[misc] lint (#9593)

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

.github/workflows/tests_npu.yml

@@ -84,4 +84,4 @@ jobs:
           make test
         env:
           HF_HOME: /root/.cache/huggingface
-          HF_HUB_OFFLINE: "${{ steps.hf-hub-cache.outputs.cache-hit == 'true' && '1' || '0' }}"
+          HF_HUB_OFFLINE: "${{ steps.hf-hub-cache.outputs.cache-hit == 'true' && '1' || '0' }}"

examples/accelerate/fsdp2_config.yaml

@@ -19,4 +19,4 @@ same_network: true
 tpu_env: []
 tpu_use_cluster: false
 tpu_use_sudo: false
-use_cpu: false
+use_cpu: false

examples/ascend/qwen3vlmoe_lora_sft_fsdp.yaml

@@ -39,4 +39,4 @@ warmup_ratio: 0.1
 bf16: true
 ddp_timeout: 180000000
 resume_from_checkpoint: null
-seed: 1234
+seed: 1234

src/llamafactory/model/model_utils/checkpointing.py

@@ -18,8 +18,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import os
 import inspect
+import os
 from functools import WRAPPER_ASSIGNMENTS, partial, wraps
 from types import MethodType
 from typing import TYPE_CHECKING, Any, Callable, Optional, Union
@@ -156,11 +156,9 @@ def prepare_model_for_training(model: "PreTrainedModel", model_args: "ModelArgum
     if (
         os.environ.get("ACCELERATE_USE_FSDP", "false").lower() == "true"
         and int(os.environ.get("FSDP_VERSION", "1")) == 2
-        ):
+    ):
         model_args.use_reentrant_gc = False
-        logger.warning_rank0(
+        logger.warning_rank0("You are using fsdp2, `use_reentrant_gc` has been set to False.")
-            "You are using fsdp2, `use_reentrant_gc` has been set to False. "
-        )
 
     if not model_args.disable_gradient_checkpointing:
         if not getattr(model, "supports_gradient_checkpointing", False):

src/llamafactory/train/sft/workflow.py

@@ -28,6 +28,7 @@ from ..trainer_utils import create_modelcard_and_push
 from .metric import ComputeAccuracy, ComputeSimilarity, eval_logit_processor
 from .trainer import CustomSeq2SeqTrainer
 
+
 if TYPE_CHECKING:
     from transformers import Seq2SeqTrainingArguments, TrainerCallback
 
@@ -144,4 +145,4 @@ def run_sft(
         trainer.save_predictions(dataset_module["eval_dataset"], predict_results, generating_args.skip_special_tokens)
 
     # Create model card
-    create_modelcard_and_push(trainer, model_args, data_args, training_args, finetuning_args)
+    create_modelcard_and_push(trainer, model_args, data_args, training_args, finetuning_args)

src/llamafactory/v1/extras/__init__.py

@@ -1 +0,0 @@
-

tests/conftest.py

@@ -40,9 +40,7 @@ def pytest_configure(config):
     config.addinivalue_line(
         "markers", "require_device: test requires specific device, e.g., @pytest.mark.require_device('cuda')"
     )
-    config.addinivalue_line(
+    config.addinivalue_line("markers", "runs_on: test requires specific device, e.g., @pytest.mark.runs_on(['cpu'])")
-        "markers", "runs_on: test requires specific device, e.g., @pytest.mark.runs_on(['cpu'])"
-    )
 
 
 def _handle_runs_on(items):
@@ -64,14 +62,12 @@ def _handle_runs_on(items):
             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(
+                    pytest.mark.skip(reason=f"test requires one of {runs_on_devices} (current: {CURRENT_DEVICE})")
-                        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.
 

tests/data/processor/test_feedback.py

@@ -42,7 +42,7 @@ TRAIN_ARGS = {
 }
 
 
-@pytest.mark.runs_on(["cpu","npu"])
+@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/test_template.py

@@ -284,7 +284,6 @@ 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 = (

tests/e2e/test_train.py

@@ -48,6 +48,7 @@ INFER_ARGS = {
 
 OS_NAME = os.getenv("OS_NAME", "")
 
+
 @pytest.mark.runs_on(["cpu"])
 @pytest.mark.parametrize(
     "stage,dataset",

tests/eval/test_eval_template.py

@@ -55,6 +55,7 @@ def test_eval_template_en():
         {"role": "assistant", "content": "C"},
     ]
 
+
 @pytest.mark.runs_on(["cpu"])
 def test_eval_template_zh():
     support_set = [

tests/model/model_utils/test_add_tokens.py

@@ -25,7 +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.runs_on(["cpu", "npu"])
 @pytest.mark.parametrize("special_tokens", [False, True])
 def test_add_tokens(special_tokens: bool):
     if special_tokens:

tests/model/model_utils/test_attention.py

@@ -17,13 +17,16 @@ import os
 import pytest
 from transformers.utils import is_flash_attn_2_available
 
