diff --git a/docs/source/Instruction/Command-line-parameters.md b/docs/source/Instruction/Command-line-parameters.md
index 52b9baa44d..16064f04af 100644
--- a/docs/source/Instruction/Command-line-parameters.md
+++ b/docs/source/Instruction/Command-line-parameters.md
@@ -248,6 +248,7 @@ gradient_checkpointing: true
 - 🔥neftune_noise_alpha: neftune添加的噪声系数。默认为0，通常可以设置为5、10、15。
 - 🔥use_liger_kernel: 是否启用[Liger](https://github.com/linkedin/Liger-Kernel)内核加速训练并减少显存消耗。默认为False。示例shell参考[这里](https://github.com/modelscope/ms-swift/blob/main/examples/train/liger)。
   - 注意：liger_kernel不支持device_map，请使用DDP/DeepSpeed进行多卡训练。liger_kernel目前只支持`task_type='causal_lm'`。
+- use_cce: 是否启用[cut-cross-entropy](https://github.com/apple/ml-cross-entropy)融合算子降低显存并加速训练。默认为False。示例shell参考[这里](https://github.com/modelscope/ms-swift/blob/main/examples/train/cce)。
 - average_tokens_across_devices: 是否在设备之间进行token数平均。如果设置为True，将使用all_reduce同步`num_tokens_in_batch`以进行精确的损失计算。默认为False。
 - max_grad_norm: 梯度裁剪。默认为1.。
   - 注意：日志中的grad_norm记录的是裁剪前的值。
diff --git a/docs/source_en/Instruction/Command-line-parameters.md b/docs/source_en/Instruction/Command-line-parameters.md
index 52176fe193..44b571d959 100644
--- a/docs/source_en/Instruction/Command-line-parameters.md
+++ b/docs/source_en/Instruction/Command-line-parameters.md
@@ -250,6 +250,7 @@ Other important parameters:
 - 🔥neftune_noise_alpha: Noise magnitude for NEFTune. Default is 0. Common values: 5, 10, 15.
 - 🔥use_liger_kernel: Whether to enable the [Liger](https://github.com/linkedin/Liger-Kernel) kernel to accelerate training and reduce GPU memory consumption. Defaults to False. Example shell script can be found [here](https://github.com/modelscope/ms-swift/blob/main/examples/train/liger).
   - Note: Liger kernel does not support `device_map`. Use DDP or DeepSpeed for multi-GPU training. Currently, liger_kernel only supports `task_type='causal_lm'`.
+- use_cce: Whether to enable the [cut-cross-entropy](https://github.com/apple/ml-cross-entropy) fused operator to reduce GPU memory usage and accelerate training. Defaults to `False`. Example shell script can be found [here](https://github.com/modelscope/ms-swift/blob/main/examples/train/cce).
 - average_tokens_across_devices: Whether to average token counts across devices. If `True`, `num_tokens_in_batch` is synchronized via `all_reduce` for accurate loss computation. Default is `False`.
 - max_grad_norm: Gradient clipping. Default is 1.
   - Note: The logged `grad_norm` reflects the value **before** clipping.
diff --git a/examples/train/cce/sft.sh b/examples/train/cce/sft.sh
new file mode 100644
index 0000000000..5c34b74db3
--- /dev/null
+++ b/examples/train/cce/sft.sh
@@ -0,0 +1,17 @@
+# test env: 1 * A10
+# Using use_cce: 2.62GB
+# Not using use_cce: 16.24G
+
+# Install CCE dependency
+pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@f643b88"
+
+# Run ms-swift (example)
+swift sft \
+  --model Qwen/Qwen2.5-0.5B-Instruct \
+  --dataset gsm8k#1024 \
+  --train_type lora \
+  --per_device_train_batch_size 64 \
+  --per_device_eval_batch_size 64 \
+  --use_hf true \
+  --use_cce true \
+  "$@"
diff --git a/swift/llm/train/tuner.py b/swift/llm/train/tuner.py
index 286a9f4b04..e18c8af3aa 100644
--- a/swift/llm/train/tuner.py
+++ b/swift/llm/train/tuner.py
@@ -86,6 +86,76 @@ def apply_liger(model_type: str):
                           'by running `pip install -U liger-kernel`')
 
 
+def apply_cce(model_type: str):
+    try:
+        from cut_cross_entropy.transformers import cce_patch
+        from swift.llm import ModelType
+    except ImportError:
+        raise ImportError('Please upgrade cut-cross-entropy to apply cce kernels to this model '
+                          'by running `pip install "cut-cross-entropy[transformers] @ '
+                          'git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@f643b88"`')
+
+    model_type_map = {
+        # llama family
+        ModelType.llama: 'llama',
+        ModelType.llama3: 'llama',
+        ModelType.llama3_1: 'llama',
+        ModelType.llama3_2: 'llama',
+        ModelType.llama4: 'llama4',
+        ModelType.llama3_2_vision: 'mllama',
+        # mistral & mixtral family
+        ModelType.mistral: 'mistral',
+        ModelType.mixtral: 'mixtral',
+        # phi
+        ModelType.phi3: 'phi3',
+        # gemma family
+        ModelType.gemma: 'gemma',
+        ModelType.gemma2: 'gemma2',
+        ModelType.gemma3_text: 'gemma3_text',
+        ModelType.gemma3_vision: 'gemma3',
+        ModelType.gemma3n: 'gemma3n',
+        # glm4 family
+        ModelType.glm4: 'glm4',
+        ModelType.glm4_0414: 'glm4',
+        ModelType.glm4_5: 'glm4_moe',
+        ModelType.glm4_z1_rumination: 'glm4_moe',
+        ModelType.glm4v: 'glm4v',
+        ModelType.glm4_1v: 'glm4v',
+        ModelType.glm4_5v: 'glm4v_moe',
+        # llava
+        ModelType.llava1_5_hf: 'llava',
+        ModelType.llava_llama3_hf: 'llava',
+        # qwen2 family
+        ModelType.qwen2: 'qwen2',
+        ModelType.qwen2_5: 'qwen2',
+        ModelType.qwen2_vl: 'qwen2_vl',
+        ModelType.qwen2_5_vl: 'qwen2_5_vl',
+        # qwen3 family
+        ModelType.qwen3: 'qwen3',
+        ModelType.qwen3_guard: 'qwen3',
+        ModelType.qwen3_thinking: 'qwen3',
+        ModelType.qwen3_nothinking: 'qwen3',
+        ModelType.qwen3_coder: 'qwen3',
+        ModelType.qwen3_moe: 'qwen3_moe',
+        ModelType.qwen3_moe_thinking: 'qwen3_moe',
+        ModelType.qwen3_next: 'qwen3_next',
+        ModelType.qwen3_next_thinking: 'qwen3_next',
+        ModelType.qwen3_vl: 'qwen3_vl',
+        ModelType.qwen3_moe_vl: 'qwen3_vl_moe',
+        # deepseek family
+        ModelType.deepseek_r1: 'deepseek_v3',
+        ModelType.deepseek_v3_1: 'deepseek_v3',
+    }
+
+    cce_model_type = model_type_map.get(model_type)
+    if cce_model_type:
+        cce_patch(cce_model_type)
+        return
+
+    supported_models = ', '.join(sorted(set(model_type_map.values())))
+    raise ValueError(f'Unsupported cce model_type: {model_type}. Supported types: {supported_models}')
+
+
 def get_multimodal_target_regex(
     model,
     *,
@@ -375,6 +445,9 @@ def prepare_model(cls, args, model, *, template=None, train_dataset=None, task_t
             # Apply liger
             apply_liger(args.model_type)
 
+        if args.use_cce and 'use_cce' not in inspect.signature(TrainingArguments).parameters:
+            apply_cce(args.model_type)
+
         if args.is_adapter:
             if args.tuner_backend != 'unsloth' and args.train_type not in extra_tuners:
                 # Fix the name of the layer in xcomposer that contains Plora.
diff --git a/swift/trainers/arguments.py b/swift/trainers/arguments.py
index f356439bd8..a8f69072c0 100644
--- a/swift/trainers/arguments.py
+++ b/swift/trainers/arguments.py
@@ -9,7 +9,7 @@
 from transformers.training_args_seq2seq import Seq2SeqTrainingArguments as HfSeq2SeqTrainingArguments
 
 from swift.plugin import loss_mapping
-from swift.utils import get_dist_setting, get_logger, is_liger_available, is_mp, json_parse_to_dict
+from swift.utils import get_dist_setting, get_logger, is_cce_available, is_liger_available, is_mp, json_parse_to_dict
 from .optimizers.galore import GaLoreConfig
 
 logger = get_logger()
@@ -53,6 +53,7 @@ class TrainArgumentsMixin:
         dataloader_prefetch_factor (Optional[int]): The number of batches loaded in advance by each worker. Defaults
             to None.
         use_liger_kernel (bool): Whether to use the Liger kernel for optimization. Defaults to False.
+        use_cce (bool): Whether to use ml-cross-entropy fused kernels for optimization. Defaults to False.
         check_model (bool): If True, checks local model files for corruption or modification and provides a warning.
             Should be set to False in an offline environment. Defaults to True.
         acc_strategy (Literal['token', 'seq']): The strategy for calculating accuracy during training and validation.
@@ -115,6 +116,7 @@ class TrainArgumentsMixin:
     dataloader_persistent_workers: bool = False
     dataloader_prefetch_factor: Optional[int] = None
     use_liger_kernel: bool = False
+    use_cce: bool = False
 
     # extra
     check_model: bool = True
@@ -163,11 +165,20 @@ def _init_liger(self):
             except Exception:
                 pass
 
+    def _init_cce(self):
+        if self.use_cce:
+            assert is_cce_available(), ('use_cce requires cut-cross-entropy, try '
+                                        '`pip install "cut-cross-entropy[transformers] @ '
+                                        'git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@f643b88"`')
+
     def __post_init__(self):
         if is_mp() and self.use_liger_kernel:
             raise ValueError('liger_kernel does not support device_map. '
                              'Please use DDP/DeepSpeed for multi-GPU training.')
 
+        if self.use_cce and self.use_liger_kernel:
+            logger.warning('Enabling both use_cce and use_liger_kernel may lead to duplicated kernel patches.')
+
         if self.optimizer is None and (self.vit_lr is not None or self.aligner_lr is not None):
             self.optimizer = 'multimodal'
         if self.gradient_accumulation_steps is None:
@@ -181,6 +192,7 @@ def __post_init__(self):
         if self.gradient_checkpointing_kwargs:
             self.gradient_checkpointing_kwargs = json_parse_to_dict(self.gradient_checkpointing_kwargs)
         self._init_liger()
+        self._init_cce()
         if self.dataloader_num_workers is None:
             if platform.system() == 'Windows':
                 self.dataloader_num_workers = 0
diff --git a/swift/trainers/trainers.py b/swift/trainers/trainers.py
index 74cd9ad481..b4d346171e 100644
--- a/swift/trainers/trainers.py
+++ b/swift/trainers/trainers.py
@@ -318,9 +318,9 @@ def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=N
         if (self.label_smoother is not None or compute_loss_func is not None or loss_scale is not None
                 or self.args.enable_dft_loss or self.args.enable_channel_loss
                 or self.template.sequence_parallel_size > 1) and 'labels' in inputs:
-            if self.args.use_liger_kernel:
-                logger.warning_once('The cross_entropy loss function defined in Liger Kernel will not '
-                                    'take effect, potentially leading to increased GPU memory consumption.')
+            if self.args.use_liger_kernel or getattr(self.args, 'use_cce', False):
+                logger.warning_once('The cross_entropy loss function defined in Liger Kernel or ml-cross-entropy will '
+                                    'not take effect, potentially leading to increased GPU memory consumption.')
             labels = inputs.pop('labels')
         outputs = model(**inputs)
         if getattr(outputs, 'aux_loss', None) is not None:
diff --git a/swift/ui/llm_grpo/llm_grpo.py b/swift/ui/llm_grpo/llm_grpo.py
index 0cd13462fe..ff6256e4ee 100644
--- a/swift/ui/llm_grpo/llm_grpo.py
+++ b/swift/ui/llm_grpo/llm_grpo.py
@@ -157,6 +157,16 @@ class LLMGRPO(LLMTrain):
                 'en': 'Liger kernel can reduce memory usage'
             }
         },
+        'use_cce': {
+            'label': {
+                'zh': '使用CCE加速',
+                'en': 'Use CCE acceleration'
+            },
+            'info': {
+                'zh': 'CCE (ml-cross-entropy) 提供融合的交叉熵算子',
+                'en': 'CCE (ml-cross-entropy) provides fused cross-entropy kernels'
+            }
+        },
         'sequence_parallel_size': {
             'label': {
                 'zh': '序列并行大小',
@@ -233,6 +243,7 @@ def do_build_ui(cls, base_tab: Type['BaseUI']):
                         gr.Textbox(elem_id='seed', scale=4)
                         gr.Dropdown(elem_id='torch_dtype', scale=4)
                         gr.Checkbox(elem_id='use_liger_kernel', scale=4)
+                        gr.Checkbox(elem_id='use_cce', scale=4)
                         gr.Textbox(elem_id='sequence_parallel_size', lines=1, scale=4)
                     with gr.Row():
                         gr.Dropdown(
diff --git a/swift/ui/llm_rlhf/llm_rlhf.py b/swift/ui/llm_rlhf/llm_rlhf.py
index d7f1f740c7..1b8c028fce 100644
--- a/swift/ui/llm_rlhf/llm_rlhf.py
+++ b/swift/ui/llm_rlhf/llm_rlhf.py
@@ -170,6 +170,16 @@ class LLMRLHF(LLMTrain):
                 'en': 'Liger kernel can reduce memory usage'
             }
         },
+        'use_cce': {
+            'label': {
+                'zh': '使用CCE加速',
+                'en': 'Use CCE acceleration'
+            },
+            'info': {
+                'zh': 'CCE (ml-cross-entropy) 提供融合的交叉熵算子',
+                'en': 'CCE (ml-cross-entropy) provides fused cross-entropy kernels'
+            }
+        },
         'sequence_parallel_size': {
             'label': {
                 'zh': '序列并行大小',
@@ -246,6 +256,7 @@ def do_build_ui(cls, base_tab: Type['BaseUI']):
                         gr.Textbox(elem_id='seed', scale=2)
                         gr.Dropdown(elem_id='torch_dtype', scale=2)
                         gr.Checkbox(elem_id='use_liger_kernel', scale=2)
+                        gr.Checkbox(elem_id='use_cce', scale=2)
                     with gr.Row():
                         gr.Dropdown(
                             elem_id='gpu_id',
diff --git a/swift/ui/llm_train/llm_train.py b/swift/ui/llm_train/llm_train.py
index b1b2b94f6d..a6c1c7dcce 100644
--- a/swift/ui/llm_train/llm_train.py
+++ b/swift/ui/llm_train/llm_train.py
@@ -177,6 +177,16 @@ class LLMTrain(BaseUI):
                 'en': 'Liger kernel can reduce memory usage'
             }
         },
+        'use_cce': {
+            'label': {
+                'zh': '使用CCE加速',
+                'en': 'Use CCE acceleration'
+            },
+            'info': {
+                'zh': 'CCE (ml-cross-entropy) 提供融合的交叉熵算子',
+                'en': 'CCE (ml-cross-entropy) provides fused cross-entropy kernels'
+            }
+        },
         'sequence_parallel_size': {
             'label': {
                 'zh': '序列并行大小',
@@ -257,6 +267,7 @@ def do_build_ui(cls, base_tab: Type['BaseUI']):
                         gr.Textbox(elem_id='seed', scale=4)
                         gr.Dropdown(elem_id='torch_dtype', scale=4)
                         gr.Checkbox(elem_id='use_liger_kernel', scale=4)
+                        gr.Checkbox(elem_id='use_cce', scale=4)
                     with gr.Row():
                         gr.Dropdown(
                             elem_id='gpu_id',
@@ -390,6 +401,9 @@ def train(cls, *args):
         use_liger_kernel = kwargs.get('use_liger_kernel', None)
         if use_liger_kernel:
             kwargs.pop('use_liger_kernel')
+        use_cce = kwargs.get('use_cce', None)
+        if use_cce:
+            kwargs.pop('use_cce')
         if other_kwargs.get('use_muon'):
             kwargs['use_muon'] = other_kwargs.pop('use_muon')
 
@@ -428,6 +442,9 @@ def train(cls, *args):
         if use_liger_kernel:
             params += f'--use_liger_kernel {cls.quote}{use_liger_kernel}{cls.quote} '
             command.extend(['--use_liger_kernel', f'{use_liger_kernel}'])
+        if use_cce:
+            params += f'--use_cce {cls.quote}{use_cce}{cls.quote} '
+            command.extend(['--use_cce', f'{use_cce}'])
         if use_muon:
             params += f'--optimizer {cls.quote}muon{cls.quote} '
             command.extend(['--optimizer', 'muon'])
diff --git a/swift/utils/__init__.py b/swift/utils/__init__.py
index 4d9216bb70..e2ff7b4ab4 100644
--- a/swift/utils/__init__.py
+++ b/swift/utils/__init__.py
@@ -2,7 +2,7 @@
 
 from .env import (get_dist_setting, get_hf_endpoint, get_node_setting, get_pai_tensorboard_dir, is_deepspeed_enabled,
                   is_dist, is_last_rank, is_local_master, is_master, is_mp, is_mp_ddp, is_pai_training_job, use_hf_hub)
-from .import_utils import (is_flash_attn_2_available, is_flash_attn_3_available, is_liger_available,
+from .import_utils import (is_cce_available, is_flash_attn_2_available, is_flash_attn_3_available, is_liger_available,
                            is_lmdeploy_available, is_megatron_available, is_swanlab_available, is_trl_available,
                            is_unsloth_available, is_vllm_ascend_available, is_vllm_available, is_wandb_available)
 from .io_utils import JsonlWriter, append_to_jsonl, download_ms_file, get_file_mm_type, read_from_jsonl, write_to_jsonl
diff --git a/swift/utils/import_utils.py b/swift/utils/import_utils.py
index 95e9ba1f47..fc46160877 100644
--- a/swift/utils/import_utils.py
+++ b/swift/utils/import_utils.py
@@ -28,6 +28,10 @@ def is_liger_available():
     return importlib.util.find_spec('liger_kernel') is not None
 
 
+def is_cce_available():
+    return importlib.util.find_spec('cut_cross_entropy') is not None
+
+
 def is_swanlab_available():
     return importlib.util.find_spec('swanlab') is not None
 
diff --git a/tests/train/test_cce.py b/tests/train/test_cce.py
new file mode 100644
index 0000000000..728aa2f76c
--- /dev/null
+++ b/tests/train/test_cce.py
@@ -0,0 +1,29 @@
+import os
+
+# os.environ['CUDA_VISIBLE_DEVICES'] = '0'
+# os.environ['HF_ENDPOINT'] = 'https://hf-mirror.com'
+
+kwargs = {
+    'per_device_train_batch_size': 64,
+    'save_steps': 30,
+    'gradient_accumulation_steps': 2,
+    'num_train_epochs': 1,
+}
+
+
+def test_sft():
+    from swift.llm import sft_main, TrainArguments, infer_main, InferArguments
+    result = sft_main(
+        TrainArguments(
+            model='Qwen/Qwen2.5-0.5B-Instruct',
+            dataset=['gsm8k#1024'],
+            split_dataset_ratio=0.01,
+            use_cce=True,
+            # use_liger_kernel=True,
+            **kwargs))
+    last_model_checkpoint = result['last_model_checkpoint']
+    infer_main(InferArguments(adapters=last_model_checkpoint, load_data_args=True))
+
+
+if __name__ == '__main__':
+    test_sft()