diff --git a/images/01.jpg b/images/01.jpg
new file mode 100755
index 000000000000..9c10070b1e73
Binary files /dev/null and b/images/01.jpg differ
diff --git a/images/02.jpg b/images/02.jpg
new file mode 100755
index 000000000000..4c4a8eef1d6a
Binary files /dev/null and b/images/02.jpg differ
diff --git a/images/03.jpg b/images/03.jpg
new file mode 100755
index 000000000000..0e125a2c5aab
Binary files /dev/null and b/images/03.jpg differ
diff --git a/images/04.jpg b/images/04.jpg
new file mode 100755
index 000000000000..3db3034c5e84
Binary files /dev/null and b/images/04.jpg differ
diff --git a/images/05.jpg b/images/05.jpg
new file mode 100755
index 000000000000..e8ba056241b4
Binary files /dev/null and b/images/05.jpg differ
diff --git a/images/06.jpg b/images/06.jpg
new file mode 100755
index 000000000000..3a545d3ec892
Binary files /dev/null and b/images/06.jpg differ
diff --git a/images/07.jpg b/images/07.jpg
new file mode 100755
index 000000000000..08dd709d9056
Binary files /dev/null and b/images/07.jpg differ
diff --git a/images/08.jpg b/images/08.jpg
new file mode 100755
index 000000000000..f1cf8ff0cc72
Binary files /dev/null and b/images/08.jpg differ
diff --git a/images/09.jpg b/images/09.jpg
new file mode 100755
index 000000000000..d0c807fb8b66
Binary files /dev/null and b/images/09.jpg differ
diff --git a/images/10.jpg b/images/10.jpg
new file mode 100755
index 000000000000..09401554a771
Binary files /dev/null and b/images/10.jpg differ
diff --git a/images/11.jpg b/images/11.jpg
new file mode 100755
index 000000000000..e5328402a251
Binary files /dev/null and b/images/11.jpg differ
diff --git a/images/12.jpg b/images/12.jpg
new file mode 100755
index 000000000000..3dc87333c030
Binary files /dev/null and b/images/12.jpg differ
diff --git a/images/13.jpg b/images/13.jpg
new file mode 100755
index 000000000000..8cf0572d6fdc
Binary files /dev/null and b/images/13.jpg differ
diff --git a/images/14.jpg b/images/14.jpg
new file mode 100755
index 000000000000..2c3bface8727
Binary files /dev/null and b/images/14.jpg differ
diff --git a/images/r01.jpg b/images/r01.jpg
new file mode 100644
index 000000000000..b1992966a74d
Binary files /dev/null and b/images/r01.jpg differ
diff --git a/images/r02.jpg b/images/r02.jpg
new file mode 100644
index 000000000000..791f76601ad0
Binary files /dev/null and b/images/r02.jpg differ
diff --git a/images/r03.jpg b/images/r03.jpg
new file mode 100644
index 000000000000..4283d8b79875
Binary files /dev/null and b/images/r03.jpg differ
diff --git a/images/r04.jpg b/images/r04.jpg
new file mode 100644
index 000000000000..c69c07bf20f7
Binary files /dev/null and b/images/r04.jpg differ
diff --git a/images/r05.jpg b/images/r05.jpg
new file mode 100644
index 000000000000..9e75e3d126d7
Binary files /dev/null and b/images/r05.jpg differ
diff --git a/images/r06.jpg b/images/r06.jpg
new file mode 100644
index 000000000000..2f99dbb87af4
Binary files /dev/null and b/images/r06.jpg differ
diff --git a/images/r07.jpg b/images/r07.jpg
new file mode 100644
index 000000000000..f13b4d66e551
Binary files /dev/null and b/images/r07.jpg differ
diff --git a/images/r08.jpg b/images/r08.jpg
new file mode 100644
index 000000000000..8620f5413a76
Binary files /dev/null and b/images/r08.jpg differ
diff --git a/images/r09.jpg b/images/r09.jpg
new file mode 100644
index 000000000000..d60af7639001
Binary files /dev/null and b/images/r09.jpg differ
diff --git a/images/r10.jpg b/images/r10.jpg
new file mode 100644
index 000000000000..6a4d2229fdbd
Binary files /dev/null and b/images/r10.jpg differ
diff --git a/images/r11.jpg b/images/r11.jpg
new file mode 100644
index 000000000000..e25b30c264b1
Binary files /dev/null and b/images/r11.jpg differ
diff --git a/images/r12.jpg b/images/r12.jpg
new file mode 100644
index 000000000000..3ea859a16042
Binary files /dev/null and b/images/r12.jpg differ
diff --git a/images/r13.jpg b/images/r13.jpg
new file mode 100644
index 000000000000..c9c523fb669b
Binary files /dev/null and b/images/r13.jpg differ
diff --git a/images/r14.jpg b/images/r14.jpg
new file mode 100644
index 000000000000..945e068e89cb
Binary files /dev/null and b/images/r14.jpg differ
diff --git a/images/r15.jpg b/images/r15.jpg
new file mode 100644
index 000000000000..4283d8b79875
Binary files /dev/null and b/images/r15.jpg differ
diff --git a/images/r16.jpg b/images/r16.jpg
new file mode 100644
index 000000000000..f13b4d66e551
Binary files /dev/null and b/images/r16.jpg differ
diff --git a/images/result.jpg b/images/result.jpg
new file mode 100644
index 000000000000..904bb872fe22
Binary files /dev/null and b/images/result.jpg differ
diff --git a/tests/v1/test_serial_utils.py b/tests/v1/test_serial_utils.py
index 0fc3b074533d..9172a946d484 100644
--- a/tests/v1/test_serial_utils.py
+++ b/tests/v1/test_serial_utils.py
@@ -1,10 +1,15 @@
 # SPDX-License-Identifier: Apache-2.0
 from collections import UserDict
 from dataclasses import dataclass
+from typing import Optional
 
