feat: add thinking_budget (version 2)

docs/backend/sampling_params.md

@@ -64,6 +64,7 @@ Please refer to our dedicated guide on [constrained decoding](./structured_outpu
 | ignore_eos                    | `bool = False`                  | Don't stop generation when EOS token is sampled.                                                                                               |
 | skip_special_tokens           | `bool = True`                   | Remove special tokens during decoding.                                                                                                         |
 | custom_params                 | `Optional[List[Optional[Dict[str, Any]]]] = None` | Used when employing `CustomLogitProcessor`. For usage, see below.                                                                              |
+| thinking_budget               | `Optional[int] = None`          | The maximum number of reasoning tokens that can be generated for a request. |
 
 ## Examples
 
@@ -296,3 +297,29 @@ response = requests.post(
 )
 print(response.json())
 ```
+
+### Thinking Budget
+
+Launch a server with `--reasoning-parser`.
+
+```bash
+python3 -m sglang.launch_server --model Qwen/Qwen3-8B --reasoning-parser qwen3
+```
+
+Send a request:
+
+```python
+import requests
+response = requests.post(
+    "http://localhost:30000/generate",
+    json={
+        "text": "9.11 and 9.8, which is greater?",
+        "sampling_params": {
+            "temperature": 0.3,
+            "max_new_tokens": 256,
+            "thinking_budget": 20,
+        },
+    },
+)
+print(response.json())
+```

python/sglang/srt/model_executor/model_runner.py

@@ -1146,7 +1146,9 @@ def sample(
                 [self.sample(values, forward_batch) for values in logits_output],
                 axis=-1,
             )
-
+        sampling_info = forward_batch.sampling_info
+        if sampling_info.thinking_budgets is not None:
+            sampling_info.apply_thinking_budgets(logits_output.next_token_logits)
         self._preprocess_logits(logits_output, forward_batch.sampling_info)
 
         # Sample the next tokens
@@ -1157,6 +1159,8 @@ def sample(
             forward_batch.top_logprobs_nums,
             forward_batch.token_ids_logprobs,
         )
+        if sampling_info.thinking_budgets is not None:
+            sampling_info.update_thinking_budgets(next_token_ids)
         return next_token_ids
 
     @property

python/sglang/srt/openai_api/adapter.py

@@ -555,6 +555,7 @@ def v1_generate_request(
                 "temperature": request.temperature,
                 "max_new_tokens": request.max_tokens,
                 "min_new_tokens": request.min_tokens,
+                "thinking_budget": request.thinking_budget,
                 "stop": request.stop,
                 "stop_token_ids": request.stop_token_ids,
                 "top_p": request.top_p,
@@ -1127,6 +1128,7 @@ def v1_chat_generate_request(
             "temperature": request.temperature,
             "max_new_tokens": request.max_tokens or request.max_completion_tokens,
             "min_new_tokens": request.min_tokens,
+            "thinking_budget": request.thinking_budget,
             "stop": stop,
             "stop_token_ids": request.stop_token_ids,
             "top_p": request.top_p,

python/sglang/srt/openai_api/protocol.py

@@ -172,6 +172,7 @@ class CompletionRequest(BaseModel):
     top_k: int = -1
     min_p: float = 0.0
     min_tokens: int = 0
+    thinking_budget: Optional[int] = None
     json_schema: Optional[str] = None
     regex: Optional[str] = None
     ebnf: Optional[str] = None
@@ -378,6 +379,7 @@ def set_tool_choice_default(cls, values):
     top_k: int = -1
     min_p: float = 0.0
     min_tokens: int = 0
+    thinking_budget: Optional[int] = None
     regex: Optional[str] = None
     ebnf: Optional[str] = None
     repetition_penalty: float = 1.0

python/sglang/srt/reasoning_parser.py

@@ -32,7 +32,7 @@ def detect_and_parse(self, text: str) -> StreamingParseResult:
         One-time parsing: Detects and parses reasoning sections in the provided text.
         Returns both reasoning content and normal text separately.
