From 5eff2f8b94f0ca07bad5f7fa2649749eaa06f860 Mon Sep 17 00:00:00 2001
From: Jintao Huang <huangjintao.hjt@alibaba-inc.com>
Date: Wed, 6 May 2026 20:40:53 +0800
Subject: [PATCH 01/13] support chat_template_kwargs

---
 swift/dataset/preprocessor/core.py | 1 +
 swift/template/template_inputs.py  | 4 +++-
 2 files changed, 4 insertions(+), 1 deletion(-)

diff --git a/swift/dataset/preprocessor/core.py b/swift/dataset/preprocessor/core.py
index 7f8d0b83d6..3d0df336de 100644
--- a/swift/dataset/preprocessor/core.py
+++ b/swift/dataset/preprocessor/core.py
@@ -34,6 +34,7 @@ class RowPreprocessor:
                                 'channel',
                                 'margin',
                                 'teacher_prompt',
+                                'chat_template_kwargs',
                             ]
 
     def __init__(self,
diff --git a/swift/template/template_inputs.py b/swift/template/template_inputs.py
index 8165f1d7b1..3108b8019b 100644
--- a/swift/template/template_inputs.py
+++ b/swift/template/template_inputs.py
@@ -28,9 +28,11 @@ class StdTemplateInputs:
     objects: Dict[str, Any] = field(default_factory=dict)
 
     margin: Optional[float] = None  # for reward modeling
-    mm_processor_kwargs: Dict[str, Any] = field(default_factory=dict)
+    chat_template_kwargs: Dict[str, Any] = field(default_factory=dict)  # from dataset
     extra_kwargs: Dict[str, Any] = field(default_factory=dict)
 
+    mm_processor_kwargs: Dict[str, Any] = field(default_factory=dict)
+
     def __post_init__(self):
         self.image_idx = 0
         self.audio_idx = 0

From 179a5f45e9b17facd86ae314b791691243098715 Mon Sep 17 00:00:00 2001
From: Jintao Huang <huangjintao.hjt@alibaba-inc.com>
Date: Thu, 7 May 2026 00:07:42 +0800
Subject: [PATCH 02/13] update

---
 swift/dataset/preprocessor/core.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/swift/dataset/preprocessor/core.py b/swift/dataset/preprocessor/core.py
index 3d0df336de..81f5c09363 100644
--- a/swift/dataset/preprocessor/core.py
+++ b/swift/dataset/preprocessor/core.py
@@ -262,6 +262,7 @@ def _new_init(self, schema=None, features=None, *args, **kwargs):
                         features[key] = messages_feature
                     features['images'] = List({'bytes': Value(dtype='binary'), 'path': Value(dtype='string')})
                     features['objects'] = Json()
+                    features['chat_template_kwargs'] = Json()
                 else:
                     messages_feature = [{
                         'role': Value(dtype='string'),

From 59dbc218f783cda8953920428e8f1c5c4c21af23 Mon Sep 17 00:00:00 2001
From: Jintao Huang <huangjintao.hjt@alibaba-inc.com>
Date: Thu, 7 May 2026 00:27:11 +0800
Subject: [PATCH 03/13] update

---
 docs/source/Customization/Custom-dataset.md    | 7 +++++++
 docs/source_en/Customization/Custom-dataset.md | 9 +++++++++
 requirements/framework.txt                     | 2 +-
 swift/template/template_inputs.py              | 2 +-
 4 files changed, 18 insertions(+), 2 deletions(-)

diff --git a/docs/source/Customization/Custom-dataset.md b/docs/source/Customization/Custom-dataset.md
index f71175a993..cb12dd3f96 100644
--- a/docs/source/Customization/Custom-dataset.md
+++ b/docs/source/Customization/Custom-dataset.md
@@ -94,6 +94,13 @@ print(inputs['loss_scale'])
 {"messages": [..., {"role": "tool_call", "content": "{\"name\": \"realtime_aqi\", \"arguments\": {\"city\": \"北京\"}}", "loss": false}, {"role": "tool_call", "content": "{\"name\": \"realtime_aqi\", \"arguments\": {\"city\": \"上海\"}}"}, ...]}
 ```
 
+- "chat_template_kwargs"字段，（需ms-swift>=4.2.1），你可以通过在数据集中传入该字段**样本级别**控制template的min_pixels, max_pixels, fps等多模态参数，以及enable_thinking（推理时）等参数。以下为不同模型支持的参数：
+  - 其中enable_thinking支持所有模型（推理时生效）；max_pixels参数支持所有多模态模型。
+  - Qwen系列多模态模型：min_pixels, max_pixels, fps等qwen_vl_utils/qwen_omni_utils支持的参数。
+```jsonl
+{"messages": [{"role": "user", "content": "<image>这是什么"}, {"role": "assistant", "content": "这是一只兔子", "loss": false}], "chat_template_kwargs": {"max_pixels": 1048576}}
+{"messages": [{"role": "user", "content": "who are you?"}], "chat_template_kwargs": {"enable_thinking": false}}
+```
 
 #### channel loss
 如果你要使用channel loss，你需要设置`--enable_channel_loss true`，并在数据集中增加"channel"字段。channel loss兼容packing/padding_free/loss_scale等技术。
diff --git a/docs/source_en/Customization/Custom-dataset.md b/docs/source_en/Customization/Custom-dataset.md
index 3b99737883..9d88f3d581 100644
--- a/docs/source_en/Customization/Custom-dataset.md
+++ b/docs/source_en/Customization/Custom-dataset.md
@@ -96,6 +96,15 @@ Note: If you set "loss"/"loss_scale" on consecutive "tool_call" messages in the
 {"messages": [..., {"role": "tool_call", "content": "{\"name\": \"realtime_aqi\", \"arguments\": {\"city\": \"Beijing\"}}", "loss": false}, {"role": "tool_call", "content": "{\"name\": \"realtime_aqi\", \"arguments\": {\"city\": \"Shanghai\"}}"}, ...]}
 ```
 
+- The "chat_template_kwargs" field (requires ms-swift>=4.2.1) allows you to control template multimodal parameters such as min_pixels, max_pixels, fps, as well as parameters like enable_thinking (during inference) at the **sample level** by passing this field in the dataset. The following parameters are supported by different models:
+  - Among them, enable_thinking is supported by all models (takes effect during inference); the max_pixels parameter is supported by all multimodal models.
+  - Qwen series multimodal models: parameters supported by qwen_vl_utils/qwen_omni_utils such as min_pixels, max_pixels, fps, etc.
+
+```jsonl
+{"messages": [{"role": "user", "content": "<image>What is this"}, {"role": "assistant", "content": "This is a rabbit", "loss": false}], "chat_template_kwargs": {"max_pixels": 1048576}}
+{"messages": [{"role": "user", "content": "who are you?"}], "chat_template_kwargs": {"enable_thinking": false}}
+```
+
 #### Channel Loss
 If you want to use channel loss, you need to set `--enable_channel_loss true` and add a "channel" field to your dataset. Channel loss is compatible with techniques such as packing, padding-free, and loss scaling.
 
diff --git a/requirements/framework.txt b/requirements/framework.txt
index cc6f80636f..66ad03f7f4 100644
--- a/requirements/framework.txt
+++ b/requirements/framework.txt
@@ -32,7 +32,7 @@ sortedcontainers>=1.5.9
 tensorboard
 tiktoken
 tqdm
-transformers>=4.33,<5.8.0
+transformers>=4.33,<5.9.0
 transformers_stream_generator
 trl>=0.15,<1.0
 uvicorn
diff --git a/swift/template/template_inputs.py b/swift/template/template_inputs.py
index 3108b8019b..3fdc44b9b0 100644
--- a/swift/template/template_inputs.py
+++ b/swift/template/template_inputs.py
@@ -59,7 +59,7 @@ def is_multimodal(self):
     def from_dict(cls, inputs: Dict[str, Any]) -> 'StdTemplateInputs':
         inputs = deepcopy(inputs)
         kwargs = {}
-        for key in ['label', 'channel', 'margin', 'rejected_response']:
+        for key in ['label', 'channel', 'margin', 'rejected_response', 'chat_template_kwargs']:
             if key in inputs:
                 kwargs[key] = inputs[key]
         messages = inputs['messages']

From 4ee8ee687e772883e628c33ff3231e36ea60aebe Mon Sep 17 00:00:00 2001
From: Jintao Huang <huangjintao.hjt@alibaba-inc.com>
Date: Thu, 7 May 2026 11:13:35 +0800
Subject: [PATCH 04/13] fix

---
 swift/template/templates/qwen.py | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/swift/template/templates/qwen.py b/swift/template/templates/qwen.py
index de0ac6c80e..00959802bb 100644
--- a/swift/template/templates/qwen.py
+++ b/swift/template/templates/qwen.py
@@ -316,7 +316,7 @@ def replace_tag(self, media_type: Literal['image', 'video', 'audio'], index: int
             # ref: https://github.com/modelscope/ms-swift/issues/8445
             inputs.mm_processor_kwargs['do_resize'] = False
         if media_type == 'image':
-            inputs.images[index] = fetch_image({'image': inputs.images[index]}, **kwargs)
+            inputs.images[index] = fetch_image({'image': inputs.images[index], **inputs.chat_template_kwargs}, **kwargs)
             if self.mode == 'lmdeploy':
                 return ['<|vision_start|>', [-100], '<|vision_end|>']
             else:
@@ -325,7 +325,7 @@ def replace_tag(self, media_type: Literal['image', 'video', 'audio'], index: int
             if self.version == 'v3':
                 kwargs['return_video_metadata'] = True
             video = inputs.videos[index]
-            video_inputs = {'video': video}
+            video_inputs = {'video': video, **inputs.chat_template_kwargs}
             if isinstance(video, list):  # image list
                 from qwen_vl_utils import vision_process
                 video_inputs['sample_fps'] = vision_process.FPS
@@ -652,7 +652,7 @@ def replace_tag(self, media_type: Literal['image', 'video', 'audio'], index: int
             # https://github.com/modelscope/ms-swift/issues/8445
             inputs.mm_processor_kwargs['do_resize'] = False
         if media_type == 'image':
-            inputs.images[index] = fetch_image({'image': inputs.images[index]}, **kwargs)
+            inputs.images[index] = fetch_image({'image': inputs.images[index], **inputs.chat_template_kwargs}, **kwargs)
             if self.version == 'omni_v2_5':
                 return ['<|vision_bos|><|IMAGE|><|vision_eos|>']
             elif self.version == 'omni_v3':
@@ -666,7 +666,7 @@ def replace_tag(self, media_type: Literal['image', 'video', 'audio'], index: int
                 return ['<|audio_start|><|audio_pad|><|audio_end|>']
         elif media_type == 'video':
             video = inputs.videos[index]
-            _video = fetch_video({'video': video}, **kwargs)
+            _video = fetch_video({'video': video, **inputs.chat_template_kwargs}, **kwargs)
             if isinstance(_video, torch.Tensor):
                 _video = _video.to(torch.uint8)
             inputs.videos[index] = _video

From 44431d660c7a9bf9b46e469ae96b2b32fa2ff828 Mon Sep 17 00:00:00 2001
From: Jintao Huang <huangjintao.hjt@alibaba-inc.com>
Date: Thu, 7 May 2026 13:58:57 +0800
Subject: [PATCH 05/13] fix

---
 swift/template/base.py           | 53 +++++++++++++++++++++-----------
 swift/template/templates/qwen.py |  6 ++++
 2 files changed, 41 insertions(+), 18 deletions(-)

diff --git a/swift/template/base.py b/swift/template/base.py
index fa33de0664..ead36f3e61 100644
--- a/swift/template/base.py
+++ b/swift/template/base.py
@@ -121,7 +121,7 @@ def __init__(
             template_meta.default_system = default_system
         if enable_thinking is None:
             enable_thinking = template_meta.is_thinking
-        self._response_prefix = response_prefix
+        self.response_prefix = response_prefix
         self.template_meta: 'TemplateMeta' = template_meta
         self.use_chat_template = use_chat_template
         self.enable_thinking = enable_thinking
@@ -156,13 +156,22 @@ def __init__(
         if processor is not None:
             self.init_processor(processor)
 
-    @property
-    def response_prefix(self):
-        if self._response_prefix is not None:
-            return self._response_prefix
+    def _get_enable_thinking(self, inputs):
+        enable_thinking = inputs.chat_template_kwargs.get('enable_thinking')
+        if enable_thinking is None:
+            enable_thinking = self.enable_thinking
+        return enable_thinking
+
+    def _get_response_prefix(self, inputs):
+        response_prefix = inputs.chat_template_kwargs.get('response_prefix')
+        if response_prefix is None:
+            response_prefix = self.response_prefix
+        if response_prefix is not None:
+            return response_prefix
         elif not self.use_chat_template:
             return ''
-        elif self.enable_thinking:
+        enable_thinking = self._get_enable_thinking(inputs)
+        if enable_thinking:
             return self.template_meta.thinking_prefix
         else:
             return self.template_meta.non_thinking_prefix
@@ -308,6 +317,12 @@ def _preprocess_function_call(self, inputs: StdTemplateInputs) -> None:
     def prepare_engine_kwargs(self) -> Dict[str, Any]:
         return {}
 
+    def _get_max_pixels(self, inputs):
+        max_pixels = inputs.chat_template_kwargs.get('max_pixels')
+        if max_pixels is None:
+            max_pixels = self.max_pixels
+        return max_pixels
+
     def _preprocess_inputs(
         self,
         inputs: StdTemplateInputs,
@@ -320,16 +335,17 @@ def _preprocess_inputs(
         images = inputs.images
         load_images = self.load_images or self.mode in {'vllm', 'lmdeploy'}
         load_images_origin = load_images
-        if self.max_pixels is not None or inputs.objects:
+        max_pixels = self._get_max_pixels(inputs)
+        if max_pixels is not None or inputs.objects:
             load_images = True
         if images:
             for i, image in enumerate(images):
                 images[i] = self._load_image(images[i], load_images)
         if inputs.objects:
             self._get_height_width(inputs)
-        if self.max_pixels is not None:
+        if max_pixels is not None:
             # Scale the image proportionally without affecting the scaled objects.
-            images = [rescale_image(img, self.max_pixels) for img in images]
+            images = [rescale_image(img, max_pixels) for img in images]
         if images and not load_images_origin:  # fix pt & qwen-vl
             for i, image in enumerate(images):
                 if isinstance(image, Image.Image):
@@ -1066,12 +1082,11 @@ def _jinja_encode(self, inputs: StdTemplateInputs):
         kwargs = {}
         if inputs.tools:
             kwargs['tools'] = inputs.tools
-        if 'thinking_budget' in inputs.extra_kwargs:
-            kwargs['thinking_budget'] = inputs.extra_kwargs.get('thinking_budget', 0)
-        if self.template_meta.is_thinking or self.enable_thinking:
-            kwargs[self.jinja_enable_thinking_key] = self.enable_thinking
-        if self.chat_template_kwargs:
-            kwargs.update(self.chat_template_kwargs)
+        enable_thinking = self._get_enable_thinking(inputs)
+        if self.template_meta.is_thinking or enable_thinking:
+            kwargs[self.jinja_enable_thinking_key] = enable_thinking
+        kwargs.update(self.chat_template_kwargs)
+        kwargs.update(inputs.chat_template_kwargs)
         text = self.tokenizer.apply_chat_template(
             messages, tokenize=False, add_generation_prompt=add_generation_prompt, **kwargs)
         answer_len = 1 if self.is_training else 0
@@ -1182,7 +1197,8 @@ def _swift_encode(self, inputs: StdTemplateInputs):
         if self.use_chat_template:
             if self.add_non_thinking_prefix:
                 self._add_non_thinking_prefix(inputs)
-            if template_meta.is_thinking or self.enable_thinking:
+            enable_thinking = self._get_enable_thinking(inputs)
+            if template_meta.is_thinking or enable_thinking:
                 self._remove_history_thinking(inputs)
             system = self._get_system(inputs)
         else:
@@ -1234,6 +1250,7 @@ def _swift_encode(self, inputs: StdTemplateInputs):
             context_list = prompt.copy()
             extra_context_list = []
             extra_context_type = None
+            response_prefix = self._get_response_prefix(inputs)
             if i < n_round - 1:
                 # Not the last round.
                 context_list.append('{{RESPONSE}}')
@@ -1264,9 +1281,9 @@ def _swift_encode(self, inputs: StdTemplateInputs):
                 if add_eos:
                     extra_context_list = template_meta.suffix
                     extra_context_type = ContextType.SUFFIX
-            elif self.response_prefix:
+            elif response_prefix:
                 # final round and during inference.
-                context_list.append(self.response_prefix)
+                context_list.append(response_prefix)
 
             self._concat_context_list(
                 context_list,
diff --git a/swift/template/templates/qwen.py b/swift/template/templates/qwen.py
index f751108974..34ef62c4f4 100644
--- a/swift/template/templates/qwen.py
+++ b/swift/template/templates/qwen.py
@@ -306,6 +306,9 @@ def init_env_args(self):
         self.transformers_version = version.parse(transformers.__version__)
         self.bbox_format = get_env_args('QWENVL_BBOX_FORMAT', str, 'legacy')
 
+    def _get_max_pixels(self, inputs):
+        return self.max_pixels
+
     def replace_tag(self, media_type: Literal['image', 'video', 'audio'], index: int,
                     inputs: StdTemplateInputs) -> List[Context]:
         from qwen_vl_utils import fetch_image, fetch_video
@@ -672,6 +675,9 @@ def init_processor(self, processor) -> None:
         self.use_audio_in_video = get_env_args('use_audio_in_video', bool, False)
         self.sampling_rate = get_env_args('sampling_rate', int, self.processor.feature_extractor.sampling_rate)
 
+    def _get_max_pixels(self, inputs):
+        return self.max_pixels
+
     def replace_tag(self, media_type: Literal['image', 'video', 'audio'], index: int,
                     inputs: StdTemplateInputs) -> List[Context]:
         from qwen_omni_utils import fetch_image, fetch_video

From ffc96d45482d9b904a89e273a3fb7da93ae6ee23 Mon Sep 17 00:00:00 2001
From: Jintao Huang <huangjintao.hjt@alibaba-inc.com>
Date: Thu, 7 May 2026 14:54:14 +0800
Subject: [PATCH 06/13] update

---
 docs/source/Customization/Custom-dataset.md    | 2 +-
 docs/source_en/Customization/Custom-dataset.md | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/docs/source/Customization/Custom-dataset.md b/docs/source/Customization/Custom-dataset.md
index cb12dd3f96..ff3a40087d 100644
--- a/docs/source/Customization/Custom-dataset.md
+++ b/docs/source/Customization/Custom-dataset.md
@@ -95,7 +95,7 @@ print(inputs['loss_scale'])
 ```
 
 - "chat_template_kwargs"字段，（需ms-swift>=4.2.1），你可以通过在数据集中传入该字段**样本级别**控制template的min_pixels, max_pixels, fps等多模态参数，以及enable_thinking（推理时）等参数。以下为不同模型支持的参数：
-  - 其中enable_thinking支持所有模型（推理时生效）；max_pixels参数支持所有多模态模型。
+  - 其中"enable_thinking", "response_prefix"支持所有模型（推理时生效）；"max_pixels"参数支持所有多模态模型。
   - Qwen系列多模态模型：min_pixels, max_pixels, fps等qwen_vl_utils/qwen_omni_utils支持的参数。
 ```jsonl
 {"messages": [{"role": "user", "content": "<image>这是什么"}, {"role": "assistant", "content": "这是一只兔子", "loss": false}], "chat_template_kwargs": {"max_pixels": 1048576}}
diff --git a/docs/source_en/Customization/Custom-dataset.md b/docs/source_en/Customization/Custom-dataset.md
index 9d88f3d581..6e5c6f7788 100644
--- a/docs/source_en/Customization/Custom-dataset.md
+++ b/docs/source_en/Customization/Custom-dataset.md
@@ -97,7 +97,7 @@ Note: If you set "loss"/"loss_scale" on consecutive "tool_call" messages in the
 ```
 
 - The "chat_template_kwargs" field (requires ms-swift>=4.2.1) allows you to control template multimodal parameters such as min_pixels, max_pixels, fps, as well as parameters like enable_thinking (during inference) at the **sample level** by passing this field in the dataset. The following parameters are supported by different models:
-  - Among them, enable_thinking is supported by all models (takes effect during inference); the max_pixels parameter is supported by all multimodal models.
+  - Among them, "enable_thinking" and "response_prefix" are supported by all models (takes effect during inference); the "max_pixels" parameter is supported by all multimodal models.
   - Qwen series multimodal models: parameters supported by qwen_vl_utils/qwen_omni_utils such as min_pixels, max_pixels, fps, etc.
 
 ```jsonl

From 2fb5e363f4aaf1ca396369e066f67ccc52234515 Mon Sep 17 00:00:00 2001
From: Jintao Huang <huangjintao.hjt@alibaba-inc.com>
Date: Thu, 7 May 2026 16:04:12 +0800
Subject: [PATCH 07/13] update

---
 swift/infer_engine/grpo_vllm_engine.py    |  2 +-
 swift/infer_engine/lmdeploy_engine.py     |  7 ++++---
 swift/infer_engine/protocol.py            |  1 +
 swift/infer_engine/sglang_engine.py       |  6 +++---
 swift/infer_engine/transformers_engine.py |  8 ++++----
 swift/infer_engine/vllm_engine.py         | 14 +++++++++----
 swift/template/base.py                    | 25 +++++++++++++++--------
 swift/template/template_inputs.py         |  4 +++-
 8 files changed, 42 insertions(+), 25 deletions(-)

diff --git a/swift/infer_engine/grpo_vllm_engine.py b/swift/infer_engine/grpo_vllm_engine.py
index 26470720ea..653e0d2de7 100644
--- a/swift/infer_engine/grpo_vllm_engine.py
+++ b/swift/infer_engine/grpo_vllm_engine.py
@@ -108,7 +108,7 @@ def _create_chat_completion_response(self, result, inputs, request_config, reque
         choices = []
         for output in result.outputs:
             output.token_ids = list(output.token_ids)
-            response = self.template.decode(output.token_ids)
+            response = self.template.decode(output.token_ids, template_inputs=inputs['template_inputs'])
             logprobs = self._get_logprobs(output.logprobs, output.token_ids, request_config.top_logprobs)
             toolcall = self._get_toolcall(response)
 
diff --git a/swift/infer_engine/lmdeploy_engine.py b/swift/infer_engine/lmdeploy_engine.py
index a15d67b395..694ff4df6e 100644
--- a/swift/infer_engine/lmdeploy_engine.py
+++ b/swift/infer_engine/lmdeploy_engine.py
@@ -202,7 +202,7 @@ async def _infer_stream_async(
         else:
             context = self.engine.safe_run(session_id)
 
-        infer_streamer = InferStreamer(self.template)
+        infer_streamer = InferStreamer(self.template, template_inputs=inputs['template_inputs'])
         token_idx = 0
         async with context as gen:
             if version.parse(lmdeploy.__version__) < version.parse('0.6.5'):
@@ -225,7 +225,8 @@ async def _infer_stream_async(
                 usage_info = self._get_usage_info(len(inputs['input_ids']), output.num_token)
                 toolcall = None
                 if is_finished:
-                    toolcall = self._get_toolcall(self.template.decode(output.token_ids))
+                    toolcall = self._get_toolcall(
+                        self.template.decode(output.token_ids, template_inputs=inputs['template_inputs']))
                 finish_reason = self._get_finish_reason(generation_config.max_new_tokens, output.num_token,
                                                         output.status.name == 'FINISH')
                 choices = [
@@ -260,7 +261,7 @@ async def _infer_full_async(
                 async for output in generator.async_stream_infer(session_id=session_id, **inputs, **kwargs):
                     pass
 
-        response = self.template.decode(output.token_ids)
+        response = self.template.decode(output.token_ids, template_inputs=inputs['template_inputs'])
         logprobs = self._get_logprobs(output.logprobs, output.token_ids, request_config.top_logprobs)
 
         usage_info = self._get_usage_info(len(inputs['input_ids']), output.num_token)
diff --git a/swift/infer_engine/protocol.py b/swift/infer_engine/protocol.py
index c4b5a1a22d..9f1b8cf821 100644
--- a/swift/infer_engine/protocol.py
+++ b/swift/infer_engine/protocol.py
@@ -91,6 +91,7 @@ class InferRequest:
 
     tools: Optional[List[Tool]] = None
     objects: Dict[str, Any] = field(default_factory=dict)
+    chat_template_kwargs: Dict[str, Any] = field(default_factory=dict)
 
     def __post_init__(self):
         for key in ['images', 'audios', 'videos']:
diff --git a/swift/infer_engine/sglang_engine.py b/swift/infer_engine/sglang_engine.py
index e19dd13d41..3d923a3988 100644
--- a/swift/infer_engine/sglang_engine.py
+++ b/swift/infer_engine/sglang_engine.py
@@ -185,7 +185,7 @@ def _create_chat_completion_response(self, output, inputs, return_details: bool
         assert output is not None
         meta_info = output['meta_info']
         usage_info = self._get_usage_info(meta_info['prompt_tokens'], meta_info['completion_tokens'])
-        response = self.template.decode(output['output_ids'])
+        response = self.template.decode(output['output_ids'], template_inputs=inputs['template_inputs'])
         toolcall = self._get_toolcall(response)
         token_ids = output['output_ids'] if return_details else None
         choice = ChatCompletionResponseChoice(
@@ -271,7 +271,7 @@ async def _infer_stream_async(self, inputs: Dict[str, Any], generation_config: D
         engine_inputs = {k: v for k, v in inputs.items() if k != 'template_inputs'}
         result_generator = await self.engine.async_generate(
             **engine_inputs, sampling_params=generation_config, stream=True)
-        infer_streamer = InferStreamer(self.template)
+        infer_streamer = InferStreamer(self.template, template_inputs=inputs['template_inputs'])
         async for output in result_generator:
             res = self._create_chat_completion_stream_response(output, infer_streamer)
             if res is None:
@@ -289,7 +289,7 @@ def _create_chat_completion_stream_response(self, output, infer_streamer) -> Opt
         toolcall = None
         if is_finished:
             finish_reason = finish_reason['type']
-            toolcall = self._get_toolcall(self.template.decode(output['output_ids']))
+            toolcall = self._get_toolcall(self.template.decode(output['output_ids'], **infer_streamer.decode_kwargs))
         meta_info = output['meta_info']
         usage_info = self._get_usage_info(meta_info['prompt_tokens'], meta_info['completion_tokens'])
         # TODO: logprobs
diff --git a/swift/infer_engine/transformers_engine.py b/swift/infer_engine/transformers_engine.py
index ac7e963717..8af16deebc 100644
--- a/swift/infer_engine/transformers_engine.py
+++ b/swift/infer_engine/transformers_engine.py
@@ -217,8 +217,7 @@ def _update_batched_logprobs(batched_logprobs: List[torch.Tensor], logits_stream
 
     def _infer_stream(self, inputs: Dict[str, Any], *, generation_config: GenerationConfig,
                       adapter_request: Optional[AdapterRequest], request_config: RequestConfig,
-                      **kwargs) -> Iterator[List[Optional[ChatCompletionStreamResponse]]]:
-
+                      template_inputs) -> Iterator[List[Optional[ChatCompletionStreamResponse]]]:
         if generation_config.num_beams != 1:
             error_msg = 'Streaming generation does not support beam search.'
             raise ValueError(error_msg)
@@ -248,7 +247,7 @@ def _model_generate(**kwargs):
         batch_size = inputs['attention_mask'].shape[0]
         all_is_finished = False
         is_finished = [False] * batch_size
-        infer_streamers = [InferStreamer(self.template) for _ in range(batch_size)]
+        infer_streamers = [InferStreamer(self.template, template_inputs=template_inputs[i]) for i in range(batch_size)]
         request_id_list = [f'chatcmpl-{random_uuid()}' for _ in range(batch_size)]
         token_idxs = [0] * batch_size
 
@@ -298,7 +297,8 @@ def _model_generate(**kwargs):
                 usage_info = self._get_usage_info(num_prompt_tokens, len(generate_ids))
                 toolcall = None
                 if is_finished[i]:
-                    toolcall = self._get_toolcall(self.template.decode(generate_ids))
+                    toolcall = self._get_toolcall(
+                        self.template.decode(generate_ids, template_inputs=template_inputs[i]))
                 finish_reason = self._get_finish_reason(generation_config.max_new_tokens, usage_info.completion_tokens,
                                                         is_finished[i])
 
diff --git a/swift/infer_engine/vllm_engine.py b/swift/infer_engine/vllm_engine.py
index 4e5e66c0bb..f1f55ed7a7 100644
--- a/swift/infer_engine/vllm_engine.py
+++ b/swift/infer_engine/vllm_engine.py
@@ -494,7 +494,9 @@ async def _infer_stream_async(
     ) -> AsyncIterator[ChatCompletionStreamResponse]:
         request_id = random_uuid()
         result_generator = self._add_request(inputs, generation_config, request_id, adapter_request=adapter_request)
-        infer_streamers = [InferStreamer(self.template) for _ in range(generation_config.n)]
+        infer_streamers = [
+            InferStreamer(self.template, template_inputs=inputs['template_inputs']) for _ in range(generation_config.n)
+        ]
         token_idxs = [0 for _ in range(generation_config.n)]
         async for result in result_generator:
             res = self._create_chat_completion_stream_response(result, request_config, request_id, infer_streamers,
@@ -559,7 +561,8 @@ def _create_chat_completion_stream_response(self, result, request_config, reques
 
             toolcall = None
             if output.is_finished:
-                toolcall = self._get_toolcall(self.template.decode(output.token_ids))
+                toolcall = self._get_toolcall(
+                    self.template.decode(output.token_ids, **infer_streamers[i].decode_kwargs))
 
             choice = ChatCompletionResponseStreamChoice(
                 index=i,
@@ -593,7 +596,7 @@ def _create_chat_completion_response(
         choices = []
         for output in result.outputs:
             output.token_ids = list(output.token_ids)
-            response = self.template.decode(output.token_ids)
+            response = self.template.decode(output.token_ids, template_inputs=inputs['template_inputs'])
 
             # Extract reasoning content if reasoning_parser is enabled
             reasoning_content = None
@@ -743,7 +746,10 @@ def infer(
             if request_config.stream:
 
                 def _gen_wrapper():
-                    infer_streamers = [InferStreamer(self.template) for _ in range(generation_config.n)]
+                    infer_streamers = [
+                        InferStreamer(self.template, template_inputs=inputs['template_inputs'])
+                        for _ in range(generation_config.n)
+                    ]
                     token_idxs = [0 for _ in range(generation_config.n)]
                     while self.engine.has_unfinished_requests():
                         result = self.engine.step()
diff --git a/swift/template/base.py b/swift/template/base.py
index ead36f3e61..7dec3b1916 100644
--- a/swift/template/base.py
+++ b/swift/template/base.py
@@ -156,14 +156,14 @@ def __init__(
         if processor is not None:
             self.init_processor(processor)
 
-    def _get_enable_thinking(self, inputs):
-        enable_thinking = inputs.chat_template_kwargs.get('enable_thinking')
+    def _get_enable_thinking(self, inputs=None):
+        enable_thinking = None if inputs is None else inputs.chat_template_kwargs.get('enable_thinking')
         if enable_thinking is None:
             enable_thinking = self.enable_thinking
         return enable_thinking
 
-    def _get_response_prefix(self, inputs):
-        response_prefix = inputs.chat_template_kwargs.get('response_prefix')
+    def _get_response_prefix(self, inputs=None):
+        response_prefix = None if inputs is None else inputs.chat_template_kwargs.get('response_prefix')
         if response_prefix is None:
             response_prefix = self.response_prefix
         if response_prefix is not None:
@@ -317,8 +317,8 @@ def _preprocess_function_call(self, inputs: StdTemplateInputs) -> None:
     def prepare_engine_kwargs(self) -> Dict[str, Any]:
         return {}
 
-    def _get_max_pixels(self, inputs):
-        max_pixels = inputs.chat_template_kwargs.get('max_pixels')
+    def _get_max_pixels(self, inputs=None):
+        max_pixels = None if inputs is None else inputs.chat_template_kwargs.get('max_pixels')
         if max_pixels is None:
             max_pixels = self.max_pixels
         return max_pixels
@@ -683,13 +683,20 @@ def decode_seq_cls(self, logits: torch.Tensor, top_logprobs: int):
             logprobs = [self._get_seq_cls_logprobs(pred, logprobs[i], top_logprobs) for i, pred in enumerate(preds)]
         return preds, logprobs
 
-    def decode(self, generate_ids: List[int], *, is_finished: bool = True, first_token=True, **kwargs) -> Any:
+    def decode(self,
+               generate_ids: List[int],
+               *,
+               is_finished: bool = True,
+               first_token=True,
+               template_inputs=None,
+               **kwargs) -> Any:
         if kwargs.get('spaces_between_special_tokens') is None:
             kwargs['spaces_between_special_tokens'] = False
         generate_ids = self.skip_stop_tokens(generate_ids, is_finished)
         response = self.tokenizer.decode(generate_ids, **kwargs)
-        if first_token and self.response_prefix:
-            response = self.response_prefix + response
+        response_prefix = self._get_response_prefix(template_inputs)
+        if first_token and response_prefix:
+            response = response_prefix + response
         return response
 
     def decode_prm(self, input_ids: torch.Tensor, logits: torch.Tensor) -> Any:
diff --git a/swift/template/template_inputs.py b/swift/template/template_inputs.py
index 3fdc44b9b0..0bdeb7adab 100644
--- a/swift/template/template_inputs.py
+++ b/swift/template/template_inputs.py
@@ -59,12 +59,13 @@ def is_multimodal(self):
     def from_dict(cls, inputs: Dict[str, Any]) -> 'StdTemplateInputs':
         inputs = deepcopy(inputs)
         kwargs = {}
-        for key in ['label', 'channel', 'margin', 'rejected_response', 'chat_template_kwargs']:
+        for key in ['label', 'channel', 'margin', 'rejected_response']:
             if key in inputs:
                 kwargs[key] = inputs[key]
         messages = inputs['messages']
         tools = inputs.get('tools')
         objects = inputs.get('objects') or {}
+        chat_template_kwargs = inputs.get('chat_template_kwargs') or {}
 
         if messages and messages[0]['role'] == 'system':
             message = messages.pop(0)
@@ -98,6 +99,7 @@ def from_dict(cls, inputs: Dict[str, Any]) -> 'StdTemplateInputs':
             system=system,
             tools=tools,
             objects=objects,
+            chat_template_kwargs=chat_template_kwargs,
             extra_kwargs=extra_kwargs,
             **kwargs,
             **media_kwargs)

From 71e40135bb80b51202a6a0653f5df3e5c284d541 Mon Sep 17 00:00:00 2001
From: Jintao Huang <huangjintao.hjt@alibaba-inc.com>
Date: Fri, 8 May 2026 11:08:59 +0800
Subject: [PATCH 08/13] fix

---
 swift/infer_engine/protocol.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/swift/infer_engine/protocol.py b/swift/infer_engine/protocol.py
index 9f1b8cf821..9a59d6f195 100644
--- a/swift/infer_engine/protocol.py
+++ b/swift/infer_engine/protocol.py
@@ -234,6 +234,7 @@ class ChatCompletionRequestMixin:
     messages: Messages
     tools: Optional[List[Tool]] = None
     tool_choice: Optional[Union[str, Dict]] = None
+    chat_template_kwargs: Dict[str, Any] = field(default_factory=dict)
 
     def __post_init__(self):
         if self.tool_choice is None:

From 7258e05d942f9adac3bb7a2cfb0b623a2340b253 Mon Sep 17 00:00:00 2001
From: Jintao Huang <huangjintao.hjt@alibaba-inc.com>
Date: Fri, 8 May 2026 11:17:51 +0800
Subject: [PATCH 09/13] fix

---
 swift/template/templates/qwen.py | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/swift/template/templates/qwen.py b/swift/template/templates/qwen.py
index 34ef62c4f4..7494c6644a 100644
--- a/swift/template/templates/qwen.py
+++ b/swift/template/templates/qwen.py
@@ -306,7 +306,7 @@ def init_env_args(self):
         self.transformers_version = version.parse(transformers.__version__)
         self.bbox_format = get_env_args('QWENVL_BBOX_FORMAT', str, 'legacy')
 
-    def _get_max_pixels(self, inputs):
+    def _get_max_pixels(self, inputs=None):
         return self.max_pixels
 
     def replace_tag(self, media_type: Literal['image', 'video', 'audio'], index: int,
@@ -675,9 +675,6 @@ def init_processor(self, processor) -> None:
         self.use_audio_in_video = get_env_args('use_audio_in_video', bool, False)
         self.sampling_rate = get_env_args('sampling_rate', int, self.processor.feature_extractor.sampling_rate)
 
-    def _get_max_pixels(self, inputs):
-        return self.max_pixels
-
     def replace_tag(self, media_type: Literal['image', 'video', 'audio'], index: int,
                     inputs: StdTemplateInputs) -> List[Context]:
         from qwen_omni_utils import fetch_image, fetch_video

From 441f6e41e7269cb5fd54fb64dce81372336dda47 Mon Sep 17 00:00:00 2001
From: Jintao Huang <huangjintao.hjt@alibaba-inc.com>
Date: Fri, 8 May 2026 11:18:33 +0800
Subject: [PATCH 10/13] update

---
 docs/source/Customization/Custom-dataset.md    | 2 +-
 docs/source_en/Customization/Custom-dataset.md | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/docs/source/Customization/Custom-dataset.md b/docs/source/Customization/Custom-dataset.md
index ff3a40087d..679dad4fe3 100644
--- a/docs/source/Customization/Custom-dataset.md
+++ b/docs/source/Customization/Custom-dataset.md
@@ -94,7 +94,7 @@ print(inputs['loss_scale'])
 {"messages": [..., {"role": "tool_call", "content": "{\"name\": \"realtime_aqi\", \"arguments\": {\"city\": \"北京\"}}", "loss": false}, {"role": "tool_call", "content": "{\"name\": \"realtime_aqi\", \"arguments\": {\"city\": \"上海\"}}"}, ...]}
 ```
 
-- "chat_template_kwargs"字段，（需ms-swift>=4.2.1），你可以通过在数据集中传入该字段**样本级别**控制template的min_pixels, max_pixels, fps等多模态参数，以及enable_thinking（推理时）等参数。以下为不同模型支持的参数：
+- "chat_template_kwargs"字段，（需ms-swift>=4.3.0），你可以通过在数据集中传入该字段**样本级别**控制template的min_pixels, max_pixels, fps等多模态参数，以及enable_thinking（推理时）等参数。以下为不同模型支持的参数：
   - 其中"enable_thinking", "response_prefix"支持所有模型（推理时生效）；"max_pixels"参数支持所有多模态模型。
   - Qwen系列多模态模型：min_pixels, max_pixels, fps等qwen_vl_utils/qwen_omni_utils支持的参数。
 ```jsonl
diff --git a/docs/source_en/Customization/Custom-dataset.md b/docs/source_en/Customization/Custom-dataset.md
index 6e5c6f7788..530e98aedb 100644
--- a/docs/source_en/Customization/Custom-dataset.md
+++ b/docs/source_en/Customization/Custom-dataset.md
@@ -96,7 +96,7 @@ Note: If you set "loss"/"loss_scale" on consecutive "tool_call" messages in the
 {"messages": [..., {"role": "tool_call", "content": "{\"name\": \"realtime_aqi\", \"arguments\": {\"city\": \"Beijing\"}}", "loss": false}, {"role": "tool_call", "content": "{\"name\": \"realtime_aqi\", \"arguments\": {\"city\": \"Shanghai\"}}"}, ...]}
 ```
 
-- The "chat_template_kwargs" field (requires ms-swift>=4.2.1) allows you to control template multimodal parameters such as min_pixels, max_pixels, fps, as well as parameters like enable_thinking (during inference) at the **sample level** by passing this field in the dataset. The following parameters are supported by different models:
+- The "chat_template_kwargs" field (requires ms-swift>=4.3.0) allows you to control template multimodal parameters such as min_pixels, max_pixels, fps, as well as parameters like enable_thinking (during inference) at the **sample level** by passing this field in the dataset. The following parameters are supported by different models:
   - Among them, "enable_thinking" and "response_prefix" are supported by all models (takes effect during inference); the "max_pixels" parameter is supported by all multimodal models.
   - Qwen series multimodal models: parameters supported by qwen_vl_utils/qwen_omni_utils such as min_pixels, max_pixels, fps, etc.
 

From 0f0ceab2ca91a53a96c5bde96f26b270b10183ed Mon Sep 17 00:00:00 2001
From: Jintao Huang <huangjintao.hjt@alibaba-inc.com>
Date: Fri, 8 May 2026 11:42:19 +0800
Subject: [PATCH 11/13] fix

---
 .../deploy/client/llm/chat/openai_client.py   | 24 ++++++++++++++++---
 .../deploy/client/llm/chat/swift_client.py    |  7 +++---
 swift/template/templates/gemma.py             |  2 +-
 swift/template/templates/qwen.py              | 21 +++++++++++-----
 4 files changed, 40 insertions(+), 14 deletions(-)

diff --git a/examples/deploy/client/llm/chat/openai_client.py b/examples/deploy/client/llm/chat/openai_client.py
index bb3f208ba3..9a0472a9c4 100644
--- a/examples/deploy/client/llm/chat/openai_client.py
+++ b/examples/deploy/client/llm/chat/openai_client.py
@@ -6,7 +6,16 @@
 
 
 def infer(client, model: str, messages):
-    resp = client.chat.completions.create(model=model, messages=messages, max_tokens=512, temperature=0)
+    resp = client.chat.completions.create(
+        model=model,
+        messages=messages,
+        max_tokens=512,
+        temperature=0,
+        extra_body={
+            "chat_template_kwargs": {
+                "enable_thinking": False
+            },
+        })
     query = messages[0]['content']
     response = resp.choices[0].message.content
     print(f'query: {query}')
@@ -16,7 +25,16 @@ def infer(client, model: str, messages):
 
 # streaming
 def infer_stream(client, model: str, messages):
-    gen = client.chat.completions.create(model=model, messages=messages, stream=True, temperature=0)
+    gen = client.chat.completions.create(
+        model=model,
+        messages=messages,
+        stream=True,
+        temperature=0,
+        extra_body={
+            "chat_template_kwargs": {
+                "enable_thinking": False
+            },
+        })
     print(f'messages: {messages}\nresponse: ', end='')
     for chunk in gen:
         if chunk is None:
@@ -43,5 +61,5 @@ def run_client(host: str = '127.0.0.1', port: int = 8000):
 
 if __name__ == '__main__':
     from swift import DeployArguments, run_deploy
-    with run_deploy(DeployArguments(model='Qwen/Qwen2.5-1.5B-Instruct', verbose=False, log_interval=-1)) as port:
+    with run_deploy(DeployArguments(model='Qwen/Qwen3.5-4B', verbose=False, log_interval=-1)) as port:
         run_client(port=port)
diff --git a/examples/deploy/client/llm/chat/swift_client.py b/examples/deploy/client/llm/chat/swift_client.py
index 8e9fad399f..81deada23d 100644
--- a/examples/deploy/client/llm/chat/swift_client.py
+++ b/examples/deploy/client/llm/chat/swift_client.py
@@ -46,7 +46,7 @@ def run_client(host: str = '127.0.0.1', port: int = 8000):
     infer_batch(engine, infer_requests)
 
     messages = [{'role': 'user', 'content': 'who are you?'}]
-    infer_stream(engine, InferRequest(messages=messages))
+    infer_stream(engine, InferRequest(messages=messages, chat_template_kwargs={'enable_thinking': False}))
 
 
 if __name__ == '__main__':
@@ -54,7 +54,6 @@ def run_client(host: str = '127.0.0.1', port: int = 8000):
                        run_deploy)
 
     # NOTE: In a real deployment scenario, please comment out the context of run_deploy.
-    with run_deploy(
-            DeployArguments(model='Qwen/Qwen2.5-1.5B-Instruct', verbose=False, log_interval=-1,
-                            infer_backend='vllm')) as port:
+    with run_deploy(DeployArguments(model='Qwen/Qwen3.5-4B', verbose=False, log_interval=-1,
+                                    infer_backend='vllm')) as port:
         run_client(port=port)
diff --git a/swift/template/templates/gemma.py b/swift/template/templates/gemma.py
index 034d22bc91..49ac3dc243 100644
--- a/swift/template/templates/gemma.py
+++ b/swift/template/templates/gemma.py
@@ -265,7 +265,7 @@ def replace_tag(self, media_type: Literal['image', 'video', 'audio'], index: int
 
     def _get_system(self, inputs: StdTemplateInputs) -> Optional[str]:
         system = super()._get_system(inputs)
-        if self.enable_thinking:
+        if self._get_enable_thinking(inputs):
             system = '<|think|>\n' + (system or '')
         return system
 
diff --git a/swift/template/templates/qwen.py b/swift/template/templates/qwen.py
index 7494c6644a..891ad82502 100644
--- a/swift/template/templates/qwen.py
+++ b/swift/template/templates/qwen.py
@@ -270,10 +270,13 @@ def replace_tag(self, media_type: Literal['image', 'video', 'audio'], index: int
 
     def _encode(self, inputs: StdTemplateInputs) -> Dict[str, Any]:
         encoded = super()._encode(inputs)
+        sampling_rate = inputs.chat_template_kwargs.get('sampling_rate')
+        if sampling_rate is None:
+            sampling_rate = self.sampling_rate
         if inputs.audios:
-            audios = load_batch(inputs.audios, load_func=partial(load_audio, sampling_rate=self.sampling_rate))
+            audios = load_batch(inputs.audios, load_func=partial(load_audio, sampling_rate=sampling_rate))
             audio_inputs = self.processor.feature_extractor(
-                audios, sampling_rate=self.sampling_rate, return_attention_mask=True, return_tensors='pt')
+                audios, sampling_rate=sampling_rate, return_attention_mask=True, return_tensors='pt')
             audio_inputs['feature_attention_mask'] = audio_inputs.pop('attention_mask')
             encoded.update(audio_inputs)
         return encoded
@@ -679,6 +682,9 @@ def replace_tag(self, media_type: Literal['image', 'video', 'audio'], index: int
                     inputs: StdTemplateInputs) -> List[Context]:
         from qwen_omni_utils import fetch_image, fetch_video
         kwargs = {'image_patch_size': self.processor.image_processor.patch_size} if self.version == 'omni_v3' else {}
+        sampling_rate = inputs.chat_template_kwargs.get('sampling_rate')
+        if sampling_rate is None:
+            sampling_rate = self.sampling_rate
         if self.mode == 'vllm':
             # https://github.com/modelscope/ms-swift/issues/8445
             inputs.mm_processor_kwargs['do_resize'] = False
@@ -690,7 +696,7 @@ def replace_tag(self, media_type: Literal['image', 'video', 'audio'], index: int
                 return ['<|vision_start|><|image_pad|><|vision_end|>']
         elif media_type == 'audio':
             if self.mode != 'vllm':
-                inputs.audios[index] = load_audio(inputs.audios[index], self.sampling_rate)
+                inputs.audios[index] = load_audio(inputs.audios[index], sampling_rate)
             if self.version == 'omni_v2_5':
                 return ['<|audio_bos|><|AUDIO|><|audio_eos|>']
             elif self.version == 'omni_v3':
@@ -706,7 +712,7 @@ def replace_tag(self, media_type: Literal['image', 'video', 'audio'], index: int
                 if video.startswith('http://') or video.startswith('https://'):
                     import audioread
                     video = audioread.ffdec.FFmpegAudioFile(video)
-                video = librosa.load(video, sr=self.sampling_rate)[0]
+                video = librosa.load(video, sr=sampling_rate)[0]
                 if self.mode != 'vllm':
                     inputs.audios.insert(inputs.audio_idx, (video, 'video'))
                 else:
@@ -992,11 +998,14 @@ def replace_tag(self, media_type: Literal['image', 'video', 'audio'], index: int
 
     def _encode(self, inputs: StdTemplateInputs) -> Dict[str, Any]:
         encoded = super()._encode(inputs)
+        sampling_rate = inputs.chat_template_kwargs.get('sampling_rate')
+        if sampling_rate is None:
+            sampling_rate = self.sampling_rate
         if inputs.audios:
-            audios = load_batch(inputs.audios, load_func=partial(load_audio, sampling_rate=self.sampling_rate))
+            audios = load_batch(inputs.audios, load_func=partial(load_audio, sampling_rate=sampling_rate))
             audio_inputs = self.processor.feature_extractor(
                 audios,
-                sampling_rate=self.sampling_rate,
+                sampling_rate=sampling_rate,
                 return_attention_mask=True,
                 return_tensors='pt',
                 padding=True,

From 1bc40402e61d11077f336286e5f50f8834203c68 Mon Sep 17 00:00:00 2001
From: Jintao Huang <huangjintao.hjt@alibaba-inc.com>
Date: Fri, 8 May 2026 11:58:54 +0800
Subject: [PATCH 12/13] update

---
 swift/template/templates/llm.py | 13 +++++++++++--
 1 file changed, 11 insertions(+), 2 deletions(-)

diff --git a/swift/template/templates/llm.py b/swift/template/templates/llm.py
index 0c903e3958..bc0909734f 100644
--- a/swift/template/templates/llm.py
+++ b/swift/template/templates/llm.py
@@ -349,16 +349,25 @@ def init_env_args(self):
         self.enable_thinking = self.reasoning_effort != 'no_think'
         self.chat_template_kwargs = {'reasoning_effort': self.reasoning_effort}
 
+    def _get_enable_thinking(self, inputs=None):
+        reasoning_effort = None if inputs is None else inputs.chat_template_kwargs.get('reasoning_effort')
+        if reasoning_effort is not None:
+            return reasoning_effort != 'no_think'
+        return super()._get_enable_thinking(inputs)
+
     def _get_system(self, inputs):
         system = super()._get_system(inputs)
+        reasoning_effort = inputs.chat_template_kwargs.get('reasoning_effort')
+        if reasoning_effort is None:
+            reasoning_effort = self.reasoning_effort
         if inputs.tools:
             # For tool calls, append reasoning_mode after </tool_calls> in the tool instruction
             system = system.replace(
                 'you should print </tool_calls>',
-                f'you should print </tool_calls><｜reasoning_mode｜>reasoning_effort:{self.reasoning_effort}')
+                f'you should print </tool_calls><｜reasoning_mode｜>reasoning_effort:{reasoning_effort}')
         else:
             # For non-tool calls, append reasoning_mode to the system/prefix area
-            mode_str = f'<｜reasoning_mode｜>reasoning_effort:{self.reasoning_effort}'
+            mode_str = f'<｜reasoning_mode｜>reasoning_effort:{reasoning_effort}'
             system = (system or '') + mode_str
         return system
 

From e526e33364d2ade1be5aef2922086c60dc943c96 Mon Sep 17 00:00:00 2001
From: Jintao Huang <huangjintao.hjt@alibaba-inc.com>
Date: Fri, 8 May 2026 12:01:26 +0800
Subject: [PATCH 13/13] update

---
 examples/deploy/client/llm/chat/openai_client.py | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/examples/deploy/client/llm/chat/openai_client.py b/examples/deploy/client/llm/chat/openai_client.py
index 9a0472a9c4..667dd8115f 100644
--- a/examples/deploy/client/llm/chat/openai_client.py
+++ b/examples/deploy/client/llm/chat/openai_client.py
@@ -12,8 +12,8 @@ def infer(client, model: str, messages):
         max_tokens=512,
         temperature=0,
         extra_body={
-            "chat_template_kwargs": {
-                "enable_thinking": False
+            'chat_template_kwargs': {
+                'enable_thinking': False
             },
         })
     query = messages[0]['content']
@@ -31,8 +31,8 @@ def infer_stream(client, model: str, messages):
         stream=True,
         temperature=0,
         extra_body={
-            "chat_template_kwargs": {
-                "enable_thinking": False
+            'chat_template_kwargs': {
+                'enable_thinking': False
             },
         })
     print(f'messages: {messages}\nresponse: ', end='')