diff --git a/lmdeploy/pytorch/engine/engine.py b/lmdeploy/pytorch/engine/engine.py
index 3b3fd2e117..756eab80b9 100644
--- a/lmdeploy/pytorch/engine/engine.py
+++ b/lmdeploy/pytorch/engine/engine.py
@@ -23,7 +23,6 @@
 from ..adapter.adapter import AdapterManager
 from ..config import CacheConfig, ModelConfig
 from ..messages import MessageStatus, SchedulerSequence, UpdateTokenMode
-from ..multimodal.data_type import ensure_multimodal_content_hashes
 from ..paging import Scheduler
 from ..strategies import build_strategy_factory
 from .base import EngineBase
@@ -416,8 +415,6 @@ def _on_add_message(self, reqs: list[Request], **kwargs):
 
             input_ids = result.input_ids
             input_multimodals = result.input_multimodals
-            if self.cache_config.enable_prefix_caching:
-                input_multimodals = ensure_multimodal_content_hashes(input_multimodals)
 
             req_data['token_ids'] = input_ids
             req_data['input_multimodals'] = input_multimodals
diff --git a/lmdeploy/pytorch/messages.py b/lmdeploy/pytorch/messages.py
index b1c5a59e68..76493e880a 100644
--- a/lmdeploy/pytorch/messages.py
+++ b/lmdeploy/pytorch/messages.py
@@ -10,9 +10,10 @@
 
 from lmdeploy.messages import EngineEvent, EventType, GenerationConfig, LogitsProcessor
 from lmdeploy.pytorch.disagg.conn.protocol import MigrationRequest
-from lmdeploy.pytorch.multimodal.data_type import MultiModalInputs, make_multimodal_content_hash
+from lmdeploy.pytorch.multimodal.data_type import MultiModalInputs
 from lmdeploy.utils import get_logger
 from lmdeploy.vl.constants import Modality
+from lmdeploy.vl.hasher import make_multimodal_content_hash
 
 from .block import LogicalTokenBlocks
 
@@ -1035,8 +1036,8 @@ def _update_prefix_cache_metas(self, multimodals: MultiModalInputs):
                     modality = modality.value
                 content_hash = modal_data.content_hash
                 if content_hash is None:
-                    # Most request paths precompute the hash after model
-                    # preprocessing.  Keep this fallback for unit tests and
+                    # Most request paths carry a hash from the shared VL path.
+                    # Keep this fallback for unit tests and
                     # defensive correctness if a processor omits it.
                     content_hash = make_multimodal_content_hash(modal_data.data, modal_data.meta,
                                                                 modal_data.mrope_pos_ids)
diff --git a/lmdeploy/pytorch/models/chatglm2.py b/lmdeploy/pytorch/models/chatglm2.py
index 2988f3c9ed..b506b9f276 100644
--- a/lmdeploy/pytorch/models/chatglm2.py
+++ b/lmdeploy/pytorch/models/chatglm2.py
@@ -883,7 +883,8 @@ def preprocess_input(self,
             mm_data = MultiModalData(data=pixel_values,
                                      start=offset,
                                      end=offset + num_pad,
-                                     meta=dict(image_token_id=image_token_id))
+                                     meta=dict(image_token_id=image_token_id),
+                                     content_hash=input_mm.get('content_hash'))
             input_imgs.append(mm_data)
 
         result = PreprocessInputResult(
diff --git a/lmdeploy/pytorch/models/cogvlm.py b/lmdeploy/pytorch/models/cogvlm.py
index 98d2f02583..fa6c874dd5 100644
--- a/lmdeploy/pytorch/models/cogvlm.py
+++ b/lmdeploy/pytorch/models/cogvlm.py
@@ -909,7 +909,8 @@ def preprocess_input(self, input_ids: list[int], input_multimodals=None, **kwarg
             mm_data = MultiModalData(data=pixel_values,
                                      start=offset,
                                      end=offset + num_pad,
-                                     meta=dict(image_token_id=image_token_id))
+                                     meta=dict(image_token_id=image_token_id),
+                                     content_hash=input_mm.get('content_hash'))
             input_imgs.append(mm_data)
 
