423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-12-27 09:10:35 +08:00
Code Issues Packages Projects Releases Wiki Activity

[misc] fix accelerator (#9661)

Browse Source 
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
This commit is contained in:
Yaowei Zheng
2025-12-25 02:11:04 +08:00
committed by GitHub
parent 6a2eafbae3
commit a754604c11
44 changed files with 396 additions and 448 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tests/data/processor/test_feedback.py
View File

@@ -42,7 +42,7 @@ TRAIN_ARGS = {
}
@pytest.mark.runs_on(["cpu", "npu", "cuda"])
@pytest.mark.runs_on(["cpu", "mps"])
@pytest.mark.parametrize("num_samples", [16])
def test_feedback_data(num_samples: int):
train_dataset = load_dataset_module(**TRAIN_ARGS)["train_dataset"]

2
tests/data/processor/test_pairwise.py
View File

@@ -51,7 +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.runs_on(["cpu", "mps"])
@pytest.mark.parametrize("num_samples", [16])
def test_pairwise_data(num_samples: int):
train_dataset = load_dataset_module(**TRAIN_ARGS)["train_dataset"]

2
tests/data/processor/test_processor_utils.py
View File

@@ -18,7 +18,7 @@ import pytest
from llamafactory.data.processor.processor_utils import infer_seqlen
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
@pytest.mark.parametrize(
"test_input,test_output",
[

8
tests/data/processor/test_supervised.py
View File

@@ -42,7 +42,7 @@ TRAIN_ARGS = {
}
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
@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"]
@@ -62,7 +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.runs_on(["cpu", "mps"])
@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)[
@@ -76,7 +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.runs_on(["cpu", "mps"])
@pytest.mark.parametrize("num_samples", [4])
def test_supervised_train_on_prompt(num_samples: int):
train_dataset = load_dataset_module(
@@ -91,7 +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.runs_on(["cpu", "mps"])
@pytest.mark.parametrize("num_samples", [4])
def test_supervised_mask_history(num_samples: int):
train_dataset = load_dataset_module(

2
tests/data/processor/test_unsupervised.py
View File

@@ -46,7 +46,7 @@ TRAIN_ARGS = {
}
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
@pytest.mark.parametrize("num_samples", [16])
def test_unsupervised_data(num_samples: int):
train_dataset = load_dataset_module(**TRAIN_ARGS)["train_dataset"]

4
tests/data/test_collator.py
View File

@@ -29,7 +29,7 @@ from llamafactory.model import load_tokenizer
TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
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)
@@ -73,7 +73,7 @@ def test_base_collator():
assert batch_input[k].eq(torch.tensor(expected_input[k])).all()
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
def test_multimodal_collator():
model_args, data_args, *_ = get_infer_args(
{"model_name_or_path": "Qwen/Qwen2-VL-2B-Instruct", "template": "qwen2_vl"}

4
tests/data/test_converter.py
View File

@@ -20,7 +20,7 @@ from llamafactory.data.parser import DatasetAttr
from llamafactory.hparams import DataArguments
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
def test_sharegpt_converter():
dataset_attr = DatasetAttr("hf_hub", "llamafactory/tiny-supervised-dataset")
data_args = DataArguments()

50
tests/data/test_formatter.py
View File

@@ -38,19 +38,19 @@ TOOLS = [
]
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
def test_empty_formatter():
formatter = EmptyFormatter(slots=["\n"])
assert formatter.apply() == ["\n"]
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
def test_function_formatter():
formatter = FunctionFormatter(slots=["{{content}}", "</s>"], tool_format="default")
tool_calls = json.dumps(FUNCTION)
@@ -60,7 +60,7 @@ def test_function_formatter():
]
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
def test_multi_function_formatter():
formatter = FunctionFormatter(slots=["{{content}}", "</s>"], tool_format="default")
tool_calls = json.dumps([FUNCTION] * 2)
@@ -71,7 +71,7 @@ def test_multi_function_formatter():
]
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
def test_mistral_function_formatter():
formatter = FunctionFormatter(slots=["[TOOL_CALLS] {{content}}", "</s>"], tool_format="mistral")
tool_calls = json.dumps(FUNCTION)
@@ -199,7 +199,7 @@ def test_mistral_function_formatter():
]
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
def test_mistral_multi_function_formatter():
formatter = FunctionFormatter(slots=["[TOOL_CALLS] {{content}}", "</s>"], tool_format="mistral")
tool_calls = json.dumps([FUNCTION] * 2)
@@ -211,7 +211,7 @@ def test_mistral_multi_function_formatter():
]
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
def test_qwen_multi_tool_extractor():
formatter = ToolFormatter(tool_format="qwen")
result = (

6
tests/data/test_loader.py
View File

@@ -40,21 +40,21 @@ TRAIN_ARGS = {
}
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
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

