[Model] Support Dots OCR (#24645)

Signed-off-by: Roger Wang <hey@rogerw.io>
Co-authored-by: yinz-aizip <yinz@aizip.ai>

docs/models/supported_models.md

@@ -352,6 +352,7 @@ th {
 | `DeepseekV2ForCausalLM` | DeepSeek-V2 | `deepseek-ai/DeepSeek-V2`, `deepseek-ai/DeepSeek-V2-Chat`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `DeepseekV3ForCausalLM` | DeepSeek-V3 | `deepseek-ai/DeepSeek-V3`, `deepseek-ai/DeepSeek-R1`, `deepseek-ai/DeepSeek-V3.1`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `Dots1ForCausalLM` | dots.llm1 | `rednote-hilab/dots.llm1.base`, `rednote-hilab/dots.llm1.inst`, etc. | | ✅︎ | ✅︎ |
+| `DotsOCRForCausalLM` | dots_ocr | `rednote-hilab/dots.ocr` | | ✅︎ | ✅︎ |
 | `Ernie4_5ForCausalLM` | Ernie4.5 | `baidu/ERNIE-4.5-0.3B-PT`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `Ernie4_5_MoeForCausalLM` | Ernie4.5MoE | `baidu/ERNIE-4.5-21B-A3B-PT`, `baidu/ERNIE-4.5-300B-A47B-PT`, etc. |✅︎| ✅︎ | ✅︎ |
 | `ExaoneForCausalLM` | EXAONE-3 | `LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct`, etc. | ✅︎ | ✅︎ | ✅︎ |

examples/offline_inference/vision_language.py

@@ -126,6 +126,23 @@ def run_chameleon(questions: list[str], modality: str) -> ModelRequestData:
     )
 
 
+# Dots-OCR
+def run_dots_ocr(questions: list[str], modality: str) -> ModelRequestData:
+    assert modality == "image"
+
+    prompts = [f"<|img|><|imgpad|><|endofimg|>{question}" for question in questions]
+    engine_args = EngineArgs(
+        model="rednote-hilab/dots.ocr",
+        limit_mm_per_prompt={modality: 1},
+        trust_remote_code=True,
+    )
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompts=prompts,
+    )
+
+
 def run_command_a_vision(questions: list[str], modality: str) -> ModelRequestData:
     assert modality == "image"
 
@@ -1676,6 +1693,7 @@ model_example_map = {
     "aya_vision": run_aya_vision,
     "blip-2": run_blip2,
     "chameleon": run_chameleon,
+    "dots_ocr": run_dots_ocr,
     "command_a_vision": run_command_a_vision,
     "deepseek_vl_v2": run_deepseek_vl2,
     "ernie45_vl": run_ernie45_vl,

tests/models/registry.py

@@ -448,6 +448,8 @@ _MULTIMODAL_EXAMPLE_MODELS = {
                                                 max_transformers_version="4.48",  # noqa: E501
                                                 transformers_version_reason="HF model is not compatible.",  # noqa: E501
                                                 hf_overrides={"architectures": ["DeepseekVLV2ForCausalLM"]}),  # noqa: E501
+    "DotsOCRForCausalLM": _HfExamplesInfo("rednote-hilab/dots.ocr",
+                                          trust_remote_code=True),
     "Emu3ForConditionalGeneration": _HfExamplesInfo("BAAI/Emu3-Chat-hf"),
     "Ernie4_5_VLMoeForConditionalGeneration": _HfExamplesInfo("baidu/ERNIE-4.5-VL-28B-A3B-PT",  # noqa: E501
                                                               trust_remote_code=True),

vllm/model_executor/models/registry.py

@@ -219,6 +219,7 @@ _MULTIMODAL_MODELS = {
     "ChameleonForConditionalGeneration": ("chameleon", "ChameleonForConditionalGeneration"),  # noqa: E501
     "Cohere2VisionForConditionalGeneration": ("cohere2_vision", "Cohere2VisionForConditionalGeneration"),  # noqa: E501
     "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
+    "DotsOCRForCausalLM": ("dots_ocr", "DotsOCRForCausalLM"),
     "Ernie4_5_VLMoeForConditionalGeneration": ("ernie45_vl", "Ernie4_5_VLMoeForConditionalGeneration"),  # noqa: E501
     "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
     "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501

vllm/transformers_utils/configs/__init__.py

@@ -9,6 +9,7 @@ Model configs may be defined in this directory for the following reasons:
 
 from vllm.transformers_utils.configs.chatglm import ChatGLMConfig
 from vllm.transformers_utils.configs.deepseek_vl2 import DeepseekVLV2Config
+from vllm.transformers_utils.configs.dotsocr import DotsOCRConfig
 from vllm.transformers_utils.configs.eagle import EAGLEConfig
 # RWConfig is for the original tiiuae/falcon-40b(-instruct) and
 # tiiuae/falcon-7b(-instruct) models. Newer Falcon models will use the
@@ -36,6 +37,7 @@ from vllm.transformers_utils.configs.ultravox import UltravoxConfig
 __all__ = [
     "ChatGLMConfig",
     "DeepseekVLV2Config",
+    "DotsOCRConfig",
     "EAGLEConfig",
     "RWConfig",
     "JAISConfig",

vllm/transformers_utils/configs/dotsocr.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from typing import Any, Optional
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.models.qwen2 import Qwen2Config
+
+
+class DotsVisionConfig(PretrainedConfig):
+    model_type: str = "dots_vit"
+
+    def __init__(
+        self,
+        embed_dim: int = 1536,  # vision encoder embed size
+        hidden_size: int = 1536,  # after merger hidden size
+        intermediate_size: int = 4224,
+        num_hidden_layers: int = 42,
+        num_attention_heads: int = 12,
+        num_channels: int = 3,
+        patch_size: int = 14,
+        spatial_merge_size: int = 2,
+        temporal_patch_size: int = 1,
+        rms_norm_eps: float = 1e-5,
+        use_bias: bool = False,
+        attn_implementation="flash_attention_2",
+        initializer_range=0.02,
+        init_merger_std=0.02,
+        is_causal=False,  # ve causal forward
+        post_norm=True,
+        gradient_checkpointing=False,
+        **kwargs: Any,
+    ):
+        super().__init__(**kwargs)
+        self.embed_dim = embed_dim
+        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_channels = num_channels
+        self.patch_size = patch_size
+        self.spatial_merge_size = spatial_merge_size
+        self.temporal_patch_size = temporal_patch_size
+        self.rms_norm_eps = rms_norm_eps
+        self.use_bias = use_bias
+        self.attn_implementation = attn_implementation
+        self.initializer_range = initializer_range
+        self.init_merger_std = init_merger_std
+        self.is_causal = is_causal
+        self.post_norm = post_norm
+        self.gradient_checkpointing = gradient_checkpointing
+
+
+class DotsOCRConfig(Qwen2Config):
+    model_type = "dots_ocr"
+
+    def __init__(self,
+                 image_token_id=151665,
+                 video_token_id=151656,
+                 vision_config: Optional[dict] = None,
+                 *args,
+                 **kwargs):
+        super().__init__(*args, **kwargs)
+        self.image_token_id = image_token_id
+        self.video_token_id = video_token_id
+        self.vision_config = DotsVisionConfig(**(vision_config or {}))
+
+    def save_pretrained(self, save_directory, **kwargs):
+        self._auto_class = None
+        super().save_pretrained(save_directory, **kwargs)