         result = PreprocessInputResult(
diff --git a/lmdeploy/pytorch/models/deepseek_vl2.py b/lmdeploy/pytorch/models/deepseek_vl2.py
index b556ffc2b8..b881849c0e 100644
--- a/lmdeploy/pytorch/models/deepseek_vl2.py
+++ b/lmdeploy/pytorch/models/deepseek_vl2.py
@@ -447,7 +447,8 @@ def preprocess_input(self,
                                      meta=dict(
                                          image_token_id=image_token_id,
                                          images_spatial_crop=images_spatial_crop,
-                                     ))
+                                     ),
+                                     content_hash=input_mm.get('content_hash'))
 
             input_imgs.append(mm_data)
 
diff --git a/lmdeploy/pytorch/models/gemma3_vl.py b/lmdeploy/pytorch/models/gemma3_vl.py
index 9f444e81f5..a0eb3c9241 100644
--- a/lmdeploy/pytorch/models/gemma3_vl.py
+++ b/lmdeploy/pytorch/models/gemma3_vl.py
@@ -112,7 +112,8 @@ def preprocess_input(self,
             mm_data = MultiModalData(data=pixel_values,
                                      start=offset,
                                      end=offset + num_pad,
-                                     meta=dict(image_token_id=image_token_id))
+                                     meta=dict(image_token_id=image_token_id),
+                                     content_hash=input_mm.get('content_hash'))
             input_imgs.append(mm_data)
 
         result = PreprocessInputResult(
diff --git a/lmdeploy/pytorch/models/glm4_1v.py b/lmdeploy/pytorch/models/glm4_1v.py
index ac802fe6a2..698181a5bc 100644
--- a/lmdeploy/pytorch/models/glm4_1v.py
+++ b/lmdeploy/pytorch/models/glm4_1v.py
@@ -745,7 +745,8 @@ def preprocess_input(self,
                                      start=offset[0],
                                      end=offset[1],
                                      mrope_pos_ids=mrope_pos_ids,
-                                     meta=dict(grid_thw=image_grid_thw, image_token_id=image_token_id))
+                                     meta=dict(grid_thw=image_grid_thw, image_token_id=image_token_id),
+                                     content_hash=input_mm.get('content_hash'))
             input_imgs.append(mm_data)
 
         result = PreprocessInputResult(
diff --git a/lmdeploy/pytorch/models/interns1_pro.py b/lmdeploy/pytorch/models/interns1_pro.py
index 9a03f30640..f863db270e 100644
--- a/lmdeploy/pytorch/models/interns1_pro.py
+++ b/lmdeploy/pytorch/models/interns1_pro.py
@@ -374,7 +374,8 @@ def _make_image_mm_data(self, input_mm: dict[str, Any]) -> MultiModalData:
                                  data=pixel_values,
                                  start=offset[0],
                                  end=offset[1],
-                                 meta=dict(grid_thw=image_grid_thw, image_token_id=image_token_id))
+                                 meta=dict(grid_thw=image_grid_thw, image_token_id=image_token_id),
+                                 content_hash=input_mm.get('content_hash'))
         return mm_data
 
     def _make_video_mm_data(self, input_mm: dict[str, Any]) -> MultiModalData:
@@ -391,7 +392,8 @@ def _make_video_mm_data(self, input_mm: dict[str, Any]) -> MultiModalData:
                                  meta=dict(
                                      grid_thw=video_grid_thw,
                                      video_token_id=video_token_id,
-                                 ))
+                                 ),
+                                 content_hash=input_mm.get('content_hash'))
         return mm_data
 
     def _make_time_series_mm_data(self, input_mm: dict[str, Any]) -> MultiModalData:
@@ -406,7 +408,8 @@ def _make_time_series_mm_data(self, input_mm: dict[str, Any]) -> MultiModalData:
                                  data=ts_values,
                                  start=offset[0],
                                  end=offset[1],
-                                 meta=dict(ts_lens=ts_lens, ts_sr=ts_sr, ts_token_id=ts_token_id))
+                                 meta=dict(ts_lens=ts_lens, ts_sr=ts_sr, ts_token_id=ts_token_id),
+                                 content_hash=input_mm.get('content_hash'))
         return mm_data
 
