diff --git a/tests/models/test_phi3v.py b/tests/models/test_phi3v.py
index 35ffe4ef50a85..3737dc2bd076e 100644
--- a/tests/models/test_phi3v.py
+++ b/tests/models/test_phi3v.py
@@ -81,7 +81,10 @@ def run_test(
 
     inputs_per_image = [(
         [prompt for _ in size_factors],
-        [rescale_image_size(image, factor) for factor in size_factors],
+        [
+            rescale_image_size(image, factor, transpose=idx)
+            for idx, factor in enumerate(size_factors)
+        ],
     ) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
 
     # NOTE: take care of the order. run vLLM first, and then run HF.
diff --git a/tests/tracing/test_tracing.py b/tests/tracing/test_tracing.py
index a492daf3b49ca..90f26400952b9 100644
--- a/tests/tracing/test_tracing.py
+++ b/tests/tracing/test_tracing.py
@@ -114,5 +114,5 @@ def test_traces(trace_service):
         SpanAttributes.LLM_LATENCY_TIME_TO_FIRST_TOKEN) == ttft
     e2e_time = metrics.finished_time - metrics.arrival_time
     assert attributes.get(SpanAttributes.LLM_LATENCY_E2E) == e2e_time
-    assert attributes.get(SpanAttributes.LLM_LATENCY_TIME_IN_SCHEDULER
-                          ) == metrics.scheduler_time
+    assert attributes.get(
+        SpanAttributes.LLM_LATENCY_TIME_IN_SCHEDULER) == metrics.scheduler_time
diff --git a/vllm/model_executor/models/phi3v.py b/vllm/model_executor/models/phi3v.py
index 823c34b101870..d39cf15a1bb96 100644
--- a/vllm/model_executor/models/phi3v.py
+++ b/vllm/model_executor/models/phi3v.py
@@ -189,7 +189,7 @@ def hd_feature_transform(self, image_features, image_sizes):
         global_image_features_hd_newline = self.add_image_newline(
             global_image_features_hd)
 
-        all_image_embeddings = []
+        batch_image_features_proj = []
         # need a for loop to process each image because of different image sizes
         # (patch arrangement is different for each image)
         for i, img_size in enumerate(image_sizes):
@@ -207,19 +207,17 @@ def hd_feature_transform(self, image_features, image_sizes):
                 sub_image_features_hd)
 
             # [sub features, separator, global features]
-            all_image_embeddings.append(
-                torch.cat([
-                    sub_image_features_hd_newline.squeeze(
-                        0),  # (h_crop*12*(w_crop*12+1), 4096)
-                    self.glb_GN.squeeze(0),
-                    global_image_features_hd_newline[i],
-                ]))
-
-        image_features_proj = self.img_projection(
-            torch.stack(all_image_embeddings).to(target_device, target_dtype)
-        )  # (num_images, (h_crop*12*(w_crop*12+1)+1), hidden_size)
-
-        return image_features_proj
+            image_embeddings = torch.cat([
+                sub_image_features_hd_newline.squeeze(
+                    0),  # (h_crop*12*(w_crop*12+1), 4096)
+                self.glb_GN.squeeze(0),
+                global_image_features_hd_newline[i],
+            ])
+            img_proj = self.img_projection(
+                image_embeddings.to(target_device, target_dtype))
+            batch_image_features_proj.append(img_proj)
+
+        return batch_image_features_proj
 
     def reshape_hd_patches_2x2merge(self, image_features, h_crop, w_crop):
         """
diff --git a/vllm/multimodal/utils.py b/vllm/multimodal/utils.py
index bafd208469788..8f7e613cdf90a 100644
--- a/vllm/multimodal/utils.py
+++ b/vllm/multimodal/utils.py
@@ -90,8 +90,13 @@ def load_image_from_base64(image: Union[bytes, str]) -> Image.Image:
     return _load_image_from_bytes(base64.b64decode(image))
 
 
-def rescale_image_size(image: Image.Image, size_factor: float) -> Image.Image:
+def rescale_image_size(image: Image.Image,
+                       size_factor: float,
+                       transpose: int = -1) -> Image.Image:
     """Rescale the dimensions of an image by a constant factor."""
     new_width = int(image.width * size_factor)
     new_height = int(image.height * size_factor)
-    return image.resize((new_width, new_height))
+    image = image.resize((new_width, new_height))
+    if transpose >= 0:
+        image = image.transpose(Image.Transpose(transpose))
+    return image