feat: Support deployment-level thinking defaults for chat templates (…

components/src/dynamo/common/backend/tests/test_backend_bindings.py

@@ -26,6 +26,10 @@
 
 # Import-time skip: if the extension hasn't been built, all tests below
 # are skipped rather than crashing the collection phase.
+core = pytest.importorskip(
+    "dynamo._core",
+    reason="dynamo._core not built — run `maturin develop` first",
+)
 backend = pytest.importorskip(
     "dynamo._core.backend",
     reason="dynamo._core.backend not built — run `maturin develop` first",
@@ -114,11 +118,58 @@ def test_worker_config_accepts_parser_runtime_settings():
         namespace="dynamo",
         tool_call_parser="kimi_k2",
         reasoning_parser="kimi_k25",
+        default_thinking_mode="disabled",
         exclude_tools_when_tool_choice_none=False,
         enable_local_indexer=False,
     )
 
 
+def test_worker_config_preserves_legacy_positional_argument_order():
+    """New optional fields must be appended after every existing argument."""
+    backend.WorkerConfig(
+        "dynamo",  # namespace
+        "backend",  # component
+        "generate",  # endpoint
+        "",  # model_name
+        None,  # served_model_name
+        core.ModelInput.Tokens,  # model_input
+        "chat,completions",  # endpoint_types
+        None,  # custom_jinja_template
+        None,  # tool_call_parser
+        None,  # reasoning_parser
+        False,  # exclude_tools_when_tool_choice_none
+        False,  # enable_local_indexer
+    )
+
+
+@pytest.mark.unified
+def test_python_worker_config_preserves_legacy_positional_argument_order():
+    from dynamo.common.backend.worker import WorkerConfig
+
+    config = WorkerConfig(
+        "dynamo",  # namespace
+        "backend",  # component
+        "generate",  # endpoint
+        "",  # model_name
+        None,  # served_model_name
+        core.ModelInput.Tokens,  # model_input
+        "chat,completions",  # endpoint_types
+        "etcd",  # discovery_backend
+        "tcp",  # request_plane
+        None,  # event_plane
+        False,  # use_kv_events
+        None,  # custom_jinja_template
+        None,  # tool_call_parser
+        None,  # reasoning_parser
+        False,  # exclude_tools_when_tool_choice_none
+        False,  # enable_local_indexer
+    )
+
+    assert config.exclude_tools_when_tool_choice_none is False
+    assert config.enable_local_indexer is False
+    assert config.default_thinking_mode is None
+
+
 def test_worker_config_accepts_media_configuration():
     """Unified registration can advertise frontend media decoding."""
     from dynamo.llm import MediaDecoder, MediaFetcher
@@ -157,6 +208,7 @@ def test_python_worker_config_from_runtime_config_copies_parser_settings():
     runtime_cfg.custom_jinja_template = None
     runtime_cfg.dyn_tool_call_parser = "kimi_k2"
     runtime_cfg.dyn_reasoning_parser = "kimi_k25"
+    runtime_cfg.dyn_default_thinking_mode = "disabled"
     runtime_cfg.exclude_tools_when_tool_choice_none = False
     runtime_cfg.enable_local_indexer = False
     runtime_cfg.dyn_enable_structural_tag = True
@@ -171,6 +223,7 @@ def test_python_worker_config_from_runtime_config_copies_parser_settings():
 
     assert config.tool_call_parser == "kimi_k2"
     assert config.reasoning_parser == "kimi_k25"
+    assert config.default_thinking_mode == "disabled"
     assert config.exclude_tools_when_tool_choice_none is False
     assert config.enable_local_indexer is False
     assert config.structural_tag_mode == "on"
@@ -195,6 +248,7 @@ class _BareRuntime:
     assert cfg.endpoint_types == "chat,completions"
     assert cfg.use_kv_events is False
     assert cfg.custom_jinja_template is None
+    assert cfg.default_thinking_mode is None
     assert cfg.structural_tag_mode == "off"
     assert cfg.structural_tag_scope == "auto"
     assert cfg.structural_tag_schema == "auto"

components/src/dynamo/common/backend/worker.py

@@ -146,6 +146,7 @@ class WorkerConfig:
     media_fetcher: Optional[MediaFetcher] = None
     # KV event/recovery ownership endpoint. None uses this worker's serving endpoint.
     kv_state_endpoint: Optional[str] = None
+    default_thinking_mode: Optional[str] = None
 
