Tiny fix

Former-commit-id: ae66e1a545f4cd209a57fd824f9bfb7e94436cba

src/llamafactory/data/mm_plugin.py

@@ -24,7 +24,6 @@ if TYPE_CHECKING:
     from av.stream import Stream
     from transformers import PreTrainedTokenizer, ProcessorMixin
     from transformers.image_processing_utils import BaseImageProcessor
-    from transformers.processing_utils import _validate_images_text_input_order, ProcessingKwargs
 
     class EncodedImage(TypedDict):
         path: Optional[str]
@@ -392,7 +391,7 @@ class PixtralPlugin(BasePlugin):
         mm_inputs = self._get_mm_inputs(images, videos, processor)
         if mm_inputs.get("image_sizes"):
             mm_inputs.pop("image_sizes")
-            
+
             if isinstance(mm_inputs.get("pixel_values"), list) and len(mm_inputs.get("pixel_values")[0]) >= 2:
                 raise ValueError("Now it only supports batchsize=1 on per gpu due to `List[tensor]` can not pack into BachEncoding")
             

src/llamafactory/extras/constants.py

@@ -1059,9 +1059,9 @@ register_model_group(
 
 register_model_group(
     models={
-        "Pixtral-12B-2409": {
-            DownloadSource.DEFAULT: "mistralai/Pixtral-12B-2409",
-            DownloadSource.MODELSCOPE: "LLM-Research/Pixtral-12B-2409",
+        "Pixtral-12B": {
+            DownloadSource.DEFAULT: "mistral-community/pixtral-12b",
+            DownloadSource.MODELSCOPE: "AI-ModelScope/pixtral-12b",
         }
     },
     template="mistral",

src/llamafactory/model/model_utils/visual.py

@@ -96,7 +96,6 @@ def autocast_projector_dtype(model: "PreTrainedModel", model_args: "ModelArgumen
             mm_projector: "torch.nn.Module" = getattr(model, "multi_modal_projector")
         elif model_type == "qwen2_vl":
             mm_projector: "torch.nn.Module" = getattr(getattr(model, "visual"), "merger")
-        # TODO check it
         elif model_type == "pixtral":
             mm_projector: "torch.nn.Module" = getattr(model, "vision_language_adapte")
         else:
@@ -128,7 +127,6 @@ def get_forbidden_modules(config: "PretrainedConfig", finetuning_args: "Finetuni
     if model_type in ["llava", "paligemma", "pixtral"]:
         if finetuning_args.freeze_vision_tower:
             forbidden_modules.add("vision_tower")
-            #TODO check it
             forbidden_modules.add("vision_encoder")
 
         if finetuning_args.train_mm_proj_only: