[Model][VLM] Add Qwen2-VL model support (#7905)

Co-authored-by: Roger Wang <136131678+ywang96@users.noreply.github.com>
Co-authored-by: DarkLight1337 <tlleungac@connect.ust.hk>

docs/source/models/supported_models.rst

@@ -252,6 +252,11 @@ Multimodal Language Models
     - Image\ :sup:`E`
     - :code:`Qwen/Qwen-VL`, :code:`Qwen/Qwen-VL-Chat`, etc.
     -
+  * - :code:`Qwen2VLForConditionalGeneration`
+    - Qwen2-VL (see note)
+    - Image\ :sup:`+` / Video\ :sup:`+`
+    - :code:`Qwen/Qwen2-VL-2B-Instruct`, :code:`Qwen/Qwen2-VL-7B-Instruct`, :code:`Qwen/Qwen2-VL-72B-Instruct`, etc.
+    -
   * - :code:`UltravoxModel`
     - Ultravox
     - Audio\ :sup:`E+`
@@ -265,15 +270,14 @@ Multimodal Language Models
   For :code:`openbmb/MiniCPM-V-2`, the official repo doesn't work yet, so we need to use a fork (:code:`HwwwH/MiniCPM-V-2`) for now.
   For more details, please see: https://github.com/vllm-project/vllm/pull/4087#issuecomment-2250397630
 
-  For :code:`LLaVA-NeXT-Video`, the latest release of :code:`huggingface/transformers` doesn't work yet, so we need to use a developer version (:code:`21fac7abba2a37fae86106f87fcf9974fd1e3830`) for now.
+.. note::
+  For :code:`LLaVA-NeXT-Video` and :code:`Qwen2-VL`, the latest release of :code:`huggingface/transformers` doesn't work yet, so we need to use a developer version (:code:`21fac7abba2a37fae86106f87fcf9974fd1e3830`) for now.
   This can be installed by running the following command: 
 
-
   .. code-block:: bash
     
     pip install git+https://github.com/huggingface/transformers.git@21fac7abba2a37fae86106f87fcf9974fd1e3830
 
-
 ----
 
 If your model uses one of the above model architectures, you can seamlessly run your model with vLLM.

examples/offline_inference_vision_language.py

@@ -179,6 +179,23 @@ def run_qwen_vl(question):
     return llm, prompt, stop_token_ids
 
 
+# Qwen2-VL
+def run_qwen2_vl(question):
+    model_name = "Qwen/Qwen2-VL-7B-Instruct"
+
+    llm = LLM(
+        model=model_name,
+        max_num_seqs=5,
+    )
+
+    prompt = ("<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
+              "<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>"
+              f"{question}<|im_end|>\n"
+              "<|im_start|>assistant\n")
+    stop_token_ids = None
+    return llm, prompt, stop_token_ids
+
+
 model_example_map = {
     "llava": run_llava,
     "llava-next": run_llava_next,
@@ -191,6 +208,7 @@ model_example_map = {
     "blip-2": run_blip2,
     "internvl_chat": run_internvl,
     "qwen_vl": run_qwen_vl,
+    "qwen2_vl": run_qwen2_vl,
 }
 
 

examples/offline_inference_vision_language_multi_image.py

@@ -6,7 +6,7 @@ by the model.
 from argparse import Namespace
 from typing import List
 
-from transformers import AutoTokenizer
+from transformers import AutoProcessor, AutoTokenizer
 
 from vllm import LLM, SamplingParams
 from vllm.multimodal.utils import fetch_image
@@ -30,7 +30,7 @@ def load_phi3v(question, image_urls: List[str]):
                              for i, _ in enumerate(image_urls, start=1))
     prompt = f"<|user|>\n{placeholders}\n{question}<|end|>\n<|assistant|>\n"
     stop_token_ids = None
-    return llm, prompt, stop_token_ids
+    return llm, prompt, stop_token_ids, None
 
 
 def load_internvl(question, image_urls: List[str]):
@@ -60,18 +60,72 @@ def load_internvl(question, image_urls: List[str]):
     # https://huggingface.co/OpenGVLab/InternVL2-2B#service
     stop_tokens = ["<|endoftext|>", "<|im_start|>", "<|im_end|>", "<|end|>"]
     stop_token_ids = [tokenizer.convert_tokens_to_ids(i) for i in stop_tokens]
-    return llm, prompt, stop_token_ids
+
+    return llm, prompt, stop_token_ids, None
+
+
+def load_qwen2_vl(question, image_urls: List[str]):
+    try:
+        from qwen_vl_utils import process_vision_info
+    except ModuleNotFoundError:
+        print('WARNING: `qwen-vl-utils` not installed, input images will not '
+              'be automatically resized. You can enable this functionality by '
+              '`pip install qwen-vl-utils`.')
+        process_vision_info = None
+
+    model_name = "Qwen/Qwen2-VL-7B-Instruct"
+
+    llm = LLM(
+        model=model_name,
+        max_num_seqs=5,
+        max_model_len=32768 if process_vision_info is None else 4096,
+        limit_mm_per_prompt={"image": len(image_urls)},
+    )
+
+    placeholders = [{"type": "image", "image": url} for url in image_urls]
+    messages = [{
+        "role": "system",
+        "content": "You are a helpful assistant."
+    }, {
+        "role":
+        "user",
+        "content": [
+            *placeholders,
+            {
+                "type": "text",
+                "text": question
+            },
+        ],
+    }]
+
+    processor = AutoProcessor.from_pretrained(model_name)
+
+    prompt = processor.apply_chat_template(messages,
+                                           tokenize=False,
+                                           add_generation_prompt=True)
+
+    stop_token_ids = None
+
+    if process_vision_info is None:
+        image_data = [fetch_image(url) for url in image_urls]
+    else:
+        image_data, _ = process_vision_info(messages)
+
+    return llm, prompt, stop_token_ids, image_data
 
 
 model_example_map = {
     "phi3_v": load_phi3v,
     "internvl_chat": load_internvl,
+    "qwen2_vl": load_qwen2_vl,
 }
 
 
 def run_generate(model, question: str, image_urls: List[str]):
-    llm, prompt, stop_token_ids = model_example_map[model](question,
-                                                           image_urls)
+    llm, prompt, stop_token_ids, image_data = model_example_map[model](
+        question, image_urls)
+    if image_data is None:
+        image_data = [fetch_image(url) for url in image_urls]
 
     sampling_params = SamplingParams(temperature=0.0,
                                      max_tokens=128,
@@ -81,7 +135,7 @@ def run_generate(model, question: str, image_urls: List[str]):
         {
             "prompt": prompt,
             "multi_modal_data": {
-                "image": [fetch_image(url) for url in image_urls]
+                "image": image_data
             },
         },
         sampling_params=sampling_params)
@@ -92,7 +146,7 @@ def run_generate(model, question: str, image_urls: List[str]):
 
 
 def run_chat(model: str, question: str, image_urls: List[str]):
-    llm, _, stop_token_ids = model_example_map[model](question, image_urls)
+    llm, _, stop_token_ids, _ = model_example_map[model](question, image_urls)
 
     sampling_params = SamplingParams(temperature=0.0,
                                      max_tokens=128,

requirements-common.txt

@@ -28,3 +28,4 @@ importlib_metadata
 mistral_common >= 1.3.4
 pyyaml
 six>=1.16.0; python_version > '3.11' # transitive dependency of pandas that needs to be the latest version for python 3.12
+einops # Required for Qwen2-VL.

tests/models/test_registry.py

@@ -1,9 +1,14 @@
 import pytest
+import transformers
 
 from vllm.model_executor.models import _MODELS, ModelRegistry
 
 
 @pytest.mark.parametrize("model_cls", _MODELS)
 def test_registry_imports(model_cls):
+    if (model_cls == "Qwen2VLForConditionalGeneration"
+            and transformers.__version__ < "4.45"):
+        pytest.skip("Waiting for next transformers release")
+
     # Ensure all model classes can be imported successfully
     ModelRegistry.resolve_model_cls([model_cls])

vllm/config.py

@@ -773,7 +773,7 @@ class LoadConfig:
         ignore_patterns: The list of patterns to ignore when loading the model.
             Default to "original/**/*" to avoid repeated loading of llama's 
             checkpoints.
-            
+
     """
 
     load_format: Union[str, LoadFormat, "BaseModelLoader"] = LoadFormat.AUTO
@@ -1733,8 +1733,11 @@ def _get_and_verify_max_len(
                     "with rope_scaling. Please raise an issue so we can "
                     "investigate.")
 
-            assert "factor" in rope_scaling
-            scaling_factor = rope_scaling["factor"]
+            if rope_type == "mrope":
+                scaling_factor = 1
+            else:
+                assert "factor" in rope_scaling
+                scaling_factor = rope_scaling["factor"]
             if rope_type == "yarn":
                 derived_max_model_len = rope_scaling[
                     "original_max_position_embeddings"]

vllm/entrypoints/chat_utils.py

@@ -108,7 +108,7 @@ class ConversationMessage(TypedDict, total=False):
     """The tool calls generated by the model, such as function calls."""
 
 
-ModalityStr = Literal["image", "audio"]
+ModalityStr = Literal["image", "audio", "video"]
 _T = TypeVar("_T")
 
 
@@ -158,12 +158,18 @@ class BaseMultiModalItemTracker(ABC, Generic[_T]):
                                               hf_config.image_token_index)
             if model_type in ("chameleon", "internvl_chat"):
                 return "<image>"
+            if model_type == "qwen2_vl":
+                return "<|vision_start|><|image_pad|><|vision_end|>"
 
             raise TypeError(f"Unknown model type: {model_type}")
         elif modality == "audio":
             if model_type == "ultravox":
                 return "<|reserved_special_token_0|>"
             raise TypeError(f"Unknown model type: {model_type}")
+        elif modality == "video":
+            if model_type == "qwen2_vl":
+                return "<|vision_start|><|video_pad|><|vision_end|>"
+            raise TypeError(f"Unknown model type: {model_type}")
         else:
             raise TypeError(f"Unknown modality: {modality}")
 

vllm/model_executor/layers/rotary_embedding.py

@@ -712,6 +712,179 @@ class Llama3RotaryEmbedding(RotaryEmbedding):
         return new_freqs
 