     @classmethod
     def from_runtime_config(
@@ -180,6 +181,9 @@ def from_runtime_config(
             ),
             "tool_call_parser": getattr(runtime_cfg, "dyn_tool_call_parser", None),
             "reasoning_parser": getattr(runtime_cfg, "dyn_reasoning_parser", None),
+            "default_thinking_mode": getattr(
+                runtime_cfg, "dyn_default_thinking_mode", None
+            ),
             "exclude_tools_when_tool_choice_none": getattr(
                 runtime_cfg, "exclude_tools_when_tool_choice_none", True
             ),
@@ -266,6 +270,7 @@ async def run(self) -> None:
             custom_jinja_template=self.config.custom_jinja_template,
             tool_call_parser=self.config.tool_call_parser,
             reasoning_parser=self.config.reasoning_parser,
+            default_thinking_mode=self.config.default_thinking_mode,
             exclude_tools_when_tool_choice_none=(
                 self.config.exclude_tools_when_tool_choice_none
             ),

components/src/dynamo/common/configuration/groups/runtime_args.py

@@ -37,6 +37,7 @@ class DynamoRuntimeConfig(ConfigBase):
 
     dyn_tool_call_parser: Optional[str] = None
     dyn_reasoning_parser: Optional[str] = None
+    dyn_default_thinking_mode: Optional[str] = None
     exclude_tools_when_tool_choice_none: bool = True
     dyn_enable_structural_tag: bool = False
     dyn_structural_tag_scope: str = "auto"
@@ -196,6 +197,16 @@ def add_arguments(self, parser: argparse.ArgumentParser) -> None:
             help="Reasoning parser name for the model. If not specified, no reasoning parsing is performed.",
             choices=get_reasoning_parser_names(),
         )
+        add_argument(
+            g,
+            flag_name="--dyn-default-thinking-mode",
+            env_var="DYN_DEFAULT_THINKING_MODE",
+            default=None,
+            choices=["enabled", "disabled"],
+            help="Deployment-level default thinking mode for chat templates. "
+            "Client request thinking, reasoning_effort, chat_template_args, or "
+            "chat_template_kwargs values override this default.",
+        )
         # NOTE: This flag also exists in FrontendArgGroup (frontend_args.py).
         # Both definitions are needed: this one controls the Rust-native chat
         # template path (oai.rs), while the frontend copy controls the Python

components/src/dynamo/common/tests/configuration/test_runtime_args.py

@@ -122,3 +122,28 @@ def test_fpm_trace_help_lists_flag_and_env(monkeypatch):
     assert "--fpm-trace" in help_text
     assert "--no-fpm-trace" in help_text
     assert "DYN_FPM_TRACE" in help_text
+
+
+@pytest.mark.parametrize("mode", ["enabled", "disabled"])
+def test_default_thinking_mode_cli(mode, monkeypatch):
+    monkeypatch.delenv("DYN_DEFAULT_THINKING_MODE", raising=False)
+
+    config, help_text = _parse_runtime_args(["--dyn-default-thinking-mode", mode])
+
+    assert config.dyn_default_thinking_mode == mode
+    assert "DYN_DEFAULT_THINKING_MODE" in help_text
+
+
+def test_default_thinking_mode_env(monkeypatch):
+    monkeypatch.setenv("DYN_DEFAULT_THINKING_MODE", "disabled")
+
+    config, _ = _parse_runtime_args([])
+
+    assert config.dyn_default_thinking_mode == "disabled"
+
+
+def test_default_thinking_mode_rejects_invalid_value(monkeypatch):
+    monkeypatch.delenv("DYN_DEFAULT_THINKING_MODE", raising=False)
+
+    with pytest.raises(SystemExit):
+        _parse_runtime_args(["--dyn-default-thinking-mode", "adaptive"])

components/src/dynamo/frontend/prepost.py

@@ -23,6 +23,8 @@
 from vllm.tool_parsers import ToolParser
 from vllm.utils.async_utils import make_async
 
+from .thinking import apply_default_thinking_mode_to_template_kwargs
+
 
 class _Renderer(Protocol):
     """Structural type for vLLM's chat-template renderer."""
@@ -94,6 +96,7 @@ def _prepare_request(
     exclude_tools_when_tool_choice_none: bool = True,
     enable_auto_tool_choice: bool = False,
     default_chat_template_kwargs: dict[str, Any] | None = None,
+    default_thinking_mode: str | None = None,
 ) -> tuple[ChatCompletionRequest, ToolParser | None, dict[str, Any], Any, ChatParams]:
     """Validate request and build arguments for template rendering.
 
