[Models] Kimi-K2.5 (#33131)

Signed-off-by: wanglinian <wanglinian@stu.pku.edu.cn>
Signed-off-by: wangln19 <96399074+wangln19@users.noreply.github.com>
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Signed-off-by: youkaichao <youkaichao@gmail.com>
Signed-off-by: Roger Wang <hey@rogerw.io>
Co-authored-by: wanglinian <wanglinian@stu.pku.edu.cn>
Co-authored-by: wangln19 <96399074+wangln19@users.noreply.github.com>
Co-authored-by: Zaida Zhou <58739961+zhouzaida@users.noreply.github.com>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: youkaichao <youkaichao@gmail.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
(cherry picked from commit b539f988)

docs/models/supported_models.md

@@ -686,6 +686,7 @@ These models primarily accept the [`LLM.generate`](./generative_models.md#llmgen
 | `KeyeForConditionalGeneration` | Keye-VL-8B-Preview | T + I<sup>E+</sup> + V<sup>E+</sup> | `Kwai-Keye/Keye-VL-8B-Preview` | ✅︎ | ✅︎ |
 | `KeyeVL1_5ForConditionalGeneration` | Keye-VL-1_5-8B | T + I<sup>E+</sup> + V<sup>E+</sup> | `Kwai-Keye/Keye-VL-1_5-8B` | ✅︎ | ✅︎ |
 | `KimiVLForConditionalGeneration` | Kimi-VL-A3B-Instruct, Kimi-VL-A3B-Thinking | T + I<sup>+</sup> | `moonshotai/Kimi-VL-A3B-Instruct`, `moonshotai/Kimi-VL-A3B-Thinking` | | ✅︎ |
+| `KimiK25ForConditionalGeneration` | Kimi-K2.5 | T + I<sup>+</sup> | `moonshotai/Kimi-K2.5` | | ✅︎ |
 | `LightOnOCRForConditionalGeneration`  | LightOnOCR-1B  | T + I<sup>+</sup> | `lightonai/LightOnOCR-1B`, etc | ✅︎ | ✅︎ |
 | `Lfm2VlForConditionalGeneration` | LFM2-VL | T + I<sup>+</sup> | `LiquidAI/LFM2-VL-450M`, `LiquidAI/LFM2-VL-3B`, `LiquidAI/LFM2-VL-8B-A1B`, etc. | ✅︎ | ✅︎ |
 | `Llama4ForConditionalGeneration` | Llama 4 | T + I<sup>+</sup> | `meta-llama/Llama-4-Scout-17B-16E-Instruct`, `meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8`, `meta-llama/Llama-4-Maverick-17B-128E-Instruct`, etc. | ✅︎ | ✅︎ |

tests/models/registry.py

@@ -771,6 +771,11 @@ _MULTIMODAL_EXAMPLE_MODELS = {
             )
         },
     ),
+    "KimiK25ForConditionalGeneration": _HfExamplesInfo(
+        "moonshotai/Kimi-K2.5",
+        trust_remote_code=True,
+        is_available_online=False,
+    ),
     "LightOnOCRForConditionalGeneration": _HfExamplesInfo(
         "lightonai/LightOnOCR-1B-1025"
     ),

vllm/entrypoints/chat_utils.py

@@ -46,6 +46,9 @@ from vllm.multimodal.inputs import (
     MultiModalBatchedField,
     MultiModalFlatField,
     MultiModalSharedField,
+    VisionChunk,
+    VisionChunkImage,
+    VisionChunkVideo,
 )
 from vllm.multimodal.processing import BaseMultiModalProcessor
 from vllm.multimodal.utils import MEDIA_CONNECTOR_REGISTRY, MediaConnector
@@ -336,7 +339,9 @@ ChatTemplateContentFormatOption = Literal["auto", "string", "openai"]
 ChatTemplateContentFormat = Literal["string", "openai"]
 
 
-ModalityStr = Literal["image", "audio", "video", "image_embeds", "audio_embeds"]
+ModalityStr = Literal[
+    "image", "audio", "video", "image_embeds", "audio_embeds", "vision_chunk"
+]
 _T = TypeVar("_T")
 
