configuration_moss_vl.py

# coding=utf-8
# Copyright 2025 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.
"""MossVL model configuration"""

from transformers.configuration_utils import PretrainedConfig
from transformers.modeling_rope_utils import rope_config_validation
from transformers.utils import logging

logger = logging.get_logger(__name__)


class MossVLVisionConfig(PretrainedConfig):
    """
    Configuration for MossVL Vision Model
    """

    model_type = "moss_vl_vision"
    base_config_key = "vision_config"

    def __init__(
        self,
        depth=27,
        hidden_size=1152,
        hidden_act="gelu_pytorch_tanh",
        intermediate_size=4304,
        num_heads=16,
        in_channels=3,
        patch_size=16,
        spatial_merge_size=2,
        temporal_patch_size=1,
        out_hidden_size=3584,
        num_position_embeddings=2304,
        deepstack_visual_indexes=[8, 16, 24],
        initializer_range=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.out_hidden_size = out_hidden_size
        self.num_position_embeddings = num_position_embeddings
        self.initializer_range = initializer_range
        self.deepstack_visual_indexes = deepstack_visual_indexes


class MossVLTextConfig(PretrainedConfig):
    """
    Configuration for MossVL Text Model
    """

    model_type = "moss_vl_text"
    base_config_key = "text_config"
    default_theta = 5000000.0
    ignore_keys_at_rope_validation = {"mrope_section", "mrope_interleaved"}

    def __init__(
        self,
        vocab_size=151936,
        hidden_size=4096,
        intermediate_size=22016,
        num_hidden_layers=32,
        num_attention_heads=32,
        num_key_value_heads=32,
        head_dim=128,
        hidden_act="silu",
        max_position_embeddings=128000,
        initializer_range=0.02,
        rms_norm_eps=1e-6,
        use_cache=True,
        tie_word_embeddings=False,
        rope_theta=5000000.0,
        rope_parameters=None,
        rope_scaling=None,
        attention_bias=False,
        attention_dropout=0.0,
        pad_token_id=None,
        # Cross attention specific
        cross_attention_layers=None,  # List of layer indices to insert cross attention
        **kwargs,
    ):

        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads

        # for backward compatibility
        if num_key_value_heads is None:
            num_key_value_heads = num_attention_heads

        self.num_key_value_heads = num_key_value_heads
        self.head_dim = head_dim
        self.hidden_act = hidden_act
        self.initializer_range = initializer_range
        self.rms_norm_eps = rms_norm_eps
        self.use_cache = use_cache
        self.rope_theta = rope_theta
        if rope_parameters is None:
            if rope_scaling is not None:
                rope_parameters = dict(rope_scaling)
            else:
                rope_parameters = {"rope_type": "default"}
        else:
            rope_parameters = dict(rope_parameters)

        if "type" in rope_parameters and "rope_type" not in rope_parameters:
            rope_parameters["rope_type"] = rope_parameters.pop("type")
        rope_parameters.setdefault("rope_type", "default")
        rope_parameters.setdefault("rope_theta", rope_theta)

        self.rope_parameters = rope_parameters
        self.rope_scaling = rope_scaling if rope_scaling is not None else dict(rope_parameters)
        self.attention_bias = attention_bias
        self.attention_dropout = attention_dropout
        self.pad_token_id = pad_token_id

        if hasattr(self, "standardize_rope_params"):
            self.standardize_rope_params()
        if hasattr(self, "validate_rope"):
            self.validate_rope()
        else:
            rope_config_validation(self, ignore_keys=self.ignore_keys_at_rope_validation)
        self.cross_attention_layers = cross_attention_layers or [2, 6, 10, 14, 18, 22, 26, 30, 34, 38, 42, 46]
        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)

class MossVLConfig(PretrainedConfig):
    """
    Configuration for MossVL Model
    """

    model_type = "moss_vl"
    sub_configs = {"vision_config": MossVLVisionConfig, "text_config": MossVLTextConfig}
    keys_to_ignore_at_inference = ["past_key_values"]

    def __init__(
        self,
        text_config=None,
        vision_config=None,
        image_token_id=151655,
        video_token_id=151656,
        vision_start_token_id=151652,
        vision_end_token_id=151653,
        vision_seq_pad_multiple=8,
        tie_word_embeddings=False,
        **kwargs,
    ):
        if isinstance(vision_config, dict):
            self.vision_config = self.sub_configs["vision_config"](**vision_config)
        elif vision_config is None:
            self.vision_config = self.sub_configs["vision_config"]()

        if isinstance(text_config, dict):
            self.text_config = self.sub_configs["text_config"](**text_config)
        elif text_config is None:
            self.text_config = self.sub_configs["text_config"]()

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

__all__ = ["MossVLConfig", "MossVLTextConfig"]