     def preprocess_input(self,
diff --git a/lmdeploy/pytorch/models/internvl.py b/lmdeploy/pytorch/models/internvl.py
index 2a2f53dae3..9243e62639 100644
--- a/lmdeploy/pytorch/models/internvl.py
+++ b/lmdeploy/pytorch/models/internvl.py
@@ -946,7 +946,8 @@ def preprocess_input(self,
             mm_data = MultiModalData(data=pixel_values,
                                      start=offset,
                                      end=offset + num_pad,
-                                     meta=dict(image_token_id=image_token_id))
+                                     meta=dict(image_token_id=image_token_id),
+                                     content_hash=input_mm.get('content_hash'))
             input_imgs.append(mm_data)
 
         result = PreprocessInputResult(
diff --git a/lmdeploy/pytorch/models/internvl3_hf.py b/lmdeploy/pytorch/models/internvl3_hf.py
index 09f159c22a..6c0f3ad00e 100644
--- a/lmdeploy/pytorch/models/internvl3_hf.py
+++ b/lmdeploy/pytorch/models/internvl3_hf.py
@@ -740,7 +740,8 @@ def preprocess_input(self,
             mm_data = MultiModalData(data=pixel_values,
                                      start=offset,
                                      end=offset + num_pad,
-                                     meta=dict(image_token_id=image_token_id))
+                                     meta=dict(image_token_id=image_token_id),
+                                     content_hash=input_mm.get('content_hash'))
             input_imgs.append(mm_data)
 
         result = PreprocessInputResult(
diff --git a/lmdeploy/pytorch/models/llama4.py b/lmdeploy/pytorch/models/llama4.py
index 2552467218..7c507be9ad 100644
--- a/lmdeploy/pytorch/models/llama4.py
+++ b/lmdeploy/pytorch/models/llama4.py
@@ -1041,7 +1041,8 @@ def preprocess_input(self,
             mm_data = MultiModalData(data=pixel_values,
                                      start=offset,
                                      end=offset + num_pad,
-                                     meta=dict(image_token_id=image_token_id))
+                                     meta=dict(image_token_id=image_token_id),
+                                     content_hash=input_mm.get('content_hash'))
             input_imgs.append(mm_data)
 
         result = PreprocessInputResult(
diff --git a/lmdeploy/pytorch/models/llava.py b/lmdeploy/pytorch/models/llava.py
index bf0618ec9a..71150df4cd 100644
--- a/lmdeploy/pytorch/models/llava.py
+++ b/lmdeploy/pytorch/models/llava.py
@@ -559,7 +559,8 @@ def preprocess_input(self,
             mm_data = MultiModalData(data=pixel_values,
                                      start=offset,
                                      end=offset + num_pad,
-                                     meta=dict(image_token_id=image_token_id))
+                                     meta=dict(image_token_id=image_token_id),
+                                     content_hash=input_mm.get('content_hash'))
             input_imgs.append(mm_data)
 
         result = PreprocessInputResult(
@@ -838,7 +839,8 @@ def preprocess_input(self,
             mm_data = MultiModalData(data=pixel_values,
                                      start=offset,
                                      end=offset + num_pad,
-                                     meta=dict(image_sizes=image_sizes, image_token_id=image_token_id))
+                                     meta=dict(image_sizes=image_sizes, image_token_id=image_token_id),
+                                     content_hash=input_mm.get('content_hash'))
             input_imgs.append(mm_data)
 
         result = PreprocessInputResult(
diff --git a/lmdeploy/pytorch/models/phi3_v.py b/lmdeploy/pytorch/models/phi3_v.py
index 589e04e20d..3d4cb618a2 100644
--- a/lmdeploy/pytorch/models/phi3_v.py
+++ b/lmdeploy/pytorch/models/phi3_v.py
@@ -383,7 +383,8 @@ def preprocess_input(self,
             mm_data = MultiModalData(data=pixel_values,
                                      start=offset,
                                      end=offset + num_pad,
-                                     meta=dict(image_sizes=image_sizes, image_token_id=image_token_id))
+                                     meta=dict(image_sizes=image_sizes, image_token_id=image_token_id),
+                                     content_hash=input_mm.get('content_hash'))
             input_imgs.append(mm_data)
 