         """
-        text = text.replace(self.think_start_token, "").strip()
+        text = text.replace(self.think_start_token, "")
         if self.think_end_token not in text:
             # Assume reasoning was truncated before `</think>` token
             return StreamingParseResult(reasoning_text=text)
@@ -73,7 +73,7 @@ def parse_streaming_increment(self, new_text: str) -> StreamingParseResult:
             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

python/sglang/srt/sampling/sampling_batch_info.py

@@ -30,8 +30,13 @@ class SamplingBatchInfo:
     # Whether any request needs min_p sampling
     need_min_p_sampling: bool
 
+    # Use thinking_budget to truncate thinking
+    num_thinking_tokens: Optional[torch.Tensor] = None
+    think_end_ids: Optional[torch.Tensor] = None
+    thinking_budgets: Optional[torch.Tensor] = None
+
     # Masking tensors for grammar-guided structured outputs
-    vocab_size: int
+    vocab_size: int = 0
     grammars: Optional[List] = None
     vocab_mask: Optional[torch.Tensor] = None
     apply_mask_func: Optional[Callable[[torch.Tensor, torch.Tensor], None]] = None
@@ -76,7 +81,22 @@ def from_schedule_batch(cls, batch: ScheduleBatch, vocab_size: int):
         min_ps = torch.tensor(
             [r.sampling_params.min_p for r in reqs], dtype=torch.float
         ).to(device, non_blocking=True)
-
+        if any(hasattr(r.tokenizer, "think_end_id") for r in reqs):
+            think_end_ids = torch.tensor(
+                [getattr(r.tokenizer, "think_end_id", -1) for r in reqs],
+                dtype=torch.int64,
+            ).to(device, non_blocking=True)
+            num_thinking_tokens = torch.tensor([0 for _ in reqs], dtype=torch.int64).to(
+                device, non_blocking=True
+            )
+            thinking_budgets = torch.tensor(
+                [r.sampling_params.thinking_budget or -1 for r in reqs],
+                dtype=torch.int64,
+            ).to(device, non_blocking=True)
+        else:
+            think_end_ids = None
+            num_thinking_tokens = None
+            thinking_budgets = None
         # Check if any request has custom logit processor
         has_custom_logit_processor = (
             batch.enable_custom_logit_processor  # check the flag first.
@@ -132,6 +152,9 @@ def from_schedule_batch(cls, batch: ScheduleBatch, vocab_size: int):
             top_ps=top_ps,
             top_ks=top_ks,
             min_ps=min_ps,
+            think_end_ids=think_end_ids,
+            num_thinking_tokens=num_thinking_tokens,
+            thinking_budgets=thinking_budgets,
             is_all_greedy=all(r.sampling_params.top_k <= 1 for r in reqs),
             need_min_p_sampling=any(r.sampling_params.min_p > 0 for r in reqs),
             vocab_size=vocab_size,
@@ -146,6 +169,35 @@ def from_schedule_batch(cls, batch: ScheduleBatch, vocab_size: int):
     def __len__(self):
         return len(self.temperatures)
 
+    def apply_thinking_budgets(self, next_token_logits: torch.Tensor):
+        has_budget = self.thinking_budgets > 0
+        if not has_budget.any():
+            return
+        torch.where(
+            has_budget,
+            self.num_thinking_tokens + 1,
+            self.num_thinking_tokens,
+            out=self.num_thinking_tokens,
+        )
+        should_stop = has_budget & (
+            self.num_thinking_tokens - 1 > self.thinking_budgets
+        )
+        next_token_logits.masked_fill_(should_stop.unsqueeze(0), float("-inf"))
+        batch_indices = torch.nonzero(should_stop, as_tuple=True)[0]
+        if len(batch_indices) > 0:
+            end_token_indices = self.think_end_ids[batch_indices]
+            next_token_logits[batch_indices, end_token_indices] = 0.0
+
+    def update_thinking_budgets(self, next_token_ids: torch.Tensor):
+        if not torch.any(self.thinking_budgets > 0):
+            return
+        torch.where(
+            next_token_ids == self.think_end_ids,
+            torch.tensor(-1, device=self.thinking_budgets.device),
+            self.thinking_budgets,
+            out=self.thinking_budgets,
+        )
+
     def update_regex_vocab_mask(self):
         if not self.grammars:
             self.vocab_mask = None