@@ -151,10 +154,20 @@ def _prepare_request(
             request_for_sampling.chat_template_kwargs or raw_template_args or {},
         )
     )
-    # Don't let an absent top-level field clobber a nested reasoning_effort.
+    # reasoning_effort is a request-level thinking control. Put an explicit
+    # value into the kwargs before applying the deployment default so the two
+    # cannot produce contradictory template controls.
     if request_for_sampling.reasoning_effort is not None:
         chat_template_kwargs["reasoning_effort"] = request_for_sampling.reasoning_effort
-    else:
+    chat_template_kwargs = apply_default_thinking_mode_to_template_kwargs(
+        chat_template_kwargs,
+        default_thinking_mode,
+        request_has_root_thinking=(
+            isinstance(request, dict) and request.get("thinking") is not None
+        ),
+    )
+    # Don't let an absent top-level field clobber a nested reasoning_effort.
+    if request_for_sampling.reasoning_effort is None:
         chat_template_kwargs.setdefault("reasoning_effort", None)
 
     # Mistral warns that tokenize=False is unsafe for chat templates.
@@ -201,6 +214,7 @@ async def preprocess_chat_request(
     exclude_tools_when_tool_choice_none: bool = True,
     enable_auto_tool_choice: bool = False,
     default_chat_template_kwargs: dict[str, Any] | None = None,
+    default_thinking_mode: str | None = None,
 ) -> PreprocessResult:
     (
         request_for_sampling,
@@ -215,6 +229,7 @@ async def preprocess_chat_request(
         exclude_tools_when_tool_choice_none=exclude_tools_when_tool_choice_none,
         enable_auto_tool_choice=enable_auto_tool_choice,
         default_chat_template_kwargs=default_chat_template_kwargs,
+        default_thinking_mode=default_thinking_mode,
     )
 
     _, engine_prompt = await renderer.render_messages_async(messages, chat_params)

components/src/dynamo/frontend/sglang_prepost.py

@@ -28,6 +28,7 @@
 )
 from sglang.srt.parser.reasoning_parser import ReasoningParser
 
+from .thinking import apply_default_thinking_mode_to_template_kwargs
 from .utils import PreprocessError, random_call_id
 
 logger = logging.getLogger(__name__)
@@ -410,6 +411,7 @@ def _flatten_message_content(content: Any) -> Any:
 
 def _normalize_openai_thinking_template_kwargs(
     request: dict[str, Any],
+    default_thinking_mode: str | None = None,
 ) -> dict[str, Any]:
     request = copy.copy(request)
     chat_template_kwargs = dict(
@@ -434,9 +436,18 @@ def setdefault_reasoning(enabled: bool) -> None:
         elif thinking_type == "disabled":
             setdefault_reasoning(False)
 
-    if request.get("reasoning_effort") == "none":
+    reasoning_effort = request.get("reasoning_effort")
+    if reasoning_effort is not None:
+        chat_template_kwargs["reasoning_effort"] = reasoning_effort
+    if reasoning_effort == "none":
         setdefault_reasoning(False)
 
+    chat_template_kwargs = apply_default_thinking_mode_to_template_kwargs(
+        chat_template_kwargs,
+        default_thinking_mode,
+        request_has_root_thinking=request.get("thinking") is not None,
+    )
+
     if chat_template_kwargs:
         request["chat_template_kwargs"] = chat_template_kwargs
     return request
@@ -703,6 +714,7 @@ def preprocess_chat_request(
     reasoning_parser_name: str | None,
     exclude_tools_when_tool_choice_none: bool = True,
     template_force_reasoning: bool = False,
+    default_thinking_mode: str | None = None,
 ) -> SglangPreprocessResult:
     """Preprocess a chat request using SGLang tokenizer and parser APIs.
 
