Transformers v5 (#2408)

* Uses AutoModelForImageTextToText instead of AutoModelForVision2Seq

* image-to-text pipelines is deprecated in v5

Author: echarlaix

optimum/exporters/tasks.py

@@ -29,7 +29,7 @@
 from transformers import AutoConfig, PretrainedConfig
 from transformers.utils import SAFE_WEIGHTS_NAME, WEIGHTS_NAME, http_user_agent
 
-from ..utils.import_utils import is_diffusers_available, is_torch_available
+from ..utils.import_utils import is_diffusers_available, is_torch_available, is_transformers_version
 from ..utils.logging import get_logger
 
 
@@ -152,8 +152,10 @@ class TasksManager:
                 "AutoModelForUniversalSegmentation",
             ),
             "image-to-image": "AutoModelForImageToImage",
-            # TODO: AutoModelForVision2Seq is deprecated and will be removed in Transformers v5
-            "image-to-text": ("AutoModelForVision2Seq", "AutoModel"),
+            "image-to-text": (
+                "AutoModelForVision2Seq" if is_transformers_version("<", "4.54") else "AutoModelForImageTextToText",
+                "AutoModel",
+            ),
             "image-text-to-text": "AutoModelForImageTextToText",
             "mask-generation": "AutoModel",
             "masked-im": "AutoModelForMaskedImageModeling",

tests/pipelines/test_pipelines.py

@@ -17,12 +17,14 @@
 from typing import Any, Dict
 
 import numpy as np
+import pytest
 from huggingface_hub.constants import HF_HUB_CACHE
 from PIL import Image
 from transformers import AutoTokenizer
 from transformers.pipelines import Pipeline
 
 from optimum.pipelines import pipeline as optimum_pipeline
+from optimum.utils import is_transformers_version
 from optimum.utils.testing_utils import remove_directory
 
 
@@ -192,6 +194,10 @@ def test_image_segmentation_pipeline(self):
             self.assertIn("score", result[0])
             self.assertIn("mask", result[0])
 
+    @pytest.mark.skipif(
+        is_transformers_version(">=", "5"),
+        reason="requires transformers < v5 since image-to-text pipelines is deprecated",
+    )
     def test_image_to_text_pipeline(self):
         """Test image to text ORT pipeline"""
         pipe = optimum_pipeline(task="image-to-text", accelerator="ort")