Add GLM4.1V model (Draft) (#19331)

Signed-off-by: zRzRzRzRzRzRzR <2448370773@qq.com>
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>

docs/models/supported_models.md

@@ -553,6 +553,7 @@ Specified using `--task generate`.
 | `FuyuForCausalLM`                            | Fuyu                                                                     | T + I                                                                 | `adept/fuyu-8b` etc.                                                                                                                                    |                        | ✅︎                          | ✅︎                    |
 | `Gemma3ForConditionalGeneration`             | Gemma 3                                                                  | T + I<sup>+</sup>                                                     | `google/gemma-3-4b-it`, `google/gemma-3-27b-it`, etc.                                                                                                   | ✅︎                     | ✅︎                          | ⚠️                    |
 | `GLM4VForCausalLM`<sup>^</sup>               | GLM-4V                                                                   | T + I                                                                 | `THUDM/glm-4v-9b`, `THUDM/cogagent-9b-20241220` etc.                                                                                                    | ✅︎                     | ✅︎                          | ✅︎                    |
+| `Glm4vForConditionalGeneration`              | GLM-4.1V-Thinking                                                        | T + I<sup>E+</sup> + V<sup>E+</sup>                                   | `THUDM/GLM-4.1V-9B-Thinkg`,  etc.                                                                                                                       | ✅︎                     | ✅︎                          | ✅︎                    |
 | `GraniteSpeechForConditionalGeneration`      | Granite Speech                                                           | T + A                                                                 | `ibm-granite/granite-speech-3.3-8b`                                                                                                                     | ✅︎                     | ✅︎                          | ✅︎                    |
 | `H2OVLChatModel`                             | H2OVL                                                                    | T + I<sup>E+</sup>                                                    | `h2oai/h2ovl-mississippi-800m`, `h2oai/h2ovl-mississippi-2b`, etc.                                                                                      |                        | ✅︎                          | ✅︎\*                  |
 | `Idefics3ForConditionalGeneration`           | Idefics3                                                                 | T + I                                                                 | `HuggingFaceM4/Idefics3-8B-Llama3` etc.                                                                                                                 | ✅︎                     |                             |  ✅︎                   |

examples/offline_inference/vision_language.py

@@ -248,6 +248,42 @@ def run_glm4v(questions: list[str], modality: str) -> ModelRequestData:
     )
 
 
+# GLM-4.1V
+def run_glm4_1v(questions: list[str], modality: str) -> ModelRequestData:
+    model_name = "THUDM/GLM-4.1V-9B-Thinking"
+
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=4096,
+        max_num_seqs=2,
+        mm_processor_kwargs={
+            "size": {"shortest_edge": 12544, "longest_edge": 47040000},
+            "fps": 1,
+        },
+        limit_mm_per_prompt={modality: 1},
+        enforce_eager=True,
+    )
+
+    if modality == "image":
+        placeholder = "<|begin_of_image|><|image|><|end_of_image|>"
+    elif modality == "video":
+        placeholder = "<|begin_of_video|><|video|><|end_of_video|>"
+
+    prompts = [
+        (
+            "[gMASK]<sop><|system|>\nYou are a helpful assistant.<|user|>\n"
+            f"{placeholder}"
+            f"{question}<|assistant|>assistant\n"
+        )
+        for question in questions
+    ]
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompts=prompts,
+    )
+
+
 # H2OVL-Mississippi
 def run_h2ovl(questions: list[str], modality: str) -> ModelRequestData:
     assert modality == "image"
@@ -1114,6 +1150,7 @@ model_example_map = {
     "fuyu": run_fuyu,
     "gemma3": run_gemma3,
     "glm4v": run_glm4v,
+    "glm4_1v": run_glm4_1v,
     "h2ovl_chat": run_h2ovl,
     "idefics3": run_idefics3,
     "internvl_chat": run_internvl,
@@ -1172,10 +1209,11 @@ def get_multi_modal_input(args):
     if args.modality == "video":
         # Input video and question
         video = VideoAsset(name="baby_reading", num_frames=args.num_frames).np_ndarrays
+        metadata = VideoAsset(name="baby_reading", num_frames=args.num_frames).metadata
         vid_questions = ["Why is this video funny?"]
 
         return {
-            "data": video,
+            "data": [(video, metadata)] if args.model_type == "glm4_1v" else video,
             "questions": vid_questions,
         }
 

tests/entrypoints/openai/test_video.py

@@ -50,7 +50,7 @@ async def client(server):
 @pytest.fixture(scope="session")
 def base64_encoded_video() -> dict[str, str]:
     return {
-        video_url: encode_video_base64(fetch_video(video_url))
+        video_url: encode_video_base64(fetch_video(video_url)[0])
         for video_url in TEST_VIDEO_URLS
     }
 

tests/models/multimodal/generation/test_common.py

@@ -309,6 +309,34 @@ VLM_TEST_SETTINGS = {
         num_logprobs=10,
         marks=[large_gpu_mark(min_gb=32)],
     ),
+    "glm4_1v": VLMTestInfo(
+        models=["THUDM/GLM-4.1V-9B-Thinking"],
+        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
+        prompt_formatter=lambda img_prompt: f"<|user|>\n{img_prompt}<|assistant|>",  # noqa: E501
+        img_idx_to_prompt=lambda idx: "<|begin_of_image|><|image|><|end_of_image|>", # noqa: E501
+        video_idx_to_prompt=lambda idx: "<|begin_of_video|><|video|><|end_of_video|>", # noqa: E501
+        max_model_len=2048,
+        max_num_seqs=2,
+        get_stop_token_ids=lambda tok: [151329, 151336, 151338],
+        num_logprobs=10,
+        image_size_factors=[(), (0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
+        auto_cls=AutoModelForImageTextToText,
+    ),
+    "glm4_1v-video": VLMTestInfo(
+        models=["THUDM/GLM-4.1V-9B-Thinking"],
+        # GLM4.1V require include video metadata for input
+        test_type=VLMTestType.CUSTOM_INPUTS,
+        max_model_len=4096,
+        max_num_seqs=2,
+        auto_cls=AutoModelForImageTextToText,
+        patch_hf_runner=model_utils.glm4_1v_patch_hf_runner,
+        custom_test_opts=[CustomTestOptions(
+            inputs=custom_inputs.video_with_metadata_glm4_1v(),
+            limit_mm_per_prompt={"video": 1},
+        )],
+        # This is needed to run on machine with 24GB VRAM
+        vllm_runner_kwargs={"gpu_memory_utilization": 0.95},
+    ),
     "h2ovl": VLMTestInfo(
         models = [
             "h2oai/h2ovl-mississippi-800m",

tests/models/multimodal/generation/vlm_utils/custom_inputs.py

@@ -129,3 +129,23 @@ def windows_attention_image_qwen2_5_vl():
 
     wrapped_sf = ImageSizeWrapper(type=SizeType.SIZE_FACTOR, data=[0.5])
     return build_single_image_inputs([image], [prompt], wrapped_sf)
+
+
+def video_with_metadata_glm4_1v():
+    video_array = VIDEO_ASSETS[0].np_ndarrays
+    metadata = VIDEO_ASSETS[0].metadata
+    question = "Describe the video."
+    video_prompt = "<|begin_of_video|><|video|><|end_of_video|>"
+    formatted_prompt = f"<|user|>\n{video_prompt}{question}<|assistant|>\n"
+
+    scales = [0.1, 0.2, 0.25]
+    video_input = [[(rescale_video_size(video_array, scale), metadata)]
+                   for scale in scales]
+    prompts = [formatted_prompt] * len(video_input)
+
+    return [
+        PromptWithMultiModalInput(
+            prompts=prompts,
+            video_data=video_input,
+        )
+    ]

tests/models/multimodal/generation/vlm_utils/model_utils.py

@@ -16,9 +16,11 @@ import torch
 from PIL.Image import Image
 from transformers import (AutoConfig, AutoTokenizer, BatchFeature,
                           GenerationConfig, GenerationMixin)
+from transformers.video_utils import VideoMetadata
 
 from vllm.sequence import SampleLogprobs
 from vllm.transformers_utils.tokenizer import patch_padding_side
+from vllm.utils import is_list_of
 
 from .....conftest import HfRunner, ImageAsset, ImageTestAssets
 from .types import RunnerOutput
@@ -373,6 +375,28 @@ def glm4v_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
     return hf_model
 
 
+def glm4_1v_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
+    """Patches and returns an instance of the HfRunner to use for GLM4.1V."""
+    hf_processor = hf_model.processor
+
+    def processor(*args, videos=None, **kwargs):
+        if videos is not None and is_list_of(videos, tuple):
+            # If videos is a list of tuples, we assume each tuple contains
+            # (video_array, metadata) as in the case of GLM4.1V.
+            video_metadata = [[VideoMetadata(**video[1])] for video in videos]
+            videos = [[video[0]] for video in videos]
+        else:
+            video_metadata = None
+
+        return hf_processor(*args,
+                            videos=videos,
+                            video_metadata=video_metadata,
+                            **kwargs)
+
+    hf_model.processor = processor
+    return hf_model
+
+
 def h2ovl_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
     """Patches and returns an instance of the HfRunner to use for H2OVL."""
 

tests/models/multimodal/processing/test_common.py

@@ -24,6 +24,22 @@ from ....multimodal.utils import random_audio, random_image, random_video
 from ...registry import HF_EXAMPLE_MODELS
 
 
+def glm4_1v_patch_mm_data(mm_data: MultiModalDataDict) -> MultiModalDataDict:
+    """
+    Patch the multimodal data for GLM4.1V model.
+    """
+    # Ensure video metadata is included
+    if "video" in mm_data:
+        video = mm_data["video"]
+        mm_data["video"] = (video, {
+            "total_num_frames": len(video),
+            "fps": len(video),
+            "duration": 1,
+            "video_backend": "opencv"
+        })
+    return mm_data
+
+
 def _test_processing_correctness(
     model_id: str,
     hit_rate: float,
@@ -154,6 +170,11 @@ _IGNORE_MM_KEYS = {
     "ultravox": {"audio_features"},
 }
 
