fix image-to-text batch incorrect output issue (#29342)

* fix image-to-text batch incorrect output issue

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

* add ci test

Signed-off-by: Wang, Yi <yi.a.wang@intel.com>

* update ci test

Signed-off-by: Wang, Yi <yi.a.wang@intel.com>

---------

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
Signed-off-by: Wang, Yi <yi.a.wang@intel.com>

src/transformers/pipelines/pt_utils.py

@@ -73,7 +73,7 @@ def loader_batch_item(self):
         """
         if isinstance(self._loader_batch_data, torch.Tensor):
             # Batch data is simple tensor, just fetch the slice
-            result = self._loader_batch_data[self._loader_batch_index]
+            result = self._loader_batch_data[self._loader_batch_index].unsqueeze(0)
         else:
             # Batch data is assumed to be BaseModelOutput (or dict)
             loader_batched = {}

tests/pipelines/test_pipelines_image_to_text.py

@@ -142,6 +142,35 @@ def test_small_model_pt_conditional(self):
         outputs = pipe(image, prompt=prompt)
         self.assertTrue(outputs[0]["generated_text"].startswith(prompt))
 
+    @require_torch
+    def test_consistent_batching_behaviour(self):
+        pipe = pipeline("image-to-text", model="hf-internal-testing/tiny-random-BlipForConditionalGeneration")
+        image = "./tests/fixtures/tests_samples/COCO/000000039769.png"
+        prompt = "a photo of"
+
+        outputs = pipe([image, image], prompt=prompt)
+        self.assertTrue(outputs[0][0]["generated_text"].startswith(prompt))
+        self.assertTrue(outputs[1][0]["generated_text"].startswith(prompt))
+
+        outputs = pipe([image, image], prompt=prompt, batch_size=2)
+        self.assertTrue(outputs[0][0]["generated_text"].startswith(prompt))
+        self.assertTrue(outputs[1][0]["generated_text"].startswith(prompt))
+
+        from torch.utils.data import Dataset
+
+        class MyDataset(Dataset):
+            def __len__(self):
+                return 5
+
+            def __getitem__(self, i):
+                return "./tests/fixtures/tests_samples/COCO/000000039769.png"
+
+        dataset = MyDataset()
+        for batch_size in (1, 2, 4):
+            outputs = pipe(dataset, prompt=prompt, batch_size=batch_size if batch_size > 1 else None)
+            self.assertTrue(list(outputs)[0][0]["generated_text"].startswith(prompt))
+            self.assertTrue(list(outputs)[1][0]["generated_text"].startswith(prompt))
+
     @slow
     @require_torch
     def test_large_model_pt(self):