diff --git a/tests/data/processor/test_feedback.py b/tests/data/processor/test_feedback.py
index f2c69d08a26..f384d8dac05 100644
--- a/tests/data/processor/test_feedback.py
+++ b/tests/data/processor/test_feedback.py
@@ -43,7 +43,7 @@
}
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@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 17d56098455..628c25c63de 100644
--- a/tests/data/processor/test_pairwise.py
+++ b/tests/data/processor/test_pairwise.py
@@ -52,7 +52,7 @@ def _convert_sharegpt_to_openai(messages: list[dict[str, str]]) -> list[dict[str
return new_messages
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@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 a2a3b7ebe05..97d18d4b334 100644
--- a/tests/data/processor/test_processor_utils.py
+++ b/tests/data/processor/test_processor_utils.py
@@ -18,7 +18,7 @@
from llamafactory.data.processor.processor_utils import infer_seqlen
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@pytest.mark.parametrize(
"test_input,test_output",
[
diff --git a/tests/data/processor/test_supervised.py b/tests/data/processor/test_supervised.py
index f515852e198..edc7f75e3d8 100644
--- a/tests/data/processor/test_supervised.py
+++ b/tests/data/processor/test_supervised.py
@@ -43,7 +43,7 @@
}
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@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"]
@@ -72,7 +72,7 @@ def test_supervised_single_turn(num_samples: int):
assert train_dataset["labels"][index] == ref_label_ids
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@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)[
@@ -90,7 +90,7 @@ def test_supervised_multi_turn(num_samples: int):
assert train_dataset["input_ids"][index] == ref_input_ids
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@pytest.mark.parametrize("num_samples", [4])
def test_supervised_train_on_prompt(num_samples: int):
train_dataset = load_dataset_module(
@@ -108,7 +108,7 @@ def test_supervised_train_on_prompt(num_samples: int):
assert train_dataset["labels"][index] == ref_input_ids
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@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 2d06219fc58..ef527ad2ef1 100644
--- a/tests/data/processor/test_unsupervised.py
+++ b/tests/data/processor/test_unsupervised.py
@@ -47,7 +47,7 @@
}
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@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 0cc7d7bd4a4..4c801fde627 100644
--- a/tests/data/test_collator.py
+++ b/tests/data/test_collator.py
@@ -32,7 +32,7 @@
TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
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)
@@ -76,7 +76,7 @@ def test_base_collator():
assert batch_input[k].eq(torch.tensor(expected_input[k])).all()
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_multimodal_collator():
model_args, data_args, *_ = get_infer_args(
{"model_name_or_path": "Qwen/Qwen2-VL-2B-Instruct", "template": "qwen2_vl"}
@@ -270,7 +270,7 @@ def _get_expected_position_ids(
return torch.cat(all_position_ids, dim=-1)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_multimodal_collator_with_packing():
model_args, data_args, *_ = get_infer_args(
{"model_name_or_path": "Qwen/Qwen2-VL-2B-Instruct", "template": "qwen2_vl"}
@@ -322,7 +322,7 @@ def test_multimodal_collator_with_packing():
assert batch_input["position_ids"][1:, :, :valid_len].eq(expected_position_ids).all()
-@pytest.mark.runs_on(["cpu"])
+@pytest.mark.runs_on(["cpu", "xpu"])
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 6b411aed53c..0ceb5a661a4 100644
--- a/tests/data/test_converter.py
+++ b/tests/data/test_converter.py
@@ -20,7 +20,7 @@
from llamafactory.hparams import DataArguments
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_alpaca_converter():
dataset_attr = DatasetAttr("hf_hub", "llamafactory/tiny-supervised-dataset")
data_args = DataArguments()
@@ -41,7 +41,7 @@ def test_alpaca_converter():
}
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
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 3aaa6f991da..9b054c29c80 100644
--- a/tests/data/test_formatter.py
+++ b/tests/data/test_formatter.py
@@ -38,19 +38,19 @@
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_empty_formatter():
formatter = EmptyFormatter(slots=["\n"])
assert formatter.apply() == ["\n"]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_string_formatter():
formatter = StringFormatter(slots=["", "Human: {{content}}\nAssistant:"])
assert formatter.apply(content="Hi") == ["", "Human: Hi\nAssistant:"]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_function_formatter():