+MM_DATA_PATCHES = {
+    # GLM4.1V requires video metadata to be included in the input
+    "glm4v": glm4_1v_patch_mm_data,
+}
+
 
 def _test_processing_correctness_one(
     model_config: ModelConfig,
@@ -166,6 +187,8 @@ def _test_processing_correctness_one(
 ):
     model_type = model_config.hf_config.model_type
     ignore_mm_keys = _IGNORE_MM_KEYS.get(model_type, set[str]())
+    if model_type in MM_DATA_PATCHES:
+        mm_data = MM_DATA_PATCHES[model_type](mm_data)
 
     if isinstance(prompt, str):
         text_prompt = prompt
@@ -245,6 +268,7 @@ def _test_processing_correctness_one(
     "adept/fuyu-8b",
     "google/gemma-3-4b-it",
     "THUDM/glm-4v-9b",
+    "THUDM/GLM-4.1V-9B-Thinking",
     "ibm-granite/granite-speech-3.3-2b",
     "h2oai/h2ovl-mississippi-800m",
     "OpenGVLab/InternVL2-1B",

tests/models/registry.py

@@ -338,6 +338,7 @@ _MULTIMODAL_EXAMPLE_MODELS = {
     "GLM4VForCausalLM": _HfExamplesInfo("THUDM/glm-4v-9b",
                                         trust_remote_code=True,
                                         hf_overrides={"architectures": ["GLM4VForCausalLM"]}),  # noqa: E501
+    "Glm4vForConditionalGeneration": _HfExamplesInfo("THUDM/GLM-4.1V-9B-Thinking", min_transformers_version="4.53"),  # noqa: E501
     "H2OVLChatModel": _HfExamplesInfo("h2oai/h2ovl-mississippi-800m",
                                       extras={"2b": "h2oai/h2ovl-mississippi-2b"},  # noqa: E501
                                       max_transformers_version="4.48",  # noqa: E501

tests/multimodal/test_utils.py

@@ -172,7 +172,9 @@ async def test_fetch_video_http(video_url: str, num_frames: int):
     video_sync = connector.fetch_video(video_url, num_frames=num_frames)
     video_async = await connector.fetch_video_async(video_url,
                                                     num_frames=num_frames)
-    assert np.array_equal(video_sync, video_async)
+    # Check that the video frames are equal and metadata are same
+    assert np.array_equal(video_sync[0], video_async[0])
+    assert video_sync[1] == video_async[1]
 
 
 # Used for the next two tests related to `merge_and_sort_multimodal_metadata`.

vllm/assets/video.py

@@ -3,7 +3,7 @@
 
 from dataclasses import dataclass
 from functools import lru_cache
-from typing import ClassVar, Literal, Optional
+from typing import Any, ClassVar, Literal, Optional
 
 import cv2
 import numpy as np
@@ -77,6 +77,24 @@ def video_to_pil_images_list(path: str,
     ]
 
 
+def video_get_metadata(path: str) -> dict[str, Any]:
+    cap = cv2.VideoCapture(path)
+    if not cap.isOpened():
+        raise ValueError(f"Could not open video file {path}")
+
+    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    fps = cap.get(cv2.CAP_PROP_FPS)
+    duration = total_frames / fps if fps > 0 else 0
+
+    metadata = {
+        "total_num_frames": total_frames,
+        "fps": fps,
+        "duration": duration,
+        "video_backend": "opencv"
+    }
+    return metadata
+
+
 VideoAssetName = Literal["baby_reading"]
 
 
@@ -105,6 +123,12 @@ class VideoAsset:
         ret = video_to_ndarrays(video_path, self.num_frames)
         return ret
 
+    @property
+    def metadata(self) -> dict[str, Any]:
+        video_path = download_video_asset(self.filename)
+        ret = video_get_metadata(video_path)
+        return ret
+
     def get_audio(self, sampling_rate: Optional[float] = None) -> npt.NDArray:
         """
         Read audio data from the video asset, used in Qwen2.5-Omni examples.

vllm/entrypoints/chat_utils.py

@@ -515,6 +515,8 @@ class BaseMultiModalItemTracker(ABC, Generic[_T]):
         if modality in ("image", "image_embeds"):
             if model_type == "chatglm":
                 return "<|begin_of_image|><|endoftext|><|end_of_image|>"
+            if model_type == "glm4v":
+                return "<|begin_of_image|><|image|><|end_of_image|>"
             if model_type in ("phi3_v", "phi4mm"):
                 return f"<|image_{current_count}|>"
             if model_type in ("minicpmo", "minicpmv"):
@@ -563,6 +565,8 @@ class BaseMultiModalItemTracker(ABC, Generic[_T]):
         elif modality == "video":
             if model_type == "internvl_chat":
                 return "<video>"
+            if model_type == "glm4v":
+                return "<|begin_of_video|><|video|><|end_of_video|>"
             if model_type in ("qwen2_vl", "qwen2_5_vl"):
                 return "<|vision_start|><|video_pad|><|vision_end|>"
             if model_type == "qwen2_5_omni":

vllm/model_executor/layers/rotary_embedding.py

@@ -23,6 +23,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Rotary Positional Embeddings."""
+import itertools
 import math
 from typing import Any, Optional, Union
 
@@ -1118,6 +1119,15 @@ class MRotaryEmbedding(RotaryEmbedding):
                 audio_feature_lengths=audio_feature_lengths,
                 use_audio_in_video=use_audio_in_video,
             )
+        elif "glm4v" in hf_config.model_type:
+            return cls._glm4v_get_input_positions_tensor(
+                input_tokens=input_tokens,
+                hf_config=hf_config,
+                image_grid_thw=image_grid_thw,
+                video_grid_thw=video_grid_thw,
+                context_len=context_len,
+                seq_len=seq_len,
+            )
         else:
             return cls._vl_get_input_positions_tensor(
                 input_tokens=input_tokens,
@@ -1129,6 +1139,115 @@ class MRotaryEmbedding(RotaryEmbedding):
                 seq_len=seq_len,
             )
 
+    @classmethod
+    def _glm4v_get_input_positions_tensor(
+        cls,
+        input_tokens: list[int],
+        hf_config: PretrainedConfig,
+        image_grid_thw: Union[list[list[int]], torch.Tensor],
+        video_grid_thw: Union[list[list[int]], torch.Tensor],
+        context_len: int = 0,
+        seq_len: Optional[int] = None,
+    ) -> tuple[torch.Tensor, int]:
+        """Get mrope input positions and delta value for GLM4V."""
+
+        image_token_id = hf_config.image_token_id
+        video_start_token_id = hf_config.video_start_token_id
+        video_end_token_id = hf_config.video_end_token_id
+        spatial_merge_size = hf_config.vision_config.spatial_merge_size
+        llm_pos_ids_list: list = []
+
+        if not (image_grid_thw is None and video_grid_thw is None):
+            if isinstance(image_grid_thw, torch.Tensor):
+                image_grid_thw = image_grid_thw.tolist()
+
+            input_token_type: list[str] = []
+            video_check_flg = False
+            for token in input_tokens:
+                if token == video_start_token_id:
+                    video_check_flg = True
+                elif token == video_end_token_id:
+                    video_check_flg = False
+
+                if (token == image_token_id) and (video_check_flg is False):
+                    input_token_type.append("image")
+                elif (token == image_token_id) and (video_check_flg is True):
+                    input_token_type.append("video")
+                else:
+                    input_token_type.append("text")
+
+            input_type_group: list[tuple[str, int, int]] = []
+            for key, group_iter in itertools.groupby(
+                    enumerate(input_token_type), lambda x: x[1]):
+                group_list = list(group_iter)
+                start_index = group_list[0][0]
+                end_index = group_list[-1][0] + 1
+                input_type_group.append((key, start_index, end_index))
+
+            video_frame_num = 1
+            mm_data_idx = 0
+            for modality_type, start_idx, end_idx in input_type_group:
+                st_idx = llm_pos_ids_list[-1].max() + 1 if len(
+                    llm_pos_ids_list) > 0 else 0
+                if modality_type == "image":
+                    t, h, w = (
+                        image_grid_thw[mm_data_idx][0],
+                        image_grid_thw[mm_data_idx][1],
+                        image_grid_thw[mm_data_idx][2],
+                    )
+                    llm_grid_t, llm_grid_h, llm_grid_w = \
+                        t, h // spatial_merge_size, w // spatial_merge_size
+
+                    t_index = torch.arange(llm_grid_t).view(-1, 1).expand(
+                        -1, llm_grid_h * llm_grid_w).flatten()
+                    h_index = torch.arange(llm_grid_h).view(1, -1, 1).expand(
+                        llm_grid_t, -1, llm_grid_w).flatten()
+                    w_index = torch.arange(llm_grid_w).view(1, 1, -1).expand(
+                        llm_grid_t, llm_grid_h, -1).flatten()
+                    llm_pos_ids_list.append(
+                        torch.stack([t_index, h_index, w_index]) + st_idx)
+                    mm_data_idx += 1
+
+                elif modality_type == "video":
+                    t, h, w = (
+                        video_frame_num,
+                        image_grid_thw[mm_data_idx][1],
+                        image_grid_thw[mm_data_idx][2],
+                    )
+                    llm_grid_t, llm_grid_h, llm_grid_w = \
+                        t, h // spatial_merge_size, w // spatial_merge_size
+
+                    for t_idx in range(llm_grid_t):
+                        t_index = torch.tensor(t_idx).view(-1, 1).expand(
+                            -1, llm_grid_h * llm_grid_w).flatten()
+                        h_index = torch.arange(llm_grid_h).view(
+                            1, -1, 1).expand(1, -1, llm_grid_w).flatten()
+                        w_index = torch.arange(llm_grid_w).view(
+                            1, 1, -1).expand(1, llm_grid_h, -1).flatten()
+                        llm_pos_ids_list.append(
+                            torch.stack([t_index, h_index, w_index]) + st_idx)
+
+                    mm_data_idx += 1
+                    video_frame_num += 1
+
+                else:
+                    text_len = end_idx - start_idx
+                    llm_pos_ids_list.append(
+                        torch.arange(text_len).view(1, -1).expand(3, -1) +
+                        st_idx)
+                    video_frame_num = 1
+
+        else:
+            text_len = len(input_tokens)
+            llm_pos_ids_list.append(
+                torch.arange(text_len).view(1, -1).expand(3, -1))
+
+        llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)
+        llm_positions = llm_positions[:, context_len:seq_len]
+        mrope_position_delta = (llm_positions.max() + 1 -
+                                len(input_tokens)).item()
+        return llm_positions, mrope_position_delta
+
     @classmethod
     def _vl_get_input_positions_tensor(
         cls,

