diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml
index c809f6d5..06a5c6c0 100644
--- a/.github/workflows/docker.yml
+++ b/.github/workflows/docker.yml
@@ -27,14 +27,18 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        device:
-          - "cuda"
-          - "npu"
+        include:
+          - device: "cuda"
+            npu_type: ""
+          - device: "npu"
+            npu_type: "a2"
+          - device: "npu"
+            npu_type: "a3"
 
     runs-on: ubuntu-latest
 
     concurrency:
-      group: ${{ github.workflow }}-${{ github.ref }}-${{ matrix.device }}
+      group: ${{ github.workflow }}-${{ github.ref }}-${{ matrix.device }}-${{ matrix.npu_type }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
 
     environment:
@@ -76,7 +80,7 @@ jobs:
           password: ${{ secrets.DOCKERHUB_TOKEN }}
 
       - name: Login to Quay
-        if: ${{ github.event_name != 'pull_request' && matrix.device == 'npu' }}
+        if: ${{ github.event_name != 'pull_request' && matrix.device == 'npu'}}
         uses: docker/login-action@v3
         with:
           registry: quay.io
@@ -97,8 +101,8 @@ jobs:
           cache-from: type=gha
           cache-to: type=gha,mode=max
 
-      - name: Build and push Docker image (NPU)
-        if: ${{ matrix.device == 'npu' }}
+      - name: Build and push Docker image (NPU-A2)
+        if: ${{ matrix.device == 'npu' && matrix.npu_type == 'a2' }}
         uses: docker/build-push-action@v6
         with:
           context: .
@@ -110,3 +114,19 @@ jobs:
             quay.io/ascend/llamafactory:${{ steps.version.outputs.tag }}-npu-a2
           cache-from: type=gha
           cache-to: type=gha,mode=max
+
+      - name: Build and push Docker image (NPU-A3)
+        if: ${{ matrix.device == 'npu' && matrix.npu_type == 'a3' }}
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          platforms: linux/amd64,linux/arm64
+          file: ./docker/docker-npu/Dockerfile
+          build-args: |
+            BASE_IMAGE=quay.io/ascend/cann:8.3.rc2-a3-ubuntu22.04-py3.11
+          push: ${{ github.event_name != 'pull_request' }}
+          tags: |
+            docker.io/hiyouga/llamafactory:${{ steps.version.outputs.tag }}-npu-a3
+            quay.io/ascend/llamafactory:${{ steps.version.outputs.tag }}-npu-a3
+          cache-from: type=gha
+          cache-to: type=gha,mode=max
diff --git a/.github/workflows/tests_npu.yml b/.github/workflows/tests_npu.yml
new file mode 100644
index 00000000..316ed2f7
--- /dev/null
+++ b/.github/workflows/tests_npu.yml
@@ -0,0 +1,87 @@
+name: tests_npu
+
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - "main"
+    paths:
+      - "**/*.py"
+      - "requirements.txt"
+      - "Makefile"
+      - ".github/workflows/*.yml"
+  pull_request:
+    branches:
+      - "main"
+    paths:
+      - "**/*.py"
+      - "requirements.txt"
+      - "Makefile"
+      - ".github/workflows/*.yml"
+
+jobs:
+  tests:
+    strategy:
+      fail-fast: false
+      matrix:
+        python:
+          - "3.11"
+        os:
+          - "linux-aarch64-a2-4"
+        pytorch_npu:
+          - "2.7.1"
+
+    runs-on: ${{ matrix.os }}
+
+    concurrency:
+      group: ${{ github.workflow }}-${{ github.ref }}-${{ matrix.os }}-${{ matrix.python }}
+      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
+
+    container:
+      image: ascendai/cann:8.3.rc2-910b-ubuntu22.04-py3.11
+      env:
+        HF_ENDPOINT: https://hf-mirror.com
+        HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        OS_NAME: ${{ matrix.os }}
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install ".[torch-npu,dev]" torch-npu==${{matrix.pytorch_npu}}
+
+      - name: Install node
+        run: |
+          apt-get update || true
+          apt-get install -y curl
+          curl -fsSL https://deb.nodesource.com/setup_20.x | bash -
+          apt-get install -y nodejs
+
+      - name: Cache files
+        id: hf-hub-cache
+        uses: actions/cache@v4
+        with:
+          path: ${{ runner.temp }}/huggingface
+          key: huggingface-${{ matrix.os }}-${{ matrix.python }}-${{ hashFiles('tests/version.txt') }}
+
+      - name: Check quality
+        run: |
+          make style && make quality
+
+      - name: Check license
+        run: |
+          make license
+
+      - name: Check build
+        run: |
+          make build
+
+      - name: Test with pytest
+        run: |
+          make test
+        env:
+          HF_HOME: /root/.cache/huggingface
+          HF_HUB_OFFLINE: "${{ steps.hf-hub-cache.outputs.cache-hit == 'true' && '1' || '0' }}"
\ No newline at end of file
diff --git a/docker/docker-npu/Dockerfile b/docker/docker-npu/Dockerfile
index 7539008b..825662b2 100644
--- a/docker/docker-npu/Dockerfile
+++ b/docker/docker-npu/Dockerfile
@@ -1,9 +1,6 @@
 # https://hub.docker.com/r/ascendai/cann/tags
 
-# default base image build for A2, if build for A3, using this image:
-# ARG BASE_IMAGE=ascendai/cann:8.3.rc1-a3-ubuntu22.04-py3.11
-
-ARG BASE_IMAGE=ascendai/cann:8.3.rc1-910b-ubuntu22.04-py3.11
+ARG BASE_IMAGE=quay.io/ascend/cann:8.3.rc2-910b-ubuntu22.04-py3.11
 FROM ${BASE_IMAGE}
 
