[Bugfix] Replace assert with ValueError for response_format validation in completions endpoint (#35456)

Signed-off-by: umut-polat <52835619+umut-polat@users.noreply.github.com>

tests/entrypoints/openai/test_completion_error.py

@@ -221,6 +221,19 @@ async def test_completion_error_stream():
     assert chunks[-1] == "data: [DONE]\n\n"
 
 
+def test_json_schema_response_format_missing_schema():
+    """When response_format type is 'json_schema' but the json_schema field
+    is not provided, request construction should raise a validation error
+    so the API returns 400 instead of 500."""
+    with pytest.raises(Exception, match="json_schema.*must be provided"):
+        CompletionRequest(
+            model=MODEL_NAME,
+            prompt="Test prompt",
+            max_tokens=10,
+            response_format={"type": "json_schema"},
+        )
+
+
 def test_negative_prompt_token_ids_nested():
     """Negative token IDs in prompt (nested list) should raise validation error."""
     with pytest.raises(Exception, match="greater than or equal to 0"):

vllm/entrypoints/openai/completion/protocol.py

@@ -259,7 +259,7 @@ class CompletionRequest(OpenAIBaseModel):
                 structured_outputs_kwargs["json"] = json_schema.json_schema
             elif response_format.type == "structural_tag":
                 structural_tag = response_format
-                assert structural_tag is not None and isinstance(
+                assert isinstance(
                     structural_tag,
                     (
                         LegacyStructuralTagResponseFormat,
@@ -313,6 +313,34 @@ class CompletionRequest(OpenAIBaseModel):
             skip_clone=True,  # Created fresh per request, safe to skip clone
         )
 
+    @model_validator(mode="before")
+    @classmethod
+    def validate_response_format(cls, data):
+        response_format = data.get("response_format")
+        if response_format is None:
+            return data
+
+        rf_type = (
+            response_format.get("type")
+            if isinstance(response_format, dict)
+            else getattr(response_format, "type", None)
+        )
+
+        if rf_type == "json_schema":
+            json_schema = (
+                response_format.get("json_schema")
+                if isinstance(response_format, dict)
+                else getattr(response_format, "json_schema", None)
+            )
+            if json_schema is None:
+                raise VLLMValidationError(
+                    "When response_format type is 'json_schema', the "
+                    "'json_schema' field must be provided.",
+                    parameter="response_format",
+                )
+
+        return data
+
     @model_validator(mode="before")
     @classmethod
     def check_structured_outputs_count(cls, data):