         result = PreprocessInputResult(
diff --git a/lmdeploy/pytorch/models/qwen2_vl.py b/lmdeploy/pytorch/models/qwen2_vl.py
index ff2bdb2e7d..7fede4d37c 100644
--- a/lmdeploy/pytorch/models/qwen2_vl.py
+++ b/lmdeploy/pytorch/models/qwen2_vl.py
@@ -814,7 +814,8 @@ def preprocess_input(self,
                                      start=start,
                                      end=start + num_pad,
                                      mrope_pos_ids=mrope_pos_ids,
-                                     meta=dict(grid_thw=image_grid_thw, image_token_id=image_token_id))
+                                     meta=dict(grid_thw=image_grid_thw, image_token_id=image_token_id),
+                                     content_hash=input_mm.get('content_hash'))
             input_imgs.append(mm_data)
 
         result = PreprocessInputResult(
diff --git a/lmdeploy/pytorch/models/qwen3_omni_moe_thinker.py b/lmdeploy/pytorch/models/qwen3_omni_moe_thinker.py
index 733c323dbf..7a4c6cf7f3 100644
--- a/lmdeploy/pytorch/models/qwen3_omni_moe_thinker.py
+++ b/lmdeploy/pytorch/models/qwen3_omni_moe_thinker.py
@@ -846,7 +846,8 @@ def _make_image_mm_data(self, input_mm: dict[str, Any]) -> MultiModalData:
                                  start=offset[0],
                                  end=offset[1],
                                  mrope_pos_ids=mrope_pos_ids,
-                                 meta=dict(grid_thw=image_grid_thw, image_token_id=image_token_id))
+                                 meta=dict(grid_thw=image_grid_thw, image_token_id=image_token_id),
+                                 content_hash=input_mm.get('content_hash'))
         return mm_data
 
     def _make_video_mm_data(self, input_mm: dict[str, Any]) -> MultiModalData:
@@ -867,7 +868,8 @@ def _make_video_mm_data(self, input_mm: dict[str, Any]) -> MultiModalData:
                                      grid_thw=video_grid_thw,
                                      video_token_id=video_token_id,
                                      second_per_grid=input_mm.get('second_per_grid'),
-                                 ))
+                                 ),
+                                 content_hash=input_mm.get('content_hash'))
         return mm_data
 
     def _make_audio_mm_data(self, input_mm: dict[str, Any]) -> MultiModalData:
@@ -896,7 +898,8 @@ def _make_audio_mm_data(self, input_mm: dict[str, Any]) -> MultiModalData:
                                  meta=dict(
                                      audio_token_id=audio_token_id,
                                      audio_feature_lengths=audio_feature_lengths,
-                                 ))
+                                 ),
+                                 content_hash=input_mm.get('content_hash'))
         return mm_data
 
     def preprocess_input(self,
diff --git a/lmdeploy/pytorch/models/qwen3_vl.py b/lmdeploy/pytorch/models/qwen3_vl.py
index db3f69af55..d9053a8088 100644
--- a/lmdeploy/pytorch/models/qwen3_vl.py
+++ b/lmdeploy/pytorch/models/qwen3_vl.py
@@ -682,7 +682,8 @@ def _make_image_mm_data(self, input_mm: dict[str, Any]) -> MultiModalData:
                                  start=offset[0],
                                  end=offset[1],
                                  mrope_pos_ids=mrope_pos_ids,
-                                 meta=dict(grid_thw=image_grid_thw, image_token_id=image_token_id))
+                                 meta=dict(grid_thw=image_grid_thw, image_token_id=image_token_id),
+                                 content_hash=input_mm.get('content_hash'))
         return mm_data
 
     def _make_video_mm_data(self, input_mm: dict[str, Any]) -> MultiModalData:
@@ -702,7 +703,8 @@ def _make_video_mm_data(self, input_mm: dict[str, Any]) -> MultiModalData:
                                  meta=dict(
                                      grid_thw=video_grid_thw,
                                      video_token_id=video_token_id,
-                                 ))
+                                 ),
+                                 content_hash=input_mm.get('content_hash'))
         return mm_data
 
     def _make_time_series_mm_data(self, input_mm: dict[str, Any]) -> MultiModalData:
@@ -717,7 +719,8 @@ def _make_time_series_mm_data(self, input_mm: dict[str, Any]) -> MultiModalData:
                                  data=ts_values,
                                  start=offset[0],
                                  end=offset[1],
-                                 meta=dict(ts_lens=ts_lens, ts_sr=ts_sr, ts_token_id=ts_token_id))
+                                 meta=dict(ts_lens=ts_lens, ts_sr=ts_sr, ts_token_id=ts_token_id),
+                                 content_hash=input_mm.get('content_hash'))
         return mm_data
 