 
+class MRotaryEmbedding(RotaryEmbedding):
+    """Rotary Embedding with Multimodal Sections."""
+
+    def __init__(
+        self,
+        head_size: int,
+        rotary_dim: int,
+        max_position_embeddings: int,
+        base: int,
+        is_neox_style: bool,
+        dtype: torch.dtype,
+        mrope_section: Optional[List[int]] = None,
+    ) -> None:
+        super().__init__(head_size, rotary_dim, max_position_embeddings, base,
+                         is_neox_style, dtype)
+
+        self.mrope_section = mrope_section
+        if self.mrope_section:
+            assert sum(self.mrope_section) == rotary_dim // 2
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        query: torch.Tensor,
+        key: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """PyTorch-native implementation equivalent to forward().
+
+        Args:
+            positions:
+                [num_tokens,] (text only) or
+                [3, num_tokens] (T/H/W positions with multimodal inputs)
+            query: [num_tokens, num_heads * head_size]
+            key: [num_tokens, num_kv_heads * head_size]
+        """
+        assert positions.ndim == 1 or positions.ndim == 2
+
+        num_tokens = positions.shape[-1]
+        cos_sin = self.cos_sin_cache[positions]
+        cos, sin = cos_sin.chunk(2, dim=-1)
+        if positions.ndim == 2:
+            assert self.mrope_section
+
+            cos = torch.cat([
+                m[i]
+                for i, m in enumerate(cos.split(self.mrope_section, dim=-1))
+            ],
+                            dim=-1)
+            sin = torch.cat([
+                m[i]
+                for i, m in enumerate(sin.split(self.mrope_section, dim=-1))
+            ],
+                            dim=-1)
+
+        query_shape = query.shape
+        query = query.view(num_tokens, -1, self.head_size)
+        query_rot = query[..., :self.rotary_dim]
+        query_pass = query[..., self.rotary_dim:]
+        query_rot = _apply_rotary_emb(query_rot, cos, sin, self.is_neox_style)
+        query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)
+
+        key_shape = key.shape
+        key = key.view(num_tokens, -1, self.head_size)
+        key_rot = key[..., :self.rotary_dim]
+        key_pass = key[..., self.rotary_dim:]
+        key_rot = _apply_rotary_emb(key_rot, cos, sin, self.is_neox_style)
+        key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
+        return query, key
+
+    @staticmethod
+    def get_input_positions(
+        input_tokens: List[int],
+        image_grid_thw: Union[List[List[int]], torch.Tensor],
+        video_grid_thw: Union[List[List[int]], torch.Tensor],
+        image_token_id: int,
+        video_token_id: int,
+        vision_start_token_id: int,
+        vision_end_token_id: int,
+        spatial_merge_size: int,
+        context_len: int = 0,
+    ) -> Tuple[List[List[int]], int]:
+        """Get mrope input positions and delta value."""
+
+        if isinstance(image_grid_thw, torch.Tensor):
+            image_grid_thw = image_grid_thw.tolist()
+        if isinstance(video_grid_thw, torch.Tensor):
+            video_grid_thw = video_grid_thw.tolist()
+
+        input_tokens_tensor = torch.tensor(input_tokens)
+        vision_start_indices = torch.argwhere(
+            input_tokens_tensor == vision_start_token_id).squeeze(1)
+        vision_tokens = input_tokens_tensor[vision_start_indices + 1]
+        image_nums = (vision_tokens == image_token_id).sum()
+        video_nums = (vision_tokens == video_token_id).sum()
+        llm_pos_ids_list: list = []
+
+        st = 0
+        remain_images, remain_videos = image_nums, video_nums
+
+        image_index, video_index = 0, 0
+        for _ in range(image_nums + video_nums):
+            if image_token_id in input_tokens and remain_images > 0:
+                ed_image = input_tokens.index(image_token_id, st)
+            else:
+                ed_image = len(input_tokens) + 1
+            if video_token_id in input_tokens and remain_videos > 0:
+                ed_video = input_tokens.index(video_token_id, st)
+            else:
+                ed_video = len(input_tokens) + 1
+            if ed_image < ed_video:
+                t, h, w = (
+                    image_grid_thw[image_index][0],
+                    image_grid_thw[image_index][1],
+                    image_grid_thw[image_index][2],
+                )
+                image_index += 1
+                remain_images -= 1
+                ed = ed_image
+            else:
+                t, h, w = (
+                    video_grid_thw[video_index][0],
+                    video_grid_thw[video_index][1],
+                    video_grid_thw[video_index][2],
+                )
+                video_index += 1
+                remain_videos -= 1
+                ed = ed_video
+            llm_grid_t, llm_grid_h, llm_grid_w = \
+                t, h // spatial_merge_size, w // spatial_merge_size
+            text_len = ed - st
+
+            st_idx = llm_pos_ids_list[-1].max() + 1 if len(
+                llm_pos_ids_list) > 0 else 0
+            llm_pos_ids_list.append(
+                torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)
+
+            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]) + text_len + st_idx)
+            st = ed + llm_grid_t * llm_grid_h * llm_grid_w
+
+        if st < len(input_tokens):
+            st_idx = llm_pos_ids_list[-1].max() + 1 if len(
+                llm_pos_ids_list) > 0 else 0
+            text_len = len(input_tokens) - st
+            llm_pos_ids_list.append(
+                torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)
+
+        llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)
+        llm_positions = llm_positions[:, context_len:]
+        mrope_position_delta = (llm_positions.max() + 1 -
+                                len(input_tokens)).item()
+
+        return llm_positions.tolist(), mrope_position_delta
+
+    @staticmethod
+    def get_next_input_positions(
+        mrope_position_delta: int,
+        context_len: int,
+        seq_len: int,
+    ) -> List[List[int]]:
+        return [
+            list(
+                range(context_len + mrope_position_delta,
+                      seq_len + mrope_position_delta)) for _ in range(3)
+        ]
+
+
 _ROPE_DICT: Dict[Tuple, RotaryEmbedding] = {}
 
 
@@ -752,7 +925,7 @@ def get_rope(
         # The correct one should be "longrope" but keep "su" here
         # for backward compatible
         if scaling_type not in {"su", "longrope"}:
-            scaling_factor = rope_scaling["factor"]
+            scaling_factor = rope_scaling.get("factor", 1.0)
         if scaling_type == "llama3":
             low_freq_factor = rope_scaling["low_freq_factor"]
             high_freq_factor = rope_scaling["high_freq_factor"]
@@ -816,6 +989,16 @@ def get_rope(
                 head_size, rotary_dim, max_position, original_max_position,
                 base, is_neox_style, dtype, short_factor, long_factor,
                 **extra_kwargs)
+        elif scaling_type == "mrope":
+            return MRotaryEmbedding(
+                head_size,
+                rotary_dim,
+                max_position,
+                base,
+                is_neox_style,
+                dtype,
+                mrope_section=rope_scaling["mrope_section"],
+            )
         else:
             raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
     _ROPE_DICT[key] = rotary_emb