 
@@ -449,6 +454,78 @@ def _get_embeds_data(
     raise NotImplementedError(type(data_items))
 
 
+def rebuild_mm_uuids_from_mm_data(
+    mm_uuids: MultiModalUUIDDict,
+    mm_data: MultiModalDataDict,
+) -> MultiModalUUIDDict:
+    """Rebuild mm_uuids after vision_chunk processing.
+
+    When videos are split into chunks, the original UUIDs need to be updated
+    to reflect the new UUIDs generated for each chunk.
+
+    Args:
+        mm_uuids: Original UUIDs dictionary
+        mm_data: Processed multimodal data with vision_chunk items
+
+    Returns:
+        Updated UUIDs dictionary with chunk UUIDs
+    """
+    vision_chunks = mm_data.get("vision_chunk")
+    if vision_chunks is None:
+        return mm_uuids
+
+    new_uuids = dict(mm_uuids)
+    vision_chunk_uuids = []
+
+    for item in vision_chunks:
+        # vision_chunk items are always dicts (VisionChunkImage/VisionChunkVideo)
+        assert isinstance(item, dict)
+        uuid_val = item.get("uuid")
+        if uuid_val is not None:
+            vision_chunk_uuids.append(uuid_val)
+
+    if vision_chunk_uuids:
+        new_uuids["vision_chunk"] = vision_chunk_uuids
+
+    return new_uuids
+
+
+def build_video_prompts_from_mm_data(
+    mm_data: MultiModalDataDict,
+) -> list[str]:
+    """Build video prompts from vision_chunk data.
+
+    Collects prompts from video chunks and groups them by video_idx.
+
+    Args:
+        mm_data: Processed multimodal data with vision_chunk items
+
+    Returns:
+        List of video prompts, one per video.
+    """
+    vision_chunks = mm_data.get("vision_chunk")
+    if vision_chunks is None:
+        return []
+
+    # Group chunks by video_idx
+    video_prompts_dict: dict[int, list[str]] = defaultdict(list)
+
+    for item in vision_chunks:
+        # vision_chunk items are always dicts (VisionChunkImage/VisionChunkVideo)
+        assert isinstance(item, dict)
+        if item.get("type") == "video_chunk":
+            video_idx = item.get("video_idx", 0)
+            prompt = item.get("prompt", "")
+            video_prompts_dict[video_idx].append(prompt)
+
+    # Build prompts in video order
+    video_prompts = []
+    for video_idx in sorted(video_prompts_dict.keys()):
+        video_prompts.append("".join(video_prompts_dict[video_idx]))
+
+    return video_prompts
+
+
 class BaseMultiModalItemTracker(ABC, Generic[_T]):
     """
     Tracks multi-modal items in a given request and ensures that the number
@@ -462,6 +539,13 @@ class BaseMultiModalItemTracker(ABC, Generic[_T]):
         self._model_config = model_config
 
         self._items_by_modality = defaultdict[str, list[_T]](list)
+        # Track original modality for each vision_chunk item (image or video)
+        self._modality_order = defaultdict[str, list[str]](list)
+
+    @cached_property
+    def use_unified_vision_chunk_modality(self) -> bool:
+        """Check if model uses unified vision_chunk modality for images/videos."""
+        return getattr(self._model_config.hf_config, "use_unified_vision_chunk", False)
 
     @property
     def model_config(self) -> ModelConfig:
@@ -499,11 +583,31 @@ class BaseMultiModalItemTracker(ABC, Generic[_T]):
         media.
         """
         input_modality = modality.replace("_embeds", "")
-        num_items = len(self._items_by_modality[modality]) + 1
+        original_modality = modality
+        use_vision_chunk = (
+            self.use_unified_vision_chunk_modality
+            and original_modality in ["video", "image"]
+        )
+
+        # If use_unified_vision_chunk_modality is enabled,
+        # map image/video to vision_chunk
+        if use_vision_chunk:
+            # To avoid validation fail
+            # because models with use_unified_vision_chunk_modality=True
+            # will only accept vision_chunk modality.
+            input_modality = "vision_chunk"
+            num_items = len(self._items_by_modality[input_modality]) + 1
+        else:
+            num_items = len(self._items_by_modality[original_modality]) + 1
 
         self.mm_processor.validate_num_items(input_modality, num_items)
 
