fix(reasoning): let --enable-strict-thinking works for DeepSeek-V4 (#32400)

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>

python/sglang/srt/parser/reasoning_parser.py

@@ -2,6 +2,14 @@ import inspect
 import re
 from typing import Dict, List, Optional, Tuple, Type
 
+from sglang.srt.entrypoints.openai.encoding_dsv4 import dsml_token as dsv4_dsml_token
+from sglang.srt.entrypoints.openai.encoding_dsv4 import eos_token as dsv4_eos_token
+from sglang.srt.entrypoints.openai.encoding_dsv4 import (
+    thinking_end_token as dsv4_thinking_end_token,
+)
+from sglang.srt.entrypoints.openai.encoding_dsv4 import (
+    thinking_start_token as dsv4_thinking_start_token,
+)
 from sglang.srt.entrypoints.openai.protocol import ChatCompletionRequest
 from sglang.srt.function_call.hunyuan_detector import resolve_hunyuan_tokens
 from sglang.srt.parser.harmony_parser import HarmonyParser
@@ -897,6 +905,29 @@ class _DeepSeekV3Detector(Qwen3Detector):
         self.reasoning_default = "explicit_thinking"
 
 
+class DeepSeekV4Detector(BaseReasoningFormatDetector):
+    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__(
+            dsv4_thinking_start_token,
+            dsv4_thinking_end_token,
+            think_excluded_tokens=[dsv4_eos_token, dsv4_dsml_token],
+            force_reasoning=force_reasoning,
+            stream_reasoning=stream_reasoning,
+            continue_final_message=continue_final_message,
+            previous_content=previous_content,
+            thinks_internally=True,
+            reasoning_default="explicit_thinking",
+            force_nonempty_content=force_nonempty_content,
+        )
+
+
 class _MimoDetector(Qwen3Detector):
     """MIMO reuses Qwen3 tokens but requires explicit enable_thinking=True to enable."""
 
@@ -1385,7 +1416,7 @@ class ReasoningParser:
         "apertus2509": Apertus2509Detector,
         "deepseek-r1": DeepSeekR1Detector,
         "deepseek-v3": _DeepSeekV3Detector,
-        "deepseek-v4": _DeepSeekV3Detector,
+        "deepseek-v4": DeepSeekV4Detector,
         "glm45": Glm45Detector,
         "hunyuan": HunyuanDetector,
         "gpt-oss": GptOssDetector,

test/registered/unit/parser/test_reasoning_parser.py

@@ -6,6 +6,7 @@ from sglang.srt.parser.reasoning_parser import (
     Apertus2509Detector,
     BaseReasoningFormatDetector,
     DeepSeekR1Detector,
+    DeepSeekV4Detector,
     Gemma4Detector,
     Glm45Detector,
     HunyuanDetector,
@@ -167,6 +168,21 @@ class TestQwen3Detector(CustomTestCase):
         self.assertEqual(result.reasoning_text, "")
 
 
+class TestDeepSeekV4Detector(CustomTestCase):
+    def test_strict_thinking_excludes_deepseek_control_tokens(self):
+        detector = ReasoningParser(model_type="deepseek-v4").detector
+        self.assertIsInstance(detector, DeepSeekV4Detector)
+        self.assertEqual(
+            detector.think_excluded_tokens,
+            ["<｜end▁of▁sentence｜>", "｜DSML｜"],
+        )
+
+    def test_thinking_stays_explicit_opt_in(self):
+        detector = ReasoningParser(model_type="deepseek-v4").detector
+        self.assertEqual(detector.reasoning_default, "explicit_thinking")
+        self.assertTrue(detector.thinks_internally)
+
+
 class TestInklingDetector(CustomTestCase):
     def test_streaming_routes_blocks_across_all_string_boundaries(self):
         detector = InklingDetector()