python/sglang/srt/sampling/sampling_params.py

@@ -30,6 +30,7 @@ class SamplingParams:
     def __init__(
         self,
         max_new_tokens: int = 128,
+        thinking_budget: Optional[int] = None,
         stop: Optional[Union[str, List[str]]] = None,
         stop_token_ids: Optional[List[int]] = None,
         temperature: float = 1.0,
@@ -58,6 +59,7 @@ def __init__(
             self.stop_token_ids = set(stop_token_ids)
         else:
             self.stop_token_ids = None
+        self.thinking_budget = thinking_budget
         self.temperature = temperature
         self.top_p = top_p
         self.top_k = top_k

test/srt/run_suite.py

@@ -61,6 +61,7 @@ class TestFile:
         TestFile("test_radix_attention.py", 167),
         TestFile("test_reasoning_content.py", 89),
         TestFile("test_enable_thinking.py", 70),
+        TestFile("test_thinking_budget.py", 60),
         TestFile("test_regex_constrained.py", 64),
         TestFile("test_release_memory_occupation.py", 44),
         TestFile("test_request_length_validation.py", 31),

test/srt/test_thinking_budget.py

@@ -0,0 +1,95 @@
+"""
+Usage:
+python3 -m unittest test_thinking_budget.TestThinkingBudget.test_chat_completion_with_thinking_budget_20
+python3 -m unittest test_thinking_budget.TestThinkingBudget.test_chat_completion_with_thinking_budget_200
+"""
+
+import unittest
+
+import requests
+from transformers import AutoTokenizer
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    popen_launch_server,
+)
+
+
+class TestThinkingBudget(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "Qwen/Qwen3-8B"
+        cls.tokenizer = AutoTokenizer.from_pretrained(cls.model)
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.api_key = "sk-1234"
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            api_key=cls.api_key,
+            other_args=[
+                "--reasoning-parser",
+                "qwen3",
+            ],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_chat_completion_with_thinking_budget_20(self):
+        response = requests.post(
+            f"{self.base_url}/v1/chat/completions",
+            headers={"Authorization": f"Bearer {self.api_key}"},
+            json={
+                "model": self.model,
+                "messages": [
+                    {"role": "user", "content": "9.11 and 9.8, which is greater?"}
+                ],
+                "temperature": 0,
+                "separate_reasoning": True,
+                "chat_template_kwargs": {"enable_thinking": True},
+                "thinking_budget": 20,
+            },
+        )
+        self.assertEqual(response.status_code, 200, f"Failed with: {response.text}")
+        data = response.json()
+        reasoning_content = data["choices"][0]["message"]["reasoning_content"]
+        tokens = self.tokenizer.encode(reasoning_content)
+        self.assertEqual(
+            len(tokens),
+            20,
+            f"Reasoning content length: {len(tokens)} not equal to 20, tokens: {tokens}, reasoning_content: {reasoning_content}",
+        )
+
+    def test_chat_completion_with_thinking_budget_200(self):
+        response = requests.post(
+            f"{self.base_url}/v1/chat/completions",
+            headers={"Authorization": f"Bearer {self.api_key}"},
+            json={
+                "model": self.model,
+                "messages": [
+                    {"role": "user", "content": "9.11 and 9.8, which is greater?"}
+                ],
+                "temperature": 0,
+                "separate_reasoning": True,
+                "chat_template_kwargs": {"enable_thinking": True},
+                "thinking_budget": 200,
+            },
+        )
+        self.assertEqual(response.status_code, 200, f"Failed with: {response.text}")
+        data = response.json()
+        reasoning_content = data["choices"][0]["message"]["reasoning_content"]
+        tokens = self.tokenizer.encode(reasoning_content)
+        self.assertEqual(
+            len(tokens),
+            200,
+            f"Reasoning content length {len(tokens)} not equal to 200, tokens: {tokens}, reasoning_content: {reasoning_content}",
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()