[Frontend] Support configurable mm placeholder strings & flexible video sampling policies via CLI flags.

tests/async_engine/test_async_llm_engine.py

@@ -6,8 +6,8 @@
 import uuid
 from asyncio import CancelledError
 from copy import copy
-from dataclasses import dataclass
-from typing import Optional
+from dataclasses import dataclass, field
+from typing import Any, Optional
 
 import pytest
 import pytest_asyncio
@@ -32,6 +32,8 @@ class RequestOutput:
 @dataclass
 class MockModelConfig:
     use_async_output_proc = True
+    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
+    mm_placeholder_str_override: dict[str, str] = field(default_factory=dict)
 
 
 class MockEngine:

tests/engine/test_arg_utils.py

@@ -231,6 +231,58 @@ def test_limit_mm_per_prompt_parser(arg, expected):
     assert args.limit_mm_per_prompt == expected
 
 
+@pytest.mark.parametrize(
+    ("arg", "expected"),
+    [
+        (None, dict()),
+        ('{"video": {"num_frames": 123} }', {
+            "video": {
+                "num_frames": 123
+            }
+        }),
+        (
+            '{"video": {"num_frames": 123, "fps": 1.0, "foo": "bar"}, "image": {"foo": "bar"} }',  # noqa
+            {
+                "video": {
+                    "num_frames": 123,
+                    "fps": 1.0,
+                    "foo": "bar"
+                },
+                "image": {
+                    "foo": "bar"
+                }
+            }),
+    ])
+def test_media_io_kwargs_parser(arg, expected):
+    parser = EngineArgs.add_cli_args(FlexibleArgumentParser())
+    if arg is None:
+        args = parser.parse_args([])
+    else:
+        args = parser.parse_args(["--media-io-kwargs", arg])
+
+    assert args.media_io_kwargs == expected
+
+
+@pytest.mark.parametrize(("arg", "expected"), [
+    (None, dict()),
+    ('{"video":"<|video_placeholder|>"}', {
+        "video": "<|video_placeholder|>"
+    }),
+    ('{"video":"<|video_placeholder|>", "image": "<|image_placeholder|>"}', {
+        "video": "<|video_placeholder|>",
+        "image": "<|image_placeholder|>"
+    }),
+])
+def test_mm_placeholder_str_override_parser(arg, expected):
+    parser = EngineArgs.add_cli_args(FlexibleArgumentParser())
+    if arg is None:
+        args = parser.parse_args([])
+    else:
+        args = parser.parse_args(["--mm-placeholder-str-override", arg])
+
+    assert args.mm_placeholder_str_override == expected
+
+
 def test_compilation_config():
     parser = EngineArgs.add_cli_args(FlexibleArgumentParser())
 

tests/entrypoints/openai/test_serving_chat.py

@@ -3,8 +3,8 @@
 
 import asyncio
 from contextlib import suppress
-from dataclasses import dataclass
-from typing import Optional
+from dataclasses import dataclass, field
+from typing import Any, Optional
 from unittest.mock import MagicMock
 
 from vllm.config import MultiModalConfig
@@ -40,6 +40,8 @@ class MockModelConfig:
     allowed_local_media_path: str = ""
     encoder_config = None
     generation_config: str = "auto"
+    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
+    mm_placeholder_str_override: dict[str, str] = field(default_factory=dict)
 
     def get_diff_sampling_param(self):
         return self.diff_sampling_param or {}

tests/multimodal/test_utils.py

@@ -169,9 +169,8 @@ async def test_fetch_image_error_conversion():
 async def test_fetch_video_http(video_url: str, num_frames: int):
     connector = MediaConnector()
 
-    video_sync = connector.fetch_video(video_url, num_frames=num_frames)
-    video_async = await connector.fetch_video_async(video_url,
-                                                    num_frames=num_frames)
+    video_sync = connector.fetch_video(video_url)
+    video_async = await connector.fetch_video_async(video_url)
     assert np.array_equal(video_sync, video_async)
 
 

tests/multimodal/test_video.py

@@ -4,7 +4,10 @@
 import numpy.typing as npt
 import pytest
 
-from vllm.multimodal.video import VIDEO_LOADER_REGISTRY, VideoLoader
+from vllm import envs
+from vllm.multimodal.image import ImageMediaIO
+from vllm.multimodal.video import (VIDEO_LOADER_REGISTRY, VideoLoader,
+                                   VideoMediaIO)
 