formatter = FunctionFormatter(slots=["{{content}}", ""], tool_format="default")
tool_calls = json.dumps(FUNCTION)
@@ -60,7 +60,7 @@ def test_function_formatter():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_multi_function_formatter():
formatter = FunctionFormatter(slots=["{{content}}", ""], tool_format="default")
tool_calls = json.dumps([FUNCTION] * 2)
@@ -71,7 +71,7 @@ def test_multi_function_formatter():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_default_tool_formatter():
formatter = ToolFormatter(tool_format="default")
assert formatter.apply(content=json.dumps(TOOLS)) == [
@@ -90,14 +90,14 @@ def test_default_tool_formatter():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
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", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_default_multi_tool_extractor():
formatter = ToolFormatter(tool_format="default")
result = (
@@ -110,14 +110,14 @@ def test_default_multi_tool_extractor():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
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", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_glm4_tool_formatter():
formatter = ToolFormatter(tool_format="glm4")
assert formatter.apply(content=json.dumps(TOOLS)) == [
@@ -128,14 +128,14 @@ def test_glm4_tool_formatter():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
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", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_llama3_function_formatter():
formatter = FunctionFormatter(slots=["{{content}}<|eot_id|>"], tool_format="llama3")
tool_calls = json.dumps(FUNCTION)
@@ -144,7 +144,7 @@ def test_llama3_function_formatter():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_llama3_multi_function_formatter():
formatter = FunctionFormatter(slots=["{{content}}<|eot_id|>"], tool_format="llama3")
tool_calls = json.dumps([FUNCTION] * 2)
@@ -155,7 +155,7 @@ def test_llama3_multi_function_formatter():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_llama3_tool_formatter():
formatter = ToolFormatter(tool_format="llama3")
date = datetime.now().strftime("%d %b %Y")
@@ -169,14 +169,14 @@ def test_llama3_tool_formatter():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
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", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_llama3_multi_tool_extractor():
formatter = ToolFormatter(tool_format="llama3")
result = (
@@ -189,7 +189,7 @@ def test_llama3_multi_tool_extractor():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_mistral_function_formatter():
formatter = FunctionFormatter(slots=["[TOOL_CALLS] {{content}}", ""], tool_format="mistral")
tool_calls = json.dumps(FUNCTION)
@@ -199,7 +199,7 @@ def test_mistral_function_formatter():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_mistral_multi_function_formatter():
formatter = FunctionFormatter(slots=["[TOOL_CALLS] {{content}}", ""], tool_format="mistral")
tool_calls = json.dumps([FUNCTION] * 2)
@@ -211,7 +211,7 @@ def test_mistral_multi_function_formatter():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_mistral_tool_formatter():
formatter = ToolFormatter(tool_format="mistral")
wrapped_tool = {"type": "function", "function": TOOLS[0]}
@@ -220,14 +220,14 @@ def test_mistral_tool_formatter():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
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", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_mistral_multi_tool_extractor():
formatter = ToolFormatter(tool_format="mistral")
result = (
@@ -240,7 +240,7 @@ def test_mistral_multi_tool_extractor():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_qwen_function_formatter():
formatter = FunctionFormatter(slots=["{{content}}<|im_end|>\n"], tool_format="qwen")
tool_calls = json.dumps(FUNCTION)
@@ -249,7 +249,7 @@ def test_qwen_function_formatter():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_qwen_multi_function_formatter():
formatter = FunctionFormatter(slots=["{{content}}<|im_end|>\n"], tool_format="qwen")
tool_calls = json.dumps([FUNCTION] * 2)
@@ -260,7 +260,7 @@ def test_qwen_multi_function_formatter():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_qwen_tool_formatter():
formatter = ToolFormatter(tool_format="qwen")
wrapped_tool = {"type": "function", "function": TOOLS[0]}
@@ -274,14 +274,14 @@ def test_qwen_tool_formatter():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_qwen_tool_extractor():
formatter = ToolFormatter(tool_format="qwen")
