[Bugfix] Map reasoning_effort=low to Nemotron-3 Super low_effort + warn on unsupported levels (#30463)

Co-authored-by: EazyReal <8047065+EazyReal@users.noreply.github.com>

python/sglang/srt/entrypoints/openai/protocol.py

@@ -800,6 +800,7 @@ class ChatCompletionRequest(BaseModel):
     @classmethod
     def normalize_reasoning_inputs(cls, values: Dict):
         r = values.get("reasoning")
+        thinking = None
 
         if r is not None and isinstance(r, dict):
             effort = r.get("effort") or r.get("reasoning_effort")
@@ -814,25 +815,21 @@ class ChatCompletionRequest(BaseModel):
             if isinstance(enabled, str):
                 enabled = enabled.strip().lower() in {"1", "true", "yes", "y", "on"}
             if enabled:
-                ctk = values.get("chat_template_kwargs")
-                if not isinstance(ctk, dict):
-                    ctk = {}
-                # different models check different keys:
-                # - "thinking" for deepseek-v3, kimi_k2
-                # - "enable_thinking" for qwen3, glm45, nemotron_3, interns1, mimo
-                ctk.setdefault("thinking", True)
-                ctk.setdefault("enable_thinking", True)
-                values["chat_template_kwargs"] = ctk
+                thinking = True
 
-        if values.get("reasoning_effort") == "none":
+        effort = values.get("reasoning_effort")
+        if effort is not None:
+            thinking = effort != "none"
+
+        if thinking is not None:
             ctk = values.get("chat_template_kwargs")
             if not isinstance(ctk, dict):
                 ctk = {}
             # different models check different keys:
             # - "thinking" for deepseek-v3, kimi_k2
             # - "enable_thinking" for qwen3, glm45, nemotron_3, interns1
-            ctk.setdefault("thinking", False)
-            ctk.setdefault("enable_thinking", False)
+            ctk.setdefault("thinking", thinking)
+            ctk.setdefault("enable_thinking", thinking)
             values["chat_template_kwargs"] = ctk
 
         return values

python/sglang/srt/entrypoints/openai/serving_chat.py

@@ -851,6 +851,18 @@ class OpenAIServingChat(OpenAIServingBase):
             if request.chat_template_kwargs:
                 extra_template_kwargs.update(request.chat_template_kwargs)
 
+            rc = self.template_manager.reasoning_config
+            if rc is not None and rc.effort_kwarg is not None:
+                if request.reasoning_effort == "low":
+                    extra_template_kwargs.setdefault(rc.effort_kwarg, True)
+                elif request.reasoning_effort in ("medium", "high", "max"):
+                    logger.warning(
+                        "Model '%s' supports only 'low' reasoning effort; "
+                        "requested '%s' treated as default thinking",
+                        self.tokenizer_manager.server_args.served_model_name,
+                        request.reasoning_effort,
+                    )
+
             # Split apply_chat_template(tokenize=True) into render + encode so we
             # can skip add_special_tokens=False on tokenizers that don't auto-add
             # specials (Kimi-like, OpenAI-chat analogue of #25265). Chat

python/sglang/srt/managers/template_detection.py

@@ -64,6 +64,7 @@ class ReasoningToggleConfig:
     toggle_param: Optional[str] = None
     default_enabled: Optional[bool] = None
     special_case: Optional[str] = None
+    effort_kwarg: Optional[str] = None
 
     @property
     def always_on(self) -> bool:
@@ -104,6 +105,16 @@ REASONING_MODE_RULES = (
             re.DOTALL,
         ),
     ),
+    DetectionRule(
+        name="nemotron_3_super_low_effort",
+        value=ReasoningToggleConfig(
+            toggle_param="enable_thinking",
+            default_enabled=True,
+            effort_kwarg="low_effort",
+        ),
+        predicate=lambda ctx: ctx.has_text("low_effort")
+        and ctx.has_text("truncate_history_thinking"),
+    ),
     DetectionRule(
         name="enable_thinking_default_true",
         value=ReasoningToggleConfig(
@@ -193,8 +204,10 @@ def _is_kimi_k2(ctx):
 
 
 def _is_nemotron_3(ctx):
-    return ctx.has_text("truncate_history_thinking") and ctx.reasoning_config == (
-        ReasoningToggleConfig(toggle_param="enable_thinking", default_enabled=True)
+    return ctx.has_text("truncate_history_thinking") and (
+        ctx.reasoning_config is not None
+        and ctx.reasoning_config.toggle_param == "enable_thinking"
+        and ctx.reasoning_config.default_enabled is True
     )