-        self._items_by_modality[modality].append(item)
+        # Track original modality for vision_chunk items
+        if use_vision_chunk:
+            self._items_by_modality[input_modality].append(item)  # type: ignore
+            self._modality_order["vision_chunk"].append(original_modality)
+        else:
+            self._items_by_modality[original_modality].append(item)
 
         return self.model_cls.get_placeholder_str(modality, num_items)
 
@@ -515,6 +619,7 @@ class BaseMultiModalItemTracker(ABC, Generic[_T]):
 def _resolve_items(
     items_by_modality: dict[str, list[tuple[object, str | None]]],
     mm_processor: BaseMultiModalProcessor,
+    vision_chunk_modality_order: dict[str, list[str]],
 ) -> tuple[MultiModalDataDict, MultiModalUUIDDict]:
     if "image" in items_by_modality and "image_embeds" in items_by_modality:
         raise ValueError("Mixing raw image and embedding inputs is not allowed")
@@ -546,6 +651,74 @@ def _resolve_items(
     if "video" in items_by_modality:
         mm_data["video"] = [data for data, uuid in items_by_modality["video"]]
         mm_uuids["video"] = [uuid for data, uuid in items_by_modality["video"]]
+    if "vision_chunk" in items_by_modality:
+        # Process vision_chunk items - extract from (data, modality) tuples
+        # and convert to VisionChunk types with proper UUID handling
+        vision_chunk_items = items_by_modality["vision_chunk"]
+        modality_order = vision_chunk_modality_order.get("vision_chunk", [])
+        mm_uuids["vision_chunk"] = [
+            uuid for data, uuid in items_by_modality["vision_chunk"]
+        ]
+
+        # Filter out None items (from asyncio.sleep(0) placeholders)
+        filtered_items = [
+            (idx, item)
+            for idx, item in enumerate(vision_chunk_items)
+            if item is not None
+        ]
+
+        assert len(filtered_items) == len(modality_order), (
+            f"vision_chunk items ({len(filtered_items)}) and "
+            f"modality_order ({len(modality_order)}) must have same length"
+        )
+
+        processed_chunks: list[VisionChunk] = []
+        video_idx = 0
+        for i, (idx, item) in enumerate(filtered_items):
+            inner_modality = modality_order[i]
+            data, uuid = item
+            uuid_val = uuid if idx < len(mm_uuids["vision_chunk"]) else None
+            if inner_modality == "image":
+                # Cast data to proper type for image
+                # Use .media (PIL.Image) directly to avoid redundant
+                # bytes→PIL conversion in media_processor
+                if hasattr(data, "media"):
+                    image_data = data.media  # type: ignore[union-attr]
+                    processed_chunks.append(
+                        VisionChunkImage(type="image", image=image_data, uuid=uuid_val)
+                    )
+                else:
+                    processed_chunks.append(data)  # type: ignore[arg-type]
+            elif inner_modality == "video":
+                # For video, we may need to split into chunks
+                # if processor supports it
+                # For now, just wrap as a video chunk placeholder
+                if hasattr(mm_processor, "split_video_chunks") and data is not None:
+                    try:
+                        video_uuid = uuid_val or random_uuid()
+                        # video await result is (video_data, video_meta) tuple
+                        if isinstance(data, tuple) and len(data) >= 1:
+                            video_data = data[0]
+                        else:
+                            video_data = data
+                        video_chunks = mm_processor.split_video_chunks(video_data)
+                        for i, vc in enumerate(video_chunks):
+                            processed_chunks.append(
+                                VisionChunkVideo(
+                                    type="video_chunk",
+                                    video_chunk=vc["video_chunk"],
+                                    uuid=f"{video_uuid}-{i}",
+                                    video_idx=video_idx,
+                                    prompt=vc["prompt"],
+                                )
+                            )
+                        video_idx += 1
+                    except Exception as e:
+                        logger.warning("Failed to split video chunks: %s", e)
+                        processed_chunks.append(data)  # type: ignore[arg-type]
+                else:
+                    processed_chunks.append(data)  # type: ignore[arg-type]
+        mm_data["vision_chunk"] = processed_chunks
 