+
 # Compatible with Transformers v4 and Transformers v5
 try:
     from transformers.utils import is_torch_sdpa_available
 except ImportError:
+
     def is_torch_sdpa_available():
         return True
 
+
 from llamafactory.extras.packages import is_transformers_version_greater_than
 from llamafactory.train.test_utils import load_infer_model
 
@@ -36,7 +39,7 @@ INFER_ARGS = {
 }
 
 
-@pytest.mark.runs_on(["cpu","npu"])
+@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"]

tests/model/model_utils/test_checkpointing.py

@@ -39,7 +39,7 @@ TRAIN_ARGS = {
 }
 
 
-@pytest.mark.runs_on(["cpu","npu"])
+@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)
@@ -47,14 +47,14 @@ def test_vanilla_checkpointing(disable_gradient_checkpointing: bool):
         assert getattr(module, "gradient_checkpointing") != disable_gradient_checkpointing
 
 
-@pytest.mark.runs_on(["cpu","npu"])
+@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"])
+@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():
@@ -62,7 +62,7 @@ def test_upcast_layernorm():
             assert param.dtype == torch.float32
 
 
-@pytest.mark.runs_on(["cpu","npu"])
+@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())

tests/model/model_utils/test_misc.py

@@ -24,7 +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.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")

tests/model/model_utils/test_packing.py

@@ -18,7 +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.runs_on(["cpu", "npu"])
 @pytest.mark.parametrize(
     "attention_mask,golden_seq_lens",
     [

tests/model/model_utils/test_visual.py

@@ -23,7 +23,7 @@ from llamafactory.hparams import FinetuningArguments, ModelArguments
 from llamafactory.model.adapter import init_adapter
 
 
-@pytest.mark.runs_on(["cpu","npu"])
+@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))
@@ -49,7 +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.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")
@@ -82,7 +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"])
+@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")

tests/model/test_base.py

@@ -29,14 +29,16 @@ INFER_ARGS = {
     "infer_dtype": "float16",
 }
 
-@pytest.mark.runs_on(["cpu","npu"])
+
+@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.runs_on(["cpu", "npu"])
 @pytest.mark.skip_on_devices("npu")
 @pytest.mark.usefixtures("fix_valuehead_cpu_loading")
 def test_valuehead():

tests/model/test_freeze.py

@@ -44,7 +44,7 @@ INFER_ARGS = {
 }
 
 
-@pytest.mark.runs_on(["cpu","npu"])
+@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():
@@ -56,7 +56,7 @@ def test_freeze_train_all_modules():
             assert param.dtype == torch.float16
 
 
-@pytest.mark.runs_on(["cpu","npu"])
+@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():
@@ -68,7 +68,7 @@ def test_freeze_train_extra_modules():
             assert param.dtype == torch.float16
 
 
-@pytest.mark.runs_on(["cpu","npu"])
+@pytest.mark.runs_on(["cpu", "npu"])
 def test_freeze_inference():
     model = load_infer_model(**INFER_ARGS)
     for param in model.parameters():

tests/model/test_full.py

@@ -43,14 +43,16 @@ INFER_ARGS = {
     "infer_dtype": "float16",
 }
 
-@pytest.mark.runs_on(["cpu","npu"])
+
+@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"])
+
+@pytest.mark.runs_on(["cpu", "npu"])
 def test_full_inference():
     model = load_infer_model(**INFER_ARGS)
     for param in model.parameters():

tests/model/test_lora.py

@@ -55,35 +55,35 @@ INFER_ARGS = {
 }
 
 
-@pytest.mark.runs_on(["cpu","npu"])
+@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"])
+@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"])
+@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"])
+@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"])
+@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)
@@ -92,7 +92,7 @@ def test_lora_train_new_adapters():
     )
 
 
-@pytest.mark.runs_on(["cpu","npu"])
+@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)
@@ -102,7 +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.runs_on(["cpu", "npu"])
 @pytest.mark.skip_on_devices("npu")
 def test_lora_inference():
     model = load_infer_model(**INFER_ARGS)

tests/model/test_pissa.py

@@ -49,14 +49,15 @@ INFER_ARGS = {
 }
 
 
-@pytest.mark.runs_on(["cpu","npu"])
+@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.runs_on(["cpu", "npu"])
 @pytest.mark.xfail(reason="Known connection error.")
 def test_pissa_inference():
     model = load_infer_model(**INFER_ARGS)

tests/train/test_sft_trainer.py

@@ -59,7 +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.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(