 NUM_FRAMES = 10
 FAKE_OUTPUT_1 = np.random.rand(NUM_FRAMES, 1280, 720, 3)
@@ -40,3 +43,46 @@ def test_video_loader_registry():
 def test_video_loader_type_doesnt_exist():
     with pytest.raises(AssertionError):
         VIDEO_LOADER_REGISTRY.load("non_existing_video_loader")
+
+
+@VIDEO_LOADER_REGISTRY.register("assert_10_frames_1_fps")
+class Assert10Frames1FPSVideoLoader(VideoLoader):
+
+    @classmethod
+    def load_bytes(cls,
+                   data: bytes,
+                   num_frames: int = -1,
+                   fps: float = -1.0,
+                   **kwargs) -> npt.NDArray:
+        assert num_frames == 10, "bad num_frames"
+        assert fps == 1.0, "bad fps"
+        return FAKE_OUTPUT_2
+
+
+def test_video_media_io_kwargs():
+    envs.VLLM_VIDEO_LOADER_BACKEND = "assert_10_frames_1_fps"
+    imageio = ImageMediaIO()
+
+    # Verify that different args pass/fail assertions as expected.
+    videoio = VideoMediaIO(imageio, **{"num_frames": 10, "fps": 1.0})
+    _ = videoio.load_bytes(b"test")
+
+    videoio = VideoMediaIO(
+        imageio, **{
+            "num_frames": 10,
+            "fps": 1.0,
+            "not_used": "not_used"
+        })
+    _ = videoio.load_bytes(b"test")
+
+    with pytest.raises(AssertionError, match="bad num_frames"):
+        videoio = VideoMediaIO(imageio, **{})
+        _ = videoio.load_bytes(b"test")
+
+    with pytest.raises(AssertionError, match="bad num_frames"):
+        videoio = VideoMediaIO(imageio, **{"num_frames": 9, "fps": 1.0})
+        _ = videoio.load_bytes(b"test")
+
+    with pytest.raises(AssertionError, match="bad fps"):
+        videoio = VideoMediaIO(imageio, **{"num_frames": 10, "fps": 2.0})
+        _ = videoio.load_bytes(b"test")

vllm/config.py

@@ -356,6 +356,12 @@ class ModelConfig:
     limit_mm_per_prompt: dict[str, int] = field(default_factory=dict)
     """Maximum number of data items per modality per prompt. Only applicable
     for multimodal models."""
+    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
+    """Additional args passed to process media inputs, keyed by modalities. 
+    For example, to set num_frames for video, set 
+    `--media-io-kwargs '{"video": {"num_frames": 40} }'` """
+    mm_placeholder_str_override: dict[str, str] = field(default_factory=dict)
+    """Optionally override placeholder string for given modalities."""
     use_async_output_proc: bool = True
     """Whether to use async output processor."""
     config_format: Union[str, ConfigFormat] = ConfigFormat.AUTO.value
@@ -690,6 +696,8 @@ def _init_multimodal_config(self) -> Optional["MultiModalConfig"]:
         if self.registry.is_multimodal_model(self.architectures):
             return MultiModalConfig(
                 limit_per_prompt=self.limit_mm_per_prompt,
+                media_io_kwargs=self.media_io_kwargs,
+                mm_placeholder_str_override=self.mm_placeholder_str_override,
                 mm_processor_kwargs=self.mm_processor_kwargs,
                 disable_mm_preprocessor_cache=self.
                 disable_mm_preprocessor_cache)
@@ -3000,6 +3008,14 @@ class MultiModalConfig:
     `{"images": 16, "videos": 2}`
     """
 
+    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
+    """Additional args passed to process media inputs, keyed by modalities. 
+    For example, to set num_frames for video, set 
+    `--media-io-kwargs '{"video": {"num_frames": 40} }'` """
+
+    mm_placeholder_str_override: dict[str, str] = field(default_factory=dict)
+    """Optionally override placeholder string for given modalities."""
+
     mm_processor_kwargs: Optional[dict[str, object]] = None
     """
     Overrides for the multi-modal processor obtained from