result = """\n{"name": "test_tool", "arguments": {"foo": "bar", "size": 10}}\n"""
assert formatter.extract(result) == [("test_tool", """{"foo": "bar", "size": 10}""")]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_qwen_multi_tool_extractor():
formatter = ToolFormatter(tool_format="qwen")
result = (
@@ -294,7 +294,7 @@ def test_qwen_multi_tool_extractor():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_lfm2_function_formatter():
formatter = FunctionFormatter(slots=["{{content}}<|im_end|>\n"], tool_format="lfm2")
tool_calls = json.dumps(FUNCTION)
@@ -303,7 +303,7 @@ def test_lfm2_function_formatter():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_lfm2_multi_function_formatter():
formatter = FunctionFormatter(slots=["{{content}}<|im_end|>\n"], tool_format="lfm2")
tool_calls = json.dumps([FUNCTION] * 2)
@@ -313,7 +313,7 @@ def test_lfm2_multi_function_formatter():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_lfm2_tool_formatter():
formatter = ToolFormatter(tool_format="lfm2")
assert formatter.apply(content=json.dumps(TOOLS)) == [
@@ -321,14 +321,14 @@ def test_lfm2_tool_formatter():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_lfm2_tool_extractor():
formatter = ToolFormatter(tool_format="lfm2")
result = """<|tool_call_start|>[test_tool(foo="bar", size=10)]<|tool_call_end|>"""
assert formatter.extract(result) == [("test_tool", """{"foo": "bar", "size": 10}""")]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_lfm2_multi_tool_extractor():
formatter = ToolFormatter(tool_format="lfm2")
result = """<|tool_call_start|>[test_tool(foo="bar", size=10), another_tool(foo="job", size=2)]<|tool_call_end|>"""
@@ -338,7 +338,7 @@ def test_lfm2_multi_tool_extractor():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_lfm2_tool_extractor_with_nested_dict():
formatter = ToolFormatter(tool_format="lfm2")
result = """<|tool_call_start|>[search(query="test", options={"limit": 10, "offset": 0})]<|tool_call_end|>"""
@@ -350,7 +350,7 @@ def test_lfm2_tool_extractor_with_nested_dict():
assert args["options"] == {"limit": 10, "offset": 0}
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_lfm2_tool_extractor_with_list_arg():
formatter = ToolFormatter(tool_format="lfm2")
result = """<|tool_call_start|>[batch_process(items=[1, 2, 3], enabled=True)]<|tool_call_end|>"""
@@ -362,7 +362,7 @@ def test_lfm2_tool_extractor_with_list_arg():
assert args["enabled"] is True
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_lfm2_tool_extractor_no_match():
formatter = ToolFormatter(tool_format="lfm2")
result = "This is a regular response without tool calls."
@@ -370,7 +370,7 @@ def test_lfm2_tool_extractor_no_match():
assert extracted == result
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_lfm2_tool_round_trip():
formatter = FunctionFormatter(slots=["{{content}}"], tool_format="lfm2")
tool_formatter = ToolFormatter(tool_format="lfm2")
diff --git a/tests/data/test_loader.py b/tests/data/test_loader.py
index 907bda347d1..f589cc85858 100644
--- a/tests/data/test_loader.py
+++ b/tests/data/test_loader.py
@@ -40,21 +40,21 @@
}
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
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", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
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", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
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 e63c866db8b..3ba2f37c463 100644
--- a/tests/data/test_mm_plugin.py
+++ b/tests/data/test_mm_plugin.py
@@ -179,7 +179,7 @@ def _check_plugin(
)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_base_plugin():
tokenizer_module = _load_tokenizer_module(model_name_or_path=TINY_LLAMA3)
base_plugin = get_mm_plugin(name="base")
@@ -187,7 +187,7 @@ def test_base_plugin():
_check_plugin(**check_inputs)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@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():
@@ -210,7 +210,7 @@ def test_gemma3_plugin():
_check_plugin(**check_inputs)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@pytest.mark.skipif(not is_transformers_version_greater_than("5.6.0"), reason="Requires transformers>=5.6.0")
def test_gemma4_plugin():
tokenizer_module = _load_tokenizer_module(model_name_or_path="google/gemma-4-31B-it")
@@ -243,7 +243,7 @@ def test_gemma4_plugin():
_check_plugin(**check_inputs)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@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
@@ -262,7 +262,7 @@ def test_internvl_plugin():
