From dbac83218f45db4f2ae4e057ee38636c70a971e5 Mon Sep 17 00:00:00 2001
From: Rakhee Singh <rakhisingh2811@gmail.com>
Date: Tue, 31 Mar 2026 13:25:45 +0530
Subject: [PATCH] fix(vllm): forward response_format to OpenAI-compatible API
 (#4608)

Co-authored-by: rasingh5 <rasingh@demandbase.com>
---
 mem0/llms/vllm.py       |  2 ++
 tests/llms/test_vllm.py | 44 +++++++++++++++++++++++++++++++++++++++++
 2 files changed, 46 insertions(+)

diff --git a/mem0/llms/vllm.py b/mem0/llms/vllm.py
index f7cbfbc58..ce2b61956 100644
--- a/mem0/llms/vllm.py
+++ b/mem0/llms/vllm.py
@@ -99,6 +99,8 @@ class VllmLLM(LLMBase):
             }
         )
 
+        if response_format:
+            params["response_format"] = response_format
         if tools:
             params["tools"] = tools
             params["tool_choice"] = tool_choice
diff --git a/tests/llms/test_vllm.py b/tests/llms/test_vllm.py
index b0acf4631..25b31fc3e 100644
--- a/tests/llms/test_vllm.py
+++ b/tests/llms/test_vllm.py
@@ -88,6 +88,50 @@ def test_generate_response_with_tools(mock_vllm_client):
 
 
 
+def test_generate_response_with_response_format(mock_vllm_client):
+    config = BaseLlmConfig(model="Qwen/Qwen2.5-32B-Instruct", temperature=0.7, max_tokens=100, top_p=1.0)
+    llm = VllmLLM(config)
+    messages = [
+        {"role": "system", "content": "You are a memory extraction assistant."},
+        {"role": "user", "content": "I like hiking on weekends."},
+    ]
+
+    mock_response = Mock()
+    mock_response.choices = [Mock(message=Mock(content='{"facts": ["User likes hiking on weekends"]}'))]
+    mock_vllm_client.chat.completions.create.return_value = mock_response
+
+    response = llm.generate_response(messages, response_format={"type": "json_object"})
+
+    mock_vllm_client.chat.completions.create.assert_called_once_with(
+        model="Qwen/Qwen2.5-32B-Instruct",
+        messages=messages,
+        temperature=0.7,
+        max_tokens=100,
+        top_p=1.0,
+        response_format={"type": "json_object"},
+    )
+    assert response == '{"facts": ["User likes hiking on weekends"]}'
+
+
+def test_generate_response_without_response_format(mock_vllm_client):
+    config = BaseLlmConfig(model="Qwen/Qwen2.5-32B-Instruct", temperature=0.7, max_tokens=100, top_p=1.0)
+    llm = VllmLLM(config)
+    messages = [
+        {"role": "system", "content": "You are a helpful assistant."},
+        {"role": "user", "content": "Tell me a joke."},
+    ]
+
+    mock_response = Mock()
+    mock_response.choices = [Mock(message=Mock(content="Why did the chicken cross the road?"))]
+    mock_vllm_client.chat.completions.create.return_value = mock_response
+
+    response = llm.generate_response(messages)
+
+    call_kwargs = mock_vllm_client.chat.completions.create.call_args[1]
+    assert "response_format" not in call_kwargs
+    assert response == "Why did the chicken cross the road?"
+
+
 def create_mocked_memory():
     """Create a fully mocked Memory instance for testing."""
     with patch('mem0.utils.factory.LlmFactory.create') as mock_llm_factory, \