vllm/model_executor/models/__init__.py

@@ -53,6 +53,8 @@ _GENERATION_MODELS = {
     "PhiMoEForCausalLM": ("phimoe", "PhiMoEForCausalLM"),
     "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
     "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
+    "Qwen2VLForConditionalGeneration":
+    ("qwen2_vl", "Qwen2VLForConditionalGeneration"),
     "RWForCausalLM": ("falcon", "FalconForCausalLM"),
     "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
     "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
@@ -90,6 +92,8 @@ _MULTIMODAL_MODELS = {
     "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
     "UltravoxModel": ("ultravox", "UltravoxModel"),
     "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
+    "Qwen2VLForConditionalGeneration": ("qwen2_vl",
+                                        "Qwen2VLForConditionalGeneration"),
 }
 _CONDITIONAL_GENERATION_MODELS = {
     "BartModel": ("bart", "BartForConditionalGeneration"),

vllm/multimodal/base.py

@@ -79,14 +79,12 @@ class MultiModalInputs(_MultiModalInputsBase):
         if len(inputs_list) == 0:
             return {}
 
-        keys = inputs_list[0].keys()
-
         item_lists: Dict[str, List[NestedTensors]] = defaultdict(list)
 
         for inputs in inputs_list:
-            if inputs.keys() != keys:
-                msg = f"Inputs do not share the same keys ({keys})"
-                raise ValueError(msg)
+            # For models that supports multiple modalities (e.g. Qwen2-VL),
+            # different modalities will return different data keys,
+            # so batch() should skip the same key check.
 
             for k, v in inputs.items():
                 item_lists[k].append(v)

vllm/sequence.py

@@ -165,6 +165,9 @@ class SequenceData(msgspec.Struct,
     # is called.
     _new_appended_tokens: List[int] = msgspec.field(default_factory=list)
 
+    # It is used to compute mrope_position_ids.
+    _mrope_position_delta: Optional[int] = None
+
     def __post_init__(self) -> None:
         assert self._prompt_token_ids.typecode == "l"
         assert self._output_token_ids.typecode == "l"
@@ -219,6 +222,14 @@ class SequenceData(msgspec.Struct,
         assert isinstance(self._output_token_ids, array)
         return self._output_token_ids
 
+    @property
+    def mrope_position_delta(self) -> Optional[int]:
+        return self._mrope_position_delta
+
+    @mrope_position_delta.setter
+    def mrope_position_delta(self, new_mrope_position_delta):
+        self._mrope_position_delta = new_mrope_position_delta
+
     def append_token_id(self, token_id: int, logprob: float) -> None:
         self._output_token_ids.append(token_id)
         self._new_appended_tokens.append(token_id)

vllm/transformers_utils/processor.py

@@ -0,0 +1,37 @@
+from typing import cast
+
+
+def get_processor(
+    processor_name: str,
+    *args,
+    trust_remote_code: bool = False,
+    **kwargs,
+):
+    """Gets a processor for the given model name via HuggingFace."""
+    # don't put this import at the top level
+    # it will call torch.cuda.device_count()
+    from transformers import AutoProcessor
+    from transformers.processing_utils import ProcessorMixin
+
+    try:
+        processor = AutoProcessor.from_pretrained(
+            processor_name,
+            *args,
+            trust_remote_code=trust_remote_code,
+            **kwargs)
+    except ValueError as e:
+        # If the error pertains to the processor class not existing or not
+        # currently being imported, suggest using the --trust-remote-code flag.
+        # Unlike AutoTokenizer, AutoProcessor does not separate such errors
+        if not trust_remote_code:
+            err_msg = (
+                "Failed to load the processor. If the processor is "
+                "a custom processor not yet available in the HuggingFace "
+                "transformers library, consider setting "
+                "`trust_remote_code=True` in LLM or using the "
+                "`--trust-remote-code` flag in the CLI.")
+            raise RuntimeError(err_msg) from e
+        else:
+            raise e
+
+    return cast(ProcessorMixin, processor)