     def preprocess_input(self,
diff --git a/lmdeploy/pytorch/multimodal/data_type.py b/lmdeploy/pytorch/multimodal/data_type.py
index f778c2aeb5..070f35867e 100644
--- a/lmdeploy/pytorch/multimodal/data_type.py
+++ b/lmdeploy/pytorch/multimodal/data_type.py
@@ -1,11 +1,8 @@
 # Copyright (c) OpenMMLab. All rights reserved.
-import enum
-import hashlib
 from dataclasses import dataclass, fields
 from typing import Any
 
 import numpy as np
-import torch
 from torch import Tensor
 
 from lmdeploy.vl.constants import Modality
@@ -13,46 +10,6 @@
 NestedTensor = Tensor | list[Tensor]
 
 
-def _hash_multimodal_value(hasher: 'hashlib._Hash', value: Any):
-    """Update a hash with a deterministic multimodal value representation."""
-    if isinstance(value, Tensor):
-        tensor = value.detach().cpu().contiguous()
-        hasher.update(f'tensor:{tensor.dtype}:{tuple(tensor.shape)}:'.encode())
-        hasher.update(tensor.view(torch.uint8).numpy().tobytes())
-    elif isinstance(value, np.ndarray):
-        array = np.ascontiguousarray(value)
-        hasher.update(f'ndarray:{array.dtype}:{array.shape}:'.encode())
-        hasher.update(array.tobytes())
-    elif isinstance(value, dict):
-        hasher.update(b'dict:{')
-        for key in sorted(value, key=lambda x: repr(x)):
-            _hash_multimodal_value(hasher, key)
-            hasher.update(b':')
-            _hash_multimodal_value(hasher, value[key])
-            hasher.update(b',')
-        hasher.update(b'}')
-    elif isinstance(value, (list, tuple)):
-        hasher.update(f'{type(value).__name__}:['.encode())
-        for item in value:
-            _hash_multimodal_value(hasher, item)
-            hasher.update(b',')
-        hasher.update(b']')
-    elif isinstance(value, enum.Enum):
-        _hash_multimodal_value(hasher, value.value)
-    else:
-        hasher.update(f'{type(value).__name__}:{repr(value)}'.encode())
-
-
-def make_multimodal_content_hash(data: Any, meta: dict[str, Any] | None,
-                                 mrope_pos_ids: np.ndarray | None = None) -> str:
-    """Create a stable content hash for prefix-cache multimodal matching."""
-    hasher = hashlib.sha256()
-    _hash_multimodal_value(hasher, data)
-    _hash_multimodal_value(hasher, meta)
-    _hash_multimodal_value(hasher, mrope_pos_ids)
-    return hasher.hexdigest()
-
-
 @dataclass
 class MultiModalData:
     data: NestedTensor
@@ -101,16 +58,3 @@ def to_device(self, device: str, non_blocking: bool = False):
 
 
 MultiModalInputs = dict[str, list[MultiModalData]]
-
-
-def ensure_multimodal_content_hashes(input_mms: MultiModalInputs | None):
-    """Populate missing multimodal content hashes in-place."""
-    if input_mms is None:
-        return input_mms
-
-    for modal_datas in input_mms.values():
-        for modal_data in modal_datas:
-            if modal_data.content_hash is None:
-                modal_data.content_hash = make_multimodal_content_hash(modal_data.data, modal_data.meta,
-                                                                       modal_data.mrope_pos_ids)
-    return input_mms
diff --git a/lmdeploy/vl/engine.py b/lmdeploy/vl/engine.py
index f2a5f62ccf..5ce37fda69 100644
--- a/lmdeploy/vl/engine.py
+++ b/lmdeploy/vl/engine.py
@@ -10,6 +10,7 @@
 from lmdeploy.messages import PytorchEngineConfig, TurbomindEngineConfig, VisionConfig
 from lmdeploy.utils import is_bf16_supported
 from lmdeploy.vl.model.builder import load_vl_model
+from lmdeploy.vl.model.preprocess_utils import attach_multimodal_content_hashes
 