_check_plugin(**check_inputs)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@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)
@@ -284,7 +284,7 @@ def test_llama4_plugin():
_check_plugin(**check_inputs)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_llava_plugin():
image_seqlen = 576
tokenizer_module = _load_tokenizer_module(model_name_or_path="llava-hf/llava-1.5-7b-hf")
@@ -298,7 +298,7 @@ def test_llava_plugin():
_check_plugin(**check_inputs)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
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")
@@ -312,7 +312,7 @@ def test_llava_next_plugin():
_check_plugin(**check_inputs)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
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")
@@ -326,7 +326,7 @@ def test_llava_next_video_plugin():
_check_plugin(**check_inputs)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
def test_paligemma_plugin():
image_seqlen = 256
@@ -346,7 +346,7 @@ def test_paligemma_plugin():
_check_plugin(**check_inputs)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@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
@@ -369,7 +369,7 @@ def test_pixtral_plugin():
_check_plugin(**check_inputs)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@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
@@ -400,7 +400,7 @@ def test_qwen2_omni_plugin():
_check_plugin(**check_inputs)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_qwen2_vl_plugin():
image_seqlen = 4
tokenizer_module = _load_tokenizer_module(model_name_or_path="Qwen/Qwen2-VL-7B-Instruct")
@@ -417,7 +417,7 @@ def test_qwen2_vl_plugin():
_check_plugin(**check_inputs)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@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
@@ -439,7 +439,7 @@ def test_qwen3_vl_plugin():
_check_plugin(**check_inputs)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@pytest.mark.skipif(not is_transformers_version_greater_than("4.57.0"), reason="Requires transformers>=4.57.0")
@pytest.mark.skipif(not is_pyav_available(), reason="Requires pyav")
def test_qwen3_vl_plugin_video_path():
@@ -488,7 +488,7 @@ def test_video_llava_plugin():
_check_plugin(**check_inputs)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_lfm2_vl_plugin():
"""Test LFM2.5-VL plugin instantiation."""
# Test plugin can be instantiated with correct tokens
diff --git a/tests/data/test_template.py b/tests/data/test_template.py
index ddd7b17da8e..8069524b2a9 100644
--- a/tests/data/test_template.py
+++ b/tests/data/test_template.py
@@ -91,7 +91,7 @@ def _check_template(
_check_tokenization(tokenizer, (prompt_ids, answer_ids), (prompt_str, answer_str))
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_encode_oneturn():
tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
template = get_template_and_fix_tokenizer(tokenizer, DataArguments(template="llama3"))
@@ -106,7 +106,7 @@ def test_encode_oneturn():
_check_tokenization(tokenizer, (prompt_ids, answer_ids), (prompt_str, answer_str))
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_encode_multiturn():
tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
template = get_template_and_fix_tokenizer(tokenizer, DataArguments(template="llama3"))
@@ -127,7 +127,7 @@ def test_encode_multiturn():
)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@pytest.mark.parametrize("cot_messages", [True, False])
@pytest.mark.parametrize("enable_thinking", [True, False, None])
def test_reasoning_encode_oneturn(cot_messages: bool, enable_thinking: bool):
@@ -153,7 +153,7 @@ def test_reasoning_encode_oneturn(cot_messages: bool, enable_thinking: bool):
_check_tokenization(tokenizer, (prompt_ids, answer_ids), (prompt_str, answer_str))
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@pytest.mark.parametrize("cot_messages", [True, False])
@pytest.mark.parametrize("enable_thinking", [True, False, None])
def test_reasoning_encode_multiturn(cot_messages: bool, enable_thinking: bool):
@@ -182,7 +182,7 @@ def test_reasoning_encode_multiturn(cot_messages: bool, enable_thinking: bool):
)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@pytest.mark.parametrize("enable_thinking", [True, False, None])
@pytest.mark.parametrize("discarding_history_cot", [True, False])
def test_reasoning_encode_multiturn_discarding_history_cot(enable_thinking: bool, discarding_history_cot: bool):
@@ -218,7 +218,7 @@ def test_reasoning_encode_multiturn_discarding_history_cot(enable_thinking: bool
)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_jinja_template():
tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