+import msgspec
 import numpy as np
 import torch
 
+from vllm.multimodal.inputs import (MultiModalBatchedField,
+                                    MultiModalFieldElem, MultiModalKwargs,
+                                    MultiModalKwargsItem, NestedTensors)
 from vllm.v1.serial_utils import MsgpackDecoder, MsgpackEncoder
 
 
@@ -36,7 +41,7 @@ def test_encode_decode():
         list_of_tensors=[
             torch.rand((1, 10), dtype=torch.float32),
             torch.rand((3, 5, 4000), dtype=torch.float64),
-            torch.tensor(1984),  # test scalar too
+            torch.tensor(1984),  # test scalar too,
         ],
         numpy_array=np.arange(512),
         unrecognized=UnrecognizedType(33),
@@ -70,6 +75,99 @@ def test_encode_decode():
     assert_equal(decoded2, obj)
 
 
+class MyRequest(msgspec.Struct):
+    mm: Optional[list[MultiModalKwargs]]
+
+
+def test_multimodal_kwargs():
+    d = {
+        "foo":
+        torch.zeros(1000, dtype=torch.float16),
+        "bar": [torch.zeros(i * 1000, dtype=torch.int8) for i in range(3)],
+        "baz": [
+            torch.rand((256), dtype=torch.float16),
+            [
+                torch.rand((1, 12), dtype=torch.float32),
+                torch.rand((3, 5, 7), dtype=torch.float64),
+            ], [torch.rand((4, 4), dtype=torch.float16)]
+        ],
+    }
+
+    # pack mm kwargs into a mock request so that it can be decoded properly
+    req = MyRequest(mm=[MultiModalKwargs(d)])
+
+    encoder = MsgpackEncoder()
+    decoder = MsgpackDecoder(MyRequest)
+
+    encoded = encoder.encode(req)
+
+    # 8 total tensors + top level buffer
+    assert len(encoded) == 9
+
+    total_len = sum([len(x) for x in encoded])
+
+    # expected total encoding length, should be 4362, +-20 for minor changes
+    assert total_len >= 4342 and total_len <= 4382
+    decoded: MultiModalKwargs = decoder.decode(encoded).mm[0]
+    assert all(nested_equal(d[k], decoded[k]) for k in d)
+
+
+def test_multimodal_items_by_modality():
+    e1 = MultiModalFieldElem("audio", "a0", torch.zeros(1000,
+                                                        dtype=torch.int16),
+                             MultiModalBatchedField())
+    e2 = MultiModalFieldElem(
+        "video",
+        "v0",
+        [torch.zeros(1000, dtype=torch.int8) for _ in range(4)],
+        MultiModalBatchedField(),
+    )
+    e3 = MultiModalFieldElem("image", "i0", torch.zeros(1000,
+                                                        dtype=torch.int32),
+                             MultiModalBatchedField())
+    e4 = MultiModalFieldElem("image", "i1", torch.zeros(1000,
+                                                        dtype=torch.int32),
+                             MultiModalBatchedField())
+    audio = MultiModalKwargsItem.from_elems([e1])
+    video = MultiModalKwargsItem.from_elems([e2])
+    image = MultiModalKwargsItem.from_elems([e3, e4])
+    mm = MultiModalKwargs.from_items([audio, video, image])
+
+    # pack mm kwargs into a mock request so that it can be decoded properly
+    req = MyRequest([mm])
+
+    encoder = MsgpackEncoder()
+    decoder = MsgpackDecoder(MyRequest)
+
+    encoded = encoder.encode(req)
+
+    # 5 total tensors + top level buffer
+    assert len(encoded) == 8
+
+    total_len = sum([len(x) for x in encoded])
+
+    # expected total encoding length, should be 7507, +-20 for minor changes
+    assert total_len >= 7487 and total_len <= 7527
+    decoded: MultiModalKwargs = decoder.decode(encoded).mm[0]
+
+    # check all modalities were recovered and do some basic sanity checks
+    assert len(decoded.modalities) == 3
+    images = decoded.get_items("image")
+    assert len(images) == 1
+    assert len(images[0].items()) == 2
+    assert list(images[0].keys()) == ["i0", "i1"]
+
+    # check the tensor contents and layout in the main dict
+    assert all(nested_equal(mm[k], decoded[k]) for k in mm)
+
+
+def nested_equal(a: NestedTensors, b: NestedTensors):
+    if isinstance(a, torch.Tensor):
+        return torch.equal(a, b)
+    else:
+        return all([nested_equal(x, y) for (x, y) in zip(a, b)])
+
+
 def assert_equal(obj1: MyType, obj2: MyType):
     assert torch.equal(obj1.tensor1, obj2.tensor1)
     assert obj1.a_string == obj2.a_string
diff --git a/vllm/v1/serial_utils.py b/vllm/v1/serial_utils.py
index 99b352fdef80..2651540021fc 100644
--- a/vllm/v1/serial_utils.py
+++ b/vllm/v1/serial_utils.py
@@ -2,7 +2,9 @@
 
