[Bugfix] Return HTTP 400 instead of 501 for unknown chat roles in DeepSeek encoders (#53071)

Signed-off-by: JC-ut0 <809602657@qq.com>

tests/test_request_input_bounds.py

@@ -271,3 +271,19 @@ def test_encode_messages_preserves_small_chat_prompt(encoding_module):
         "<｜begin▁of▁sentence｜><｜User｜>Hello<｜Assistant｜></think>"
         "Hi<｜end▁of▁sentence｜>Again<｜end▁of▁sentence｜>"
     )
+
+
+@pytest.mark.parametrize(
+    "encoding_module",
+    ENCODING_MODULES,
+    ids=["deepseek_v32", "deepseek_v4"],
+)
+def test_encode_messages_unknown_role_raises_value_error(encoding_module):
+    # An invalid role (e.g. uppercase "SYSTEM") is a client error and must be
+    # raised as ValueError so the OpenAI serving layer maps it to HTTP 400
+    # instead of NotImplementedError, which would map to HTTP 501.
+    with pytest.raises(ValueError, match="Invalid role: SYSTEM"):
+        encoding_module.encode_messages(
+            [{"role": "SYSTEM", "content": "Hello"}],
+            thinking_mode="chat",
+        )

tests/tokenizers_/test_deepseek_v4.py

@@ -113,6 +113,17 @@ def test_deepseek_v4_explicitly_disables_thinking(kwargs):
     assert prompt == ("<｜begin▁of▁sentence｜><｜User｜>Hello<｜Assistant｜></think>")
 
 
+def test_deepseek_v4_unknown_role_raises_value_error():
+    # Invalid roles are client errors: they must surface as ValueError
+    # (mapped to HTTP 400 by the OpenAI serving layer), not
+    # NotImplementedError (mapped to HTTP 501).
+    with pytest.raises(ValueError, match="Invalid role: SYSTEM"):
+        _tokenizer().apply_chat_template(
+            [{"role": "SYSTEM", "content": "Hello"}],
+            tokenize=False,
+        )
+
+
 def test_deepseek_v4_uses_v4_tool_prompt_from_request_tools():
     tools = [
         {

vllm/tokenizers/deepseek_v32_encoding.py

@@ -258,7 +258,7 @@ def render_message(
                 tool_calls=tool_calls_content,
             )
     else:
-        raise NotImplementedError(f"Unknown role: {role}")
+        raise ValueError(f"Invalid role: {role}")
 
     return prompt
 

vllm/tokenizers/deepseek_v4_encoding.py

@@ -353,7 +353,7 @@ def render_message(
                 tool_calls=tc_content,
             )
     else:
-        raise NotImplementedError(f"Unknown role: {role}")
+        raise ValueError(f"Invalid role: {role}")
 
     # Append transition tokens based on what follows
     if index + 1 < len(messages) and messages[index + 1].get("role") not in ["assistant", "latest_reminder"]: