Add granite_thinking_parser reasoning parser for Granite 4.2 (#38693)

Signed-off-by: Yousaf Shah <yousaf.shah@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com>

python/sglang/srt/parser/reasoning_parser.py

@@ -1043,6 +1043,71 @@ class Nemotron3Detector(BaseReasoningFormatDetector):
         )
 
 
+class GraniteThinkingDetector(BaseReasoningFormatDetector):
+    """Detector for Granite 4.2 thinking models (ibm-granite/granite-4.2-*).
+
+    Strips leading newlines from content after </think> — the Granite chat
+    template writes ``\\n</think>\\n``, so content starts with ``\\n``.
+    Matches the behavior of the HF plugin ``granite_thinking_parser.py``.
+    """
+
+    def __init__(
+        self,
+        stream_reasoning: bool = True,
+        force_reasoning: bool = False,
+        continue_final_message: bool = False,
+        previous_content: str = "",
+        force_nonempty_content: bool = False,
+    ):
+        super().__init__(
+            "<think>",
+            "</think>",
+            force_reasoning=force_reasoning,
+            stream_reasoning=stream_reasoning,
+            tool_start_token="<tool_call>",
+            continue_final_message=continue_final_message,
+            previous_content=previous_content,
+            reasoning_default="enable_thinking",
+            force_nonempty_content=force_nonempty_content,
+        )
+        self._content_started = False
+        self._reasoning_seen = False
+
+    def detect_and_parse(self, text: str) -> StreamingParseResult:
+        ret = self._detect_and_parse_impl(text)
+        think_end_present = self.think_end_token in text
+        # HF plugin swaps when parsed content is absent (text ends at </think>);
+        # newline-only content the plugin strips itself, without swapping.
+        content_absent = think_end_present and not ret.normal_text
+        if think_end_present and ret.normal_text:
+            ret.normal_text = ret.normal_text.lstrip("\n")
+        if (
+            self._force_nonempty_content
+            and not ret.normal_text
+            and (content_absent or not think_end_present)
+        ):
+            ret.normal_text, ret.reasoning_text = ret.reasoning_text, ret.normal_text
+        return ret
+
+    def parse_streaming_increment(self, new_text: str) -> StreamingParseResult:
+        was_in_reasoning = self._in_reasoning
+        ret = super().parse_streaming_increment(new_text)
+        # Sampling only the pre-call state misses a whole think block arriving
+        # in one chunk; post-call evidence keeps stripping chunk-independent.
+        if (
+            was_in_reasoning
+            or self._in_reasoning
+            or self.stripped_think_start
+            or ret.reasoning_text
+        ):
+            self._reasoning_seen = True
+        if self._reasoning_seen and not self._content_started and ret.normal_text:
+            ret.normal_text = ret.normal_text.lstrip("\n")
+            if ret.normal_text:
+                self._content_started = True
+        return ret
+
+
 class MiniMaxM3Detector(BaseReasoningFormatDetector):
     """MiniMax-M3 detector. Format: (<mm:think>)*(.*)</mm:think>.
 
@@ -2123,6 +2188,7 @@ class ReasoningParser:
         "step3p5": DeepSeekR1Detector,
         "mistral": MistralDetector,
         "nemotron_3": Nemotron3Detector,
+        "granite_thinking_parser": GraniteThinkingDetector,
         "interns1": Qwen3Detector,
         "gemma4": Gemma4Detector,
         "inkling": InklingDetector,
@@ -2177,6 +2243,10 @@ class ReasoningParser:
         if chat_template_kwargs.get("force_nonempty_content") is True:
             kwargs["force_nonempty_content"] = True
 
+        if model_type.lower() == "granite_thinking_parser":
+            if chat_template_kwargs.get("enable_thinking") is False:
+                kwargs["force_nonempty_content"] = True
+
         if model_type.lower() == "k2_horizon":
             # Template kwargs are the final values passed to Jinja and therefore
             # take precedence over the convenience fields on API requests.

python/sglang/srt/parser/template_detection.py

@@ -314,6 +314,18 @@ def _is_k2_v3(ctx):
     )
 
 
+def _is_granite_thinking_parser(ctx):
+    # Nemotron-3 templates share the same <parameter= tool-call block, so it
+    # cannot discriminate; defer_loading is Granite's deferred tool loading.
+    return (
+        ctx.has_text("truncate_history_thinking")
+        and ctx.has_text("defer_loading")
+        and ctx.reasoning_config is not None
+        and ctx.reasoning_config.toggle_param == "enable_thinking"
+        and ctx.reasoning_config.default_enabled is True
+    )
+
+
 def _is_nemotron_3(ctx):
     return ctx.has_text("truncate_history_thinking") and (
         ctx.reasoning_config is not None
@@ -502,6 +514,11 @@ REASONING_PARSER_RULES = (
     DetectionRule(name="mistral", value="mistral", predicate=_is_mistral),
     DetectionRule(name="gpt_oss", value="gpt-oss", predicate=_is_gpt_oss),
     DetectionRule(name="kimi_k2", value="kimi_k2", predicate=_is_kimi_k2),
+    DetectionRule(
+        name="granite_thinking_parser",
+        value="granite_thinking_parser",
+        predicate=_is_granite_thinking_parser,
+    ),
     DetectionRule(name="nemotron_3", value="nemotron_3", predicate=_is_nemotron_3),
     DetectionRule(name="glm45", value="glm45", predicate=_is_glm_family),
     DetectionRule(name="hunyuan", value="hunyuan", predicate=_is_hunyuan),

test/registered/unit/parser/test_reasoning_parser.py

@@ -10,6 +10,7 @@ from sglang.srt.parser.reasoning_parser import (
     DeepSeekV4Detector,
     Gemma4Detector,
     Glm45Detector,
+    GraniteThinkingDetector,
     HunyuanDetector,
     InklingDetector,
     KimiDetector,
@@ -1595,5 +1596,135 @@ class TestCohereCommand4DetectorFinish(CustomTestCase):
         self.assertEqual(end.reasoning_text, "")
 
 
+class TestGraniteThinkingDetector(CustomTestCase):
+    def setUp(self):
+        self.detector = GraniteThinkingDetector()
+
+    def test_leading_newline_stripped(self):
+        text = "<think>reasoning</think>\nHello"
+        result = self.detector.detect_and_parse(text)
+        self.assertEqual(result.reasoning_text, "reasoning")
+        self.assertEqual(result.normal_text, "Hello")
+
+    def test_reasoning_only(self):
+        text = "<think>reasoning</think>"
+        result = self.detector.detect_and_parse(text)
+        self.assertEqual(result.reasoning_text, "reasoning")
+        self.assertEqual(result.normal_text, "")
+
+    def test_force_nonempty_no_swap_when_think_end_present(self):
+        """When </think> is present, force_nonempty_content does NOT swap
+        even if content is empty after lstrip. Matches HF plugin behavior."""
+        detector = GraniteThinkingDetector(force_nonempty_content=True)
+        text = "<think>reasoning</think>\n\n"
+        result = detector.detect_and_parse(text)
+        self.assertEqual(result.reasoning_text, "reasoning")
+        self.assertEqual(result.normal_text, "")
+
+    def test_force_nonempty_swaps_when_text_ends_at_think_end(self):
+        """Content absent right after </think> (e.g. max_tokens cut there) swaps
+        like the truncated case; newline-only content still does not."""
+        detector = GraniteThinkingDetector(force_nonempty_content=True)
+        result = detector.detect_and_parse("<think>reasoning</think>")
+        self.assertEqual(result.reasoning_text, "")
+        self.assertEqual(result.normal_text, "reasoning")
+
+    def test_force_nonempty_content_truncated_reasoning(self):
+        detector = GraniteThinkingDetector(force_nonempty_content=True)
+        text = "<think>truncated reasoning"
+        result = detector.detect_and_parse(text)
+        self.assertEqual(result.normal_text, "truncated reasoning")
+        self.assertEqual(result.reasoning_text, "")
+
+    def test_plain_text_no_think_tags(self):
+        text = "Hello"
+        result = self.detector.detect_and_parse(text)
+        self.assertEqual(result.normal_text, "Hello")
+        self.assertEqual(result.reasoning_text, "")
+
+    def test_tool_interrupt(self):
+        text = "<think>reasoning<tool_call>get_weather</tool_call>"
+        result = self.detector.detect_and_parse(text)
+        self.assertEqual(result.reasoning_text, "reasoning")
+        self.assertEqual(result.normal_text, "<tool_call>get_weather</tool_call>")
+
+    def test_multiline_reasoning_and_content(self):
+        text = "<think>line1\nline2</think>\nresult1\nresult2"
+        result = self.detector.detect_and_parse(text)
+        self.assertEqual(result.reasoning_text, "line1\nline2")
+        self.assertEqual(result.normal_text, "result1\nresult2")
+
+    def test_streaming_newlines_preserved_after_content_starts(self):
+        self.detector.parse_streaming_increment("<think>")
+        self.detector.parse_streaming_increment("r")
+        self.detector.parse_streaming_increment("</think>")
+        self.detector.parse_streaming_increment("\n")
+        self.detector.parse_streaming_increment("Hello")
+        result = self.detector.parse_streaming_increment("\nworld")
+        self.assertEqual(result.normal_text, "\nworld")
+
+    def test_streaming_no_strip_without_reasoning(self):
+        result = self.detector.parse_streaming_increment("\nHello")
+        self.assertEqual(result.normal_text, "\nHello")
+
+    def test_streaming_result_is_chunking_independent(self):
+        # The empty think block only trips stripped_think_start evidence:
+        # reasoning text and pre/post _in_reasoning are all empty/False there.
+        for text, exp_r, exp_c in (
+            ("<think>r</think>\nHello", "r", "Hello"),
+            ("<think></think>\nHello", "", "Hello"),
+        ):
+            for stream_reasoning in (True, False):
+                for chunks in (
+                    [text],
+                    [text[: text.index("</think>") + len("</think>")], "\nHello"],
+                    [
+                        "<think>",
+                        text[len("<think>") : text.index("</think>")],
+                        "</think>",
+                        "\nHello",
+                    ],
+                    list(text),
+                ):
+                    with self.subTest(
+                        text=text, stream_reasoning=stream_reasoning, chunks=chunks
+                    ):
+                        detector = GraniteThinkingDetector(
+                            stream_reasoning=stream_reasoning
+                        )
+                        all_r = all_c = ""
+                        for chunk in chunks:
+                            ret = detector.parse_streaming_increment(chunk)
+                            all_r += ret.reasoning_text
+                            all_c += ret.normal_text
+                        end = detector.finish()
+                        all_r += end.reasoning_text
+                        all_c += end.normal_text
+                        self.assertEqual(all_r, exp_r)
+                        self.assertEqual(all_c, exp_c)
+
+    def test_reasoning_parser_integration(self):
+        parser = ReasoningParser("granite_thinking_parser")
+        self.assertIsInstance(parser.detector, GraniteThinkingDetector)
+        reasoning, normal = parser.parse_non_stream(
+            "<think>thinking</think>\nThe answer"
+        )
+        self.assertEqual(reasoning, "thinking")
+        self.assertEqual(normal, "The answer")
+
+    def test_enable_thinking_false_swaps_truncated_reasoning(self):
+        from sglang.srt.entrypoints.openai.protocol import ChatCompletionRequest
+
+        request = ChatCompletionRequest(
+            model="granite-4.2-30b",
+            messages=[{"role": "user", "content": "hi"}],
+            chat_template_kwargs={"enable_thinking": False},
+        )
+        parser = ReasoningParser("granite_thinking_parser", request=request)
+        reasoning, normal = parser.parse_non_stream("<think>truncated")
+        self.assertEqual(reasoning, "")
+        self.assertEqual(normal, "truncated")
+
+
 if __name__ == "__main__":
     unittest.main()

test/registered/unit/parser/test_template_manager.py

@@ -278,6 +278,38 @@ class TestTemplateManagerReasoningDetection(unittest.TestCase):
         )
         self.assertEqual(parser, "nemotron_3")
 
+    def test_nemotron_with_shared_parameter_block_not_misclassified_as_granite(self):
+        # Granite 4.2 and Nemotron-3 templates share the same
+        # <function=name>/<parameter=key> tool-call instruction block; without
+        # a Granite-only signature (defer_loading) this must stay nemotron_3.
+        template = """
+        {% set enable_thinking = enable_thinking if enable_thinking is defined else True %}
+        {% set truncate_history_thinking = truncate_history_thinking if truncate_history_thinking is defined else True %}
+        {{- '<tool_call>\\n<function=example_function_name>\\n<parameter=example_parameter_1>\\nvalue_1\\n</parameter>\\n</function>\\n</tool_call>' }}
+        """
+        _, config, parser = self._detect(template, ["<|endoftext|>"])
+
+        self.assertEqual(
+            config,
+            ReasoningToggleConfig(toggle_param="enable_thinking", default_enabled=True),
+        )
+        self.assertEqual(parser, "nemotron_3")
+
+    def test_granite_detected_via_defer_loading_signature(self):
+        template = """
+        {% set enable_thinking = enable_thinking if enable_thinking is defined else True %}
+        {% set truncate_history_thinking = truncate_history_thinking if truncate_history_thinking is defined else True %}
+        {%- if tool.defer_loading is not defined or not tool.defer_loading %}{%- endif %}
+        {{- '<tool_call>\\n<function=example_function_name>\\n<parameter=example_parameter_1>\\nvalue_1\\n</parameter>\\n</function>\\n</tool_call>' }}
+        """
+        _, config, parser = self._detect(template, [])
+
+        self.assertEqual(
+            config,
+            ReasoningToggleConfig(toggle_param="enable_thinking", default_enabled=True),
+        )
+        self.assertEqual(parser, "granite_thinking_parser")
+
     def test_minimax_uses_template_signature_without_toggle_config(self):
         template = """
         {%- set toolcall_begin_token = '<minimax:tool_call>' -%}