[misc] lint (#9636)

.gitignore

@@ -85,7 +85,7 @@ ipython_config.py
 # pyenv
 #   For a library or package, you might want to ignore these files since the code is
 #   intended to run in multiple environments; otherwise, check them in:
-# .python-version
+.python-version
 
 # pipenv
 #   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.

src/llamafactory/data/mm_plugin.py

@@ -1624,7 +1624,12 @@ class Qwen3VLPlugin(Qwen2VLPlugin):
                 for video, duration in zip(videos["videos"], videos["durations"])
             ]
             mm_inputs.update(
-                video_processor(videos=videos["videos"], video_metadata=video_metadata, fps=getattr(processor, "video_fps", 2.0), return_metadata=True)
+                video_processor(
+                    videos=videos["videos"],
+                    video_metadata=video_metadata,
+                    fps=getattr(processor, "video_fps", 2.0),
+                    return_metadata=True,
+                )
             )
             temporal_patch_size: int = getattr(image_processor, "temporal_patch_size", 2)
             if "second_per_grid_ts" in processor.model_input_names:

src/llamafactory/train/dpo/ktrainer.py

@@ -15,32 +15,24 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import TYPE_CHECKING, Optional
+from typing import TYPE_CHECKING
+
 import torch
-import torch.nn.functional as F
+from ktransformers.sft.lora import KTrainer  # type: ignore
-from transformers import Trainer
-from trl import DPOTrainer
-from trl.trainer import disable_dropout_in_model
 from typing_extensions import override
 
-from ...extras.constants import IGNORE_INDEX
+from ..trainer_utils import get_batch_logps, nested_detach
-from ...extras.packages import is_transformers_version_greater_than
+from .trainer import CustomDPOTrainer
-from ..callbacks import SaveProcessorCallback
-from ..trainer_utils import create_custom_optimizer, create_custom_scheduler, get_batch_logps, nested_detach
-from .trainer import CustomDPOTrainer as BaseDPOTrainer
-from ktransformers.sft.lora import KTrainer
 
 
 if TYPE_CHECKING:
-    from transformers import PreTrainedModel, ProcessorMixin
+    from transformers import PreTrainedModel
-
-    from ...hparams import FinetuningArguments
 
 
-class CustomDPOTrainer(KTrainer, BaseDPOTrainer):
+class KDPOTrainer(KTrainer, CustomDPOTrainer):
     @override
     def concatenated_forward(
-            self, model: "PreTrainedModel", batch: dict[str, "torch.Tensor"], is_ref_model: bool = False
+        self, model: "PreTrainedModel", batch: dict[str, "torch.Tensor"], is_ref_model: bool = False
     ) -> tuple["torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor"]:
         r"""Compute the sum log probabilities of the labels under given logits if loss_type is not IPO, ORPO or SimPO.
 
@@ -48,9 +40,8 @@ class CustomDPOTrainer(KTrainer, BaseDPOTrainer):
         """
         if self.finetuning_args.use_ref_model:
             batch = nested_detach(batch, clone=True)  # avoid error
-        labels = batch["labels"]
+
-        # dpo not need compute loss in forward, waste mem
+        labels = batch.pop("labels")  # dpo do not need compute loss in forward
-        del batch["labels"]
         all_logits: torch.Tensor = model(**batch, return_dict=True, use_cache=False).logits.to(torch.float32)
         all_logits = all_logits.to("cpu")
         labels = labels.to(all_logits.device)

src/llamafactory/train/dpo/trainer.py

@@ -218,9 +218,10 @@ class CustomDPOTrainer(DPOTrainer):
         if self.finetuning_args.use_ref_model:
             batch = nested_detach(batch, clone=True)  # avoid error
 
+        labels = batch.pop("labels")  # dpo do not need compute loss in forward
         all_logits: torch.Tensor = model(**batch, return_dict=True, use_cache=False).logits.to(torch.float32)
         all_logps, valid_length = get_batch_logps(
-            logits=all_logits, labels=batch["labels"], ld_alpha=(self.ld_alpha if not is_ref_model else None)
+            logits=all_logits, labels=labels, ld_alpha=(self.ld_alpha if not is_ref_model else None)
         )
         if self.loss_type in ["ipo", "orpo", "simpo"]:
             all_logps = all_logps / valid_length

src/llamafactory/train/dpo/workflow.py

@@ -62,13 +62,13 @@ def run_dpo(
     else:
         ref_model = None
 
-        
     if model_args.use_kt:
-        from ktransformers.util.globals import GLOBAL_CONFIG
+        from ktransformers.util.globals import GLOBAL_CONFIG  # type: ignore
+
+        from .ktrainer import KDPOTrainer as CustomDPOTrainer
 
         GLOBAL_CONFIG._config["mod"] = "sft"
 
-        from .ktrainer import CustomDPOTrainer
     else:
         from .trainer import CustomDPOTrainer
 

src/llamafactory/train/sft/workflow.py

@@ -99,8 +99,8 @@ def run_sft(
 
     # Initialize our Trainer
     if model_args.use_kt:
-        from ktransformers.util.globals import GLOBAL_CONFIG
+        from ktransformers.sft.lora import KTrainer  # type: ignore
-        from ktransformers.sft.lora import KTrainer
+        from ktransformers.util.globals import GLOBAL_CONFIG  # type: ignore
 
         GLOBAL_CONFIG._config["mod"] = "sft"