vllm/model_executor/models/glm4_1v.py

+# SPDX-License-Identifier: Apache-2.0
+
+# Adapted from
+# https://github.com/huggingface/transformers/blob/main/src/transformers/models/Glm4v/modeling_Glm4v.py
+# Copyright 2025 The vLLM team.
+# Copyright 2025 The ZhipuAI Team.
+# Copyright 2025 The HuggingFace Inc. team.
+# All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# 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.
+"""Inference-only GLM-4V model compatible with HuggingFace weights."""
+
+import math
+from collections.abc import Iterable, Mapping, Sequence
+from functools import partial
+from typing import Any, Callable, Literal, Optional, TypedDict, Union
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from transformers import BatchFeature
+from transformers.models.glm4v.configuration_glm4v import (Glm4vConfig,
+                                                           Glm4vVisionConfig)
+from transformers.models.glm4v.image_processing_glm4v import (
+    Glm4vImageProcessor, smart_resize)
+from transformers.models.glm4v.video_processing_glm4v import (
+    Glm4vVideoProcessor)
+from transformers.video_utils import VideoMetadata
+
+from vllm.config import VllmConfig
+from vllm.distributed import parallel_state
+from vllm.distributed import utils as dist_utils
+from vllm.logger import init_logger
+from vllm.model_executor import SamplingMetadata
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               MergedColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.layers.quantization.gptq import GPTQConfig
+from vllm.model_executor.layers.quantization.gptq_marlin import (
+    GPTQMarlinConfig)
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+from vllm.model_executor.models.module_mapping import MultiModelKeys
+from vllm.multimodal import MULTIMODAL_REGISTRY
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    MultiModalKwargs, VideoItem)
+from vllm.multimodal.parse import (ImageSize, MultiModalDataItems,
+                                   MultiModalDataParser)
+from vllm.multimodal.processing import (BaseMultiModalProcessor,
+                                        BaseProcessingInfo, PromptReplacement,
+                                        PromptUpdate)
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
+from vllm.platforms import _Backend
+from vllm.sequence import IntermediateTensors
+from vllm.transformers_utils.config import uses_mrope
+
+from ..layers.activation import SiluAndMul
+from .interfaces import (MultiModalEmbeddings, SupportsLoRA,
+                         SupportsMultiModal, SupportsPP)
+from .qwen2_vl import _qwen2vl_field_config, apply_rotary_pos_emb_vision
+from .utils import (AutoWeightsLoader, WeightsMapper,
+                    init_vllm_registered_model, maybe_prefix,
+                    merge_multimodal_embeddings)
+from .vision import get_vit_attn_backend
+
+logger = init_logger(__name__)
+
+# For profile run
+_MAX_FRAMES_PER_VIDEO = 600
+
+# === Vision Inputs === #
+
+
+class Glm4vImagePixelInputs(TypedDict):
+    type: Literal["pixel_values"]
+    pixel_values: torch.Tensor
+    """Shape:
+    `(num_patches, num_channels * patch_size * patch_size)`
+    """
+
+    image_grid_thw: torch.Tensor
+    """Shape: `(num_images, 3)`
+    This should be in `(grid_t, grid_h, grid_w)` format.
+    """
+
+
+class Glm4vImageEmbeddingInputs(TypedDict):
+    type: Literal["image_embeds"]
+    image_embeds: torch.Tensor
+    """Supported types:
+    - List[`torch.Tensor`]: A list of tensors holding all images' features.
+        Each tensor holds an image's features.
+    - `torch.Tensor`: A tensor holding all images' features
+        (concatenation of all images' feature tensors).
+
+    Tensor shape: `(num_image_features, hidden_size)`
+    - `num_image_features` varies based on
+        the number and resolution of the images.
+    - `hidden_size` must match the hidden size of language model backbone.
+    """
+
+    image_grid_thw: torch.Tensor
+    """Shape: `(num_images, 3)`
+    This should be in `(grid_t, grid_h, grid_w)` format.
+    """
+
+
+Glm4vImageInputs = Union[Glm4vImagePixelInputs, Glm4vImageEmbeddingInputs]
+
+
+class Glm4vVideoPixelInputs(TypedDict):
+    type: Literal["pixel_values_videos"]
+    pixel_values_videos: torch.Tensor
+    """Shape:
+    `(num_patches,
+      num_channels * temporal_patch_size * patch_size * patch_size)`
+    """
+    # video_metadata: Union[list[VideoMetadata], list[dict]]
+    video_grid_thw: Union[list[torch.Tensor], torch.Tensor]
+    """Shape: `(num_videos, num_frames, 3)` or `(1, num_frames, 3)` 
+    for single video.
+    Each entry represents [grid_t, grid_h, grid_w] format where:
+    - grid_t: Temporal grid size (usually 1 for processed video)
+    - grid_h: Height grid size  
+    - grid_w: Width grid size
+    This describes the grid structure of the video patches.
+    """
+
+
+class Glm4vVideoEmbeddingInputs(TypedDict):
+    type: Literal["video_embeds"]
+
+    video_embeds: torch.Tensor
+    """
+    Tensor shape: `(num_video_patches, hidden_size)`
+    - `num_video_patches`: Total number of video patches across all frames
+    - `hidden_size`: Must match the hidden size of language model backbone
+    """
+
+    video_grid_thw: torch.Tensor
+    """Shape: `(num_videos, 1, 3)` or `(1, 1, 3)` for single video
+    Each entry represents [grid_t, grid_h, grid_w] format where:
+    - grid_t: Temporal grid size (usually 1 for processed video)
+    - grid_h: Height grid size  
+    - grid_w: Width grid size
+    This describes the grid structure of the video patches.
+    """
+
+
+Glm4vVideoInputs = Union[Glm4vVideoPixelInputs, Glm4vVideoEmbeddingInputs]
+
+# === Vision Encoder === #
+
+
+class Glm4vVisionMLP(nn.Module):
+
+    def __init__(
+        self,
+        in_features: int,
+        hidden_features: int,
+        bias: bool = False,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            input_size=in_features,
+            output_sizes=[hidden_features] * 2,
+            bias=bias,
+            quant_config=quant_config,
+        )
+        self.down_proj = RowParallelLinear(
+            hidden_features,
+            in_features,
+            bias=bias,
+            quant_config=quant_config,
+        )
+        self.act_fn = SiluAndMul()
+
+    def forward(self, x: torch.Tensor):
+        x, _ = self.gate_up_proj(x)
+        x = self.act_fn(x)
+        x, _ = self.down_proj(x)
+        return x
+
+
+def all_gather_interleave(local_tensor, hidden_size: int, tp_size: int):
+    """All-gather the input tensor interleavely across model parallel group."""
+    import torch.distributed as dist
+
+    gathered_tensors = [torch.zeros_like(local_tensor) for _ in range(tp_size)]
+    dist.all_gather(
+        gathered_tensors,
+        local_tensor,
+        group=parallel_state.get_tp_group().device_group,
+    )
+
+    gathered_tensors_split = [
+        torch.split(tensor, hidden_size // tp_size, -1)
+        for tensor in gathered_tensors
+    ]
+    ordered_tensors = [
+        tensor for pair in zip(*gathered_tensors_split) for tensor in pair
+    ]
+    result_tensor = torch.cat(ordered_tensors, dim=-1)
+    return result_tensor
+
+
+class Glm4vVisionAttention(nn.Module):
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        projection_size: int,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        # Per attention head and per partition values.
+        self.tp_size = parallel_state.get_tensor_model_parallel_world_size()
+        self.tp_rank = parallel_state.get_tensor_model_parallel_rank()
+        self.hidden_size_per_attention_head = dist_utils.divide(
+            projection_size, num_heads)
+        self.num_attention_heads_per_partition = dist_utils.divide(
+            num_heads, self.tp_size)
+
+        self.qkv = QKVParallelLinear(
+            hidden_size=embed_dim,
+            head_size=self.hidden_size_per_attention_head,
+            total_num_heads=num_heads,
+            total_num_kv_heads=num_heads,
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.qkv",
+        )
+        self.proj = RowParallelLinear(
+            input_size=projection_size,
+            output_size=embed_dim,
+            quant_config=quant_config,
+            prefix=f"{prefix}.proj",
+            bias=False,
+        )
+
+        # Detect attention implementation.
+        self.attn_backend: _Backend = get_vit_attn_backend(support_fa=True)
+        if self.attn_backend not in {
+                _Backend.FLASH_ATTN,
+                _Backend.TORCH_SDPA,
+                _Backend.XFORMERS,
+        }:
+            raise RuntimeError(
+                f"GLM-4V does not support {self.attn_backend} backend now.")
+
+    def split_qkv(self, qkv: torch.Tensor) -> tuple[torch.Tensor, ...]:
+        # [s, b, 3 * head * head_dim]
+        seq_len, bs, _ = qkv.shape
+        if self.tp_size > 1:
+            qkv = all_gather_interleave(qkv, self.qkv.hidden_size,
+                                        self.tp_size)
+
+        # [s, b, 3 * head * head_dim] -> 3 * [s, b, head * head_dim]
+        q, k, v = qkv.chunk(3, dim=2)
+
+        # 3 * [s, b, head * head_dim]
+        if self.tp_size > 1:
+            splitter = partial(
+                dist_utils.split_tensor_along_last_dim,
+                num_partitions=self.tp_size,
+            )
+            q = splitter(q)[self.tp_rank]
+            k = splitter(k)[self.tp_rank]
+            v = splitter(v)[self.tp_rank]
+
+        # 3 * [s, b, head * head_dim] -> 3 * [s, b, head, head_dim]
+        new_shape = (
+            seq_len,
+            bs,
+            self.num_attention_heads_per_partition,
+            self.hidden_size_per_attention_head,
+        )
+        q, k, v = (x.view(*new_shape) for x in (q, k, v))
+        return q, k, v
+
+    def forward(
+            self,
+            x: torch.Tensor,
+            cu_seqlens: torch.Tensor,
+            rotary_pos_emb: torch.Tensor,
+            max_seqlen: Optional[int] = None,  # Only used for Flash Attention
+            seqlens: Optional[list[int]] = None,  # Only used for xFormers
+    ) -> torch.Tensor:
+        # [s, b, c] --> [s, b, head * 3 * head_dim]
+        x, _ = self.qkv(x)
+
+        # [s, b, 3 * head * head_dim] -> 3 * [s, b, head, head_dim]
+        q, k, v = self.split_qkv(x)
+        batch_size = q.shape[1]
+
+        q, k, v = (rearrange(x, "s b ... -> b s ...").contiguous()
+                   for x in (q, k, v))
+        if rotary_pos_emb is not None:
+            q = apply_rotary_pos_emb_vision(q, rotary_pos_emb)
+            k = apply_rotary_pos_emb_vision(k, rotary_pos_emb)
+
+        if self.attn_backend == _Backend.FLASH_ATTN:
+            # from vllm_flash_attn.flash_attn_interface import (
+            #   flash_attn_varlen_func)
+            from flash_attn import flash_attn_varlen_func
+
+            q, k, v = (rearrange(x, "b s ... -> (b s) ...") for x in [q, k, v])
+
+            output = flash_attn_varlen_func(
+                q,
+                k,
+                v,
+                cu_seqlens_q=cu_seqlens,
+                cu_seqlens_k=cu_seqlens,
+                max_seqlen_q=max_seqlen,
+                max_seqlen_k=max_seqlen,
+                dropout_p=0,
+                causal=False,
+            )
+
+            context_layer = rearrange(output,
+                                      "(b s) ... -> b s ...",
+                                      b=batch_size)
+        elif self.attn_backend == _Backend.TORCH_SDPA:
+            # Execute attention entry by entry for speed & less VRAM.
+            outputs = []
+            for i in range(1, len(cu_seqlens)):
+                start_idx = cu_seqlens[i - 1]
+                end_idx = cu_seqlens[i]
+                q_i = q[:, start_idx:end_idx]
+                k_i = k[:, start_idx:end_idx]
+                v_i = v[:, start_idx:end_idx]
