Fix chat completion 500 on non-object JSON bodies (#51654)

Signed-off-by: Tarun Kumar <[email protected]>

tests/entrypoints/openai/chat_completion/test_non_object_body_validation.py

@@ -0,0 +1,38 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""Non-object JSON bodies must fail validation cleanly (4xx), not AttributeError (500).
+
+mode=before validators that call data.get(...) without an isinstance(data, dict)
+guard raise AttributeError for string/list/scalar bodies and surface as HTTP 500.
+"""
+
+import pytest
+from pydantic import ValidationError
+
+from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
+
+
+@pytest.mark.parametrize(
+    "payload",
+    [
+        "this is not valid json{{{",
+        ["not", "an", "object"],
+        42,
+        None,
+    ],
+)
+def test_chat_completion_request_rejects_non_object_body(payload):
+    with pytest.raises(ValidationError):
+        ChatCompletionRequest.model_validate(payload)
+
+
+def test_chat_completion_request_cache_salt_still_validated_on_dict():
+    with pytest.raises(ValidationError, match="cache_salt"):
+        ChatCompletionRequest.model_validate(
+            {
+                "model": "qwen",
+                "messages": [{"role": "user", "content": "hello"}],
+                "cache_salt": "",
+            }
+        )

vllm/entrypoints/openai/chat_completion/protocol.py

@@ -745,6 +745,8 @@ class ChatCompletionRequest(OpenAIBaseModel):
     @model_validator(mode="before")
     @classmethod
     def validate_response_format(cls, data):
+        if not isinstance(data, dict):
+            return data
         response_format = data.get("response_format")
         if response_format is None:
             return data
@@ -776,6 +778,8 @@ class ChatCompletionRequest(OpenAIBaseModel):
     @model_validator(mode="before")
     @classmethod
     def validate_stream_options(cls, data):
+        if not isinstance(data, dict):
+            return data
         if data.get("stream_options") and not data.get("stream"):
             raise VLLMValidationError(
                 "Stream options can only be defined when `stream=True`.",
@@ -787,6 +791,8 @@ class ChatCompletionRequest(OpenAIBaseModel):
     @model_validator(mode="before")
     @classmethod
     def check_logprobs(cls, data):
+        if not isinstance(data, dict):
+            return data
         if data.get("logprob_token_ids") and data.get("use_beam_search"):
             raise VLLMValidationError(
                 "`logprob_token_ids` is not supported with beam search.",
@@ -845,6 +851,8 @@ class ChatCompletionRequest(OpenAIBaseModel):
     def check_structured_outputs_count(cls, data):
         if isinstance(data, ValueError):
             raise data
+        if not isinstance(data, dict):
+            return data
 
         if data.get("structured_outputs", None) is None:
             return data
@@ -970,6 +978,8 @@ class ChatCompletionRequest(OpenAIBaseModel):
     @model_validator(mode="before")
     @classmethod
     def check_generation_prompt(cls, data):
+        if not isinstance(data, dict):
+            return data
         if data.get("continue_final_message") and data.get("add_generation_prompt"):
             raise VLLMValidationError(
                 "Cannot set both `continue_final_message` and "
@@ -1105,6 +1115,8 @@ class BatchChatCompletionRequest(OpenAIBaseModel):
     def check_batch_mode(cls, data: Any) -> Any:
         if isinstance(data, BatchChatCompletionRequest):
             data = data.model_dump(exclude_unset=True)
+        if not isinstance(data, dict):
+            return data
         if data.get("use_beam_search"):
             raise VLLMValidationError(
                 "Batch chat completions do not support beam search. "
@@ -1117,13 +1129,14 @@ class BatchChatCompletionRequest(OpenAIBaseModel):
                 parameter="logprob_token_ids",
             )
         response_format = data.get("response_format")
-        rf_type = (
-            response_format.get("type")
-            if isinstance(response_format, dict)
-            else getattr(response_format, "type", None)
-        )
-        if rf_type == "structural_tag":
-            validate_structural_tag_response_format(response_format)
+        if response_format is not None:
+            rf_type = (
+                response_format.get("type")
+                if isinstance(response_format, dict)
+                else getattr(response_format, "type", None)
+            )
+            if rf_type == "structural_tag":
+                validate_structural_tag_response_format(response_format)
         if (structured_outputs := data.get("structured_outputs")) is not None:
             validate_structured_outputs_structural_tag(structured_outputs)
         n = data.get("n", 1)