[test] add npu test yaml and add ascend a3 docker file (#9547)

Co-authored-by: jiaqiw09 <jiaqiw960714@gmail.com>

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"]

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"]

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",
     [

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(

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"]

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

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()

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 = (

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

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

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)