     return mm_data, mm_uuids
 
@@ -557,7 +730,9 @@ class MultiModalItemTracker(BaseMultiModalItemTracker[tuple[object, str | None]]
         if not self._items_by_modality:
             return None, None
 
-        return _resolve_items(dict(self._items_by_modality), self.mm_processor)
+        return _resolve_items(
+            dict(self._items_by_modality), self.mm_processor, self._modality_order
+        )
 
     def create_parser(self) -> "BaseMultiModalContentParser":
         return MultiModalContentParser(self)
@@ -577,7 +752,9 @@ class AsyncMultiModalItemTracker(
             for modality, coros in self._items_by_modality.items()
         }
 
-        return _resolve_items(resolved_items_by_modality, self.mm_processor)
+        return _resolve_items(
+            resolved_items_by_modality, self.mm_processor, self._modality_order
+        )
 
     def create_parser(self) -> "BaseMultiModalContentParser":
         return AsyncMultiModalContentParser(self)

vllm/envs.py

@@ -782,6 +782,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
     ),
     # Backend for Video IO
     # - "opencv": Default backend that uses OpenCV stream buffered backend.
+    # - "identity": Returns raw video bytes for model processor to handle.
     #
     # Custom backend implementations can be registered
     # via `@VIDEO_LOADER_REGISTRY.register("my_custom_video_loader")` and

vllm/model_executor/models/registry.py

@@ -359,6 +359,7 @@ _MULTIMODAL_MODELS = {
     ),
     "RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
     "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),  # noqa: E501
+    "KimiK25ForConditionalGeneration": ("kimi_k25", "KimiK25ForConditionalGeneration"),  # noqa: E501
     "LightOnOCRForConditionalGeneration": (
         "lightonocr",
         "LightOnOCRForConditionalGeneration",

vllm/multimodal/inputs.py

@@ -20,6 +20,7 @@ from typing import (
 )
 
 import numpy as np
+from PIL.Image import Image
 from typing_extensions import NotRequired, TypeVar
 
 from vllm.utils.collection_utils import full_groupby, is_list_of
@@ -29,7 +30,6 @@ from vllm.utils.jsontree import json_map_leaves
 if TYPE_CHECKING:
     import torch
     import torch.types
-    from PIL.Image import Image
     from transformers.feature_extraction_utils import BatchFeature
 
     from .media import MediaWithBytes
@@ -105,6 +105,28 @@ The number of data items allowed per modality is restricted by
 """
 
 
+class VisionChunkImage(TypedDict):
+    """Represents an image wrapped as a vision chunk."""
+
+    type: Literal["image"]
+    image: Image
+    uuid: str | None
+
+
+class VisionChunkVideo(TypedDict):
+    """Represents a video chunk with metadata."""
+
+    type: Literal["video_chunk"]
+    video_chunk: list[Image]
+    uuid: str | None
+    prompt: str
+    video_idx: int
+
+
+VisionChunk = VisionChunkImage | VisionChunkVideo
+"""A vision chunk is either an image or a video chunk."""
+
+
 @final
 class MultiModalDataBuiltins(TypedDict, total=False):
     """Type annotations for modality types predefined by vLLM."""
@@ -118,6 +140,9 @@ class MultiModalDataBuiltins(TypedDict, total=False):
     audio: ModalityData[AudioItem]
     """The input audio(s)."""
 