vllm/worker/model_runner.py

@@ -30,6 +30,7 @@ from vllm.lora.layers import LoRAMapping
 from vllm.lora.request import LoRARequest
 from vllm.lora.worker_manager import LRUCacheWorkerLoRAManager
 from vllm.model_executor import SamplingMetadata, SamplingMetadataCache
+from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
 from vllm.model_executor.layers.sampler import SamplerOutput
 from vllm.model_executor.model_loader import get_model
 from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
@@ -181,6 +182,7 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
         def simple_reinit(self):
             self.input_tokens[0].clear()  # type: ignore
             self.input_positions[0].clear()  # type: ignore
+            self.mrope_input_positions = None  # type: ignore
             self.seq_lens[0] = 0  # type: ignore
             self.orig_seq_lens[0] = 0  # type: ignore
             self.query_lens[0] = 0  # type: ignore
@@ -206,6 +208,7 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
             # Input tokens and positions.
             input_tokens: Optional[List[List[int]]] = None,
             input_positions: Optional[List[List[int]]] = None,
+            mrope_input_positions: Optional[List[List[List[int]]]] = None,
 
             # The sequence length (may be capped to the sliding window).
             seq_lens: Optional[List[int]] = None,
@@ -266,6 +269,8 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
                         for seq_id in range(len(self.seq_ids)):
                             self.input_positions[seq_id].clear()
 
+                    self.mrope_input_positions = None
+
                     if seq_lens:
                         self.seq_lens = seq_lens
                     else:
@@ -327,6 +332,7 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
             else:
                 self.input_tokens = input_tokens or []
                 self.input_positions = input_positions or []
+                self.mrope_input_positions = mrope_input_positions or None
                 self.seq_lens = seq_lens or []
                 self.orig_seq_lens = orig_seq_lens or []
                 self.query_lens = query_lens or []
@@ -357,6 +363,7 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
 
             self.input_tokens = [[] for _ in range(self.n_seqs)]
             self.input_positions = [[] for _ in range(self.n_seqs)]
+            self.mrope_input_positions = None
             self.seq_lens = [0] * self.n_seqs
             self.orig_seq_lens = [0] * self.n_seqs
             self.query_lens = [0] * self.n_seqs
@@ -493,6 +500,17 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
         inter_data.query_lens[
             seq_idx] = seq_len - context_len if inter_data.is_prompt else 1
 
+        if seq_data.mrope_position_delta is not None:
+            if inter_data.mrope_input_positions is None:
+                inter_data.mrope_input_positions = [None] * inter_data.n_seqs
+
+            inter_data.mrope_input_positions[
+                seq_idx] = MRotaryEmbedding.get_next_input_positions(
+                    seq_data.mrope_position_delta,
+                    context_len,
+                    seq_len,
+                )
+
     def _compute_for_prefix_cache_hit(
             self, inter_data: InterDataForSeqGroup, seq_idx: int,
             seq_group_metadata: SequenceGroupMetadata):
@@ -636,6 +654,40 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
         mm_kwargs = self.multi_modal_input_mapper(mm_data)
         inter_data.multi_modal_inputs = mm_kwargs
 
+        # special processing for mrope position deltas.
+        if self.runner.model_is_mrope:
+            image_grid_thw = mm_kwargs.get("image_grid_thw", None)
+            video_grid_thw = mm_kwargs.get("video_grid_thw", None)
+            assert image_grid_thw is not None or video_grid_thw is not None, (
+                "mrope embedding type requires multi-modal input mapper "
+                "returns 'image_grid_thw' or 'video_grid_thw'.")
+
+            hf_config = self.runner.model_config.hf_config
+
+            inter_data.mrope_input_positions = [None] * inter_data.n_seqs
+            for seq_idx in range(inter_data.n_seqs):
+                seq_data = seq_group_metadata.seq_data[
+                    inter_data.seq_ids[seq_idx]]
+                token_ids = seq_data.get_token_ids()
+
+                mrope_input_positions, mrope_position_delta = \
+                    MRotaryEmbedding.get_input_positions(
+                        token_ids,
+                        image_grid_thw=image_grid_thw,
+                        video_grid_thw=video_grid_thw,
+                        image_token_id=hf_config.image_token_id,
+                        video_token_id=hf_config.video_token_id,
+                        vision_start_token_id=hf_config.vision_start_token_id,
+                        vision_end_token_id=hf_config.vision_end_token_id,
+                        spatial_merge_size=hf_config.vision_config.
+                        spatial_merge_size,
+                        context_len=inter_data.context_lens[seq_idx],
+                    )
+
+                seq_data.mrope_position_delta = mrope_position_delta
+                inter_data.mrope_input_positions[
+                    seq_idx] = mrope_input_positions
+
     def add_seq_group(self, seq_group_metadata: SequenceGroupMetadata):
         """Add a sequence group to the builder."""
         seq_ids = seq_group_metadata.seq_data.keys()
@@ -684,10 +736,27 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
             # prefix caching and there is no decode request.
             return self.model_input_cls()
 
