From 6af693de364166cb9c2b5056c3db26b61a3b5cb1 Mon Sep 17 00:00:00 2001
From: Dominic Feliton <37809476+dominicfeliton@users.noreply.github.com>
Date: Thu, 4 Dec 2025 19:24:28 -0800
Subject: [PATCH] (fix): empty response + vllm streaming (#17516)

* Fix empty response + vllm streaming

* Add unit test
---
 litellm/llms/openai/openai.py                 |  4 +--
 .../llms/openai/test_openai_empty_response.py | 30 +++++++++++++++++++
 2 files changed, 32 insertions(+), 2 deletions(-)

diff --git a/litellm/llms/openai/openai.py b/litellm/llms/openai/openai.py
index 20842525e59..bb9225fc79b 100644
--- a/litellm/llms/openai/openai.py
+++ b/litellm/llms/openai/openai.py
@@ -444,7 +444,7 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM):
             else:
                 headers = {}
             response = raw_response.parse()
-            if not hasattr(response, "model_dump"):
+            if not data.get("stream") and not hasattr(response, "model_dump"):
                 raise OpenAIError(
                     status_code=500,
                     message=f"Empty or invalid response from LLM endpoint. Received: {response!r}. Check the reverse proxy or model server configuration.",
@@ -482,7 +482,7 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM):
             else:
                 headers = {}
             response = raw_response.parse()
-            if not hasattr(response, "model_dump"):
+            if not data.get("stream") and not hasattr(response, "model_dump"):
                 raise OpenAIError(
                     status_code=500,
                     message=f"Empty or invalid response from LLM endpoint. Received: {response!r}. Check the reverse proxy or model server configuration.",
diff --git a/tests/test_litellm/llms/openai/test_openai_empty_response.py b/tests/test_litellm/llms/openai/test_openai_empty_response.py
index fb42918f381..d1692bdf9bd 100644
--- a/tests/test_litellm/llms/openai/test_openai_empty_response.py
+++ b/tests/test_litellm/llms/openai/test_openai_empty_response.py
@@ -95,3 +95,33 @@ class TestEmptyResponseHandling:
 
         assert response == mock_response
         assert headers == {"x-request-id": "123"}
+
+    def test_sync_streaming_response_passes_through_without_model_dump(self):
+        """
+        Test that streaming responses (which don't have model_dump) pass through
+        correctly without raising an error. This validates the fix for VLLM streaming.
+        """
+        openai_chat = OpenAIChatCompletion()
+
+        # Create a mock response WITHOUT model_dump (like an AsyncStream/Iterator)
+        mock_stream = MagicMock(spec=[])  # spec=[] means no attributes
+
+        mock_raw_response = MagicMock()
+        mock_raw_response.headers = {"x-request-id": "123"}
+        mock_raw_response.parse.return_value = mock_stream
+
+        mock_client = MagicMock()
+        mock_client.chat.completions.with_raw_response.create.return_value = (
+            mock_raw_response
+        )
+
+        # Key: data has stream=True - this should bypass the model_dump check
+        headers, response = openai_chat.make_sync_openai_chat_completion_request(
+            openai_client=mock_client,
+            data={"messages": [{"role": "user", "content": "test"}], "stream": True},
+            timeout=30,
+            logging_obj=MagicMock(),
+        )
+
+        assert response == mock_stream
+        assert headers == {"x-request-id": "123"}