+    vision_chunk: ModalityData[VisionChunk]
+    """The input visual atom(s) - unified modality for images and video chunks."""
+
 
 MultiModalDataDict: TypeAlias = Mapping[str, ModalityData[Any]]
 """

vllm/multimodal/parse.py

@@ -384,6 +384,13 @@ class VideoEmbeddingItems(EmbeddingItems):
         super().__init__(data, "video", expected_hidden_size)
 
 
+class VisionChunkProcessorItems(ProcessorBatchItems[Any]):
+    """Processor items for vision chunks (unified image and video chunks)."""
+
+    def __init__(self, data: Sequence[Any]) -> None:
+        super().__init__(data, "vision_chunk")
+
+
 _D = TypeVar("_D", bound=ModalityDataItems[Any, Any])
 
 
@@ -652,11 +659,23 @@ class MultiModalDataParser:
 
         return VideoProcessorItems(new_videos, metadata=metadata_lst)
 
+    def _parse_vision_chunk_data(
+        self,
+        data: ModalityData[Any],
+    ) -> ModalityDataItems[Any, Any] | None:
+        """Parse vision chunk data (unified image and video chunks)."""
+        if data is None or self._is_empty(data):
+            return None
+        if self.is_embeddings(data):
+            raise ValueError("Do not support embedding data for vision_chunk right now")
+        return VisionChunkProcessorItems(data)
+
     def _get_subparsers(self) -> Mapping[str, ModalityDataParser]:
         return {
             "audio": self._parse_audio_data,
             "image": self._parse_image_data,
             "video": self._parse_video_data,
+            "vision_chunk": self._parse_vision_chunk_data,
         }
 
     def parse_mm_data(self, mm_data: MultiModalDataDict) -> MultiModalDataItems:

vllm/multimodal/video.py

@@ -235,6 +235,27 @@ class VideoLoader:
 VIDEO_LOADER_REGISTRY = ExtensionManager()
 
 
+@VIDEO_LOADER_REGISTRY.register("identity")
+class IdentityVideoLoader(VideoLoader):
+    """IdentityVideoLoader returns raw video bytes without decoding.
+
+    This allows the model processor to handle video decoding and
+    is required for models like Kimi-K2.5 that need custom video chunk splitting.
+
+    NOTE: This is temporary for Kimi-K2.5 testing. Remember to change back
+    to opencv before release if needed.
+    """
+
+    @classmethod
+    def load_bytes(
+        cls,
+        data: bytes,
+        num_frames: int = -1,
+        **kwargs: Any,
+    ) -> tuple[Any, Any]:
+        return data, None
+
+
 @VIDEO_LOADER_REGISTRY.register("opencv")
 class OpenCVVideoBackend(VideoLoader):
     def get_cv2_video_api(self):

vllm/reasoning/__init__.py

@@ -53,8 +53,8 @@ _REASONING_PARSERS_TO_REGISTER = {
         "HunyuanA13BReasoningParser",
     ),
     "kimi_k2": (
-        "deepseek_r1_reasoning_parser",
-        "DeepSeekR1ReasoningParser",
+        "kimi_k2_reasoning_parser",
+        "KimiK2ReasoningParser",
     ),
     "minimax_m2": (
         "minimax_m2_reasoning_parser",

