fix: Properly handle multiple text components from request (#5196)

components/src/dynamo/vllm/multimodal_handlers/preprocessor_handler.py

@@ -24,6 +24,7 @@ from ..multimodal_utils import (
     MultiModalRequest,
     MyRequestOutput,
     ProcessMixIn,
+    extract_user_text,
     vLLMMultimodalRequest,
 )
 
@@ -157,10 +158,7 @@ class ProcessorHandler(ProcessMixIn):
             raise ValueError("prompt_template must contain '<prompt>' placeholder")
 
         # Safely extract user text
-        try:
-            user_text = raw_request.messages[0].content[0].text
-        except (IndexError, AttributeError) as e:
-            raise ValueError(f"Invalid message structure: {e}")
+        user_text = extract_user_text(raw_request.messages)
 
         prompt = template.replace("<prompt>", user_text)
 

components/src/dynamo/vllm/multimodal_utils/__init__.py

 # SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
+from dynamo.vllm.multimodal_utils.chat_message_utils import extract_user_text
 from dynamo.vllm.multimodal_utils.chat_processor import (
     ChatProcessor,
     CompletionsProcessor,
@@ -34,6 +35,7 @@ __all__ = [
     "CompletionsProcessor",
     "ProcessMixIn",
     "encode_image_embeddings",
+    "extract_user_text",
     "get_encoder_components",
     "get_http_client",
     "ImageLoader",

components/src/dynamo/vllm/multimodal_utils/chat_message_utils.py

+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Utility functions for processing chat messages."""
+
+from typing import List
+
+from dynamo.vllm.multimodal_utils.protocol import ChatMessage
+
+
+def extract_user_text(messages: List[ChatMessage]) -> str:
+    """Extract and concatenate text content from user messages."""
+
+    # This function finds all text content items from "user" role messages,
+    # and concatenates them. For multi-turn conversation, it adds a newline
+    # between each turn. This is not a perfect solution as we encode multi-turn
+    # conversation as a single turn. However, multi-turn conversation in a
+    # single request is not well defined in the spec.
+
+    # TODO: Revisit this later when adding multi-turn conversation support.
+    user_texts = []
+    for message in messages:
+        if message.role == "user":
+            # Collect all text content items from this user message
+            text_parts = []
+            for item in message.content:
+                if item.type == "text" and item.text:
+                    text_parts.append(item.text)
+            # If this user message has text content, join it and add to user_texts
+            if text_parts:
+                user_texts.append("".join(text_parts))
+
+    if not user_texts:
+        raise ValueError("No text content found in user messages")
+
+    # Join all user turns with newline separator
+    return "\n".join(user_texts)

components/src/dynamo/vllm/multimodal_utils/chat_processor.py

@@ -59,12 +59,6 @@ class ProcessMixIn(ProcessMixInRequired):
     Mixin for pre and post processing for vLLM
     """
 
-    engine_args: AsyncEngineArgs
-    chat_processor: "ChatProcessor | None"
-    completions_processor: "CompletionsProcessor | None"
-    model_config: ModelConfig
-    default_sampling_params: SamplingParams
-
     def __init__(self):
         pass
 

components/src/dynamo/vllm/multimodal_utils/protocol.py

@@ -18,7 +18,7 @@ import json
 from typing import Any, List, Literal, Optional, Tuple, Union
 
 import msgspec
-from pydantic import BaseModel, ConfigDict, Field, field_validator
+from pydantic import BaseModel, ConfigDict, Field, field_serializer, field_validator
 from pydantic_core import core_schema
 from typing_extensions import NotRequired
 from vllm.inputs.data import TokensPrompt
@@ -89,9 +89,13 @@ class vLLMGenerateRequest(BaseModel):
             return SamplingParams(**v)
         return v
 
+    @field_serializer("sampling_params")
+    def serialize_sampling_params(self, value: SamplingParams) -> dict[str, Any]:
+        """Serialize SamplingParams using msgspec and return as dict."""
+        return json.loads(msgspec.json.encode(value))
+
     model_config = ConfigDict(
         arbitrary_types_allowed=True,
-        json_encoders={SamplingParams: lambda v: json.loads(msgspec.json.encode(v))},
     )
 
 