 import pickle
 from collections.abc import Sequence
+from dataclasses import asdict
 from inspect import isclass
+from itertools import chain
 from types import FunctionType
 from typing import Any, Optional, Union
 
@@ -12,12 +14,17 @@
 import zmq
 from msgspec import msgpack
 
+from vllm.logger import init_logger
+from vllm.multimodal.inputs import (MultiModalFieldElem, MultiModalKwargs,
+                                    MultiModalKwargsItem, NestedTensors)
+
 CUSTOM_TYPE_PICKLE = 1
 CUSTOM_TYPE_CLOUDPICKLE = 2
 
 # TODO calibrate this size
 INLINE_BUF_SIZE_THRESHOLD = 256
 
+logger = init_logger(__name__)
 bytestr = Union[bytes, bytearray, memoryview, zmq.Frame]
 
 
@@ -64,6 +71,25 @@ def enc_hook(self, obj: Any) -> Any:
         if isinstance(obj, np.ndarray) and obj.dtype.kind not in ('O', 'V'):
             return self._encode_ndarray(obj)
 
+        if isinstance(obj, MultiModalKwargs):
+            mm: MultiModalKwargs = obj
+            if mm.modalities:
+                # ignore the main dict, it will be re-indexed.
+                # pass a list of MultiModalKwargsItem, then see below
+                # Any tensors *not* indexed by modality will be ignored.
+                return [mm.get_items(m) for m in mm.modalities]
+            # just return the main dict if there are no modalities
+            return {k: v for k, v in obj.items()}
+
+        if isinstance(obj, MultiModalKwargsItem):
+            rd = {}
+            for k, v in obj.items():
+                vv = asdict(v)
+                vv['field'] = pickle.dumps(v.field,
+                                           protocol=pickle.HIGHEST_PROTOCOL)
+                rd[k] = vv
+            return rd
+
         if isinstance(obj, FunctionType):
             # `pickle` is generally faster than cloudpickle, but can have
             # problems serializing methods.
@@ -87,7 +113,7 @@ def _encode_ndarray(
         # We serialize the ndarray as a tuple of native types.
         # The data is either inlined if small, or an index into a list of
         # backing buffers that we've stashed in `aux_buffers`.
-        return obj.dtype.str, obj.shape, data
+        return (obj.dtype.str, obj.shape, data)
 
 
 class MsgpackDecoder:
@@ -121,15 +147,39 @@ def dec_hook(self, t: type, obj: Any) -> Any:
         if isclass(t):
             if issubclass(t, np.ndarray):
                 return self._decode_ndarray(obj)
+            if issubclass(t, MultiModalKwargs) and isinstance(obj, dict):
+                return MultiModalKwargs(
+                    {k: self._decode_nested(obj[k])
+                     for k in obj})
+            if issubclass(t, MultiModalKwargs) and isinstance(obj, list):
+                return MultiModalKwargs.from_items(self._decode_items(obj))
             if issubclass(t, torch.Tensor):
                 return torch.from_numpy(self._decode_ndarray(obj))
         return obj
 
-    def _decode_ndarray(self, arr: Any) -> np.ndarray:
-        dtype, shape, data = arr
+    def _decode_ndarray(self, obj: Any) -> np.ndarray:
+        (dtype, shape, data) = obj
         buffer = self.aux_buffers[data] if isinstance(data, int) else data
         return np.ndarray(buffer=buffer, dtype=np.dtype(dtype), shape=shape)
 
+    def _decode_items(self, obj: list) -> list[MultiModalKwargsItem]:
+        all = []
+        for item in chain.from_iterable(obj):
+            elems = []
+            for v in item.values():
+                v['data'] = self._decode_nested(v['data'])
+                v['field'] = pickle.loads(v['field'])
+                elems.append(MultiModalFieldElem(**v))
+            all.append(MultiModalKwargsItem.from_elems(elems))
+        return all
+
+    def _decode_nested(self, obj: Any) -> NestedTensors:
+        if isinstance(obj, list) and isinstance(obj[0], str):
+            return torch.from_numpy(self._decode_ndarray(obj))
+        if isinstance(obj, list):
+            return [self._decode_nested(x) for x in obj]
+        raise TypeError(f"Unexpected NestedArray contents: {obj}")
+
     def ext_hook(self, code: int, data: memoryview) -> Any:
         if code == CUSTOM_TYPE_PICKLE:
             return pickle.loads(data)