@@ -228,7 +228,7 @@ def test_jinja_template():
assert tokenizer.apply_chat_template(MESSAGES) == ref_tokenizer.apply_chat_template(MESSAGES)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_ollama_modelfile():
tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
template = get_template_and_fix_tokenizer(tokenizer, DataArguments(template="llama3"))
@@ -246,14 +246,14 @@ def test_ollama_modelfile():
)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
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", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
def test_gemma_template():
prompt_str = (
@@ -266,7 +266,7 @@ def test_gemma_template():
_check_template("google/gemma-3-4b-it", "gemma", prompt_str, answer_str)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
def test_gemma2_template():
prompt_str = (
@@ -279,7 +279,7 @@ def test_gemma2_template():
_check_template("google/gemma-2-2b-it", "gemma2", prompt_str, answer_str)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
def test_llama3_template():
prompt_str = (
@@ -292,7 +292,7 @@ def test_llama3_template():
_check_template("meta-llama/Meta-Llama-3-8B-Instruct", "llama3", prompt_str, answer_str)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_llama4_template():
prompt_str = (
f"<|begin_of_text|><|header_start|>user<|header_end|>\n\n{MESSAGES[0]['content']}<|eot|>"
@@ -304,7 +304,7 @@ def test_llama4_template():
_check_template(TINY_LLAMA4, "llama4", prompt_str, answer_str)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_phi4_template():
prompt_str = (
f"<|im_start|>user<|im_sep|>{MESSAGES[0]['content']}<|im_end|>"
@@ -316,7 +316,7 @@ def test_phi4_template():
_check_template("microsoft/phi-4", "phi4", prompt_str, answer_str)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@pytest.mark.xfail(not HF_TOKEN, reason="Authorization.")
def test_qwen2_5_template():
prompt_str = (
@@ -330,7 +330,7 @@ def test_qwen2_5_template():
_check_template("Qwen/Qwen2.5-7B-Instruct", "qwen", prompt_str, answer_str)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@pytest.mark.parametrize("cot_messages", [True, False])
def test_qwen3_template(cot_messages: bool):
prompt_str = (
@@ -349,7 +349,7 @@ def test_qwen3_template(cot_messages: bool):
_check_template("Qwen/Qwen3-8B", "qwen3", prompt_str, answer_str, messages=messages)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_parse_llama3_template():
tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
template = parse_template(tokenizer)
@@ -363,7 +363,7 @@ def test_parse_llama3_template():
assert template.default_system == ""
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@pytest.mark.xfail(not HF_TOKEN, reason="Authorization.")
def test_parse_qwen_template():
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
@@ -376,7 +376,7 @@ 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", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@pytest.mark.xfail(not HF_TOKEN, reason="Authorization.")
def test_parse_qwen3_template():
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")
diff --git a/tests/e2e/test_chat.py b/tests/e2e/test_chat.py
index e33f32c5646..3ba41c97941 100644
--- a/tests/e2e/test_chat.py
+++ b/tests/e2e/test_chat.py
@@ -37,13 +37,13 @@
EXPECTED_RESPONSE = "_rho"
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_chat():
chat_model = ChatModel(INFER_ARGS)
assert chat_model.chat(MESSAGES)[0].response_text == EXPECTED_RESPONSE
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_stream_chat():
chat_model = ChatModel(INFER_ARGS)
response = ""
diff --git a/tests/e2e/test_train.py b/tests/e2e/test_train.py
index dabb888b9f0..16b7002ef79 100644
--- a/tests/e2e/test_train.py
+++ b/tests/e2e/test_train.py
@@ -49,7 +49,7 @@
OS_NAME = os.getenv("OS_NAME", "")
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
@pytest.mark.parametrize(
"stage,dataset",
[
@@ -66,7 +66,7 @@ def test_run_exp(stage: str, dataset: str):
assert os.path.exists(output_dir)
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
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 783d0b9e3d5..d0a477c09a9 100644
--- a/tests/eval/test_eval_template.py
+++ b/tests/eval/test_eval_template.py
@@ -17,7 +17,7 @@
from llamafactory.eval.template import get_eval_template
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_eval_template_en():
support_set = [
{
@@ -56,7 +56,7 @@ def test_eval_template_en():
]
-@pytest.mark.runs_on(["cpu", "mps"])
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
def test_eval_template_zh():
support_set = [
{