components/src/dynamo/vllm/tests/test_chat_message_utils.py

+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Unit tests for chat message utility functions."""
+
+import pytest
+
+from dynamo.vllm.multimodal_utils.chat_message_utils import extract_user_text
+from dynamo.vllm.multimodal_utils.protocol import (
+    ChatMessage,
+    ImageContent,
+    ImageURLDetail,
+    TextContent,
+)
+
+pytestmark = [
+    pytest.mark.unit,
+    pytest.mark.vllm,
+    pytest.mark.gpu_0,
+    pytest.mark.pre_merge,
+]
+
+
+def test_extract_user_text_single_message():
+    """Test extracting text from a single user message with one text content."""
+    messages = [
+        ChatMessage(
+            role="user", content=[TextContent(type="text", text="Hello, world!")]
+        )
+    ]
+
+    result = extract_user_text(messages)
+    assert result == "Hello, world!"
+
+
+def test_extract_user_text_multiple_text_parts():
+    """Test extracting text from a user message with multiple text content items."""
+    messages = [
+        ChatMessage(
+            role="user",
+            content=[
+                TextContent(type="text", text="First part "),
+                ImageContent(
+                    type="image_url",
+                    image_url=ImageURLDetail(url="http://example.com/image.jpg"),
+                ),
+                TextContent(type="text", text="second part"),
+            ],
+        )
+    ]
+
+    result = extract_user_text(messages)
+    assert result == "First part second part"
+
+
+def test_extract_user_text_multi_turn():
+    """Test extracting text from multi-turn conversation."""
+    messages = [
+        ChatMessage(
+            role="user", content=[TextContent(type="text", text="First question")]
+        ),
+        ChatMessage(
+            role="assistant", content=[TextContent(type="text", text="First answer")]
+        ),
+        ChatMessage(
+            role="user", content=[TextContent(type="text", text="Second question")]
+        ),
+    ]
+
+    result = extract_user_text(messages)
+    assert result == "First question\nSecond question"
+
+
+def test_extract_user_text_only_images():
+    """Test that ValueError is raised when messages contain only images."""
+    messages = [
+        ChatMessage(
+            role="user",
+            content=[
+                ImageContent(
+                    type="image_url",
+                    image_url=ImageURLDetail(url="http://example.com/image.jpg"),
+                )
+            ],
+        )
+    ]
+
+    with pytest.raises(ValueError, match="No text content found in user messages"):
+        extract_user_text(messages)
+
+
+def test_extract_user_text_empty_messages():
+    """Test that ValueError is raised when messages list is empty."""
+    messages: list[ChatMessage] = []
+
+    with pytest.raises(ValueError, match="No text content found in user messages"):
+        extract_user_text(messages)
+
+
+def test_extract_user_text_no_user_messages():
+    """Test that ValueError is raised when there are no user role messages."""
+    messages = [
+        ChatMessage(
+            role="assistant",
+            content=[TextContent(type="text", text="Just an assistant message")],
+        )
+    ]
+
+    with pytest.raises(ValueError, match="No text content found in user messages"):
+        extract_user_text(messages)
+
+
+def test_extract_user_text_mixed_roles():
+    """Test extracting text only from user messages, ignoring other roles."""
+    messages = [
+        ChatMessage(
+            role="system", content=[TextContent(type="text", text="System prompt")]
+        ),
+        ChatMessage(
+            role="user", content=[TextContent(type="text", text="User message 1")]
+        ),
+        ChatMessage(
+            role="assistant",
+            content=[TextContent(type="text", text="Assistant response")],
+        ),
+        ChatMessage(
+            role="user", content=[TextContent(type="text", text="User message 2")]
+        ),
+    ]
+
+    result = extract_user_text(messages)
+    assert result == "User message 1\nUser message 2"
+
+
+def test_extract_user_text_empty_text_content():
+    """Test that empty text content items are ignored."""
+    messages = [
+        ChatMessage(
+            role="user",
+            content=[
+                TextContent(type="text", text=""),
+                TextContent(type="text", text="Valid text"),
+                TextContent(type="text", text=""),
+            ],
+        )
+    ]
+
+    result = extract_user_text(messages)
+    assert result == "Valid text"

examples/multimodal/components/processor.py

@@ -27,6 +27,7 @@ from dynamo.runtime.logging import configure_dynamo_logging
 # To import example local module
 sys.path.append(os.path.join(os.path.dirname(os.path.abspath(__file__)), ".."))
 from utils.args import Config, base_parse_args, parse_endpoint
+from utils.chat_message_utils import extract_user_text
 from utils.chat_processor import ChatProcessor, CompletionsProcessor, ProcessMixIn
 from utils.protocol import (
     MultiModalInput,
@@ -203,15 +204,7 @@ class Processor(ProcessMixIn):
         if "<prompt>" not in template:
             raise ValueError("prompt_template must contain '<prompt>' placeholder")
 
-        # Safely extract user text - find the text content item
-        user_text = None
-        for message in raw_request.messages:
-            for item in message.content:
-                if item.type == "text":
-                    user_text = item.text
-                    break
-        if user_text is None:
-            raise ValueError("No text content found in the request messages")
+        user_text = extract_user_text(raw_request.messages)
 
         prompt = template.replace("<prompt>", user_text)
 

examples/multimodal/utils/chat_message_utils.py

+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Utility functions for processing chat messages."""
+
+
+def extract_user_text(messages) -> str:
+    """Extract and concatenate text content from user messages."""
+    user_texts = []
+    for message in messages:
+        if message.role == "user":
+            # Collect all text content items from this user message
+            text_parts = []
+            for item in message.content:
+                if item.type == "text" and item.text:
+                    text_parts.append(item.text)
+            # If this user message has text content, join it and add to user_texts
+            if text_parts:
+                user_texts.append("".join(text_parts))
+
+    if not user_texts:
+        raise ValueError("No text content found in user messages")
+
+    # Join all user turns with newline separator
+    return "\n".join(user_texts)

examples/multimodal/utils/protocol.py

@@ -18,7 +18,7 @@ import json
 from typing import Any, List, Literal, Optional, Tuple, Union
 
 import msgspec
-from pydantic import BaseModel, ConfigDict, Field, field_validator
+from pydantic import BaseModel, ConfigDict, Field, field_serializer, field_validator
 from pydantic_core import core_schema
 from typing_extensions import NotRequired
 from vllm.inputs.data import TokensPrompt
@@ -89,9 +89,13 @@ class vLLMGenerateRequest(BaseModel):
             return SamplingParams(**v)
         return v
 
+    @field_serializer("sampling_params")
+    def serialize_sampling_params(self, value: SamplingParams) -> dict[str, Any]:
+        """Serialize SamplingParams using msgspec and return as dict."""
+        return json.loads(msgspec.json.encode(value))
+
     model_config = ConfigDict(
         arbitrary_types_allowed=True,
-        json_encoders={SamplingParams: lambda v: msgspec.json.encode(v)},
     )