+                q_i, k_i, v_i = (rearrange(x, "b s h d -> b h s d")
+                                 for x in [q_i, k_i, v_i])
+                output_i = F.scaled_dot_product_attention(q_i,
+                                                          k_i,
+                                                          v_i,
+                                                          dropout_p=0.0)
+                output_i = rearrange(output_i, "b h s d -> b s h d ")
+                outputs.append(output_i)
+            context_layer = torch.cat(outputs, dim=1)
+        elif self.attn_backend == _Backend.XFORMERS:
+            from xformers import ops as xops
+            from xformers.ops.fmha.attn_bias import BlockDiagonalMask
+
+            attn_bias = BlockDiagonalMask.from_seqlens(q_seqlen=seqlens,
+                                                       kv_seqlen=None,
+                                                       device=q.device)
+
+            context_layer = xops.memory_efficient_attention_forward(
+                q, k, v, attn_bias=attn_bias, p=0, scale=None)
+
+        context_layer = rearrange(context_layer,
+                                  "b s h d -> s b (h d)").contiguous()
+
+        output, _ = self.proj(context_layer)
+        return output
+
+
+class Glm4vVisionBlock(nn.Module):
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        mlp_hidden_dim: int,
+        norm_layer: Optional[Callable[[int], nn.Module]] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        if norm_layer is None:
+            norm_layer = partial(nn.LayerNorm, eps=1e-6)
+        self.norm1 = norm_layer(dim)
+        self.norm2 = norm_layer(dim)
+        self.attn = Glm4vVisionAttention(
+            embed_dim=dim,
+            num_heads=num_heads,
+            projection_size=dim,
+            quant_config=quant_config,
+            prefix=f"{prefix}.attn",
+        )
+        self.mlp = Glm4vVisionMLP(
+            dim,
+            mlp_hidden_dim,
+            bias=False,
+            quant_config=quant_config,
+        )
+
+    def forward(
+            self,
+            x: torch.Tensor,
+            cu_seqlens: torch.Tensor,
+            rotary_pos_emb: torch.Tensor,
+            max_seqlen: Optional[int] = None,  # Only used for Flash Attention
+            seqlens: Optional[list[int]] = None,  # Only used for xFormers
+    ) -> torch.Tensor:
+        x = x + self.attn(
+            self.norm1(x),
+            cu_seqlens=cu_seqlens,
+            rotary_pos_emb=rotary_pos_emb,
+            max_seqlen=max_seqlen,
+            seqlens=seqlens,
+        )
+
+        x = x + self.mlp(self.norm2(x))
+        return x
+
+
+class Glm4vVisionPatchEmbed(nn.Module):
+
+    def __init__(
+        self,
+        patch_size: int = 14,
+        temporal_patch_size: int = 1,
+        in_channels: int = 3,
+        hidden_size: int = 1536,
+    ) -> None:
+        super().__init__()
+        self.patch_size = patch_size
+        self.temporal_patch_size = temporal_patch_size
+        self.hidden_size = hidden_size
+
+        kernel_size = (temporal_patch_size, patch_size, patch_size)
+        self.proj = nn.Conv3d(
+            in_channels,
+            hidden_size,
+            kernel_size=kernel_size,
+            stride=kernel_size,
+            bias=True,
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        L, C = x.shape
+        x = x.view(L, -1, self.temporal_patch_size, self.patch_size,
+                   self.patch_size)
+        x = self.proj(x).view(L, self.hidden_size)
+        return x
+
+
+class Glm4vPatchMerger(nn.Module):
+
+    def __init__(
+        self,
+        d_model: int,
+        context_dim: int,
+        quant_config: Optional[QuantizationConfig] = None,
+        bias: bool = False,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = d_model
+        self.proj = ColumnParallelLinear(self.hidden_size,
+                                         self.hidden_size,
+                                         bias=bias,
+                                         gather_output=True)
+        self.post_projection_norm = nn.LayerNorm(self.hidden_size)
+        self.gate_up_proj = MergedColumnParallelLinear(
+            input_size=self.hidden_size,
+            output_sizes=[context_dim] * 2,
+            bias=bias,
+            quant_config=quant_config,
+        )
+        self.down_proj = RowParallelLinear(
+            context_dim,
+            self.hidden_size,
+            bias=bias,
+            quant_config=quant_config,
+        )
+        self.act_fn = SiluAndMul()
+        self.extra_activation_func = nn.GELU()
+
+    def forward(self, x: torch.Tensor):
+        x, _ = self.proj(x)
+        x = self.extra_activation_func(self.post_projection_norm(x))
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x, _ = self.down_proj(x)
+        return x
+
+
+class Glm4vVisionEmbeddings(nn.Module):
+
+    def __init__(self, config: Glm4vVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.num_patches = (self.image_size // self.patch_size)**2
+        self.num_positions = self.num_patches
+        self.position_embedding = nn.Embedding(self.num_positions,
+                                               self.embed_dim)
+        self.register_buffer(
+            "position_ids",
+            torch.arange(self.num_positions).expand((1, -1)),
+            persistent=False,
+        )
+
+    def forward(self, embeddings, lengths, image_shapes, h_coords,
+                w_coords) -> torch.Tensor:
+        pos_embed_weight = self.position_embedding.weight
+        hidden_size = pos_embed_weight.shape[1]
+        total_seq = h_coords.shape[0]
+        device = pos_embed_weight.device
+
+        # Move coordinates to correct device
+        h_coords, w_coords = h_coords.to(device), w_coords.to(device)
+
+        # Handle empty sequence case
+        if total_seq == 0:
+            adapted_pos_embed = torch.empty(0,
+                                            hidden_size,
+                                            device=device,
+                                            dtype=pos_embed_weight.dtype)
+        else:
+            # Convert inputs to tensors if needed
+            if isinstance(lengths, list):
+                lengths = torch.tensor(lengths,
+                                       device=device,
+                                       dtype=torch.long)
+            if not isinstance(image_shapes, torch.Tensor):
+                image_shapes = torch.tensor(image_shapes,
+                                            device=device,
+                                            dtype=torch.long)
+
+            # Prepare 2D position embedding
+            orig_size_sq = pos_embed_weight.shape[0]
+            orig_size = int(orig_size_sq**0.5)
+            pos_embed_2d = (pos_embed_weight.view(
+                orig_size, orig_size,
+                hidden_size).permute(2, 0,
+                                     1).unsqueeze(0).to(device=device,
+                                                        dtype=torch.float32))
+
+            # Calculate target dimensions for each patch
+            target_h = torch.cat([
+                image_shapes[i, 1].repeat(lengths[i])
+                for i in range(len(lengths))
+            ]).to(device=device, dtype=torch.float32)
+            target_w = torch.cat([
+                image_shapes[i, 2].repeat(lengths[i])
+                for i in range(len(lengths))
+            ]).to(device=device, dtype=torch.float32)
+
+            # Normalize coordinates to [-1, 1] range for grid_sample
+            h_coords = h_coords.to(device=device, dtype=torch.float32)
+            w_coords = w_coords.to(device=device, dtype=torch.float32)
+            norm_w = ((w_coords + 0.5) / target_w) * 2 - 1
+            norm_h = ((h_coords + 0.5) / target_h) * 2 - 1
+
+            # Create sampling grid
+            grid = (torch.stack((norm_w, norm_h),
+                                dim=-1).unsqueeze(0).unsqueeze(2))
+
+            # Perform bicubic interpolation
+            interpolated_embed_fp32 = F.grid_sample(
+                pos_embed_2d,
+                grid,
+                mode="bicubic",
+                align_corners=False,
+                padding_mode="border",
+            )
+
+            # Reshape and convert back to original dtype
+            adapted_pos_embed_fp32 = (
+                interpolated_embed_fp32.squeeze(0).squeeze(-1).permute(1, 0))
+            adapted_pos_embed = adapted_pos_embed_fp32.to(
+                pos_embed_weight.dtype).to(embeddings.device)
+
+        # Add adapted position encoding to embeddings
+        embeddings = embeddings + adapted_pos_embed
+        return embeddings
+
+
+class Glm4vVisionRotaryEmbedding(nn.Module):
+
+    def __init__(self, dim: int, theta: float = 10000.0) -> None:
+        super().__init__()
+        self.dim = dim
+        self.theta = theta
+        inv_freq = 1.0 / (theta
+                          **(torch.arange(0, dim, 2, dtype=torch.float) / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self._seq_len_cached = 0
+        self._freqs_cached = None
+
+    def update_freqs_cache(self, seqlen: int) -> None:
+        if seqlen > self._seq_len_cached:
+            seqlen *= 2
+            self._seq_len_cached = seqlen
+            self.inv_freq = 1.0 / (self.theta**(torch.arange(
+                0,
+                self.dim,
+                2,
+                dtype=torch.float,
+                device=self.inv_freq.device,
+            ) / self.dim))
+            seq = torch.arange(seqlen,
+                               device=self.inv_freq.device,
+                               dtype=self.inv_freq.dtype)
+            freqs = torch.outer(seq, self.inv_freq)
+            self._freqs_cached = freqs
+
+    def forward(self, seqlen: int) -> torch.Tensor:
+        self.update_freqs_cache(seqlen)
+        return self._freqs_cached[:seqlen]
+
+
+class Glm4vVisionTransformer(nn.Module):
+
+    def __init__(
+        self,
+        vision_config: Glm4vVisionConfig,
+        norm_eps: float = 1e-6,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+
+        patch_size = vision_config.patch_size
+        temporal_patch_size = vision_config.temporal_patch_size
+        in_channels = vision_config.in_channels
+        depth = vision_config.depth
+        self.hidden_size = vision_config.hidden_size
+        self.num_heads = vision_config.num_heads
+
+        self.patch_size = vision_config.patch_size
+        self.spatial_merge_size = vision_config.spatial_merge_size
+        self.out_hidden_size = vision_config.out_hidden_size
+
+        self.patch_embed = Glm4vVisionPatchEmbed(
+            patch_size=patch_size,
+            temporal_patch_size=temporal_patch_size,
+            in_channels=in_channels,
+            hidden_size=self.hidden_size,
+        )
+
+        norm_layer = partial(RMSNorm, eps=norm_eps)
+        head_dim = self.hidden_size // self.num_heads
+        self.rotary_pos_emb = Glm4vVisionRotaryEmbedding(head_dim // 2)
+        self.blocks = nn.ModuleList([
+            Glm4vVisionBlock(
+                dim=self.hidden_size,
+                num_heads=self.num_heads,
+                mlp_hidden_dim=vision_config.out_hidden_size,
+                norm_layer=norm_layer,
+                quant_config=quant_config,
+                prefix=f"{prefix}.blocks.{layer_idx}",
+            ) for layer_idx in range(depth)
+        ])
+        self.merger = Glm4vPatchMerger(
+            d_model=vision_config.out_hidden_size,
+            context_dim=vision_config.intermediate_size,
+            quant_config=quant_config,
+            bias=False,
+        )
+        self.embeddings = Glm4vVisionEmbeddings(vision_config)
+
+        self.post_conv_layernorm = RMSNorm(vision_config.hidden_size,
+                                           eps=vision_config.rms_norm_eps)