vllm/reasoning/kimi_k2_reasoning_parser.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from collections.abc import Sequence
+from typing import TYPE_CHECKING, Any
+
+from transformers import PreTrainedTokenizerBase
+
+from vllm.entrypoints.openai.engine.protocol import DeltaMessage
+from vllm.logger import init_logger
+from vllm.reasoning import ReasoningParser
+from vllm.reasoning.deepseek_r1_reasoning_parser import DeepSeekR1ReasoningParser
+
+from .identity_reasoning_parser import IdentityReasoningParser
+
+if TYPE_CHECKING:
+    from vllm.entrypoints.openai.chat_completion.protocol import (
+        ChatCompletionRequest,
+    )
+else:
+    ChatCompletionRequest = Any
+
+
+logger = init_logger(__name__)
+
+
+class KimiK2ReasoningParser(ReasoningParser):
+    """
+    Kimi K2 parser that delegates to either DeepSeekR1ReasoningParser or
+    IdentityReasoningParser based on `thinking` and `separate_reasoning`.
+
+    Unlike DeepSeekV3ReasoningParser which defaults to NOT thinking,
+    KimiK2ReasoningParser defaults to thinking mode (uses DeepSeekR1ReasoningParser).
+    """
+
+    def __init__(self, tokenizer: PreTrainedTokenizerBase, *args, **kwargs):
+        super().__init__(tokenizer, *args, **kwargs)
+
+        chat_kwargs = kwargs.pop("chat_template_kwargs", {}) or {}
+        # Key difference: default to True instead of False
+        thinking = bool(chat_kwargs.pop("thinking", True))
+
+        if thinking:
+            self._parser = DeepSeekR1ReasoningParser(tokenizer, *args, **kwargs)
+        else:
+            self._parser = IdentityReasoningParser(tokenizer, *args, **kwargs)
+
+    def is_reasoning_end(self, input_ids: Sequence[int]) -> bool:
+        return self._parser.is_reasoning_end(input_ids)
+
+    def is_reasoning_end_streaming(
+        self, input_ids: list[int], delta_ids: list[int]
+    ) -> bool:
+        return self._parser.is_reasoning_end_streaming(input_ids, delta_ids)
+
+    def extract_content_ids(self, input_ids: list[int]) -> list[int]:
+        return self._parser.extract_content_ids(input_ids)
+
+    def extract_reasoning(
+        self, model_output: str, request: "ChatCompletionRequest"
+    ) -> tuple[str | None, str | None]:
+        return self._parser.extract_reasoning(model_output, request)
+
+    def extract_reasoning_streaming(
+        self,
+        previous_text: str,
+        current_text: str,
+        delta_text: str,
+        previous_token_ids: Sequence[int],
+        current_token_ids: Sequence[int],
+        delta_token_ids: Sequence[int],
+    ) -> DeltaMessage | None:
+        return self._parser.extract_reasoning_streaming(
+            previous_text,
+            current_text,
+            delta_text,
+            previous_token_ids,
+            current_token_ids,
+            delta_token_ids,
+        )

vllm/renderers/hf.py

@@ -20,9 +20,11 @@ from vllm.entrypoints.chat_utils import (
     ChatTemplateContentFormatOption,
     ChatTemplateResolutionError,
     ConversationMessage,
+    build_video_prompts_from_mm_data,
     load_chat_template,
     parse_chat_messages,
     parse_chat_messages_async,
+    rebuild_mm_uuids_from_mm_data,
 )
 from vllm.inputs import TextPrompt, TokensPrompt
 from vllm.logger import init_logger
@@ -547,6 +549,40 @@ class HfRenderer(RendererLike):
             **kwargs,
         )
 
+        # NOTE: use_unified_vision_chunk is currently specific to Kimi-K2.5
+        # model which uses unified vision chunks for both images and videos.
+        if (
+            getattr(model_config.hf_config, "use_unified_vision_chunk", False)
+            and mm_uuids is not None
+            and mm_data is not None
+        ):
+            mm_uuids = rebuild_mm_uuids_from_mm_data(mm_uuids, mm_data)
+
+            # get video placehoder, replace it with runtime video-chunk prompts
+            video_placeholder = getattr(
+                model_config.hf_config, "video_placeholder", None
+            )
+            if video_placeholder and isinstance(prompt_raw, str):
+                video_prompts = build_video_prompts_from_mm_data(mm_data)
+
+                # replace in order
+                prompt_raw_parts = prompt_raw.split(video_placeholder)
+                if len(prompt_raw_parts) == len(video_prompts) + 1:
+                    prompt_raw = "".join(
+                        [
+                            prompt_raw_parts[i] + video_prompts[i]
+                            for i in range(len(video_prompts))
+                        ]
+                    )
+                    prompt_raw += prompt_raw_parts[-1]
+                else:
+                    logger.warning(
+                        "Number of video placeholders (%d) does not match "
+                        "number of videos (%d) in the request.",
+                        len(prompt_raw_parts) - 1,
+                        len(video_prompts),
+                    )
+
         prompt = (
             TextPrompt(prompt=prompt_raw)
             if isinstance(prompt_raw, str)
@@ -587,6 +623,40 @@ class HfRenderer(RendererLike):
             **kwargs,
         )
 