26
tests/data/test_mm_plugin.py
View File

@@ -179,7 +179,7 @@ def _check_plugin(
)
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
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"])
@pytest.mark.runs_on(["cpu", "mps"])
@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"])
@pytest.mark.runs_on(["cpu", "mps"])
@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
@@ -229,7 +229,7 @@ def test_internvl_plugin():
_check_plugin(**check_inputs)
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
@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)
@@ -251,7 +251,7 @@ def test_llama4_plugin():
_check_plugin(**check_inputs)
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
def test_llava_plugin():
image_seqlen = 576
tokenizer_module = _load_tokenizer_module(model_name_or_path="llava-hf/llava-1.5-7b-hf")
@@ -265,7 +265,7 @@ def test_llava_plugin():
_check_plugin(**check_inputs)
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
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")
@@ -279,7 +279,7 @@ def test_llava_next_plugin():
_check_plugin(**check_inputs)
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
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")
@@ -293,7 +293,7 @@ def test_llava_next_video_plugin():
_check_plugin(**check_inputs)
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
@pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
def test_paligemma_plugin():
image_seqlen = 256
@@ -313,7 +313,7 @@ def test_paligemma_plugin():
_check_plugin(**check_inputs)
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
@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
@@ -336,7 +336,7 @@ def test_pixtral_plugin():
_check_plugin(**check_inputs)
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
@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
@@ -367,7 +367,7 @@ def test_qwen2_omni_plugin():
_check_plugin(**check_inputs)
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
def test_qwen2_vl_plugin():
image_seqlen = 4
tokenizer_module = _load_tokenizer_module(model_name_or_path="Qwen/Qwen2-VL-7B-Instruct")
@@ -384,7 +384,7 @@ def test_qwen2_vl_plugin():
_check_plugin(**check_inputs)
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
@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
@@ -406,7 +406,7 @@ def test_qwen3_vl_plugin():
_check_plugin(**check_inputs)
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
@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

34
tests/data/test_template.py
View File

@@ -89,7 +89,7 @@ def _check_template(
_check_tokenization(tokenizer, (prompt_ids, answer_ids), (prompt_str, answer_str))
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
@pytest.mark.parametrize("use_fast", [True, False])
def test_encode_oneturn(use_fast: bool):
tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3, use_fast=use_fast)
@@ -105,7 +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.runs_on(["cpu", "mps"])
@pytest.mark.parametrize("use_fast", [True, False])
def test_encode_multiturn(use_fast: bool):
tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3, use_fast=use_fast)
@@ -127,7 +127,7 @@ def test_encode_multiturn(use_fast: bool):
)
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
@pytest.mark.parametrize("use_fast", [True, False])
@pytest.mark.parametrize("cot_messages", [True, False])
@pytest.mark.parametrize("enable_thinking", [True, False, None])
@@ -154,7 +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.runs_on(["cpu", "mps"])
@pytest.mark.parametrize("use_fast", [True, False])
@pytest.mark.parametrize("cot_messages", [True, False])
@pytest.mark.parametrize("enable_thinking", [True, False, None])
@@ -184,7 +184,7 @@ def test_reasoning_encode_multiturn(use_fast: bool, cot_messages: bool, enable_t
)
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
@pytest.mark.parametrize("use_fast", [True, False])
def test_jinja_template(use_fast: bool):
tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3, use_fast=use_fast)
@@ -195,7 +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"])
@pytest.mark.runs_on(["cpu", "mps"])
def test_ollama_modelfile():
tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
template = get_template_and_fix_tokenizer(tokenizer, DataArguments(template="llama3"))
@@ -213,14 +213,14 @@ def test_ollama_modelfile():
)
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
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.runs_on(["cpu", "mps"])
@pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
@pytest.mark.parametrize("use_fast", [True, False])
def test_gemma_template(use_fast: bool):
@@ -234,7 +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.runs_on(["cpu", "mps"])
@pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
@pytest.mark.parametrize("use_fast", [True, False])
def test_gemma2_template(use_fast: bool):
@@ -248,7 +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.runs_on(["cpu", "mps"])
@pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
@pytest.mark.parametrize("use_fast", [True, False])
def test_llama3_template(use_fast: bool):
@@ -262,7 +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.runs_on(["cpu", "mps"])
@pytest.mark.parametrize(
"use_fast", [True, pytest.param(False, marks=pytest.mark.xfail(reason="Llama 4 has no slow tokenizer."))]
)
@@ -284,7 +284,7 @@ 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"])
@pytest.mark.runs_on(["cpu", "mps"])
def test_phi4_template(use_fast: bool):
prompt_str = (
f"<|im_start|>user<|im_sep|>{MESSAGES[0]['content']}<|im_end|>"
@@ -296,7 +296,7 @@ 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.runs_on(["cpu", "mps"])
@pytest.mark.xfail(not HF_TOKEN, reason="Authorization.")
@pytest.mark.parametrize("use_fast", [True, False])
def test_qwen2_5_template(use_fast: bool):
@@ -311,7 +311,7 @@ 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.runs_on(["cpu", "mps"])
@pytest.mark.parametrize("use_fast", [True, False])
@pytest.mark.parametrize("cot_messages", [True, False])
def test_qwen3_template(use_fast: bool, cot_messages: bool):
@@ -331,7 +331,7 @@ 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"])
@pytest.mark.runs_on(["cpu", "mps"])
def test_parse_llama3_template():
tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3, token=HF_TOKEN)
template = parse_template(tokenizer)
@@ -345,7 +345,7 @@ def test_parse_llama3_template():
assert template.default_system == ""
@pytest.mark.runs_on(["cpu"])
@pytest.mark.runs_on(["cpu", "mps"])
@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)
@@ -358,7 +358,7 @@ 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.runs_on(["cpu", "mps"])
@pytest.mark.xfail(not HF_TOKEN, reason="Authorization.")
def test_parse_qwen3_template():
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B", token=HF_TOKEN)