 
 def _get_hf_config_mm_feature_dtype(hf_config) -> torch.dtype | None:
@@ -119,6 +120,7 @@ async def preprocess(self,
                 self.executor, self.model.preprocess, messages)
         future.add_done_callback(_raise_exception_on_finish)
         outputs = await future
+        attach_multimodal_content_hashes(outputs)
         return outputs
 
     async def async_infer(self, messages: list[dict]) -> list[dict]:
diff --git a/lmdeploy/vl/hasher.py b/lmdeploy/vl/hasher.py
new file mode 100644
index 0000000000..500449bb08
--- /dev/null
+++ b/lmdeploy/vl/hasher.py
@@ -0,0 +1,50 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import enum
+import hashlib
+from collections.abc import Mapping
+from typing import Any
+
+import numpy as np
+import torch
+from torch import Tensor
+
+
+def _hash_multimodal_value(hasher: 'hashlib._Hash', value: Any):
+    """Update a hash with a deterministic multimodal value representation."""
+    if isinstance(value, Tensor):
+        tensor = value.detach().cpu().contiguous()
+        hasher.update(f'tensor:{tensor.dtype}:{tuple(tensor.shape)}:'.encode())
+        hasher.update(tensor.view(torch.uint8).numpy().tobytes())
+    elif isinstance(value, np.ndarray):
+        array = np.ascontiguousarray(value)
+        hasher.update(f'ndarray:{array.dtype}:{array.shape}:'.encode())
+        hasher.update(array.tobytes())
+    elif isinstance(value, Mapping):
+        hasher.update(b'dict:{')
+        for key in sorted(value, key=lambda x: repr(x)):
+            _hash_multimodal_value(hasher, key)
+            hasher.update(b':')
+            _hash_multimodal_value(hasher, value[key])
+            hasher.update(b',')
+        hasher.update(b'}')
+    elif isinstance(value, (list, tuple)):
+        hasher.update(f'{type(value).__name__}:['.encode())
+        for item in value:
+            _hash_multimodal_value(hasher, item)
+            hasher.update(b',')
+        hasher.update(b']')
+    elif isinstance(value, enum.Enum):
+        _hash_multimodal_value(hasher, value.value)
+    else:
+        hasher.update(f'{type(value).__name__}:{repr(value)}'.encode())
+
+
+def make_multimodal_content_hash(data: Any,
+                                 meta: Mapping[str, Any] | None = None,
+                                 mrope_pos_ids: np.ndarray | None = None) -> str:
+    """Create a stable content hash for multimodal content identity."""
+    hasher = hashlib.sha256()
+    _hash_multimodal_value(hasher, data)
+    _hash_multimodal_value(hasher, meta)
+    _hash_multimodal_value(hasher, mrope_pos_ids)
+    return hasher.hexdigest()
diff --git a/lmdeploy/vl/model/preprocess_utils.py b/lmdeploy/vl/model/preprocess_utils.py
index f267722ff7..8e6346625f 100644
--- a/lmdeploy/vl/model/preprocess_utils.py
+++ b/lmdeploy/vl/model/preprocess_utils.py
@@ -6,6 +6,7 @@
 
 from lmdeploy.utils import get_logger
 from lmdeploy.vl.constants import Modality
+from lmdeploy.vl.hasher import make_multimodal_content_hash
 
 if TYPE_CHECKING:
     from lmdeploy.vl.model.base import MultimodalSpecialTokens
@@ -203,6 +204,26 @@ def _expand_bundled_audio_items(item: dict, token_id: int) -> list[dict]:
     return expanded_mm_items
 