 # Installation arguments
diff --git a/docker/docker-npu/docker-compose.yml b/docker/docker-npu/docker-compose.yml
index 659f8d1b..8530efaf 100644
--- a/docker/docker-npu/docker-compose.yml
+++ b/docker/docker-npu/docker-compose.yml
@@ -1,12 +1,13 @@
 services:
-  llamafactory:
+  llamafactory-a2:
     build:
       dockerfile: ./docker/docker-npu/Dockerfile
       context: ../..
       args:
         PIP_INDEX: https://pypi.org/simple
         EXTRAS: torch-npu,metrics
-    container_name: llamafactory
+    container_name: llamafactory-a2
+    image: llamafactory:npu-a2
     volumes:
       - /usr/local/dcmi:/usr/local/dcmi
       - /usr/local/bin/npu-smi:/usr/local/bin/npu-smi
@@ -26,3 +27,34 @@ services:
       - /dev/devmm_svm
       - /dev/hisi_hdc
     restart: unless-stopped
+
+  llamafactory-a3:
+    profiles: ["a3"]
+    build:
+      dockerfile: ./docker/docker-npu/Dockerfile
+      context: ../..
+      args:
+        BASE_IMAGE: quay.io/ascend/cann:8.3.rc2-a3-ubuntu22.04-py3.11
+        PIP_INDEX: https://pypi.org/simple
+        EXTRAS: torch-npu,metrics
+    container_name: llamafactory-a3
+    image: llamafactory:npu-a3
+    volumes:
+      - /usr/local/dcmi:/usr/local/dcmi
+      - /usr/local/bin/npu-smi:/usr/local/bin/npu-smi
+      - /usr/local/Ascend/driver:/usr/local/Ascend/driver
+      - /etc/ascend_install.info:/etc/ascend_install.info
+    ports:
+      - "7861:7860"
+      - "8001:8000"
+    ipc: host
+    tty: true
+    # shm_size: "16gb"  # ipc: host is set
+    stdin_open: true
+    command: bash
+    devices:
+      - /dev/davinci0
+      - /dev/davinci_manager
+      - /dev/devmm_svm
+      - /dev/hisi_hdc
+    restart: unless-stopped
diff --git a/tests/conftest.py b/tests/conftest.py
index f5811c9d..ddcaf22f 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -40,8 +40,38 @@ def pytest_configure(config):
     config.addinivalue_line(
         "markers", "require_device: test requires specific device, e.g., @pytest.mark.require_device('cuda')"
     )
+    config.addinivalue_line(
+        "markers", "runs_on: test requires specific device, e.g., @pytest.mark.runs_on(['cpu'])"
+    )
 
 
+def _handle_runs_on(items):
+    """Skip tests on specified devices based on runs_on marker.
+
+    Usage:
+        # Skip tests on specified devices
+        @pytest.mark.runs_on(['cpu'])
+        def test_something():
+            pass
+    """
+    for item in items:
+        runs_on_marker = item.get_closest_marker("runs_on")
+        if runs_on_marker:
+            runs_on_devices = runs_on_marker.args[0]
+
+            # Compatibility handling: Allow a single string instead of a list
+            # Example: @pytest.mark.("cpu")
+            if isinstance(runs_on_devices, str):
+                runs_on_devices = [runs_on_devices]
+
+
+            if CURRENT_DEVICE not in runs_on_devices:
+                item.add_marker(
+                    pytest.mark.skip(
+                        reason=f"test requires one of {runs_on_devices} (current: {CURRENT_DEVICE})"
+                    )
+                )
+
 def _handle_slow_tests(items):
     """Skip slow tests unless RUN_SLOW environment variable is set.
 
@@ -104,6 +134,7 @@ def pytest_collection_modifyitems(config, items):
     _handle_slow_tests(items)
     _handle_device_skips(items)
     _handle_device_requirements(items)
+    _handle_runs_on(items)
 
 
 @pytest.fixture
diff --git a/tests/data/processor/test_feedback.py b/tests/data/processor/test_feedback.py
index 355e7fe0..73b06675 100644
--- a/tests/data/processor/test_feedback.py
+++ b/tests/data/processor/test_feedback.py
@@ -42,6 +42,7 @@ TRAIN_ARGS = {
 }
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 @pytest.mark.parametrize("num_samples", [16])
 def test_feedback_data(num_samples: int):
     train_dataset = load_dataset_module(**TRAIN_ARGS)["train_dataset"]
diff --git a/tests/data/processor/test_pairwise.py b/tests/data/processor/test_pairwise.py
index 1040ba82..d3b8dbce 100644
--- a/tests/data/processor/test_pairwise.py
+++ b/tests/data/processor/test_pairwise.py
@@ -51,6 +51,7 @@ def _convert_sharegpt_to_openai(messages: list[dict[str, str]]) -> list[dict[str
     return new_messages
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.parametrize("num_samples", [16])
 def test_pairwise_data(num_samples: int):
     train_dataset = load_dataset_module(**TRAIN_ARGS)["train_dataset"]
diff --git a/tests/data/processor/test_processor_utils.py b/tests/data/processor/test_processor_utils.py
index e004cb06..256f5a6e 100644
--- a/tests/data/processor/test_processor_utils.py
+++ b/tests/data/processor/test_processor_utils.py
@@ -18,6 +18,7 @@ import pytest
 from llamafactory.data.processor.processor_utils import infer_seqlen
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.parametrize(
     "test_input,test_output",
     [
diff --git a/tests/data/processor/test_supervised.py b/tests/data/processor/test_supervised.py
index 6eaa34d3..903930b8 100644
--- a/tests/data/processor/test_supervised.py
+++ b/tests/data/processor/test_supervised.py
@@ -42,6 +42,7 @@ TRAIN_ARGS = {
 }
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.parametrize("num_samples", [16])
 def test_supervised_single_turn(num_samples: int):
     train_dataset = load_dataset_module(dataset_dir="ONLINE", dataset=TINY_DATA, **TRAIN_ARGS)["train_dataset"]
@@ -61,6 +62,7 @@ def test_supervised_single_turn(num_samples: int):
         assert train_dataset["input_ids"][index] == ref_input_ids
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.parametrize("num_samples", [8])
 def test_supervised_multi_turn(num_samples: int):
     train_dataset = load_dataset_module(dataset_dir="REMOTE:" + DEMO_DATA, dataset="system_chat", **TRAIN_ARGS)[
@@ -74,6 +76,7 @@ def test_supervised_multi_turn(num_samples: int):
         assert train_dataset["input_ids"][index] == ref_input_ids
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.parametrize("num_samples", [4])
 def test_supervised_train_on_prompt(num_samples: int):
     train_dataset = load_dataset_module(
@@ -88,6 +91,7 @@ def test_supervised_train_on_prompt(num_samples: int):
         assert train_dataset["labels"][index] == ref_ids
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.parametrize("num_samples", [4])
 def test_supervised_mask_history(num_samples: int):
     train_dataset = load_dataset_module(
diff --git a/tests/data/processor/test_unsupervised.py b/tests/data/processor/test_unsupervised.py
index 947b2e39..d9a9c9c4 100644
--- a/tests/data/processor/test_unsupervised.py
+++ b/tests/data/processor/test_unsupervised.py
@@ -45,6 +45,7 @@ TRAIN_ARGS = {
 }
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.parametrize("num_samples", [16])
 def test_unsupervised_data(num_samples: int):
     train_dataset = load_dataset_module(**TRAIN_ARGS)["train_dataset"]
diff --git a/tests/data/test_collator.py b/tests/data/test_collator.py
index 657f280d..047354ab 100644
--- a/tests/data/test_collator.py
+++ b/tests/data/test_collator.py
@@ -14,6 +14,7 @@
 
 import os
 
+import pytest
 import torch
 from PIL import Image
 from transformers import AutoConfig, AutoModelForVision2Seq
@@ -28,6 +29,7 @@ from llamafactory.model import load_tokenizer
 TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_base_collator():
     model_args, data_args, *_ = get_infer_args({"model_name_or_path": TINY_LLAMA3, "template": "default"})
     tokenizer_module = load_tokenizer(model_args)
@@ -71,6 +73,7 @@ def test_base_collator():
         assert batch_input[k].eq(torch.tensor(expected_input[k])).all()
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_multimodal_collator():
     model_args, data_args, *_ = get_infer_args(
         {"model_name_or_path": "Qwen/Qwen2-VL-2B-Instruct", "template": "qwen2_vl"}
@@ -126,6 +129,7 @@ def test_multimodal_collator():
         assert batch_input[k].eq(torch.tensor(expected_input[k])).all()
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_4d_attention_mask():
     o = 0.0
     x = torch.finfo(torch.float16).min
diff --git a/tests/data/test_converter.py b/tests/data/test_converter.py
index 6997f75f..23929c24 100644
--- a/tests/data/test_converter.py
+++ b/tests/data/test_converter.py
@@ -12,12 +12,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import pytest
+
 from llamafactory.data import Role
 from llamafactory.data.converter import get_dataset_converter
 from llamafactory.data.parser import DatasetAttr
 from llamafactory.hparams import DataArguments
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_alpaca_converter():
     dataset_attr = DatasetAttr("hf_hub", "llamafactory/tiny-supervised-dataset")
     data_args = DataArguments()
@@ -38,6 +41,7 @@ def test_alpaca_converter():
     }
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_sharegpt_converter():
     dataset_attr = DatasetAttr("hf_hub", "llamafactory/tiny-supervised-dataset")
     data_args = DataArguments()
diff --git a/tests/data/test_formatter.py b/tests/data/test_formatter.py
index 3ccb8bb3..c2da6dbf 100644
--- a/tests/data/test_formatter.py
+++ b/tests/data/test_formatter.py
@@ -15,6 +15,8 @@
 import json
 from datetime import datetime
 
+import pytest
+
 from llamafactory.data.formatter import EmptyFormatter, FunctionFormatter, StringFormatter, ToolFormatter
 