-        input_positions = []
-        for inter_data in self.inter_data_list:
-            for cur_input_positions in inter_data.input_positions:
-                input_positions.extend(cur_input_positions)
+        mrope_input_positions: Optional[List[List[int]]] = None
+        if any(inter_data.mrope_input_positions is not None
+               for inter_data in self.inter_data_list):
+            mrope_input_positions = [[] for _ in range(3)]
+            for idx in range(3):
+                for inter_data in self.inter_data_list:
+                    msections = inter_data.mrope_input_positions
+                    if msections is None:
+                        for _seq_input_positions in inter_data.input_positions:
+                            mrope_input_positions[idx].extend(
+                                _seq_input_positions)
+                    else:
+                        for _seq_mrope_input_positions in msections:
+                            mrope_input_positions[idx].extend(
+                                _seq_mrope_input_positions[idx])
+            input_positions = None
+        else:
+            input_positions = []
+            for inter_data in self.inter_data_list:
+                for cur_input_positions in inter_data.input_positions:
+                    input_positions.extend(cur_input_positions)
 
         seq_lens = []
         max_decode_seq_len = 0
@@ -724,15 +793,24 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
         # Tokens and positions.
         if cuda_graph_pad_size:
             input_tokens.extend(itertools.repeat(0, cuda_graph_pad_size))
-            input_positions.extend(itertools.repeat(0, cuda_graph_pad_size))
         assert self.runner.device is not None
         input_tokens_tensor = async_tensor_h2d(input_tokens, torch.long,
                                                self.runner.device,
                                                self.runner.pin_memory)
-        input_positions_tensor = async_tensor_h2d(input_positions, torch.long,
-                                                  self.runner.device,
-                                                  self.runner.pin_memory)
-
+        if mrope_input_positions is not None:
+            for idx in range(3):
+                mrope_input_positions[idx].extend(
+                    itertools.repeat(0, cuda_graph_pad_size))
+            input_positions_tensor = async_tensor_h2d(mrope_input_positions,
+                                                      torch.long,
+                                                      self.runner.device,
+                                                      self.runner.pin_memory)
+        else:
+            input_positions.extend(itertools.repeat(0, cuda_graph_pad_size))
+            input_positions_tensor = async_tensor_h2d(input_positions,
+                                                      torch.long,
+                                                      self.runner.device,
+                                                      self.runner.pin_memory)
         # Sequence and query lengths.
         if cuda_graph_pad_size:
             seq_lens.extend(itertools.repeat(1, cuda_graph_pad_size))
@@ -1199,6 +1277,15 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
             raise RuntimeError("PromptAdapter is not enabled.")
         return self.prompt_adapter_manager.list_adapters()
 
+    @property
+    def model_is_mrope(self) -> bool:
+        """Detect if the model has "mrope" rope_scaling type.
+        mrope requires keep "rope_deltas" between prompt and decoding phases."""
+        rope_scaling = getattr(self.model_config.hf_config, "rope_scaling", {})
+        if rope_scaling is None:
+            return False
+        return rope_scaling.get("type", None) == "mrope"
+
     @torch.inference_mode()
     def capture_model(self, kv_caches: List[List[torch.Tensor]]) -> None:
         """Cuda graph capture a model.
@@ -1229,7 +1316,8 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
         max_batch_size = self.max_batchsize_to_capture
         input_tokens = torch.zeros(max_batch_size, dtype=torch.long).cuda()
         input_positions = torch.zeros(max_batch_size, dtype=torch.long).cuda()
-
+        if self.model_is_mrope:
+            input_positions = torch.tile(input_positions, (3, 1))
         # Prepare dummy previous_hidden_states only if needed by the model.
         # This is used by draft models such as EAGLE.
         previous_hidden_states = None
@@ -1293,7 +1381,7 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
                         "input_ids":
                         input_tokens[:batch_size],
                         "positions":
-                        input_positions[:batch_size],
+                        input_positions[..., :batch_size],
                         "hidden_or_intermediate_states":
                         hidden_or_intermediate_states[
                             virtual_engine]  # type: ignore