["DiffusionVL_Qwen2_5_VL_Config", "DiffusionVL_Qwen2_5_VL_VisionConfig"]