 
+def attach_multimodal_content_hashes(result):
+    """Attach content hashes to new-style and legacy preprocess outputs."""
+    if isinstance(result, dict):
+        mm_items = result.get('multimodal') or []
+    else:
+        mm_items = result
+        if result and 'role' in result[0]:
+            mm_items = []
+            for message in result:
+                if message.get('role') == 'preprocess':
+                    mm_items = message.get('content') or []
+                    break
+
+    for item in mm_items:
+        content_view = {key: value for key, value in item.items() if key not in ('content_hash', 'offset')}
+        item['content_hash'] = make_multimodal_content_hash(content_view)
+
+    return result
+
+
 # adapted from https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/managers/mm_utils.py
 def get_expanded_mm_items(collected_mm_items, mm_tokens: 'MultimodalSpecialTokens'):
     """When multiple mm items of the same modality present, HF processors
diff --git a/tests/pytorch/engine/test_multimodal_content_hash.py b/tests/pytorch/engine/test_multimodal_content_hash.py
new file mode 100644
index 0000000000..c29182419c
--- /dev/null
+++ b/tests/pytorch/engine/test_multimodal_content_hash.py
@@ -0,0 +1,39 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import torch
+
+from lmdeploy.pytorch.models.qwen3_vl import Qwen3VLInputProcessor
+from lmdeploy.vl.constants import Modality
+
+
+def test_qwen3vl_input_processor_preserves_image_content_hash():
+    processor = Qwen3VLInputProcessor(config=object(), dtype=torch.float32)
+    result = processor.preprocess_input(
+        input_ids=[1, 2, 3],
+        input_multimodals=[{
+            'modality': Modality.IMAGE,
+            'pixel_values': torch.ones(1, 1),
+            'image_grid_thw': torch.tensor([1, 2, 2]),
+            'offset': (1, 2),
+            'image_token_id': 99,
+            'content_hash': 'image-a',
+        }],
+    )
+
+    assert result.input_multimodals['mm_data'][0].content_hash == 'image-a'
+
+
+def test_qwen3vl_input_processor_preserves_video_content_hash():
+    processor = Qwen3VLInputProcessor(config=object(), dtype=torch.float32)
+    result = processor.preprocess_input(
+        input_ids=[1, 2, 3],
+        input_multimodals=[{
+            'modality': Modality.VIDEO,
+            'pixel_values_videos': torch.ones(1, 1),
+            'video_grid_thw': torch.tensor([1, 2, 2]),
+            'offset': (1, 2),
+            'video_token_id': 99,
+            'content_hash': 'video-a',
+        }],
+    )
+
+    assert result.input_multimodals['mm_data'][0].content_hash == 'video-a'
diff --git a/tests/test_lmdeploy/test_vl/test_hasher.py b/tests/test_lmdeploy/test_vl/test_hasher.py
new file mode 100644
index 0000000000..f0045ab3c2
--- /dev/null
+++ b/tests/test_lmdeploy/test_vl/test_hasher.py
@@ -0,0 +1,33 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import numpy as np
+import torch
+
+from lmdeploy.vl.constants import Modality
+from lmdeploy.vl.hasher import make_multimodal_content_hash
+
+
+def test_multimodal_content_hash_is_stable_for_nested_values():
+    data = {
+        'tensor': torch.arange(4, dtype=torch.float32).reshape(2, 2),
+        'array': np.arange(3, dtype=np.int64),
+        'modality': Modality.IMAGE,
+    }
+    meta = {'image_token_id': 99, 'shape': [2, 2]}
+    mrope_pos_ids = np.arange(6, dtype=np.int64).reshape(2, 3)
+
+    hash1 = make_multimodal_content_hash(data, meta, mrope_pos_ids)
+    hash2 = make_multimodal_content_hash(dict(reversed(data.items())), dict(reversed(meta.items())),
+                                         mrope_pos_ids.copy())
+
+    assert hash1 == hash2
+
+
+def test_multimodal_content_hash_changes_with_payload_meta_or_mrope():
+    data = torch.arange(4, dtype=torch.float32).reshape(2, 2)
+    meta = {'image_token_id': 99}
+    mrope_pos_ids = np.arange(6, dtype=np.int64).reshape(2, 3)
+    base = make_multimodal_content_hash(data, meta, mrope_pos_ids)
+
+    assert base != make_multimodal_content_hash(data + 1, meta, mrope_pos_ids)
+    assert base != make_multimodal_content_hash(data, {'image_token_id': 100}, mrope_pos_ids)
+    assert base != make_multimodal_content_hash(data, meta, mrope_pos_ids + 1)
diff --git a/tests/test_lmdeploy/test_vl/test_preprocess_utils.py b/tests/test_lmdeploy/test_vl/test_preprocess_utils.py
index 22084f0e88..58b0e1ea43 100644
--- a/tests/test_lmdeploy/test_vl/test_preprocess_utils.py
+++ b/tests/test_lmdeploy/test_vl/test_preprocess_utils.py
@@ -3,7 +3,8 @@
 import torch
 