vllm/engine/arg_utils.py

@@ -348,6 +348,11 @@ class EngineArgs:
         get_field(TokenizerPoolConfig, "extra_config")
     limit_mm_per_prompt: dict[str, int] = \
         get_field(MultiModalConfig, "limit_per_prompt")
+    media_io_kwargs: dict[str, dict[str,
+                                    Any]] = get_field(MultiModalConfig,
+                                                      "media_io_kwargs")
+    mm_placeholder_str_override: dict[str, str] = \
+        get_field(MultiModalConfig, "mm_placeholder_str_override")
     mm_processor_kwargs: Optional[Dict[str, Any]] = \
         MultiModalConfig.mm_processor_kwargs
     disable_mm_preprocessor_cache: bool = \
@@ -714,6 +719,11 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
         )
         multimodal_group.add_argument("--limit-mm-per-prompt",
                                       **multimodal_kwargs["limit_per_prompt"])
+        multimodal_group.add_argument("--media-io-kwargs",
+                                      **multimodal_kwargs["media_io_kwargs"])
+        multimodal_group.add_argument(
+            "--mm-placeholder-str-override",
+            **multimodal_kwargs["mm_placeholder_str_override"])
         multimodal_group.add_argument(
             "--mm-processor-kwargs",
             **multimodal_kwargs["mm_processor_kwargs"])
@@ -938,6 +948,8 @@ def create_model_config(self) -> ModelConfig:
             enable_prompt_embeds=self.enable_prompt_embeds,
             served_model_name=self.served_model_name,
             limit_mm_per_prompt=self.limit_mm_per_prompt,
+            media_io_kwargs=self.media_io_kwargs,
+            mm_placeholder_str_override=self.mm_placeholder_str_override,
             use_async_output_proc=not self.disable_async_output_proc,
             config_format=self.config_format,
             mm_processor_kwargs=self.mm_processor_kwargs,

vllm/entrypoints/chat_utils.py

@@ -506,6 +506,9 @@ def _cached_token_str(tokenizer: AnyTokenizer, token_index: int) -> str:
 
     def _placeholder_str(self, modality: ModalityStr,
                          current_count: int) -> Optional[str]:
+        if modality in self._model_config.mm_placeholder_str_override:
+            return self._model_config.mm_placeholder_str_override[modality]
+
         # TODO: Let user specify how to insert image tokens into prompt
         # (similar to chat template)
         hf_config = self._model_config.hf_config
@@ -720,6 +723,7 @@ def __init__(self, tracker: MultiModalItemTracker) -> None:
         self._tracker = tracker
 
         self._connector = MediaConnector(
+            media_io_kwargs=self._tracker._model_config.media_io_kwargs,
             allowed_local_media_path=tracker.allowed_local_media_path,
         )
 
@@ -758,7 +762,7 @@ def parse_input_audio(self, input_audio: InputAudio) -> None:
         return self.parse_audio(audio_url)
 
     def parse_video(self, video_url: str) -> None:
-        video = self._connector.fetch_video(video_url)
+        video = self._connector.fetch_video(video_url=video_url)
 
         placeholder = self._tracker.add("video", video)
         self._add_placeholder(placeholder)
@@ -771,7 +775,8 @@ def __init__(self, tracker: AsyncMultiModalItemTracker) -> None:
 
         self._tracker = tracker
         self._connector = MediaConnector(
-            allowed_local_media_path=tracker.allowed_local_media_path,
+            media_io_kwargs=self._tracker._model_config.media_io_kwargs,
+            allowed_local_media_path=tracker.allowed_local_media_path
         )
 
     def parse_image(self, image_url: str) -> None:
@@ -813,7 +818,7 @@ def parse_input_audio(self, input_audio: InputAudio) -> None:
         return self.parse_audio(audio_url)
 
     def parse_video(self, video_url: str) -> None:
-        video = self._connector.fetch_video_async(video_url)
+        video = self._connector.fetch_video_async(video_url=video_url)
 
         placeholder = self._tracker.add("video", video)
         self._add_placeholder(placeholder)

vllm/multimodal/audio.py

@@ -83,6 +83,16 @@ def resample(
 
 class AudioMediaIO(MediaIO[tuple[npt.NDArray, float]]):
 
+    def __init__(self, **kwargs) -> None:
+        super().__init__()
+
+        # `kwargs` contains custom arguments from
+        # --media-io-kwargs for this modality.
+        # They can be passed to the underlying
+        # media loaders (e.g. custom implementations)
+        # for flexible control.
+        self.kwargs = kwargs
+
     def load_bytes(self, data: bytes) -> tuple[npt.NDArray, float]:
         return librosa.load(BytesIO(data), sr=None)
 

vllm/multimodal/image.py

@@ -44,10 +44,16 @@ def convert_image_mode(image: Image.Image, to_mode: str):
 
 class ImageMediaIO(MediaIO[Image.Image]):
 
-    def __init__(self, *, image_mode: str = "RGB") -> None:
+    def __init__(self, image_mode: str = "RGB", **kwargs) -> None:
         super().__init__()
 
         self.image_mode = image_mode
+        # `kwargs` contains custom arguments from
+        # --media-io-kwargs for this modality.
+        # They can be passed to the underlying
+        # media loaders (e.g. custom implementations)
+        # for flexible control.
+        self.kwargs = kwargs
 
     def load_bytes(self, data: bytes) -> Image.Image:
         image = Image.open(BytesIO(data))