+        self.downsample = nn.Conv2d(
+            in_channels=vision_config.hidden_size,
+            out_channels=vision_config.out_hidden_size,
+            kernel_size=vision_config.spatial_merge_size,
+            stride=vision_config.spatial_merge_size,
+        )
+        self.post_layernorm = RMSNorm(vision_config.hidden_size,
+                                      eps=vision_config.rms_norm_eps)
+
+        self.attn_backend: _Backend = get_vit_attn_backend(support_fa=True)
+
+    @property
+    def dtype(self) -> torch.dtype:
+        return self.patch_embed.proj.weight.dtype
+
+    @property
+    def device(self) -> torch.device:
+        return self.patch_embed.proj.weight.device
+
+    def rot_pos_emb(self, grid_thw: torch.Tensor) -> torch.Tensor:
+        pos_ids = []
+        for t, h, w in grid_thw:
+            hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
+            wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
+            hpos_ids = (hpos_ids.reshape(
+                h // self.spatial_merge_size,
+                self.spatial_merge_size,
+                w // self.spatial_merge_size,
+                self.spatial_merge_size,
+            ).permute(0, 2, 1, 3).flatten())
+            wpos_ids = (wpos_ids.reshape(
+                h // self.spatial_merge_size,
+                self.spatial_merge_size,
+                w // self.spatial_merge_size,
+                self.spatial_merge_size,
+            ).permute(0, 2, 1, 3).flatten())
+            pos_ids.append(
+                torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
+        pos_ids = torch.cat(pos_ids, dim=0)
+        max_grid_size = grid_thw[:, 1:].max()
+        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
+        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
+        return rotary_pos_emb, pos_ids
+
+    def compute_attn_mask_seqlen(
+        self,
+        cu_seqlens: torch.Tensor,
+    ) -> tuple[Optional[int], Optional[list[int]]]:
+        max_seqlen, seqlens = None, None
+        seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
+        if self.attn_backend == _Backend.FLASH_ATTN:
+            max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
+        return max_seqlen, seqlens
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        grid_thw: torch.Tensor,
+    ) -> torch.Tensor:
+        # patchify
+        x = x.to(device=self.device, dtype=self.dtype)
+        x = self.patch_embed(x)
+        x = self.post_conv_layernorm(x)
+
+        # compute position embedding
+        rotary_pos_emb, image_type_ids = self.rot_pos_emb(grid_thw)
+        # compute cu_seqlens
+        cu_seqlens = torch.repeat_interleave(grid_thw[:, 1] * grid_thw[:, 2],
+                                             grid_thw[:, 0]).cumsum(
+                                                 dim=0, dtype=torch.int32)
+        cu_seqlens = F.pad(cu_seqlens, (1, 0), "constant", 0)
+
+        # pre-compute seqlens for attn mask to reduce cuMemcpy operations
+        max_seqlen, seqlens = self.compute_attn_mask_seqlen(cu_seqlens)
+        x = self.embeddings(x, seqlens, grid_thw, image_type_ids[:, 0],
+                            image_type_ids[:, 1])
+
+        # transformers
+        x = x.unsqueeze(1)
+        for blk in self.blocks:
+            x = blk(
+                x,
+                cu_seqlens=cu_seqlens,
+                rotary_pos_emb=rotary_pos_emb,
+                max_seqlen=max_seqlen,
+                seqlens=seqlens,
+            )
+
+        # adapter
+        x = self.post_layernorm(x)
+
+        x = x.view(-1, self.spatial_merge_size, self.spatial_merge_size,
+                   x.shape[-1])
+        x = x.permute(0, 3, 1, 2)
+        x = self.downsample(x).view(-1, self.out_hidden_size)
+        x = self.merger(x)
+
+        return x
+
+    def load_weights(self, weights: Iterable[tuple[str,
+                                                   torch.Tensor]]) -> set[str]:
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            ("attn.qkv.", "attn.q.", "q"),
+            ("attn.qkv.", "attn.k.", "k"),
+            ("attn.qkv.", "attn.v.", "v"),
+            ("gate_up_proj", "gate_proj", 0),
+            ("gate_up_proj", "up_proj", 1),
+        ]
+        params_dict = dict(self.named_parameters(remove_duplicate=False))
+        loaded_params: set[str] = set()
+
+        for name, loaded_weight in weights:
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+                if weight_name not in name:
+                    continue
+                name = name.replace(weight_name, param_name)
+
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params
+
+
+class Glm4vProcessingInfo(BaseProcessingInfo):
+
+    def get_hf_config(self):
+        return self.ctx.get_hf_config(Glm4vConfig)
+
+    def get_tokenizer(self):
+        return self.ctx.tokenizer
+
+    def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
+        return {"image": None, "video": 1}
+
+    def get_image_processor(self) -> Glm4vImageProcessor:
+        return self.get_hf_processor().image_processor
+
+    def get_video_processor(self) -> Glm4vVideoProcessor:
+        return self.get_hf_processor().video_processor
+
+    def _get_vision_info(
+        self,
+        *,
+        image_width: int,
+        image_height: int,
+        num_frames: int = 16,
+        do_resize: bool = True,
+        max_image_pixels: int = 28 * 28 * 2 * 30000,
+    ) -> tuple[ImageSize, int]:
+        hf_config = self.get_hf_config()
+        vision_config = hf_config.vision_config
+        patch_size = vision_config.patch_size
+        merge_size = vision_config.spatial_merge_size
+        temporal_patch_size = vision_config.temporal_patch_size
+        if do_resize:
+            resized_height, resized_width = smart_resize(
+                num_frames=num_frames
+                if num_frames > temporal_patch_size else temporal_patch_size,
+                height=image_height,
+                width=image_width,
+                factor=patch_size * merge_size,
+                max_pixels=max_image_pixels,
+            )
+            preprocessed_size = ImageSize(width=resized_width,
+                                          height=resized_height)
+        else:
+            preprocessed_size = ImageSize(width=image_width,
+                                          height=image_height)
+
+        # NOTE: Frames are padded to be divisible by `temporal_patch_size`
+        # https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/qwen2_vl/image_processing_qwen2_vl.py#L294
+        padded_num_frames = num_frames + num_frames % temporal_patch_size
+
+        grid_t = max(padded_num_frames // temporal_patch_size, 1)
+        grid_h = preprocessed_size.height // patch_size
+        grid_w = preprocessed_size.width // patch_size
+
+        num_patches = grid_t * grid_h * grid_w
+        num_vision_tokens = num_patches // (merge_size**2)
+
+        return preprocessed_size, num_vision_tokens
+
+    def get_image_size_with_most_features(self) -> ImageSize:
+        max_image_size, _ = self._get_vision_info(image_width=9999999,
+                                                  image_height=9999999)
+        return max_image_size
+
+    def get_num_image_tokens(
+        self,
+        *,
+        image_width: int,
+        image_height: int,
+    ) -> int:
+        _, num_image_tokens = self._get_vision_info(
+            image_width=image_width,
+            image_height=image_height,
+            max_image_pixels=28 * 28 * 2 * 6144,
+        )
+        return num_image_tokens
+
+    def get_max_image_tokens(self) -> int:
+        target_width, target_height = self.get_image_size_with_most_features()
+
+        return self.get_num_image_tokens(
+            image_width=target_width,
+            image_height=target_height,
+        )
+
+    def get_num_video_tokens(
+        self,
+        *,
+        image_width: int,
+        image_height: int,
+        num_frames: int,
+    ) -> int:
+        _, num_video_tokens = self._get_vision_info(
+            image_width=image_width,
+            image_height=image_height,
+            num_frames=num_frames,
+            max_image_pixels=28 * 28 * 2 * 30000,
+        )
+        return num_video_tokens
+
+    def _get_max_video_frames(self, max_tokens: int) -> int:
+        target_width, target_height = self.get_image_size_with_most_features()
+
+        num_frames = 0
+
+        while True:
+            next_num_frames = num_frames + 1
+            next_max_tokens = self.get_num_video_tokens(
+                image_width=target_width,
+                image_height=target_height,
+                num_frames=next_num_frames,
+            )
+            if next_max_tokens > max_tokens or next_max_tokens == 0:
+                break
+
+            num_frames = next_num_frames
+
+        return num_frames
+
+    def get_num_frames_with_most_features(
+        self,
+        seq_len: int,
+        mm_counts: Mapping[str, int],
+    ) -> int:
+        max_images = mm_counts.get("image", 0)
+        max_videos = mm_counts.get("video", 0)
+
+        max_image_tokens = self.get_max_image_tokens() * max_images
+        max_total_frames = self._get_max_video_frames(seq_len -
+                                                      max_image_tokens)
+        max_frames_per_video = min(max_total_frames // max(max_videos, 1),
+                                   _MAX_FRAMES_PER_VIDEO)
+
+        return max(max_frames_per_video, 1)
+
+    def _get_video_second_idx(self, metadata: dict[str, Any],
+                              total_frames: int) -> list[int]:
+        video_processor = self.get_video_processor()
+
+        video_fps = metadata.get("fps", 2.0)
+        meta_frames = metadata.get("total_num_frames", total_frames)
+        max_frame_idx = meta_frames - 1
+        duration = metadata.get("duration",
+                                round(max_frame_idx / video_fps) + 1)
+        if duration <= video_processor.max_duration:
+            n = int(math.floor(duration * video_processor.fps))
+            frame_indices = [
+                min(
+                    max_frame_idx,
+                    int(math.ceil(i * video_fps / video_processor.fps)),
+                ) for i in range(n)
+            ]
+        else:
+            num_samples = int(video_processor.max_duration *
+                              video_processor.fps)
+            if num_samples >= meta_frames:
+                frame_indices = list(range(meta_frames))
+            else:
+                target_seconds = np.linspace(0,
+                                             duration,
+                                             num_samples,
+                                             endpoint=True)
+                frame_indices = [
+                    min(max_frame_idx, int(math.ceil(t * video_fps)))
+                    for t in target_seconds
+                ]
+
+        seen, uniq = set(), []
+        for idx in frame_indices:
+            if idx not in seen:
+                seen.add(idx)
+                uniq.append(idx)
+        if len(uniq) & 1:
+            uniq.append(uniq[-1])
+        frame_indices = uniq
+
+        full_second_idxs = [int(idx / video_fps) for idx in frame_indices]
+        timestamps_list = full_second_idxs[::2]
+        selected_timestamps = []
+        for idx in range(0, len(timestamps_list)):
+            selected_timestamps.append(timestamps_list[idx])
+        return selected_timestamps
+
+
+class Glm4vDummyInputsBuilder(BaseDummyInputsBuilder[Glm4vProcessingInfo]):
+
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        num_images = mm_counts.get("image", 0)
+        num_videos = mm_counts.get("video", 0)
+
+        hf_config = self.info.get_hf_config()
+        hf_processor = self.info.get_hf_processor()
+        tokenizer = self.info.get_tokenizer()
