| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Configuration for ColModernVBERT visual document retrieval model. |
| 4 | |
| 5 | ColModernVBERT combines SigLIP vision encoder + ModernBERT text encoder |
| 6 | with a pixel shuffle connector and ColBERT-style 128-dim per-token embeddings. |
| 7 | |
| 8 | Reference: https://huggingface.co/ModernVBERT/colmodernvbert-merged |
| 9 | """ |
| 10 | |
| 11 | from transformers import ModernBertConfig, PreTrainedConfig, SiglipVisionConfig |
| 12 | |
| 13 | |
| 14 | class ColModernVBertConfig(PreTrainedConfig): |
| 15 | model_type = "colmodernvbert" |
| 16 | |
| 17 | def __init__( |
| 18 | self, |
| 19 | embedding_dim: int = 128, |
| 20 | image_token_id: int = 50407, |
| 21 | pixel_shuffle_factor: int = 4, |
| 22 | text_config: dict | None = None, |
| 23 | vision_config: dict | None = None, |
| 24 | **kwargs, |
| 25 | ): |
| 26 | self.embedding_dim = embedding_dim |
| 27 | self.image_token_id = image_token_id |
| 28 | self.pixel_shuffle_factor = pixel_shuffle_factor |
| 29 | |
| 30 | text_config = text_config or {} |
| 31 | self.hidden_size = text_config.get("hidden_size", 768) |
| 32 | |
| 33 | self.text_config = ModernBertConfig( |
| 34 | vocab_size=text_config.get("vocab_size", 50408), |
| 35 | hidden_size=text_config.get("hidden_size", 768), |
| 36 | intermediate_size=text_config.get("intermediate_size", 1152), |
| 37 | num_hidden_layers=text_config.get("num_hidden_layers", 22), |
| 38 | num_attention_heads=text_config.get("num_attention_heads", 12), |
| 39 | mlp_bias=text_config.get("mlp_bias", False), |
| 40 | max_position_embeddings=text_config.get("max_position_embeddings", 8192), |
| 41 | ) |
| 42 | |
| 43 | vision_config = vision_config or {} |
| 44 | self.vision_config = SiglipVisionConfig( |
| 45 | hidden_size=vision_config.get("hidden_size", 768), |
| 46 | image_size=vision_config.get("image_size", 512), |
| 47 | patch_size=vision_config.get("patch_size", 16), |
| 48 | num_hidden_layers=vision_config.get("num_hidden_layers", 12), |
| 49 | intermediate_size=vision_config.get("intermediate_size", 3072), |
| 50 | num_attention_heads=vision_config.get("num_attention_heads", 12), |
| 51 | ) |
| 52 | |
| 53 | # Ensure architectures is set so vLLM routes to our model class |
| 54 | kwargs.setdefault("architectures", ["ColModernVBertForRetrieval"]) |
| 55 | super().__init__(**kwargs) |
| 56 | |
| 57 | @property |
| 58 | def image_seq_len(self) -> int: |
| 59 | ps = self.vision_config.image_size // self.vision_config.patch_size |
| 60 | return (ps * ps) // (self.pixel_shuffle_factor**2) |
| 61 | |
| 62 | def get_text_config(self, **kwargs): |
| 63 | return self.text_config |