🥅 Handle template errors

tests/generative_detectors/test_granite_guardian.py

@@ -6,6 +6,7 @@
 import asyncio
 
 # Third Party
+from jinja2.exceptions import TemplateError, UndefinedError
 from vllm.config import MultiModalConfig
 from vllm.entrypoints.openai.protocol import (
     ChatCompletionLogProb,
@@ -415,3 +416,41 @@ def test_chat_detection_errors_on_stream(granite_guardian_detection):
     assert type(detection_response) == ErrorResponse
     assert detection_response.code == HTTPStatus.BAD_REQUEST.value
     assert "streaming is not supported" in detection_response.message
+
+
+def test_chat_detection_errors_on_jinja_template_error(granite_guardian_detection):
+    granite_guardian_detection_instance = asyncio.run(granite_guardian_detection)
+    chat_request = ChatDetectionRequest(
+        messages=[
+            DetectionChatMessageParam(role="user", content="How do I pick a lock?")
+        ],
+    )
+    with patch(
+        "vllm_detector_adapter.generative_detectors.granite_guardian.GraniteGuardian.create_chat_completion",
+        side_effect=TemplateError(),
+    ):
+        detection_response = asyncio.run(
+            granite_guardian_detection_instance.chat(chat_request)
+        )
+        assert type(detection_response) == ErrorResponse
+        assert detection_response.code == HTTPStatus.BAD_REQUEST.value
+        assert "Template error" in detection_response.message
+
+
+def test_chat_detection_errors_on_undefined_jinja_error(granite_guardian_detection):
+    granite_guardian_detection_instance = asyncio.run(granite_guardian_detection)
+    chat_request = ChatDetectionRequest(
+        messages=[
+            DetectionChatMessageParam(role="user", content="How do I pick a lock?")
+        ],
+    )
+    with patch(
+        "vllm_detector_adapter.generative_detectors.granite_guardian.GraniteGuardian.create_chat_completion",
+        side_effect=UndefinedError(),  # class of TemplateError
+    ):
+        detection_response = asyncio.run(
+            granite_guardian_detection_instance.chat(chat_request)
+        )
+        assert type(detection_response) == ErrorResponse
+        assert detection_response.code == HTTPStatus.BAD_REQUEST.value
+        assert "Template error" in detection_response.message

vllm_detector_adapter/generative_detectors/base.py

@@ -7,6 +7,7 @@
 
 # Third Party
 from fastapi import Request
+from jinja2.exceptions import TemplateError
 from vllm.entrypoints.openai.protocol import ChatCompletionResponse, ErrorResponse
 from vllm.entrypoints.openai.serving_chat import OpenAIServingChat
 import jinja2
@@ -157,9 +158,24 @@ async def process_chat_completion_with_scores(
         logger.debug("Request to chat completion: %s", chat_completion_request)
 
         # Call chat completion
-        chat_response = await self.create_chat_completion(
-            chat_completion_request, raw_request
-        )
+        try:
+            chat_response = await self.create_chat_completion(
+                chat_completion_request, raw_request
+            )
+        except TemplateError as e:
+            # Propagate template errors including those from raise_exception in the chat_template.
+            # UndefinedError, a subclass of TemplateError, can happen due to a variety of reasons -
+            # e.g. for Granite Guardian it is not limited but including the following
+            # for a particular risk definition: unexpected number of messages, unexpected
+            # ordering of messages, unexpected roles used for particular messages.
+            # Users _may_ be able to correct some of these errors by changing the input
+            # but the error message may not be directly user-comprehensible
+            chat_response = ErrorResponse(
+                message=e.message or "Template error",
+                type="BadRequestError",
+                code=HTTPStatus.BAD_REQUEST.value,
+            )
+
         logger.debug("Raw chat completion response: %s", chat_response)
         if isinstance(chat_response, ErrorResponse):
             # Propagate chat completion errors directly