[Bugfix] validate urls object for multimodal content parts (vllm-project#16990)

Signed-off-by: Guillaume Calmettes <[email protected]>

Author: gcalmettes

tests/entrypoints/openai/test_audio.py

@@ -104,6 +104,35 @@ async def test_single_chat_session_audio(client: openai.AsyncOpenAI,
     assert message.content is not None and len(message.content) >= 0
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("audio_url", [TEST_AUDIO_URLS[0]])
+async def test_error_on_invalid_audio_url_type(client: openai.AsyncOpenAI,
+                                               model_name: str,
+                                               audio_url: str):
+    messages = [{
+        "role":
+        "user",
+        "content": [
+            {
+                "type": "audio_url",
+                "audio_url": audio_url
+            },
+            {
+                "type": "text",
+                "text": "What's happening in this audio?"
+            },
+        ],
+    }]
+
+    # audio_url should be a dict {"url": "some url"}, not directly a string
+    with pytest.raises(openai.BadRequestError):
+        _ = await client.chat.completions.create(model=model_name,
+                                                 messages=messages,
+                                                 max_completion_tokens=10,
+                                                 temperature=0.0)
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
 @pytest.mark.parametrize("audio_url", [TEST_AUDIO_URLS[0]])

tests/entrypoints/openai/test_video.py

@@ -108,6 +108,35 @@ async def test_single_chat_session_video(client: openai.AsyncOpenAI,
     assert message.content is not None and len(message.content) >= 0
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("video_url", TEST_VIDEO_URLS)
+async def test_error_on_invalid_video_url_type(client: openai.AsyncOpenAI,
+                                               model_name: str,
+                                               video_url: str):
+    messages = [{
+        "role":
+        "user",
+        "content": [
+            {
+                "type": "video_url",
+                "video_url": video_url
+            },
+            {
+                "type": "text",
+                "text": "What's in this video?"
+            },
+        ],
+    }]
+
+    # video_url should be a dict {"url": "some url"}, not directly a string
+    with pytest.raises(openai.BadRequestError):
+        _ = await client.chat.completions.create(model=model_name,
+                                                 messages=messages,
+                                                 max_completion_tokens=10,
+                                                 temperature=0.0)
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
 @pytest.mark.parametrize("video_url", TEST_VIDEO_URLS)

tests/entrypoints/openai/test_vision.py

@@ -137,6 +137,36 @@ async def test_single_chat_session_image(client: openai.AsyncOpenAI,
     assert message.content is not None and len(message.content) >= 0
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)
+async def test_error_on_invalid_image_url_type(client: openai.AsyncOpenAI,
+                                               model_name: str,
+                                               image_url: str):
+    content_text = "What's in this image?"
+    messages = [{
+        "role":
+        "user",
+        "content": [
+            {
+                "type": "image_url",
+                "image_url": image_url
+            },
+            {
+                "type": "text",
+                "text": content_text
+            },
+        ],
+    }]
+
+    # image_url should be a dict {"url": "some url"}, not directly a string
+    with pytest.raises(openai.BadRequestError):
+        _ = await client.chat.completions.create(model=model_name,
+                                                 messages=messages,
+                                                 max_completion_tokens=10,
+                                                 temperature=0.0)
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
 @pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)

vllm/entrypoints/chat_utils.py

@@ -27,10 +27,11 @@
                                ChatCompletionToolMessageParam)
 from openai.types.chat.chat_completion_content_part_input_audio_param import (
     InputAudio)
+from pydantic import TypeAdapter
 # yapf: enable
-# pydantic needs the TypedDict from typing_extensions
 from transformers import (PreTrainedTokenizer, PreTrainedTokenizerFast,
                           ProcessorMixin)
+# pydantic needs the TypedDict from typing_extensions
 from typing_extensions import Required, TypeAlias, TypedDict
 
 from vllm.config import ModelConfig
@@ -879,12 +880,13 @@ def _get_full_multimodal_text_prompt(placeholder_counts: dict[str, int],
 
 # No need to validate using Pydantic again
 _TextParser = partial(cast, ChatCompletionContentPartTextParam)
-_ImageParser = partial(cast, ChatCompletionContentPartImageParam)
 _ImageEmbedsParser = partial(cast, ChatCompletionContentPartImageEmbedsParam)
-_AudioParser = partial(cast, ChatCompletionContentPartAudioParam)
 _InputAudioParser = partial(cast, ChatCompletionContentPartInputAudioParam)
 _RefusalParser = partial(cast, ChatCompletionContentPartRefusalParam)
-_VideoParser = partial(cast, ChatCompletionContentPartVideoParam)
+# Need to validate url objects
+_ImageParser = TypeAdapter(ChatCompletionContentPartImageParam).validate_python
+_AudioParser = TypeAdapter(ChatCompletionContentPartAudioParam).validate_python
+_VideoParser = TypeAdapter(ChatCompletionContentPartVideoParam).validate_python
 
 _ContentPart: TypeAlias = Union[str, dict[str, str], InputAudio]