+
+        image_token: str = hf_processor.image_token
+        video_token_ids = [
+            hf_config.video_start_token_id,
+            hf_processor.video_token_id,
+            hf_config.video_end_token_id,
+        ]
+        video_token = tokenizer.decode(video_token_ids)
+
+        return image_token * num_images + video_token * num_videos
+
+    def get_dummy_mm_data(
+        self,
+        seq_len: int,
+        mm_counts: Mapping[str, int],
+    ) -> MultiModalDataDict:
+        num_images = mm_counts.get("image", 0)
+        num_videos = mm_counts.get("video", 0)
+
+        target_width, target_height = (
+            self.info.get_image_size_with_most_features())
+        target_num_frames = self.info.get_num_frames_with_most_features(
+            seq_len, mm_counts)
+        return {
+            "image":
+            self._get_dummy_images(width=target_width,
+                                   height=target_height,
+                                   num_images=num_images),
+            "video":
+            self._get_dummy_videos(
+                width=target_width,
+                height=target_height,
+                num_frames=target_num_frames,
+                num_videos=num_videos,
+            ),
+        }
+
+    def _get_dummy_videos(
+        self,
+        *,
+        width: int,
+        height: int,
+        num_frames: int,
+        num_videos: int,
+    ) -> list[VideoItem]:
+        video = np.full((num_frames, width, height, 3), 255, dtype=np.uint8)
+        video_items = []
+        for i in range(num_videos):
+            video_metadata = {
+                "fps": 2.0,
+                "duration": num_frames / 2.0,
+                "total_num_frames": num_frames,
+                "video_backend": "opencv",
+            }
+            video_item = (video.copy(), video_metadata)
+            video_items.append(video_item)
+
+        return video_items
+
+
+class Glm4vMultiModalProcessor(BaseMultiModalProcessor[Glm4vProcessingInfo]):
+
+    def _get_data_parser(self) -> MultiModalDataParser:
+        return MultiModalDataParser(video_needs_metadata=True)
+
+    def _call_hf_processor(
+        self,
+        prompt: str,
+        mm_data: Mapping[str, object],
+        mm_kwargs: Mapping[str, object],
+        tok_kwargs: Mapping[str, object],
+    ) -> BatchFeature:
+        mm_data = dict(mm_data)
+        processor = self.info.get_hf_processor(**mm_kwargs)
+
+        # GLM-4.1V use `image_token_id` as video placeholder, we need to
+        # replace it with `video_token_id` for video processing. So we
+        # separate video processing from image processing.
+        if ("videos" in mm_data and isinstance(mm_data["videos"], list)
+                and len(mm_data["videos"]) > 0):
+            video_grid_thw_lst = []
+            pixel_values_videos_lst = []
+            for item in mm_data.pop("videos", []):
+                video_array, metadata = item
+
+                # FIXME(Isotr0py): Activate the below logic after we can disable
+                # resampling from video loader backend.
+                # assert metadata["total_num_frames"] == len(video_array), (
+                #     f"Total frames {metadata['total_num_frames']} does not "
+                #     f"match the length of video array {len(video_array)}.")
+
+                # NOTE: Temporary workaround for resampled videos.
+                # this can cause a divergence with HF implementation if
+                # the input video is resampled in advance.
+
+                if metadata["total_num_frames"] != len(video_array):
+                    logger.warning(
+                        "Total frames in metadata "
+                        "(%s) does not match the length of "
+                        "video array %s. This can "
+                        "be because the video is resampled "
+                        "in advance. This may cause "
+                        "a divergence with HF implementation.",
+                        metadata["total_num_frames"],
+                        len(video_array),
+                    )
+                    metadata["total_num_frames"] = len(video_array)
+                metadata = VideoMetadata(**metadata)
+
+                video_mm_data = dict()
+                video_mm_data["videos"] = [[video_array]]
+                video_mm_data["video_metadata"] = [[metadata]]
+
+                video_outputs = super()._call_hf_processor(
+                    prompt="<|begin_of_video|><|video|><|end_of_video|>",
+                    mm_data=video_mm_data,
+                    mm_kwargs=mm_kwargs,
+                    tok_kwargs=tok_kwargs,
+                )
+                input_ids = video_outputs.pop("input_ids")
+                input_ids[input_ids == processor.image_token_id] = (
+                    processor.video_token_id)
+                video_placeholder = processor.tokenizer.batch_decode(
+                    input_ids)[0]
+                prompt = prompt.replace(
+                    "<|begin_of_video|><|video|><|end_of_video|>",
+                    video_placeholder,
+                )
+
+                grid_t = len(video_outputs["video_grid_thw"])
+                _, grid_h, grid_w = video_outputs["video_grid_thw"][0]
+                grid_thw = torch.tensor([[grid_t, grid_h, grid_w]])
+
+                video_grid_thw_lst.append(grid_thw)
+                pixel_values_videos_lst.append(
+                    video_outputs["pixel_values_videos"])
+            video_outputs = dict(
+                pixel_values_videos=torch.cat(pixel_values_videos_lst),
+                video_grid_thw=torch.cat(video_grid_thw_lst),
+            )
+        else:
+            video_outputs = dict()
+
+        processed_outputs = super()._call_hf_processor(
+            prompt=prompt,
+            mm_data=mm_data,
+            mm_kwargs=mm_kwargs,
+            tok_kwargs=tok_kwargs,
+        )
+        combined_outputs = dict(
+            processed_outputs,
+            **video_outputs,
+        )
+        return BatchFeature(combined_outputs)
+
+    def _get_mm_fields_config(
+        self,
+        hf_inputs: BatchFeature,
+        hf_processor_mm_kwargs: Mapping[str, object],
+    ) -> Mapping[str, MultiModalFieldConfig]:
+        return _qwen2vl_field_config(hf_inputs)
+
+    def _get_prompt_updates(
+        self,
+        mm_items: MultiModalDataItems,
+        hf_processor_mm_kwargs: Mapping[str, Any],
+        out_mm_kwargs: MultiModalKwargs,
+    ) -> Sequence[PromptUpdate]:
+        hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
+        image_processor = self.info.get_image_processor(
+            **hf_processor_mm_kwargs)
+        tokenizer = self.info.get_tokenizer()
+        hf_config = self.info.get_hf_config()
+
+        boi_token_id = hf_config.image_start_token_id
+        eoi_token_id = hf_config.image_end_token_id
+
+        bov_token_id = hf_config.video_start_token_id
+        eov_token_id = hf_config.video_end_token_id
+
+        merge_length = image_processor.merge_size**2
+
+        def get_image_replacement_glm4v(item_idx: int):
+            grid_thw = out_mm_kwargs["image_grid_thw"][item_idx]
+            assert isinstance(grid_thw, torch.Tensor)
+
+            num_tokens = int(grid_thw.prod()) // merge_length
+            return [hf_processor.image_token_id] * num_tokens
+
+        def get_video_replacement_glm4v(item_idx: int):
+            grid_thw = out_mm_kwargs["video_grid_thw"][item_idx]
+            assert isinstance(grid_thw, torch.Tensor)
+
+            video, metadata = mm_items["video"][item_idx]
+            timestamps = self.info._get_video_second_idx(metadata, len(video))
+            frames_idx_token = [
+                tokenizer.encode(str(i), add_special_tokens=False)
+                for i in timestamps
+            ]
+            num_tokens_per_frame = int(grid_thw[1:].prod()) // merge_length
+            placeholder = []
+            placeholder.append(bov_token_id)
+            for frame_idx in frames_idx_token:
+                placeholder.append(boi_token_id)
+                placeholder.extend([hf_processor.video_token_id] *
+                                   num_tokens_per_frame)
+                placeholder.append(eoi_token_id)
+                placeholder.extend(frame_idx)
+            placeholder.append(eov_token_id)
+            return placeholder
+
+        return [
+            PromptReplacement(
+                modality="image",
+                target=hf_processor.image_token,
+                replacement=get_image_replacement_glm4v,
+            ),
+            PromptReplacement(
+                modality="video",
+                target="<|begin_of_video|><|video|><|end_of_video|>",
+                replacement=get_video_replacement_glm4v,
+            ),
+        ]
+
+
+@MULTIMODAL_REGISTRY.register_processor(
+    Glm4vMultiModalProcessor,
+    info=Glm4vProcessingInfo,
+    dummy_inputs=Glm4vDummyInputsBuilder,
+)
+class Glm4vForConditionalGeneration(nn.Module, SupportsMultiModal,
+                                    SupportsLoRA, SupportsPP):
+    packed_modules_mapping = {
+        "qkv_proj": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+        ],
+        "gate_up_proj": [
+            "gate_proj",
+            "up_proj",
+        ],
+    }
+
+    # To ensure correct weight loading and mapping.
+    hf_to_vllm_mapper = WeightsMapper(
+        orig_to_new_prefix={
+            "lm_head.": "language_model.lm_head.",
+            "model.language_model.": "language_model.model.",
+            "model.visual.": "visual.",
+        })
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        config: Glm4vConfig = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
+        multimodal_config = vllm_config.model_config.multimodal_config
+
+        self.config = config
+        self.multimodal_config = multimodal_config
+
+        self.visual = Glm4vVisionTransformer(
+            config.vision_config,
+            norm_eps=getattr(config, "rms_norm_eps", 1e-5),
+            quant_config=self._maybe_ignore_quant_config(quant_config),
+            prefix=maybe_prefix(prefix, "visual"),
+        )
+
+        self.language_model = init_vllm_registered_model(
+            vllm_config=vllm_config,
+            prefix=maybe_prefix(prefix, ""),
+            architectures=["Glm4ForCausalLM"],
+        )
+
+        self.make_empty_intermediate_tensors = (
+            self.language_model.make_empty_intermediate_tensors)
+
+    def _maybe_ignore_quant_config(self, quant_config: QuantizationConfig):
+        # GPTQ configs do not have a list of ignored modules, however AutoGPTQ
+        # seems to avoid vision encoder sections for some models.
+        if isinstance(quant_config, (GPTQConfig, GPTQMarlinConfig)):
+            return None
+        return quant_config
+
+    def _validate_and_reshape_mm_tensor(self, mm_input: object,
+                                        name: str) -> torch.Tensor:
+        if not isinstance(mm_input, (torch.Tensor, list)):
+            raise ValueError(
+                f"Incorrect type of {name}. Got type: {type(mm_input)}")
+        if isinstance(mm_input, torch.Tensor):
+            if mm_input.ndim == 2:
+                return mm_input
+            if mm_input.ndim != 3:
+                raise ValueError(f"{name} should be 2D or batched 3D tensor. "
+                                 f"Got ndim: {mm_input.ndim} "
+                                 f"(shape={mm_input.shape})")
+            return torch.concat(list(mm_input))
+        else:
+            return torch.concat(mm_input)
+
+    def _parse_and_validate_image_input(
+            self, **kwargs: object) -> Optional[Glm4vImageInputs]:
+        pixel_values = kwargs.pop("pixel_values", None)
+        image_embeds = kwargs.pop("image_embeds", None)