 
@@ -36,16 +38,19 @@ TOOLS = [
 ]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_empty_formatter():
     formatter = EmptyFormatter(slots=["\n"])
     assert formatter.apply() == ["\n"]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_string_formatter():
     formatter = StringFormatter(slots=["<s>", "Human: {{content}}\nAssistant:"])
     assert formatter.apply(content="Hi") == ["<s>", "Human: Hi\nAssistant:"]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_function_formatter():
     formatter = FunctionFormatter(slots=["{{content}}", "</s>"], tool_format="default")
     tool_calls = json.dumps(FUNCTION)
@@ -55,6 +60,7 @@ def test_function_formatter():
     ]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_multi_function_formatter():
     formatter = FunctionFormatter(slots=["{{content}}", "</s>"], tool_format="default")
     tool_calls = json.dumps([FUNCTION] * 2)
@@ -65,6 +71,7 @@ def test_multi_function_formatter():
     ]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_default_tool_formatter():
     formatter = ToolFormatter(tool_format="default")
     assert formatter.apply(content=json.dumps(TOOLS)) == [
@@ -83,12 +90,14 @@ def test_default_tool_formatter():
     ]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_default_tool_extractor():
     formatter = ToolFormatter(tool_format="default")
     result = """Action: test_tool\nAction Input: {"foo": "bar", "size": 10}"""
     assert formatter.extract(result) == [("test_tool", """{"foo": "bar", "size": 10}""")]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_default_multi_tool_extractor():
     formatter = ToolFormatter(tool_format="default")
     result = (
@@ -101,12 +110,14 @@ def test_default_multi_tool_extractor():
     ]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_glm4_function_formatter():
     formatter = FunctionFormatter(slots=["{{content}}"], tool_format="glm4")
     tool_calls = json.dumps(FUNCTION)
     assert formatter.apply(content=tool_calls) == ["""tool_name\n{"foo": "bar", "size": 10}"""]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_glm4_tool_formatter():
     formatter = ToolFormatter(tool_format="glm4")
     assert formatter.apply(content=json.dumps(TOOLS)) == [
@@ -117,12 +128,14 @@ def test_glm4_tool_formatter():
     ]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_glm4_tool_extractor():
     formatter = ToolFormatter(tool_format="glm4")
     result = """test_tool\n{"foo": "bar", "size": 10}\n"""
     assert formatter.extract(result) == [("test_tool", """{"foo": "bar", "size": 10}""")]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_llama3_function_formatter():
     formatter = FunctionFormatter(slots=["{{content}}<|eot_id|>"], tool_format="llama3")
     tool_calls = json.dumps(FUNCTION)
@@ -131,6 +144,7 @@ def test_llama3_function_formatter():
     ]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_llama3_multi_function_formatter():
     formatter = FunctionFormatter(slots=["{{content}}<|eot_id|>"], tool_format="llama3")
     tool_calls = json.dumps([FUNCTION] * 2)
@@ -141,6 +155,7 @@ def test_llama3_multi_function_formatter():
     ]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_llama3_tool_formatter():
     formatter = ToolFormatter(tool_format="llama3")
     date = datetime.now().strftime("%d %b %Y")
@@ -154,12 +169,14 @@ def test_llama3_tool_formatter():
     ]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_llama3_tool_extractor():
     formatter = ToolFormatter(tool_format="llama3")
     result = """{"name": "test_tool", "parameters": {"foo": "bar", "size": 10}}\n"""
     assert formatter.extract(result) == [("test_tool", """{"foo": "bar", "size": 10}""")]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_llama3_multi_tool_extractor():
     formatter = ToolFormatter(tool_format="llama3")
     result = (
@@ -172,6 +189,7 @@ def test_llama3_multi_tool_extractor():
     ]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_mistral_function_formatter():
     formatter = FunctionFormatter(slots=["[TOOL_CALLS] {{content}}", "</s>"], tool_format="mistral")
     tool_calls = json.dumps(FUNCTION)
@@ -181,6 +199,7 @@ def test_mistral_function_formatter():
     ]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_mistral_multi_function_formatter():
     formatter = FunctionFormatter(slots=["[TOOL_CALLS] {{content}}", "</s>"], tool_format="mistral")
     tool_calls = json.dumps([FUNCTION] * 2)
@@ -192,6 +211,7 @@ def test_mistral_multi_function_formatter():
     ]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_mistral_tool_formatter():
     formatter = ToolFormatter(tool_format="mistral")
     wrapped_tool = {"type": "function", "function": TOOLS[0]}
@@ -200,12 +220,14 @@ def test_mistral_tool_formatter():
     ]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_mistral_tool_extractor():
     formatter = ToolFormatter(tool_format="mistral")
     result = """{"name": "test_tool", "arguments": {"foo": "bar", "size": 10}}"""
     assert formatter.extract(result) == [("test_tool", """{"foo": "bar", "size": 10}""")]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_mistral_multi_tool_extractor():
     formatter = ToolFormatter(tool_format="mistral")
     result = (
@@ -218,6 +240,7 @@ def test_mistral_multi_tool_extractor():
     ]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_qwen_function_formatter():
     formatter = FunctionFormatter(slots=["{{content}}<|im_end|>\n"], tool_format="qwen")
     tool_calls = json.dumps(FUNCTION)
@@ -226,6 +249,7 @@ def test_qwen_function_formatter():
     ]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_qwen_multi_function_formatter():
     formatter = FunctionFormatter(slots=["{{content}}<|im_end|>\n"], tool_format="qwen")
     tool_calls = json.dumps([FUNCTION] * 2)
@@ -236,6 +260,7 @@ def test_qwen_multi_function_formatter():
     ]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_qwen_tool_formatter():
     formatter = ToolFormatter(tool_format="qwen")
     wrapped_tool = {"type": "function", "function": TOOLS[0]}
@@ -249,12 +274,14 @@ def test_qwen_tool_formatter():
     ]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_qwen_tool_extractor():
     formatter = ToolFormatter(tool_format="qwen")
     result = """<tool_call>\n{"name": "test_tool", "arguments": {"foo": "bar", "size": 10}}\n</tool_call>"""
     assert formatter.extract(result) == [("test_tool", """{"foo": "bar", "size": 10}""")]
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_qwen_multi_tool_extractor():
     formatter = ToolFormatter(tool_format="qwen")
     result = (
diff --git a/tests/data/test_loader.py b/tests/data/test_loader.py
index b45bdaad..9546cf4f 100644
--- a/tests/data/test_loader.py
+++ b/tests/data/test_loader.py
@@ -14,6 +14,8 @@
 
