Fix for transformers pin bump

Author: larryliu0820

.ci/scripts/test_model_e2e.sh

@@ -86,7 +86,7 @@ case "$HF_MODEL" in
     MODEL_NAME="voxtral"
     RUNNER_TARGET="voxtral_runner"
     RUNNER_PATH="voxtral"
-    EXPECTED_OUTPUT="poem"
+    EXPECTED_OUTPUT="existence"
     PREPROCESSOR="voxtral_preprocessor.pte"
     TOKENIZER_URL="https://huggingface.co/mistralai/Voxtral-Mini-3B-2507/resolve/main" # @lint-ignore
     TOKENIZER_FILE="tekken.json"

.github/workflows/trunk.yml

@@ -839,7 +839,7 @@ jobs:
           qwen3-0.6b|xnnpack|--quantize,
           qwen3-1.7b|xnnpack|--quantize,
           gemma3-1b|xnnpack|--quantize,
-          phi4-mini|xnnpack|--quantize,
+          # phi4-mini|xnnpack|--quantize, transformers v5.0.0rc0 introduces a data-dependent branching in transformers/modeling_rope_utils.py:61 
           smollm2-135m|xnnpack|--quantize,
           smollm3-3b|xnnpack|--quantize
         ]

backends/qualcomm/quantizer/annotators.py

@@ -955,6 +955,10 @@ def annotate_elu(node: Node, quantization_config: QuantizationConfig) -> None:
 def annotate_embedding(node: Node, quantization_config: QuantizationConfig) -> None:
     weight = node.args[0]
 
+    # Only quantize if input is a float tensor
+    if not _is_float_tensor(weight):
+        return
+
     input_qspec_map = {}
     input_qspec_map[weight] = quantization_config.input_activation
 

examples/qualcomm/oss_scripts/mobilevit_v1.py

@@ -24,7 +24,7 @@
 )
 from PIL import Image
 from torchvision import datasets
-from transformers import AutoModelForImageClassification, MobileViTFeatureExtractor
+from transformers import AutoImageProcessor, AutoModelForImageClassification
 
 
 def get_imagenet_dataset(dataset_path, data_size, shuffle=True):
@@ -39,15 +39,13 @@ def get_data_loader():
     # prepare input data
     inputs, targets = [], []
     data_loader = get_data_loader()
-    feature_extractor = MobileViTFeatureExtractor.from_pretrained(
-        "apple/mobilevit-xx-small"
-    )
+    image_processor = AutoImageProcessor.from_pretrained("apple/mobilevit-xx-small")
     for index, data in enumerate(data_loader.dataset.imgs):
         if index >= data_size:
             break
         data_path, target = data
         image = Image.open(data_path).convert("RGB")
-        feature = feature_extractor(images=image, return_tensors="pt")
+        feature = image_processor(images=image, return_tensors="pt")
         inputs.append((feature["pixel_values"],))
         targets.append(torch.tensor(target))
 

examples/qualcomm/oss_scripts/mobilevit_v2.py

@@ -25,7 +25,7 @@
 )
 from PIL import Image
 from torchvision import datasets
-from transformers import AutoModelForImageClassification, MobileViTFeatureExtractor
+from transformers import AutoImageProcessor, AutoModelForImageClassification
 
 
 def get_imagenet_dataset(dataset_path, data_size, shuffle=True):
@@ -40,15 +40,13 @@ def get_data_loader():
     # prepare input data
     inputs, targets = [], []
     data_loader = get_data_loader()
-    feature_extractor = MobileViTFeatureExtractor.from_pretrained(
-        "apple/mobilevit-xx-small"
-    )
+    image_processor = AutoImageProcessor.from_pretrained("apple/mobilevit-xx-small")
     for index, data in enumerate(data_loader.dataset.imgs):
         if index >= data_size:
             break
         data_path, target = data
         image = Image.open(data_path).convert("RGB")
-        feature = feature_extractor(images=image, return_tensors="pt")
+        feature = image_processor(images=image, return_tensors="pt")
         inputs.append((feature["pixel_values"],))
         targets.append(torch.tensor(target))
 

examples/qualcomm/scripts/mobilebert_fine_tune.py

@@ -131,17 +131,17 @@ def get_fine_tuned_mobilebert(artifacts_dir, pretrained_weight, batch_size):
     )
 
     # tokenize dataset
-    encoded_data_train = tokenizer.batch_encode_plus(
-        data[data.data_type == "train"].Title.values,
+    encoded_data_train = tokenizer(
+        data[data.data_type == "train"].Title.values.tolist(),
         add_special_tokens=True,
         return_attention_mask=True,
         max_length=256,
         padding="max_length",
         truncation=True,
         return_tensors="pt",
     )
-    encoded_data_val = tokenizer.batch_encode_plus(
-        data[data.data_type == "val"].Title.values,
+    encoded_data_val = tokenizer(
+        data[data.data_type == "val"].Title.values.tolist(),
         add_special_tokens=True,
         return_attention_mask=True,
         max_length=256,