[Reasoning + Structured Output] make reasoning compatible with structured output (#12551)

Signed-off-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>

python/sglang/srt/managers/io_struct.py

@@ -197,6 +197,9 @@ class GenerateReqInput(BaseReq):
     bootstrap_room: Optional[Union[List[int], int]] = None
     bootstrap_pair_key: Optional[Union[List[str], str]] = None
 
+    # For reasoning
+    reasoning: bool = False
+
     # Validation step duration
     validation_time: Optional[float] = None
 
@@ -675,6 +678,9 @@ class TokenizedGenerateReqInput(BaseReq):
     bootstrap_room: Optional[int] = None
     bootstrap_pair_key: Optional[str] = None
 
+    # For reasoning
+    reasoning: bool = False
+
     # For data parallel rank routing
     data_parallel_rank: Optional[int] = None
 

python/sglang/srt/managers/schedule_batch.py

@@ -472,6 +472,7 @@ class Req:
         token_type_ids: List[int] = None,
         session_id: Optional[str] = None,
         custom_logit_processor: Optional[str] = None,
+        reasoning: bool = False,
         return_hidden_states: bool = False,
         eos_token_ids: Optional[Set[int]] = None,
         bootstrap_host: Optional[str] = None,
@@ -517,6 +518,9 @@ class Req:
         # For multi-http worker
         self.http_worker_ipc = http_worker_ipc
 
+        # For reasoning
+        self.reasoning = reasoning
+
         # Sampling info
         if isinstance(sampling_params.custom_params, dict):
             sampling_params = copy.copy(sampling_params)

python/sglang/srt/managers/scheduler.py

@@ -1304,6 +1304,7 @@ class Scheduler(
                 lora_id=recv_req.lora_id,
                 input_embeds=recv_req.input_embeds,
                 custom_logit_processor=recv_req.custom_logit_processor,
+                reasoning=recv_req.reasoning,
                 return_hidden_states=recv_req.return_hidden_states,
                 eos_token_ids=self.model_config.hf_eos_token_id,
                 bootstrap_host=recv_req.bootstrap_host,
@@ -1431,7 +1432,9 @@ class Scheduler(
                 elif req.sampling_params.structural_tag:
                     key = ("structural_tag", req.sampling_params.structural_tag)
 
-                value, cache_hit = self.grammar_backend.get_cached_or_future_value(key)
+                value, cache_hit = self.grammar_backend.get_cached_or_future_value(
+                    key, req.reasoning
+                )
                 req.grammar = value
 
                 if not cache_hit:

python/sglang/srt/managers/tokenizer_manager.py

@@ -887,6 +887,7 @@ class TokenizerManager(TokenizerCommunicatorMixin):
                 input_embeds=input_embeds,
                 session_params=session_params,
                 custom_logit_processor=obj.custom_logit_processor,
+                reasoning=obj.reasoning,
                 return_hidden_states=obj.return_hidden_states,
                 data_parallel_rank=obj.data_parallel_rank,
                 priority=obj.priority,