Arm backend: Require safetensors for VGF DeiT export (#24025)

Require safetensors when loading the fine-tuned DeiT checkpoint so
export does not select legacy pickle-based PyTorch weights. Document the
required format in the CLI help and export instructions.

Signed-off-by: Sebastian Larsson <sebastian.larsson@arm.com>

Author: Sebastian-Larsson

examples/arm/image_classification_example_vgf/model_export/README.md

@@ -25,6 +25,9 @@ python examples/arm/image_classification_example_vgf/model_export/train_deit.py
 
 ## Export and quantize for VGF
 
+The exporter requires a safetensors checkpoint. The training script saves
+this format by default. Legacy PyTorch `.bin` checkpoints are not accepted.
+
 ```bash
 python examples/arm/image_classification_example_vgf/model_export/export_deit.py \
   --model-path ./deit-tiny-oxford-pet/final_model \

examples/arm/image_classification_example_vgf/model_export/export_deit.py

@@ -93,7 +93,7 @@ def measure_accuracy(quantized_model, test_set):
         type=str,
         default="./deit-tiny-oxford-pet/final_model",
         required=True,
-        help="Path to the fine-tuned ViT model.",
+        help="Path to the fine-tuned ViT model saved in safetensors format.",
     )
     argparser.add_argument(
         "--output-path",
@@ -120,7 +120,7 @@ def measure_accuracy(quantized_model, test_set):
         num_labels=37,
         ignore_mismatched_sizes=True,
         revision="b3428f18dcc7b543470d07f14b4a4157815d1880",
-        use_safetensors=False,
+        use_safetensors=True,
     ).eval()
     image_preprocessor = AutoImageProcessor.from_pretrained(
         "facebook/deit-tiny-patch16-224",