 import os
 
+import pytest
+
 from llamafactory.train.test_utils import load_dataset_module
 
 
@@ -38,18 +40,21 @@ TRAIN_ARGS = {
 }
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_load_train_only():
     dataset_module = load_dataset_module(**TRAIN_ARGS)
     assert dataset_module.get("train_dataset") is not None
     assert dataset_module.get("eval_dataset") is None
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_load_val_size():
     dataset_module = load_dataset_module(val_size=0.1, **TRAIN_ARGS)
     assert dataset_module.get("train_dataset") is not None
     assert dataset_module.get("eval_dataset") is not None
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_load_eval_data():
     dataset_module = load_dataset_module(eval_dataset=TINY_DATA, **TRAIN_ARGS)
     assert dataset_module.get("train_dataset") is not None
diff --git a/tests/data/test_mm_plugin.py b/tests/data/test_mm_plugin.py
index 4b702e83..6efc9e43 100644
--- a/tests/data/test_mm_plugin.py
+++ b/tests/data/test_mm_plugin.py
@@ -179,6 +179,7 @@ def _check_plugin(
     )
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_base_plugin():
     tokenizer_module = _load_tokenizer_module(model_name_or_path=TINY_LLAMA3)
     base_plugin = get_mm_plugin(name="base")
@@ -186,6 +187,7 @@ def test_base_plugin():
     _check_plugin(**check_inputs)
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
 @pytest.mark.skipif(not is_transformers_version_greater_than("4.50.0"), reason="Requires transformers>=4.50.0")
 def test_gemma3_plugin():
@@ -208,6 +210,7 @@ def test_gemma3_plugin():
     _check_plugin(**check_inputs)
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.skipif(not is_transformers_version_greater_than("4.52.0"), reason="Requires transformers>=4.52.0")
 def test_internvl_plugin():
     image_seqlen = 256
@@ -226,6 +229,7 @@ def test_internvl_plugin():
     _check_plugin(**check_inputs)
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.skipif(not is_transformers_version_greater_than("4.51.0"), reason="Requires transformers>=4.51.0")
 def test_llama4_plugin():
     tokenizer_module = _load_tokenizer_module(model_name_or_path=TINY_LLAMA4)
@@ -247,6 +251,7 @@ def test_llama4_plugin():
     _check_plugin(**check_inputs)
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_llava_plugin():
     image_seqlen = 576
     tokenizer_module = _load_tokenizer_module(model_name_or_path="llava-hf/llava-1.5-7b-hf")
@@ -260,6 +265,7 @@ def test_llava_plugin():
     _check_plugin(**check_inputs)
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_llava_next_plugin():
     image_seqlen = 1176
     tokenizer_module = _load_tokenizer_module(model_name_or_path="llava-hf/llava-v1.6-vicuna-7b-hf")
@@ -273,6 +279,7 @@ def test_llava_next_plugin():
     _check_plugin(**check_inputs)
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_llava_next_video_plugin():
     image_seqlen = 1176
     tokenizer_module = _load_tokenizer_module(model_name_or_path="llava-hf/LLaVA-NeXT-Video-7B-hf")
@@ -286,6 +293,7 @@ def test_llava_next_video_plugin():
     _check_plugin(**check_inputs)
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
 def test_paligemma_plugin():
     image_seqlen = 256
@@ -305,6 +313,7 @@ def test_paligemma_plugin():
     _check_plugin(**check_inputs)
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.skipif(not is_transformers_version_greater_than("4.50.0"), reason="Requires transformers>=4.50.0")
 def test_pixtral_plugin():
     image_slice_height, image_slice_width = 2, 2
@@ -327,6 +336,7 @@ def test_pixtral_plugin():
     _check_plugin(**check_inputs)
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.skipif(not is_transformers_version_greater_than("4.52.0"), reason="Requires transformers>=4.52.0")
 def test_qwen2_omni_plugin():
     image_seqlen, audio_seqlen = 4, 2
@@ -357,6 +367,7 @@ def test_qwen2_omni_plugin():
     _check_plugin(**check_inputs)
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_qwen2_vl_plugin():
     image_seqlen = 4
     tokenizer_module = _load_tokenizer_module(model_name_or_path="Qwen/Qwen2-VL-7B-Instruct")
@@ -373,6 +384,7 @@ def test_qwen2_vl_plugin():
     _check_plugin(**check_inputs)
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.skipif(not is_transformers_version_greater_than("4.57.0"), reason="Requires transformers>=4.57.0")
 def test_qwen3_vl_plugin():
     frame_seqlen = 1
@@ -394,6 +406,7 @@ def test_qwen3_vl_plugin():
     _check_plugin(**check_inputs)
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.skipif(not is_transformers_version_greater_than("4.47.0"), reason="Requires transformers>=4.47.0")
 def test_video_llava_plugin():
     image_seqlen = 256
diff --git a/tests/data/test_template.py b/tests/data/test_template.py
index dd2deca8..dc510172 100644
--- a/tests/data/test_template.py
+++ b/tests/data/test_template.py
@@ -89,6 +89,7 @@ def _check_template(
     _check_tokenization(tokenizer, (prompt_ids, answer_ids), (prompt_str, answer_str))
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.parametrize("use_fast", [True, False])
 def test_encode_oneturn(use_fast: bool):
     tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3, use_fast=use_fast)
@@ -104,6 +105,7 @@ def test_encode_oneturn(use_fast: bool):
     _check_tokenization(tokenizer, (prompt_ids, answer_ids), (prompt_str, answer_str))
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.parametrize("use_fast", [True, False])
 def test_encode_multiturn(use_fast: bool):
     tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3, use_fast=use_fast)
@@ -125,6 +127,7 @@ def test_encode_multiturn(use_fast: bool):
     )
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.parametrize("use_fast", [True, False])
 @pytest.mark.parametrize("cot_messages", [True, False])
 @pytest.mark.parametrize("enable_thinking", [True, False, None])
@@ -151,6 +154,7 @@ def test_reasoning_encode_oneturn(use_fast: bool, cot_messages: bool, enable_thi
     _check_tokenization(tokenizer, (prompt_ids, answer_ids), (prompt_str, answer_str))
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.parametrize("use_fast", [True, False])
 @pytest.mark.parametrize("cot_messages", [True, False])
 @pytest.mark.parametrize("enable_thinking", [True, False, None])
@@ -180,6 +184,7 @@ def test_reasoning_encode_multiturn(use_fast: bool, cot_messages: bool, enable_t
     )
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.parametrize("use_fast", [True, False])
 def test_jinja_template(use_fast: bool):
     tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3, use_fast=use_fast)
@@ -190,6 +195,7 @@ def test_jinja_template(use_fast: bool):
     assert tokenizer.apply_chat_template(MESSAGES) == ref_tokenizer.apply_chat_template(MESSAGES)
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_ollama_modelfile():
     tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
     template = get_template_and_fix_tokenizer(tokenizer, DataArguments(template="llama3"))
@@ -207,12 +213,14 @@ def test_ollama_modelfile():
     )
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_get_stop_token_ids():
     tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
     template = get_template_and_fix_tokenizer(tokenizer, DataArguments(template="llama3"))
     assert set(template.get_stop_token_ids(tokenizer)) == {128008, 128009}
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
 @pytest.mark.parametrize("use_fast", [True, False])
 def test_gemma_template(use_fast: bool):
@@ -226,6 +234,7 @@ def test_gemma_template(use_fast: bool):
     _check_template("google/gemma-3-4b-it", "gemma", prompt_str, answer_str, use_fast)
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
 @pytest.mark.parametrize("use_fast", [True, False])
 def test_gemma2_template(use_fast: bool):
@@ -239,6 +248,7 @@ def test_gemma2_template(use_fast: bool):
     _check_template("google/gemma-2-2b-it", "gemma2", prompt_str, answer_str, use_fast)
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
 @pytest.mark.parametrize("use_fast", [True, False])
 def test_llama3_template(use_fast: bool):
@@ -252,6 +262,7 @@ def test_llama3_template(use_fast: bool):
     _check_template("meta-llama/Meta-Llama-3-8B-Instruct", "llama3", prompt_str, answer_str, use_fast)
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.parametrize(
     "use_fast", [True, pytest.param(False, marks=pytest.mark.xfail(reason="Llama 4 has no slow tokenizer."))]
 )
@@ -273,6 +284,8 @@ def test_llama4_template(use_fast: bool):
         pytest.param(False, marks=pytest.mark.xfail(reason="Phi-4 slow tokenizer is broken.")),
     ],
 )
+
+@pytest.mark.runs_on(["cpu"])
 def test_phi4_template(use_fast: bool):
     prompt_str = (
         f"<|im_start|>user<|im_sep|>{MESSAGES[0]['content']}<|im_end|>"
@@ -284,6 +297,7 @@ def test_phi4_template(use_fast: bool):
     _check_template("microsoft/phi-4", "phi4", prompt_str, answer_str, use_fast)
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.xfail(not HF_TOKEN, reason="Authorization.")
 @pytest.mark.parametrize("use_fast", [True, False])
 def test_qwen2_5_template(use_fast: bool):
@@ -298,6 +312,7 @@ def test_qwen2_5_template(use_fast: bool):
     _check_template("Qwen/Qwen2.5-7B-Instruct", "qwen", prompt_str, answer_str, use_fast)
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.parametrize("use_fast", [True, False])
 @pytest.mark.parametrize("cot_messages", [True, False])
 def test_qwen3_template(use_fast: bool, cot_messages: bool):
@@ -317,6 +332,7 @@ def test_qwen3_template(use_fast: bool, cot_messages: bool):
     _check_template("Qwen/Qwen3-8B", "qwen3", prompt_str, answer_str, use_fast, messages=messages)
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_parse_llama3_template():
     tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3, token=HF_TOKEN)
     template = parse_template(tokenizer)
@@ -330,6 +346,7 @@ def test_parse_llama3_template():
     assert template.default_system == ""
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.xfail(not HF_TOKEN, reason="Authorization.")
 def test_parse_qwen_template():
     tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct", token=HF_TOKEN)
@@ -342,6 +359,7 @@ def test_parse_qwen_template():
     assert template.default_system == "You are Qwen, created by Alibaba Cloud. You are a helpful assistant."
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.xfail(not HF_TOKEN, reason="Authorization.")
 def test_parse_qwen3_template():
     tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B", token=HF_TOKEN)
diff --git a/tests/e2e/test_chat.py b/tests/e2e/test_chat.py
index 3221b6de..b05c7962 100644
--- a/tests/e2e/test_chat.py
+++ b/tests/e2e/test_chat.py
@@ -14,6 +14,8 @@
 
 import os
 
+import pytest
+
 from llamafactory.chat import ChatModel
 
 
@@ -35,11 +37,13 @@ MESSAGES = [
 EXPECTED_RESPONSE = "_rho"
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_chat():
     chat_model = ChatModel(INFER_ARGS)
     assert chat_model.chat(MESSAGES)[0].response_text == EXPECTED_RESPONSE
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_stream_chat():
     chat_model = ChatModel(INFER_ARGS)
     response = ""
diff --git a/tests/e2e/test_sglang.py b/tests/e2e/test_sglang.py
index de9a5c1c..8db1703a 100644
--- a/tests/e2e/test_sglang.py
+++ b/tests/e2e/test_sglang.py
@@ -39,6 +39,7 @@ MESSAGES = [
 ]
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.skipif(not is_sglang_available(), reason="SGLang is not installed")
 def test_chat():
     r"""Test the SGLang engine's basic chat functionality."""
@@ -48,6 +49,7 @@ def test_chat():
     print(response.response_text)
 
 
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.skipif(not is_sglang_available(), reason="SGLang is not installed")
 def test_stream_chat():
     r"""Test the SGLang engine's streaming chat functionality."""
diff --git a/tests/e2e/test_train.py b/tests/e2e/test_train.py
index 48f6f79d..0a2a1cba 100644
--- a/tests/e2e/test_train.py
+++ b/tests/e2e/test_train.py
@@ -48,7 +48,7 @@ INFER_ARGS = {
 
 OS_NAME = os.getenv("OS_NAME", "")
 
-
+@pytest.mark.runs_on(["cpu"])
 @pytest.mark.parametrize(
     "stage,dataset",
     [
@@ -65,6 +65,7 @@ def test_run_exp(stage: str, dataset: str):
     assert os.path.exists(output_dir)
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_export():
     export_dir = os.path.join("output", "llama3_export")
     export_model({"export_dir": export_dir, **INFER_ARGS})
diff --git a/tests/eval/test_eval_template.py b/tests/eval/test_eval_template.py
index eddc1640..6f61cc2d 100644
--- a/tests/eval/test_eval_template.py
+++ b/tests/eval/test_eval_template.py
@@ -12,9 +12,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import pytest
+
 from llamafactory.eval.template import get_eval_template
 
 
+@pytest.mark.runs_on(["cpu"])
 def test_eval_template_en():
     support_set = [
         {
@@ -52,7 +55,7 @@ def test_eval_template_en():
         {"role": "assistant", "content": "C"},
     ]
 
-
+@pytest.mark.runs_on(["cpu"])
 def test_eval_template_zh():
     support_set = [
         {
diff --git a/tests/model/model_utils/test_add_tokens.py b/tests/model/model_utils/test_add_tokens.py
index cb1c414a..4710819a 100644
--- a/tests/model/model_utils/test_add_tokens.py
+++ b/tests/model/model_utils/test_add_tokens.py
@@ -25,6 +25,7 @@ TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
 UNUSED_TOKEN = "<|UNUSED_TOKEN|>"
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 @pytest.mark.parametrize("special_tokens", [False, True])
 def test_add_tokens(special_tokens: bool):
     if special_tokens:
diff --git a/tests/model/model_utils/test_attention.py b/tests/model/model_utils/test_attention.py
index 0063630a..446d8063 100644
--- a/tests/model/model_utils/test_attention.py
+++ b/tests/model/model_utils/test_attention.py
@@ -29,6 +29,7 @@ INFER_ARGS = {
 }
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 @pytest.mark.xfail(is_transformers_version_greater_than("4.48"), reason="Attention refactor.")
 def test_attention():
     attention_available = ["disabled"]
diff --git a/tests/model/model_utils/test_checkpointing.py b/tests/model/model_utils/test_checkpointing.py
index 2402e6fb..63df0730 100644
--- a/tests/model/model_utils/test_checkpointing.py
+++ b/tests/model/model_utils/test_checkpointing.py
@@ -39,6 +39,7 @@ TRAIN_ARGS = {
 }
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 @pytest.mark.parametrize("disable_gradient_checkpointing", [False, True])
 def test_vanilla_checkpointing(disable_gradient_checkpointing: bool):
     model = load_train_model(disable_gradient_checkpointing=disable_gradient_checkpointing, **TRAIN_ARGS)
@@ -46,12 +47,14 @@ def test_vanilla_checkpointing(disable_gradient_checkpointing: bool):
         assert getattr(module, "gradient_checkpointing") != disable_gradient_checkpointing
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 def test_unsloth_gradient_checkpointing():
     model = load_train_model(use_unsloth_gc=True, **TRAIN_ARGS)
     for module in filter(lambda m: hasattr(m, "gradient_checkpointing"), model.modules()):
         assert module._gradient_checkpointing_func.__self__.__name__ == "UnslothGradientCheckpointing"
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 def test_upcast_layernorm():
     model = load_train_model(upcast_layernorm=True, **TRAIN_ARGS)
     for name, param in model.named_parameters():
@@ -59,6 +62,7 @@ def test_upcast_layernorm():
             assert param.dtype == torch.float32
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 def test_upcast_lmhead_output():
     model = load_train_model(upcast_lmhead_output=True, **TRAIN_ARGS)
     inputs = torch.randn((1, 16), dtype=torch.float16, device=get_current_device())
diff --git a/tests/model/model_utils/test_misc.py b/tests/model/model_utils/test_misc.py
index b2c8b3bf..537ae4f1 100644
--- a/tests/model/model_utils/test_misc.py
+++ b/tests/model/model_utils/test_misc.py
@@ -24,6 +24,7 @@ from llamafactory.model.model_utils.misc import find_expanded_modules
 HF_TOKEN = os.getenv("HF_TOKEN")
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 @pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
 def test_expanded_modules():
     config = AutoConfig.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct")
diff --git a/tests/model/model_utils/test_packing.py b/tests/model/model_utils/test_packing.py
index 81e0d66a..6dde0751 100644
--- a/tests/model/model_utils/test_packing.py
+++ b/tests/model/model_utils/test_packing.py
@@ -18,6 +18,7 @@ import torch
 from llamafactory.model.model_utils.packing import get_seqlens_in_batch, get_unpad_data
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 @pytest.mark.parametrize(
     "attention_mask,golden_seq_lens",
     [
diff --git a/tests/model/model_utils/test_visual.py b/tests/model/model_utils/test_visual.py
index fc53b69c..703bbb7f 100644
--- a/tests/model/model_utils/test_visual.py
+++ b/tests/model/model_utils/test_visual.py
@@ -23,6 +23,7 @@ from llamafactory.hparams import FinetuningArguments, ModelArguments
 from llamafactory.model.adapter import init_adapter
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 @pytest.mark.parametrize("freeze_vision_tower", (False, True))
 @pytest.mark.parametrize("freeze_multi_modal_projector", (False, True))
 @pytest.mark.parametrize("freeze_language_model", (False, True))
@@ -48,6 +49,7 @@ def test_visual_full(freeze_vision_tower: bool, freeze_multi_modal_projector: bo
             assert param.requires_grad != freeze_language_model
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 @pytest.mark.parametrize("freeze_vision_tower,freeze_language_model", ((False, False), (False, True), (True, False)))
 def test_visual_lora(freeze_vision_tower: bool, freeze_language_model: bool):
     model_args = ModelArguments(model_name_or_path="Qwen/Qwen2-VL-2B-Instruct")
@@ -80,6 +82,7 @@ def test_visual_lora(freeze_vision_tower: bool, freeze_language_model: bool):
     assert (merger_param_name in trainable_params) is False
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 def test_visual_model_save_load():
     # check VLM's state dict: https://github.com/huggingface/transformers/pull/38385
     model_args = ModelArguments(model_name_or_path="Qwen/Qwen2-VL-2B-Instruct")
diff --git a/tests/model/test_base.py b/tests/model/test_base.py
index 14afff63..382bea2f 100644
--- a/tests/model/test_base.py
+++ b/tests/model/test_base.py
@@ -29,13 +29,15 @@ INFER_ARGS = {
     "infer_dtype": "float16",
 }
 
-
+@pytest.mark.runs_on(["cpu","npu"])
+@pytest.mark.skip_on_devices("npu")
 def test_base():
     model = load_infer_model(**INFER_ARGS)
     ref_model = load_reference_model(TINY_LLAMA3)
     compare_model(model, ref_model)
 
-
+@pytest.mark.runs_on(["cpu","npu"])
+@pytest.mark.skip_on_devices("npu")
 @pytest.mark.usefixtures("fix_valuehead_cpu_loading")
 def test_valuehead():
     model = load_infer_model(add_valuehead=True, **INFER_ARGS)
diff --git a/tests/model/test_freeze.py b/tests/model/test_freeze.py
index b82ec88d..9d39ded1 100644
--- a/tests/model/test_freeze.py
+++ b/tests/model/test_freeze.py
@@ -14,6 +14,7 @@
 
 import os
 
+import pytest
 import torch
 
 from llamafactory.train.test_utils import load_infer_model, load_train_model
@@ -43,6 +44,7 @@ INFER_ARGS = {
 }
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 def test_freeze_train_all_modules():
     model = load_train_model(freeze_trainable_layers=1, **TRAIN_ARGS)
     for name, param in model.named_parameters():
@@ -54,6 +56,7 @@ def test_freeze_train_all_modules():
             assert param.dtype == torch.float16
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 def test_freeze_train_extra_modules():
     model = load_train_model(freeze_trainable_layers=1, freeze_extra_modules="embed_tokens,lm_head", **TRAIN_ARGS)
     for name, param in model.named_parameters():
@@ -65,6 +68,7 @@ def test_freeze_train_extra_modules():
             assert param.dtype == torch.float16
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 def test_freeze_inference():
     model = load_infer_model(**INFER_ARGS)
     for param in model.parameters():
diff --git a/tests/model/test_full.py b/tests/model/test_full.py
index 9058b6ac..3f55f1a0 100644
--- a/tests/model/test_full.py
+++ b/tests/model/test_full.py
@@ -14,6 +14,7 @@
 
 import os
 
+import pytest
 import torch
 
 from llamafactory.train.test_utils import load_infer_model, load_train_model
@@ -42,14 +43,14 @@ INFER_ARGS = {
     "infer_dtype": "float16",
 }
 
-
+@pytest.mark.runs_on(["cpu","npu"])
 def test_full_train():
     model = load_train_model(**TRAIN_ARGS)
     for param in model.parameters():
         assert param.requires_grad is True
         assert param.dtype == torch.float32
 
-
+@pytest.mark.runs_on(["cpu","npu"])
 def test_full_inference():
     model = load_infer_model(**INFER_ARGS)
     for param in model.parameters():
diff --git a/tests/model/test_lora.py b/tests/model/test_lora.py
index 38b6b505..3a855391 100644
--- a/tests/model/test_lora.py
+++ b/tests/model/test_lora.py
@@ -55,30 +55,35 @@ INFER_ARGS = {
 }
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 def test_lora_train_qv_modules():
     model = load_train_model(lora_target="q_proj,v_proj", **TRAIN_ARGS)
     linear_modules, _ = check_lora_model(model)
     assert linear_modules == {"q_proj", "v_proj"}
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 def test_lora_train_all_modules():
     model = load_train_model(lora_target="all", **TRAIN_ARGS)
     linear_modules, _ = check_lora_model(model)
     assert linear_modules == {"q_proj", "k_proj", "v_proj", "o_proj", "up_proj", "gate_proj", "down_proj"}
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 def test_lora_train_extra_modules():
     model = load_train_model(additional_target="embed_tokens,lm_head", **TRAIN_ARGS)
     _, extra_modules = check_lora_model(model)
     assert extra_modules == {"embed_tokens", "lm_head"}
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 def test_lora_train_old_adapters():
     model = load_train_model(adapter_name_or_path=TINY_LLAMA_ADAPTER, create_new_adapter=False, **TRAIN_ARGS)
     ref_model = load_reference_model(TINY_LLAMA3, TINY_LLAMA_ADAPTER, use_lora=True, is_trainable=True)
     compare_model(model, ref_model)
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 def test_lora_train_new_adapters():
     model = load_train_model(adapter_name_or_path=TINY_LLAMA_ADAPTER, create_new_adapter=True, **TRAIN_ARGS)
     ref_model = load_reference_model(TINY_LLAMA3, TINY_LLAMA_ADAPTER, use_lora=True, is_trainable=True)
@@ -87,6 +92,7 @@ def test_lora_train_new_adapters():
     )
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 @pytest.mark.usefixtures("fix_valuehead_cpu_loading")
 def test_lora_train_valuehead():
     model = load_train_model(add_valuehead=True, **TRAIN_ARGS)
@@ -96,7 +102,8 @@ def test_lora_train_valuehead():
     assert torch.allclose(state_dict["v_head.summary.weight"], ref_state_dict["v_head.summary.weight"])
     assert torch.allclose(state_dict["v_head.summary.bias"], ref_state_dict["v_head.summary.bias"])
 
-
+@pytest.mark.runs_on(["cpu","npu"])
+@pytest.mark.skip_on_devices("npu")
 def test_lora_inference():
     model = load_infer_model(**INFER_ARGS)
     ref_model = load_reference_model(TINY_LLAMA3, TINY_LLAMA_ADAPTER, use_lora=True).merge_and_unload()
diff --git a/tests/model/test_pissa.py b/tests/model/test_pissa.py
index 3b6101f8..6b830290 100644
--- a/tests/model/test_pissa.py
+++ b/tests/model/test_pissa.py
@@ -49,13 +49,14 @@ INFER_ARGS = {
 }
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 @pytest.mark.xfail(reason="PiSSA initialization is not stable in different platform.")
 def test_pissa_train():
     model = load_train_model(**TRAIN_ARGS)
     ref_model = load_reference_model(TINY_LLAMA_PISSA, TINY_LLAMA_PISSA, use_pissa=True, is_trainable=True)
     compare_model(model, ref_model)
 
-
+@pytest.mark.runs_on(["cpu","npu"])
 @pytest.mark.xfail(reason="Known connection error.")
 def test_pissa_inference():
     model = load_infer_model(**INFER_ARGS)
diff --git a/tests/train/test_sft_trainer.py b/tests/train/test_sft_trainer.py
index 9f6ebe41..1dd2c0e6 100644
--- a/tests/train/test_sft_trainer.py
+++ b/tests/train/test_sft_trainer.py
@@ -59,6 +59,7 @@ class DataCollatorWithVerbose(DataCollatorWithPadding):
         return {k: v[:, :1] for k, v in batch.items()}  # truncate input length
 
 
+@pytest.mark.runs_on(["cpu","npu"])
 @pytest.mark.parametrize("disable_shuffling", [False, True])
 def test_shuffle(disable_shuffling: bool):
     model_args, data_args, training_args, finetuning_args, _ = get_train_args(
diff --git a/tests/utils.py b/tests/utils.py
new file mode 100644
index 00000000..6ff92420
--- /dev/null
+++ b/tests/utils.py
@@ -0,0 +1,18 @@
+# Copyright 2025 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import pytest
+
+
+runs_on = pytest.mark.runs_on