Fix pipeline tests blip2

src/transformers/models/kosmos2/processing_kosmos2.py

@@ -59,9 +59,6 @@ class Kosmos2ProcessorKwargs(ProcessingKwargs, total=False):
             "verbose": True,
             "add_eos_token": False,
         },
-        "images_kwargs": {
-            "num_image_tokens": 64,
-        },
     }
 
 
@@ -84,11 +81,11 @@ class Kosmos2Processor(ProcessorMixin):
     """
 
     attributes = ["image_processor", "tokenizer"]
-    valid_kwargs = ["num_patch_index_tokens"]
+    valid_kwargs = ["num_patch_index_tokens", "num_image_tokens"]
     image_processor_class = "CLIPImageProcessor"
     tokenizer_class = "AutoTokenizer"
 
-    def __init__(self, image_processor, tokenizer, num_patch_index_tokens=1024, *kwargs):
+    def __init__(self, image_processor, tokenizer, num_patch_index_tokens=1024, num_image_tokens=64, *kwargs):
         tokenizer.return_token_type_ids = False
 
         self.eod_token = "</doc>"
@@ -125,6 +122,7 @@ def __init__(self, image_processor, tokenizer, num_patch_index_tokens=1024, *kwa
 
         self.num_patch_index_tokens = num_patch_index_tokens
         patch_index_tokens = [f"<patch_index_{str(x).zfill(4)}>" for x in range(self.num_patch_index_tokens)]
+        self.num_image_tokens = num_image_tokens
 
         tokens_to_add = []
         for token in self.tag_tokens + patch_index_tokens:
@@ -171,7 +169,7 @@ def __call__(
         )
 
         bboxes = output_kwargs["images_kwargs"].pop("bboxes", None)
-        num_image_tokens = output_kwargs["images_kwargs"].pop("num_image_tokens", 64)
+        num_image_tokens = output_kwargs["images_kwargs"].pop("num_image_tokens", self.num_image_tokens)
         first_image_token_id = output_kwargs["images_kwargs"].pop("first_image_token_id", None)
         add_eos_token = output_kwargs["text_kwargs"].pop("add_eos_token", False)
 

src/transformers/pipelines/image_text_to_text.py

@@ -208,6 +208,14 @@ def _sanitize_parameters(
                     " please use only one"
                 )
             forward_kwargs["generate_kwargs"]["max_new_tokens"] = max_new_tokens
+        else:
+            if "generate_kwargs" not in forward_kwargs:
+                forward_kwargs["generate_kwargs"] = {}
+            if "max_new_tokens" not in forward_kwargs["generate_kwargs"]:
+                logger.warning_once(
+                    "The `max_new_tokens` parameter is not set. By default, the model will generate up to 20 tokens."
+                )
+                forward_kwargs["generate_kwargs"]["max_new_tokens"] = 20
 
         if return_full_text is not None and return_type is None:
             if return_tensors is not None:

tests/pipelines/test_pipelines_image_text_to_text.py

@@ -43,15 +43,15 @@ class ImageTextToTextPipelineTests(unittest.TestCase):
 
     def get_test_pipeline(self, model, tokenizer, processor, image_processor, torch_dtype="float32"):
         pipe = ImageTextToTextPipeline(model=model, processor=processor, torch_dtype=torch_dtype)
-        image_token = processor.image_token if hasattr(processor, "image_token") else "<image>"
+        image_token = getattr(processor.tokenizer, "image_token", "")
         examples = [
             {
                 "images": Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png"),
-                "text": f"{image_token} This is a ",
+                "text": f"{image_token}This is a ",
             },
             {
                 "images": "./tests/fixtures/tests_samples/COCO/000000039769.png",
-                "text": f"{image_token} Here I see a ",
+                "text": f"{image_token}Here I see a ",
             },
         ]
         return pipe, examples