[RL][TITO] Preserve whitespace in reasoning parser outputs (#24251)

python/sglang/srt/function_call/deepseekv32_detector.py

@@ -216,7 +216,7 @@ class DeepSeekV32Detector(BaseFormatDetector):
         :return: ParseResult indicating success or failure, consumed text, leftover text, and parsed calls.
         """
         idx = text.find(self.bot_token)
-        normal_text = text[:idx].strip() if idx != -1 else text
+        normal_text = text[:idx].removesuffix("\n\n") if idx != -1 else text
         if self.bot_token not in text:
             return StreamingParseResult(normal_text=normal_text, calls=[])
 

python/sglang/srt/parser/reasoning_parser.py

@@ -70,9 +70,10 @@ class BaseReasoningFormatDetector:
             return StreamingParseResult(normal_text=text)
 
         # The text is considered to be in a reasoning block.
-        processed_text = text.replace(
-            self.think_start_token + self.think_start_self_label, ""
-        ).strip()
+        think_start_text = self.think_start_token + self.think_start_self_label
+        processed_text = text
+        while processed_text.startswith(think_start_text):
+            processed_text = processed_text[len(think_start_text) :]
 
         if (
             self.think_end_token not in processed_text
@@ -86,7 +87,7 @@ class BaseReasoningFormatDetector:
             ):
                 # Find the first occurrence of tool_start_token and split there
                 tool_idx = processed_text.find(self.tool_start_token)
-                reasoning_text = processed_text[:tool_idx].strip()
+                reasoning_text = processed_text[:tool_idx]
                 # Preserve tool_start_token in normal text
                 normal_text = processed_text[tool_idx:]
                 return StreamingParseResult(
@@ -99,7 +100,7 @@ class BaseReasoningFormatDetector:
         if self.think_end_token in processed_text:
             splits = processed_text.split(self.think_end_token, maxsplit=1)
             reasoning_text = splits[0]
-            normal_text = splits[1].strip()
+            normal_text = splits[1]
 
             return StreamingParseResult(
                 normal_text=normal_text, reasoning_text=reasoning_text
@@ -150,7 +151,7 @@ class BaseReasoningFormatDetector:
             normal_text = current_text[end_idx + len(self.think_end_token) :]
 
             return StreamingParseResult(
-                normal_text=normal_text, reasoning_text=reasoning_text.rstrip()
+                normal_text=normal_text, reasoning_text=reasoning_text
             )
 
         # Continue with reasoning content