+        # NOTE: use_unified_vision_chunk is currently specific to Kimi-K2.5
+        # model which uses unified vision chunks for both images and videos.
+        if (
+            getattr(model_config.hf_config, "use_unified_vision_chunk", False)
+            and mm_uuids is not None
+            and mm_data is not None
+        ):
+            mm_uuids = rebuild_mm_uuids_from_mm_data(mm_uuids, mm_data)
+
+            # get video placehoder, replace it with runtime video-chunk prompts
+            video_placeholder = getattr(
+                model_config.hf_config, "video_placeholder", None
+            )
+            if video_placeholder and isinstance(prompt_raw, str):
+                video_prompts = build_video_prompts_from_mm_data(mm_data)
+
+                # replace in order
+                prompt_raw_parts = prompt_raw.split(video_placeholder)
+                if len(prompt_raw_parts) == len(video_prompts) + 1:
+                    prompt_raw = "".join(
+                        [
+                            prompt_raw_parts[i] + video_prompts[i]
+                            for i in range(len(video_prompts))
+                        ]
+                    )
+                    prompt_raw += prompt_raw_parts[-1]
+                else:
+                    logger.warning(
+                        "Number of video placeholders (%d) does not match "
+                        "number of videos (%d) in the request.",
+                        len(prompt_raw_parts) - 1,
+                        len(video_prompts),
+                    )
+
         prompt = (
             TextPrompt(prompt=prompt_raw)
             if isinstance(prompt_raw, str)

vllm/transformers_utils/config.py

@@ -81,6 +81,7 @@ _CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(
     isaac="IsaacConfig",
     kimi_linear="KimiLinearConfig",
     kimi_vl="KimiVLConfig",
+    kimi_k25="KimiK25Config",
     RefinedWeb="RWConfig",  # For tiiuae/falcon-40b(-instruct)
     RefinedWebModel="RWConfig",  # For tiiuae/falcon-7b(-instruct)
     jais="JAISConfig",

vllm/transformers_utils/configs/__init__.py

@@ -38,6 +38,7 @@ _CLASS_TO_MODULE: dict[str, str] = {
     "MoonViTConfig": "vllm.transformers_utils.configs.moonvit",
     "KimiLinearConfig": "vllm.transformers_utils.configs.kimi_linear",
     "KimiVLConfig": "vllm.transformers_utils.configs.kimi_vl",
+    "KimiK25Config": "vllm.transformers_utils.configs.kimi_k25",
     "NemotronConfig": "vllm.transformers_utils.configs.nemotron",
     "NemotronHConfig": "vllm.transformers_utils.configs.nemotron_h",
     "Olmo3Config": "vllm.transformers_utils.configs.olmo3",
@@ -77,6 +78,7 @@ __all__ = [
     "MoonViTConfig",
     "KimiLinearConfig",
     "KimiVLConfig",
+    "KimiK25Config",
     "NemotronConfig",
     "NemotronHConfig",
     "Olmo3Config",

vllm/transformers_utils/configs/kimi_k25.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Kimi-K2.5 Model Configuration.
+
+This configuration supports video-chunk as an internal modality type.
+A video-chunk is the smallest independently processable unit of video.
+"""
+
+from transformers import DeepseekV3Config
+from transformers.configuration_utils import PretrainedConfig
+
+
+class KimiK25VisionConfig(PretrainedConfig):
+    model_type = "kimi_k25_vision"
+
+    def __init__(
+        self,
+        # Vision Tower
+        patch_size: int = 14,
+        init_pos_emb_height: int = 64,
+        init_pos_emb_width: int = 64,
+        init_pos_emb_time: int = 4,
+        pos_emb_type: str = "divided_fixed",
+        num_attention_heads: int = 16,
+        num_hidden_layers: int = 27,
+        hidden_size: int = 1152,
+        intermediate_size: int = 4304,
+        merge_kernel_size: tuple[int, int] = (2, 2),
+        video_attn_type: str = "spatial_temporal",
+        merge_type: str = "sd2_tpool",
+        # MM Projector
+        mm_projector_type: str = "patchmerger",
+        mm_hidden_size: int | None = None,
+        projector_hidden_act: str = "gelu",
+        projector_ln_eps: float = 1e-5,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        # Vision Tower
+        self.patch_size = patch_size
+        self.init_pos_emb_height = init_pos_emb_height
+        self.init_pos_emb_width = init_pos_emb_width
+        self.init_pos_emb_time = init_pos_emb_time
+        self.pos_emb_type = pos_emb_type
+        self.num_attention_heads = num_attention_heads
+        self.num_hidden_layers = num_hidden_layers
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.merge_kernel_size = merge_kernel_size
+        self.video_attn_type = video_attn_type
+        self.merge_type = merge_type
+        # MM Projector
+        self.mm_projector_type = mm_projector_type
+        if mm_hidden_size is not None:
+            self.mm_hidden_size = mm_hidden_size
+        else:
+            self.mm_hidden_size = hidden_size
+        self.projector_hidden_act = projector_hidden_act
+        self.projector_ln_eps = projector_ln_eps
+
+
+class KimiK25Config(PretrainedConfig):
+    """Kimi-K2.5 model configuration.
+
+    Kimi-K2.5 extends Kimi-K2 with vision support using video-chunks.
+    A video-chunk consists of multiple consecutive frames
+    that are processed together with temporal pooling.
+
+    Args:
+        vision_config: Configuration for the vision tower and projector.
+        text_config: Configuration for the text model (DeepseekV3).
+        ignore_index: The ignore index for the loss function.
+        media_placeholder_token_id: The token ID for media placeholders.
+        pad_token_id: The token ID for padding.
+    """
+
+    model_type = "kimi_k25"
+
+    def __init__(
+        self,
+        vision_config: dict | KimiK25VisionConfig | None = None,
+        text_config: dict | DeepseekV3Config | None = None,
+        ignore_index: int = -100,
+        media_placeholder_token_id: int = 163605,
+        pad_token_id: int = 0,
+        use_unified_vision_chunk: bool = False,
+        video_placeholder: str = "<|kimi_k25_video_placeholder|>",
+        **kwargs,
+    ):
+        # Vision config
+        if vision_config is None:
+            vision_config = KimiK25VisionConfig()
+        elif isinstance(vision_config, dict):
+            vision_config = KimiK25VisionConfig(**vision_config)
+        self.vision_config: KimiK25VisionConfig = vision_config
+
+        # Text config
+        if text_config is None:
+            text_config = DeepseekV3Config()
+        elif isinstance(text_config, dict):
+            text_config = DeepseekV3Config(**text_config)
+        self.text_config: DeepseekV3Config = text_config
+
+        # Set mm_hidden_size to text hidden size if not explicitly set
+        if self.vision_config.mm_hidden_size == self.vision_config.hidden_size:
+            self.vision_config.mm_hidden_size = self.text_config.hidden_size
+
+        # Other config
+        self.ignore_index = ignore_index
+        self.media_placeholder_token_id = media_placeholder_token_id
+        self.use_unified_vision_chunk = use_unified_vision_chunk
+        self.video_placeholder = video_placeholder
+
+        # Propagate quantization config from text model
+        if getattr(self.text_config, "quantization_config", None) is not None:
+            self.quantization_config = self.text_config.quantization_config
+
+        super().__init__(pad_token_id=pad_token_id, **kwargs)
+
+    @property
+    def hidden_size(self) -> int:
+        """Get hidden size from text config for compatibility."""
+        return self.text_config.hidden_size
+
+    @property
+    def vocab_size(self) -> int:
+        """Get vocab size from text config for compatibility."""
+        return self.text_config.vocab_size