Add a check in `ImageToTextPipeline._forward` (#24373)

* fix

* fix

* fix

* Update src/transformers/pipelines/image_to_text.py
Co-authored-by: Nicolas Patry <patry.nicolas@protonmail.com>

* fix

---------
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>
Co-authored-by: Nicolas Patry <patry.nicolas@protonmail.com>

src/transformers/pipelines/image_to_text.py

@@ -145,6 +145,15 @@ class ImageToTextPipeline(Pipeline):
         return model_inputs
 
     def _forward(self, model_inputs, generate_kwargs=None):
+        # Git model sets `model_inputs["input_ids"] = None` in `preprocess` (when `prompt=None`). In batch model, the
+        # pipeline will group them into a list of `None`, which fail `_forward`. Avoid this by checking it first.
+        if (
+            "input_ids" in model_inputs
+            and isinstance(model_inputs["input_ids"], list)
+            and all(x is None for x in model_inputs["input_ids"])
+        ):
+            model_inputs["input_ids"] = None
+
         if generate_kwargs is None:
             generate_kwargs = {}
         # FIXME: We need to pop here due to a difference in how `generation.py` and `generation.tf_utils.py`

tests/models/git/test_modeling_git.py

@@ -390,15 +390,6 @@ class GitModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin,
     fx_compatible = False
     test_torchscript = False
 
-    # `GitForCausalLM` doesn't fit into image-to-text pipeline. We might need to overwrite its `generate` function.
-    def is_pipeline_test_to_skip(
-        self, pipeline_test_casse_name, config_class, model_architecture, tokenizer_name, processor_name
-    ):
-        if pipeline_test_casse_name == "ImageToTextPipelineTests":
-            return True
-
-        return False
-
     # special case for GitForCausalLM model
     def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
         inputs_dict = super()._prepare_for_class(inputs_dict, model_class, return_labels=return_labels)