[Bugfix] Attach request-level tools to existing system message in DeepSeek V4 Python renderer (#51856)

Signed-off-by: Jason Yao <wsyjh8@gmail.com>
Co-authored-by: Chauncey <chaunceyjiang@gmail.com>

tests/tokenizers_/fixtures/deepseek_v4/test_output_1.txt

@@ -1,4 +1,4 @@
-<｜begin▁of▁sentence｜>
+<｜begin▁of▁sentence｜>You are a helpful assistant.
 
 ## Tools
 
@@ -26,7 +26,7 @@ Otherwise, output directly after </think> with tool calls or final response.
 {"name": "search", "description": "Search the web for information", "parameters": {"type": "object", "properties": {"query": {"type": "string", "description": "Search query"}, "num_results": {"type": "integer", "description": "Number of results to return"}}, "required": ["query"]}}
 
 You MUST strictly follow the above defined tool name and parameter schemas to invoke tool calls.
-You are a helpful assistant.<｜User｜>What's the weather in Beijing?<｜Assistant｜><think>The user wants to know the weather in Beijing. I should use the get_weather tool.</think>
+<｜User｜>What's the weather in Beijing?<｜Assistant｜><think>The user wants to know the weather in Beijing. I should use the get_weather tool.</think>
 
 <｜DSML｜tool_calls>
 <｜DSML｜invoke name="get_weather">

tests/tokenizers_/test_deepseek_v4.py

@@ -485,3 +485,146 @@ def test_deepseek_v4_image_sentinel_ids_match_tokenizer():
         ]
     )
     assert image_sentinel_mask(ids).tolist() == [False, True, True, False, False]
+
+
+def _request_tools():
+    return [
+        {
+            "type": "function",
+            "function": {
+                "name": "get_weather",
+                "description": "Get weather for a city",
+                "parameters": {
+                    "type": "object",
+                    "properties": {"city": {"type": "string"}},
+                    "required": ["city"],
+                },
+            },
+        }
+    ]
+
+
+def _encode_reference(messages):
+    from vllm.tokenizers.deepseek_v4_encoding import encode_messages
+
+    return encode_messages(messages, thinking_mode="thinking", reasoning_effort="low")
+
+
+def test_deepseek_v4_attaches_request_tools_to_existing_system_message():
+    messages = [
+        {"role": "system", "content": "You are helpful."},
+        {"role": "user", "content": "Weather in Paris?"},
+    ]
+
+    prompt = _tokenizer().apply_chat_template(
+        messages,
+        tools=_request_tools(),
+        tokenize=False,
+        thinking=True,
+        reasoning_effort="low",
+    )
+
+    expected = _encode_reference(
+        [{**messages[0], "tools": _request_tools()}, messages[1]]
+    )
+    assert prompt == expected
+    assert prompt.startswith("<｜begin▁of▁sentence｜>You are helpful.\n\n## Tools")
+    assert prompt.count("## Tools") == 1
+
+
+def test_deepseek_v4_synthetic_system_only_when_no_system_message():
+    """Without a system message, request tools still get a synthetic leading
+    system entry."""
+    messages = [{"role": "user", "content": "Weather in Paris?"}]
+
+    prompt = _tokenizer().apply_chat_template(
+        messages,
+        tools=_request_tools(),
+        tokenize=False,
+        thinking=True,
+        reasoning_effort="low",
+    )
+
+    expected = _encode_reference(
+        [{"role": "system", "tools": _request_tools()}, messages[0]]
+    )
+    assert prompt == expected
+    assert prompt.startswith("<｜begin▁of▁sentence｜>\n\n## Tools")
+
+
+def test_deepseek_v4_request_tools_override_system_message_tools():
+    message_tools = [
+        {
+            "type": "function",
+            "function": {
+                "name": "message_level_tool",
+                "description": "Should not be rendered",
+                "parameters": {"type": "object", "properties": {}},
+            },
+        }
+    ]
+    system_message = {
+        "role": "system",
+        "content": "You are helpful.",
+        "tools": message_tools,
+    }
+    user_message = {"role": "user", "content": "Weather in Paris?"}
+
+    prompt = _tokenizer().apply_chat_template(
+        [system_message, user_message],
+        tools=_request_tools(),
+        tokenize=False,
+        thinking=True,
+        reasoning_effort="low",
+    )
+
+    expected = _encode_reference(
+        [{**system_message, "tools": _request_tools()}, user_message]
+    )
+    assert prompt == expected
+    assert '"name": "get_weather"' in prompt
+    assert "message_level_tool" not in prompt
+
+
+def test_deepseek_v4_attaches_request_tools_to_first_system_message_only():
+    """Only the first system message, at any index, gets the request tools."""
+    messages = [
+        {"role": "user", "content": "Hi"},
+        {"role": "system", "content": "First system."},
+        {"role": "system", "content": "Second system."},
+        {"role": "user", "content": "Weather in Paris?"},
+    ]
+
+    prompt = _tokenizer().apply_chat_template(
+        messages,
+        tools=_request_tools(),
+        tokenize=False,
+        thinking=True,
+        reasoning_effort="low",
+    )
+
+    expected = _encode_reference(
+        [
+            messages[0],
+            {**messages[1], "tools": _request_tools()},
+            messages[2],
+            messages[3],
+        ]
+    )
+    assert prompt == expected
+    assert "First system.\n\n## Tools" in prompt
+    assert prompt.count("## Tools") == 1
+
+
+def test_deepseek_v4_request_tools_do_not_mutate_caller_messages():
+    import copy
+
+    messages = [
+        {"role": "system", "content": "You are helpful."},
+        {"role": "user", "content": "Weather in Paris?"},
+    ]
+    snapshot = copy.deepcopy(messages)
+
+    _tokenizer().apply_chat_template(messages, tools=_request_tools(), tokenize=False)
+
+    assert messages == snapshot

vllm/tokenizers/deepseek_v4.py

@@ -35,8 +35,18 @@ def get_deepseek_v4_tokenizer(tokenizer: HfTokenizer) -> HfTokenizer:
             conversation = kwargs.get("conversation", messages)
             messages = conversation.copy()
             if tools is not None and len(tools) > 0:
-                messages.insert(0, {"role": "system"})
-                messages[0]["tools"] = tools  # type: ignore[typeddict-unknown-key]
+                # Match the Rust renderer: request tools attach to the first
+                # system message; synthesize one only when none exists.
+                system_idx = next(
+                    (i for i, m in enumerate(messages) if m.get("role") == "system"),
+                    None,
+                )
+                if system_idx is None:
+                    messages.insert(0, {"role": "system"})
+                    system_idx = 0
+                else:
+                    messages[system_idx] = copy.copy(messages[system_idx])
+                messages[system_idx]["tools"] = tools  # type: ignore[typeddict-unknown-key]
 
             reasoning_effort = kwargs.get("reasoning_effort")
             if not isinstance(reasoning_effort, str):