+        image_grid_thw = kwargs.pop("image_grid_thw", None)
+
+        if pixel_values is None and image_embeds is None:
+            return None
+
+        if pixel_values is not None:
+            pixel_values = self._validate_and_reshape_mm_tensor(
+                pixel_values, "image pixel values")
+            image_grid_thw = self._validate_and_reshape_mm_tensor(
+                image_grid_thw, "image grid_thw")
+
+            if not isinstance(pixel_values, (torch.Tensor, list)):
+                raise ValueError("Incorrect type of image pixel values. "
+                                 f"Got type: {type(pixel_values)}")
+
+            return Glm4vImagePixelInputs(
+                type="pixel_values",
+                pixel_values=pixel_values,
+                image_grid_thw=image_grid_thw,
+            )
+
+        if image_embeds is not None:
+            image_embeds = self._validate_and_reshape_mm_tensor(
+                image_embeds, "image embeds")
+            image_grid_thw = self._validate_and_reshape_mm_tensor(
+                image_grid_thw, "image grid_thw")
+
+            if not isinstance(image_embeds, torch.Tensor):
+                raise ValueError("Incorrect type of image embeddings. "
+                                 f"Got type: {type(image_embeds)}")
+            return Glm4vImageEmbeddingInputs(
+                type="image_embeds",
+                image_embeds=image_embeds,
+                image_grid_thw=image_grid_thw,
+            )
+
+    def _parse_and_validate_video_input(
+            self, **kwargs: object) -> Optional[Glm4vVideoInputs]:
+        pixel_values_videos = kwargs.pop("pixel_values_videos", None)
+        video_embeds = kwargs.pop("video_embeds", None)
+        video_grid_thw = kwargs.pop("video_grid_thw", None)
+        if pixel_values_videos is None and video_embeds is None:
+            return None
+        if pixel_values_videos is not None:
+            pixel_values_videos = self._validate_and_reshape_mm_tensor(
+                pixel_values_videos, "video pixel values")
+            video_grid_thw = self._validate_and_reshape_mm_tensor(
+                video_grid_thw, "video grid_thw")
+
+            return Glm4vVideoPixelInputs(
+                type="pixel_values_videos",
+                # video_metadata=video_metadata,
+                pixel_values_videos=pixel_values_videos,
+                video_grid_thw=video_grid_thw,
+            )
+
+        if video_embeds is not None:
+            video_embeds = self._validate_and_reshape_mm_tensor(
+                video_embeds, "video embeds")
+            video_grid_thw = self._validate_and_reshape_mm_tensor(
+                video_grid_thw, "video grid_thw")
+
+            if not isinstance(video_embeds, torch.Tensor):
+                raise ValueError("Incorrect type of video embeddings. "
+                                 f"Got type: {type(video_embeds)}")
+            return Glm4vVideoEmbeddingInputs(
+                type="video_embeds",
+                video_embeds=video_embeds,
+                video_grid_thw=video_grid_thw,
+            )
+
+    def _process_image_input(
+            self, image_input: Glm4vImageInputs) -> tuple[torch.Tensor, ...]:
+        grid_thw = image_input["image_grid_thw"]
+        assert grid_thw.ndim == 2
+
+        if image_input["type"] == "image_embeds":
+            image_embeds = image_input["image_embeds"].type(self.visual.dtype)
+        else:
+            pixel_values = image_input["pixel_values"].type(self.visual.dtype)
+            image_embeds = self.visual(pixel_values, grid_thw=grid_thw)
+
+        merge_size = self.visual.spatial_merge_size
+        sizes = grid_thw.prod(-1) // merge_size // merge_size
+        return image_embeds.split(sizes.tolist())
+
+    def _process_video_input(
+            self, video_input: Glm4vVideoInputs) -> tuple[torch.Tensor, ...]:
+        grid_thw = video_input["video_grid_thw"]
+        assert grid_thw.ndim == 2
+
+        device = self.visual.device
+        flat_grid_thw = torch.cat([
+            torch.tensor([[1, h, w]] * t, device=device)
+            for t, h, w in grid_thw
+        ])
+        if video_input["type"] == "video_embeds":
+            video_embeds = video_input["video_embeds"].type(self.visual.dtype)
+        else:
+            pixel_values_videos = video_input["pixel_values_videos"].type(
+                self.visual.dtype)
+            video_embeds = self.visual(pixel_values_videos,
+                                       grid_thw=flat_grid_thw)
+
+        # Split concatenated embeddings for each video item.
+        merge_size = self.visual.spatial_merge_size
+        sizes = grid_thw.prod(-1) // merge_size // merge_size
+
+        return video_embeds.split(sizes.tolist())
+
+    def _parse_and_validate_multimodal_inputs(self, **kwargs: object) -> dict:
+        mm_input_by_modality = {}
+
+        # Preserve the order of modalities if there are multiple of them
+        # from the order of kwargs.
+        for input_key in kwargs:
+            if (input_key in ("pixel_values", "image_embeds")
+                    and "image" not in mm_input_by_modality):
+                mm_input_by_modality["image"] = (
+                    self._parse_and_validate_image_input(**kwargs))
+            if (input_key in ("pixel_values_videos", "video_embeds")
+                    and "video" not in mm_input_by_modality):
+                mm_input_by_modality["video"] = (
+                    self._parse_and_validate_video_input(**kwargs))
+        return mm_input_by_modality
+
+    def get_language_model(self) -> torch.nn.Module:
+        return self.language_model
+
+    def get_multimodal_embeddings(
+            self, **kwargs: object) -> Optional[MultiModalEmbeddings]:
+        mm_input_by_modality = self._parse_and_validate_multimodal_inputs(
+            **kwargs)
+        if not mm_input_by_modality:
+            return None
+
+        # The result multimodal_embeddings is tuple of tensors, with each
+        # tensor correspoending to a multimodal data item (image or video).
+        multimodal_embeddings: tuple[torch.Tensor, ...] = ()
+
+        # NOTE: It is important to iterate over the keys in this dictionary
+        # to preserve the order of the modalities.
+        for modality in mm_input_by_modality:
+            multimodal_input = mm_input_by_modality[modality]
+            if modality == "image":
+                vision_embeddings = self._process_image_input(multimodal_input)
+                multimodal_embeddings += vision_embeddings
+            if modality == "video":
+                video_embeddings = self._process_video_input(multimodal_input)
+                multimodal_embeddings += video_embeddings
+        return multimodal_embeddings
+
+    def get_input_embeddings(
+        self,
+        input_ids: torch.Tensor,
+        multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
+    ) -> torch.Tensor:
+        inputs_embeds = self.language_model.get_input_embeddings(input_ids)
+        if (multimodal_embeddings is not None
+                and len(multimodal_embeddings) != 0
+                and all(embed.numel() > 0 for embed in multimodal_embeddings)):
+            inputs_embeds = merge_multimodal_embeddings(
+                input_ids,
+                inputs_embeds,
+                multimodal_embeddings,
+                [self.config.image_token_id, self.config.video_token_id],
+            )
+        return inputs_embeds
+
+    def get_input_embeddings_v0(
+        self,
+        input_ids: torch.Tensor,
+        image_input: Optional[Glm4vImageInputs] = None,
+        video_input: Optional[Glm4vVideoInputs] = None,
+    ) -> torch.Tensor:
+        inputs_embeds = self.get_input_embeddings(input_ids)
+        if image_input is not None:
+            image_embeds = self._process_image_input(image_input)
+            inputs_embeds = merge_multimodal_embeddings(
+                input_ids,
+                inputs_embeds,
+                image_embeds,
+                placeholder_token_id=self.config.image_token_id,
+            )
+
+        if video_input is not None:
+            video_embeds = self._process_video_input(video_input)
+            inputs_embeds = merge_multimodal_embeddings(
+                input_ids,
+                inputs_embeds,
+                video_embeds,
+                placeholder_token_id=self.config.video_token_id,
+            )
+        return inputs_embeds
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        **kwargs: object,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        """Run forward pass for GLM-4V.
+
+        Args:
+            input_ids: Flattened (concatenated) input_ids corresponding to a
+                batch.
+            positions: Flattened (concatenated) position ids corresponding to a
+                batch.
+                **NOTE**: If mrope is enabled (default setting for GLM-4V
+                opensource models), the shape will be `(3, seq_len)`,
+                otherwise it will be `(seq_len,).
+            pixel_values: Pixel values to be fed to a model.
+                `None` if no images are passed.
+            image_grid_thw: Tensor `(n_images, 3)` of image 3D grid in LLM.
+                `None` if no images are passed.
+            pixel_values_videos: Pixel values of videos to be fed to a model.
+                `None` if no videos are passed.
+            video_grid_thw: Tensor `(n_videos, 3)` of video 3D grid in LLM.
+                `None` if no videos are passed.
+            second_per_grid_ts: Tensor `(num_videos)` of video time interval (
+                in seconds) for each grid along the temporal dimension in the
+                3D position IDs. `None` if no videos are passed.
+        """
+        if intermediate_tensors is not None:
+            inputs_embeds = None
+
+        # NOTE: In v1, inputs_embeds is always generated at model runner from
+        # `get_multimodal_embeddings` and `get_input_embeddings`, this
+        # condition is only for v0 compatibility.
+        elif inputs_embeds is None:
+            image_input = self._parse_and_validate_image_input(**kwargs)
+            video_input = self._parse_and_validate_video_input(**kwargs)
+
+            if image_input is None and video_input is None:
+                inputs_embeds = None
+            else:
+                if uses_mrope(self.config):
+                    assert positions.ndim == 2 and positions.size(0) == 3, (
+                        "multimodal section rotary embedding requires "
+                        f"(3, seq_len) positions, but got {positions.size()}")
+                inputs_embeds = self.get_input_embeddings_v0(
+                    input_ids,
+                    image_input=image_input,
+                    video_input=video_input)
+                input_ids = None
+
+        hidden_states = self.language_model.model(
+            input_ids=input_ids,
+            positions=positions,
+            intermediate_tensors=intermediate_tensors,
+            inputs_embeds=inputs_embeds,
+        )
+        return hidden_states
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[torch.Tensor]:
+        return self.language_model.compute_logits(hidden_states,
+                                                  sampling_metadata)
+
+    def load_weights(self, weights: Iterable[tuple[str,
+                                                   torch.Tensor]]) -> set[str]:
+        loader = AutoWeightsLoader(self)
+        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
+
+    def get_mm_mapping(self) -> MultiModelKeys:
+        """
+        Get the module prefix in multimodal models
+        """
+        return MultiModelKeys.from_string_field(
+            language_model="language_model",
+            connector="visual.merger.",
+            tower_model="visual.",
+        )

vllm/model_executor/models/registry.py

@@ -190,6 +190,7 @@ _MULTIMODAL_MODELS = {
     "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
     "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501
     "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
+    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),  # noqa: E501
     "GraniteSpeechForConditionalGeneration": ("granite_speech", "GraniteSpeechForConditionalGeneration"),  # noqa: E501
     "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
     "InternVLChatModel": ("internvl", "InternVLChatModel"),

vllm/multimodal/inputs.py

@@ -57,10 +57,12 @@ which are treated as image embeddings;
 these are directly passed to the model without HF processing.
 """
 
-VideoItem: TypeAlias = Union[HfVideoItem, "torch.Tensor"]
+VideoItem: TypeAlias = Union[HfVideoItem, "torch.Tensor",
+                             tuple[HfVideoItem, dict[str, Any]]]
 """
-A `transformers.image_utils.VideoInput` representing a single video
-item, which can be passed to a HuggingFace `VideoProcessor`.
+A `transformers.video_utils.VideoInput` representing a single video item. 
+This can be passed to a HuggingFace `VideoProcessor` 
+with `transformers.video_utils.VideoMetadata`.
 
 Alternatively, a 3-D tensor or batch of 2-D tensors,
 which are treated as video embeddings;

vllm/multimodal/parse.py

@@ -224,8 +224,14 @@ class ImageEmbeddingItems(EmbeddingItems):
 
 class VideoProcessorItems(ProcessorBatchItems[HfVideoItem]):
 
-    def __init__(self, data: Sequence[HfVideoItem]) -> None:
+    def __init__(
+        self,
+        data: Sequence[HfVideoItem],
+        metadata: Optional[Union[dict[str, Any],
+                                 list[Optional[dict[str, Any]]]]] = None,
+    ) -> None:
         super().__init__(data, "video")
+        self.metadata = metadata
 
     def get_num_frames(self, item_idx: int) -> int:
         return len(self.get(item_idx))
@@ -320,6 +326,7 @@ class MultiModalDataParser:
         *,
         target_sr: Optional[float] = None,
         audio_resample_method: Literal["librosa", "scipy"] = "librosa",
+        video_needs_metadata: bool = False,
     ) -> None:
         super().__init__()
 
@@ -327,6 +334,7 @@ class MultiModalDataParser:
             target_sr=target_sr,
             method=audio_resample_method,
         )
+        self.video_needs_metadata = video_needs_metadata
 
     def _is_embeddings(
             self, data: object
@@ -361,6 +369,21 @@ class MultiModalDataParser:
 
         assert_never(audio)
 
+    def _get_video_with_metadata(
+        self,
+        video: VideoItem,
+    ) -> tuple[np.ndarray, Optional[dict[str, Any]]]:
+        if isinstance(video, tuple):
+            return video
+        if isinstance(video, list):
+            return np.array(video), None
+        if isinstance(video, np.ndarray):
+            return video, None
+        if isinstance(video, torch.Tensor):
+            return video.numpy(), None
+
+        assert_never(video)
+
     def _parse_audio_data(
         self,
         data: ModalityData[AudioItem],
@@ -433,10 +456,25 @@ class MultiModalDataParser:
             data_items = [data]
         elif isinstance(data, (np.ndarray, torch.Tensor)):
             data_items = [elem for elem in data]
+        elif isinstance(data, tuple) and len(data) == 2:
+            data_items = [data]
         else:
             data_items = data
 
-        return VideoProcessorItems(data_items)
+        new_videos = list[tuple[np.ndarray, Optional[dict[str, Any]]]]()
+        metadata_lst: list[Optional[dict[str, Any]]] = []
+        for data_item in data_items:
+            video, metadata = self._get_video_with_metadata(data_item)
+            if self.video_needs_metadata:
+                new_videos.append((video, metadata))
+                metadata_lst.append(metadata)
+            else:
+                new_videos.append(video)
+
+        if not self.video_needs_metadata:
+            metadata = None
+
+        return VideoProcessorItems(new_videos, metadata=metadata_lst)
 
     def _get_subparsers(self) -> Mapping[str, ModalityDataParser]:
         return {

vllm/multimodal/video.py

@@ -24,6 +24,7 @@ def resize_video(frames: npt.NDArray, size: tuple[int, int]) -> npt.NDArray:
                               dtype=frames.dtype)
     # lazy import cv2 to avoid bothering users who only use text models
     import cv2
+
     for i, frame in enumerate(frames):
         resized_frame = cv2.resize(frame, (new_width, new_height))
         resized_frames[i] = resized_frame
@@ -92,14 +93,16 @@ class OpenCVVideoBackend(VideoLoader):
                 continue
             if not vr.isBackendBuiltIn(backend):
                 _, abi, api = vr.getStreamBufferedBackendPluginVersion(backend)
-                if (abi < 1 or (abi == 1 and api < 2)):
+                if abi < 1 or (abi == 1 and api < 2):
                     continue
             api_pref = backend
             break
         return api_pref
 
     @classmethod
-    def load_bytes(cls, data: bytes, num_frames: int = -1) -> npt.NDArray:
+    def load_bytes(cls,
+                   data: bytes,
+                   num_frames: int = -1) -> tuple[npt.NDArray, dict]:
         import cv2
 
         backend = cls().get_cv2_video_api()
@@ -108,6 +111,9 @@ class OpenCVVideoBackend(VideoLoader):
             raise ValueError("Could not open video stream")
 
         total_frames_num = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+        original_fps = cap.get(cv2.CAP_PROP_FPS)
+        duration = total_frames_num / original_fps if original_fps > 0 else 0
+
         full_read = num_frames == -1 or total_frames_num < num_frames
         if full_read:
             num_frames = total_frames_num
@@ -125,18 +131,27 @@ class OpenCVVideoBackend(VideoLoader):
 
         i = 0
         for idx in range(total_frames_num):
-            ok = cap.grab()  # next img
+            ok = cap.grab()
             if not ok:
                 break
-            if idx in frame_idx:  # only decompress needed
+            if idx in frame_idx:
                 ret, frame = cap.retrieve()
                 if ret:
                     frames[i] = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
                     i += 1
-        # we expect all frames loaded
+
         assert i == num_frames, (f"Expected reading {num_frames} frames, "
                                  f"but only loaded {i} frames from video.")
-        return frames
+
+        # Use transformers transformers.video_utils.VideoMetadata format
+        metadata = {
+            "total_num_frames": total_frames_num,
+            "fps": original_fps,
+            "duration": duration,
+            "video_backend": "opencv"
+        }
+
+        return frames, metadata
 
 
 class VideoMediaIO(MediaIO[npt.NDArray]):