[model] support kimi_k25 (#52)

Author: Jintao-Huang

README.md

@@ -85,7 +85,7 @@ The following is the list of models supported by MCore-Bridge:
 | DeepSeek | deepseek_v3, deepseek_v32                                    |
 | GLM      | glm4, glm4_moe, glm4_moe_lite<br />glm4v, glm4v_moe, <br />glm_moe_dsa |
 | MiniMax  | minimax_m2                                                   |
-| Kimi     | kimi_k2, kimi_vl                                             |
+| Kimi     | kimi_k2, kimi_vl, kimi_k25                                   |
 | InternLM | internlm3, internvl_chat, internvl                           |
 | Ovis     | ovis2_5                                                      |
 | Llama    | llama, llama4                                                |

README_zh.md

@@ -84,7 +84,7 @@ uv pip install -e . --torch-backend=auto
 | DeepSeek | deepseek_v3, deepseek_v32                                    |
 | GLM      | glm4, glm4_moe, glm4_moe_lite<br />glm4v, glm4v_moe, <br />glm_moe_dsa |
 | MiniMax  | minimax_m2                                                   |
-| Kimi     | kimi_k2, kimi_vl                                             |
+| Kimi     | kimi_k2, kimi_vl, kimi_k25                                   |
 | InternLM | internlm3, internvl_chat, internvl                           |
 | Ovis     | ovis2_5                                                      |
 | Llama    | llama, llama4                                                |

src/mcore_bridge/config/model_config.py

@@ -321,6 +321,7 @@ def __post_init__(self):
             self.mcore_model_type = get_mcore_model_type(self.hf_model_type)
         self.model_meta = get_model_meta(self.mcore_model_type)
         self.is_multimodal = self.model_meta.visual_cls is not None
+        self.support_multimodal = self.is_multimodal and getattr(self.model_meta.visual_cls, 'support_multimodal', True)
         if self.is_multimodal and self.hf_config is None:
             raise ValueError('Multimodal model must specify hf_config.')
         self.is_moe_model = self.num_moe_experts is not None

src/mcore_bridge/model/constant.py

@@ -27,6 +27,8 @@ class MLLMModelType:
     kimi_vl = 'kimi_vl'
     llama4 = 'llama4'
 
+    kimi_k25 = 'kimi_k25'
+
 
 class ModelType(LLMModelType, MLLMModelType):
     pass

src/mcore_bridge/model/mm_gpts/kimi_vl.py

@@ -66,3 +66,35 @@ def _merge_with_image_features(self, inputs_embeds, input_ids, image_features):
     bridge_cls=KimiVLBridge,
     visual_cls=KimiVLVit,
 ))
+
+
+class KimiK25Vit(HuggingFaceVit):
+    module_mapping = {'vision_tower': 'vision_tower', 'mm_projector': 'mm_projector'}
+    _vision_tower = ['vision_tower']
+    _aligner = ['mm_projector']
+    support_multimodal = False
+
+    def prepare_model(self, hf_config: PretrainedConfig):
+        output = []
+        for key in ['MoonViT3dPretrainedModel', 'PatchMergerMLP', 'VisionTowerConfig', 'ProjectorConfig']:
+            output.append(get_class_from_dynamic_module(f'modeling_kimi_k25.{key}', hf_config.name_or_path))
+        MoonViT3dPretrainedModel, PatchMergerMLP, VisionTowerConfig, ProjectorConfig = output
+        assert hf_config.vision_config.mm_projector_type == 'patchmerger'
+        vit_config = VisionTowerConfig(hf_config.vision_config)
+        proj_config = ProjectorConfig(hf_config.vision_config)
+        self.vision_tower = MoonViT3dPretrainedModel._from_config(vit_config)
+        self.mm_projector = PatchMergerMLP(proj_config).to(self.vision_tower.dtype)
+
+    def get_inputs_embeds(self, inputs_embeds, **kwargs):
+        pixel_values = kwargs.pop('pixel_values', None)
+        if pixel_values is not None:
+            raise NotImplementedError('Kimi-K25 currently only supports plain text training.')
+        return inputs_embeds
+
+
+register_model(ModelMeta(
+    ModelType.kimi_k25,
+    ['kimi_k25'],
+    bridge_cls=KimiVLBridge,
+    visual_cls=KimiK25Vit,
+))

src/mcore_bridge/model/mm_gpts/utils.py

@@ -26,6 +26,7 @@ def new_get_cached_module_file(pretrained_model_name_or_path, *args, **kwargs):
 
 class HuggingFaceVit(_HuggingFaceModule, ABC):
     module_mapping = {}  # hf -> mcore
+    support_multimodal = True
 
     @contextmanager
     def patch_hf_config(self):