 from lmdeploy.vl.constants import Modality
-from lmdeploy.vl.model.preprocess_utils import get_expanded_mm_items
+from lmdeploy.vl.hasher import make_multimodal_content_hash
+from lmdeploy.vl.model.preprocess_utils import attach_multimodal_content_hashes, get_expanded_mm_items
 
 
 class _Tokens:
@@ -97,3 +98,86 @@ def test_expand_audio_items_use_compact_tensor_storage():
     for entry in expanded:
         _assert_compact_storage(entry['input_features'])
         _assert_compact_storage(entry['feature_attention_mask'])
+
+
+def test_attach_multimodal_content_hashes_to_single_image():
+    items = {
+        Modality.IMAGE: {
+            'feature': torch.arange(2, dtype=torch.float32).reshape(2, 1),
+            'image_grid_thw': torch.tensor([[1, 2, 1]]),
+            'offset': [(0, 2)],
+        }
+    }
+    expanded = get_expanded_mm_items(items, _Tokens())
+
+    attach_multimodal_content_hashes(expanded)
+
+    assert len(expanded[0]['content_hash']) == 64
+    content_view = {key: value for key, value in expanded[0].items() if key not in ('content_hash', 'offset')}
+    assert expanded[0]['content_hash'] == make_multimodal_content_hash(content_view)
+
+
+def test_attach_multimodal_content_hashes_ignores_offset():
+    item1 = {
+        'modality': Modality.IMAGE,
+        'pixel_values': torch.arange(2, dtype=torch.float32).reshape(2, 1),
+        'image_grid_thw': torch.tensor([1, 2, 1]),
+        'offset': (0, 2),
+        'image_token_id': 42,
+    }
+    item2 = dict(item1, offset=(10, 12))
+
+    attach_multimodal_content_hashes([item1, item2])
+
+    assert item1['content_hash'] == item2['content_hash']
+
+
+def test_attach_multimodal_content_hashes_changes_with_processed_content():
+    item1 = {
+        'modality': Modality.IMAGE,
+        'pixel_values': torch.arange(2, dtype=torch.float32).reshape(2, 1),
+        'image_grid_thw': torch.tensor([1, 2, 1]),
+        'offset': (0, 2),
+        'image_token_id': 42,
+    }
+    item2 = dict(item1, pixel_values=item1['pixel_values'] + 1)
+
+    attach_multimodal_content_hashes([item1, item2])
+
+    assert item1['content_hash'] != item2['content_hash']
+
+
+def test_attach_multimodal_content_hashes_to_frame_split_videos():
+    items = {
+        Modality.VIDEO: {
+            'feature': torch.arange(8, dtype=torch.float32).reshape(8, 1),
+            'video_grid_thw': torch.tensor([[2, 2, 1], [2, 2, 1]]),
+            'offset': [(0, 1), (1, 2), (2, 3), (3, 4)],
+        }
+    }
+    expanded = get_expanded_mm_items(items, _Tokens())
+
+    attach_multimodal_content_hashes(expanded)
+
+    assert len({entry['content_hash'] for entry in expanded}) == len(expanded)
+
+
+def test_attach_multimodal_content_hashes_preserves_prompt_order_for_mixed_modalities():
+    items = {
+        Modality.AUDIO: {
+            'feature': torch.arange(1 * 2 * 3, dtype=torch.float32).reshape(1, 2, 3),
+            'feature_attention_mask': torch.ones(1, 3, dtype=torch.long),
+            'offset': [(5, 8)],
+        },
+        Modality.IMAGE: {
+            'feature': torch.arange(2, dtype=torch.float32).reshape(2, 1),
+            'image_grid_thw': torch.tensor([[1, 2, 1]]),
+            'offset': [(0, 2)],
+        },
+    }
+    expanded = get_expanded_mm_items(items, _Tokens())
+
+    attach_multimodal_content_hashes(expanded)
+
+    assert [entry['modality'] for entry in expanded] == [Modality.IMAGE, Modality.AUDIO]
+    assert all(len(entry['content_hash']) == 64 for entry in expanded)