@@ -713,7 +725,7 @@ def preprocess_chat_request(
 
     Synchronous -- suitable for both main-process and worker-process execution.
     """
-    request = _normalize_openai_thinking_template_kwargs(request)
+    request = _normalize_openai_thinking_template_kwargs(request, default_thinking_mode)
     messages = _materialize_messages(request.get("messages", []))
 
     # Generation mode is independent of whether the client wants reasoning

components/src/dynamo/frontend/sglang_processor.py

@@ -37,6 +37,7 @@
     detect_force_reasoning_from_template,
     preprocess_chat_request,
 )
+from .thinking import runtime_default_thinking_mode
 from .utils import (
     PreprocessError,
     extract_mm_urls,
@@ -145,6 +146,7 @@ def _map_finish_reason(raw: str | None) -> str | None:
 _w_reasoning_parser_name: str | None = None
 _w_exclude_tools_when_tool_choice_none: bool = True
 _w_template_force_reasoning: bool = False
+_w_default_thinking_mode: str | None = None
 
 
 def _load_chat_template(chat_template: str | None) -> str | None:
@@ -194,17 +196,20 @@ def _init_worker(
     trust_remote_code: bool = False,
     template_force_reasoning: bool = False,
     chat_template: str | None = None,
+    default_thinking_mode: str | None = None,
 ) -> None:
     """Initialize a worker process with its own tokenizer."""
     global _w_tokenizer, _w_tool_call_parser_name, _w_reasoning_parser_name
     global _w_exclude_tools_when_tool_choice_none, _w_template_force_reasoning
+    global _w_default_thinking_mode
     _w_tokenizer = _load_tokenizer(model_path, trust_remote_code)
     if chat_template is not None:
         _w_tokenizer.chat_template = chat_template
     _w_tool_call_parser_name = tool_call_parser_name
     _w_reasoning_parser_name = reasoning_parser_name
     _w_exclude_tools_when_tool_choice_none = exclude_tools_when_tool_choice_none
     _w_template_force_reasoning = template_force_reasoning
+    _w_default_thinking_mode = default_thinking_mode
 
 
 def _preprocess_worker(
@@ -220,6 +225,7 @@ def _preprocess_worker(
         reasoning_parser_name=_w_reasoning_parser_name,
         exclude_tools_when_tool_choice_none=_w_exclude_tools_when_tool_choice_none,
         template_force_reasoning=_w_template_force_reasoning,
+        default_thinking_mode=_w_default_thinking_mode,
     )
 
     n = request.get("n", 1)
@@ -358,6 +364,7 @@ def __init__(
         preprocess_pool: ProcessPoolExecutor | None = None,
         preprocess_workers: int = 0,
         stream_interval: int = 1,
+        default_thinking_mode: str | None = None,
     ):
         self.tokenizer = tokenizer
         # Detect force_reasoning once from the chat template, matching
@@ -381,6 +388,7 @@ def __init__(
         self.eos_token_ids = _normalize_eos_token_ids(eos_token_ids)
         self.debug_perf = debug_perf
         self.stream_interval = stream_interval
+        self.default_thinking_mode = default_thinking_mode
         self.preprocess_pool = preprocess_pool
         if preprocess_pool is not None:
             self._worker_semaphore: asyncio.Semaphore | None = asyncio.Semaphore(
@@ -437,6 +445,7 @@ async def _generator_inner(
                 reasoning_parser_name=self.reasoning_parser_name,
                 exclude_tools_when_tool_choice_none=self.exclude_tools_when_tool_choice_none,
                 template_force_reasoning=self.template_force_reasoning,
+                default_thinking_mode=self.default_thinking_mode,
             )
 
             if self.debug_perf:
@@ -789,11 +798,14 @@ async def chat_engine_factory(
             self.reasoning_parser_name
             or _runtime_config_parser_name(mdc, "reasoning_parser")
         )
+        default_thinking_mode = runtime_default_thinking_mode(mdc.runtime_config())
 
         if tool_call_parser_name:
             logger.info("SGLang tool call parser: %s", tool_call_parser_name)
         if reasoning_parser_name:
             logger.info("SGLang reasoning parser: %s", reasoning_parser_name)
+        if default_thinking_mode:
+            logger.info("SGLang default thinking mode: %s", default_thinking_mode)
 
         preprocess_pool = None
         preprocess_workers = self.config.preprocess_workers
@@ -814,6 +826,7 @@ async def chat_engine_factory(
                     self.trust_remote_code,
                     template_force_reasoning,
                     chat_template,
+                    default_thinking_mode,
                 ),
             )
             futures = [
@@ -849,6 +862,7 @@ async def chat_engine_factory(
             preprocess_pool=preprocess_pool,
             preprocess_workers=preprocess_workers,
             stream_interval=self.stream_interval,
+            default_thinking_mode=default_thinking_mode,
         )
         gen.exclude_tools_when_tool_choice_none = (
             self.config.exclude_tools_when_tool_choice_none