[Bugfix][Gemma4] Align parser enable_thinking default with template (#52430)

Signed-off-by: lxy-alexander <lxy_alexander@outlook.com>
Co-authored-by: Chauncey <chaunceyjiang@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

tests/parser/engine/test_gemma4_streaming_reasoning.py

@@ -170,7 +170,7 @@ def mock_tokenizer():
 
 @pytest.fixture
 def parser(mock_tokenizer):
-    return Gemma4Parser(mock_tokenizer)
+    return Gemma4Parser(mock_tokenizer, chat_template_kwargs={"enable_thinking": True})
 
 
 @pytest.fixture
@@ -289,7 +289,10 @@ class TestGemma4PromptOpenReasoning:
 
     @pytest.fixture
     def open_reasoning_parser(self, open_reasoning_tokenizer):
-        return Gemma4Parser(open_reasoning_tokenizer)
+        return Gemma4Parser(
+            open_reasoning_tokenizer,
+            chat_template_kwargs={"enable_thinking": True},
+        )
 
     @staticmethod
     def _prompt_ids_open_channel() -> list[int]:
@@ -394,7 +397,9 @@ class TestGemma4PreInitReasoningRobustness:
 
     @pytest.fixture
     def pre_init_parser(self, pre_init_tokenizer):
-        return Gemma4Parser(pre_init_tokenizer)
+        return Gemma4Parser(
+            pre_init_tokenizer, chat_template_kwargs={"enable_thinking": True}
+        )
 
     def test_model_emitted_channel_open_after_new_turn(
         self, pre_init_parser, pre_init_tokenizer, request_obj
@@ -489,7 +494,9 @@ class TestGemma4ChannelLessOutputConsistency:
 
     @pytest.fixture
     def plain_parser(self, plain_tokenizer):
-        return Gemma4Parser(plain_tokenizer)
+        return Gemma4Parser(
+            plain_tokenizer, chat_template_kwargs={"enable_thinking": True}
+        )
 
     def test_streaming_channel_less_output_is_content(
         self, plain_parser, plain_tokenizer, request_obj
@@ -521,7 +528,9 @@ class TestGemma4ChannelLessOutputConsistency:
         assert reasoning is None
         assert content == _PLAIN_ANSWER_TEXT
 
-        stream_parser = Gemma4Parser(plain_tokenizer)
+        stream_parser = Gemma4Parser(
+            plain_tokenizer, chat_template_kwargs={"enable_thinking": True}
+        )
         results = _stream_tokens_batched(
             stream_parser,
             plain_tokenizer,

tests/reasoning/test_gemma4_reasoning_parser.py

@@ -262,14 +262,27 @@ def test_gemma4_adjust_request(generic_tokenizer):
     assert result is request
 
 
-def test_gemma4_previous_turn_reasoning_is_reasoning_end(generic_tokenizer):
+@pytest.mark.parametrize(
+    ("chat_template_kwargs", "expected_is_reasoning_end"),
+    [
+        pytest.param({}, True, id="omitted_enable_thinking"),
+        pytest.param({"enable_thinking": False}, True, id="thinking_disabled"),
+        pytest.param({"enable_thinking": True}, False, id="thinking_enabled"),
+    ],
+)
+def test_gemma4_new_turn_reasoning_end_matches_enable_thinking(
+    generic_tokenizer,
+    chat_template_kwargs,
+    expected_is_reasoning_end,
+):
     output = (
         "<|channel>thought\n1st thought<channel|>1st content<turn|>\n"
         "<|turn>user\nThanks<|turn>model\n"
     )
     output_tokens = gemma4_encode_output(generic_tokenizer, output)
     parser: ReasoningParser = ReasoningParserManager.get_reasoning_parser(parser_name)(
-        generic_tokenizer
+        generic_tokenizer,
+        chat_template_kwargs=chat_template_kwargs,
     )
     is_reasoning_end = parser.is_reasoning_end(output_tokens)
-    assert not is_reasoning_end
+    assert is_reasoning_end is expected_is_reasoning_end

vllm/parser/gemma4.py

@@ -400,7 +400,7 @@ class Gemma4Parser(ParserEngine):
         **kwargs,
     ) -> None:
         chat_kwargs = kwargs.get("chat_template_kwargs", {}) or {}
-        self._thinking_enabled = chat_kwargs.get("enable_thinking", True)
+        self._thinking_enabled = chat_kwargs.get("enable_thinking", False)
         super().__init__(
             tokenizer,
             tools,