[Bugfix] Inconsistent token calculation compared to HF in llava family (#18479)

Signed-off-by: jaycha <[email protected]>

Author: cyr0930

vllm/model_executor/models/llava_next.py

@@ -135,11 +135,13 @@ def _get_num_unpadded_features(
         current_aspect_ratio = current_width / current_height
 
         if aspect_ratio > current_aspect_ratio:
-            new_height = (original_height * current_width) // original_width
+            new_height = int(
+                round(original_height * (current_width / original_width), 7))
             padding = (current_height - new_height) // 2
             current_height = current_height - (2 * padding)
         else:
-            new_width = (original_width * current_height) // original_height
+            new_width = int(
+                round(original_width * (current_height / original_height), 7))
             padding = (current_width - new_width) // 2
             current_width = current_width - (2 * padding)
 

vllm/model_executor/models/llava_onevision.py

@@ -116,11 +116,13 @@ def _get_num_unpadded_features(
         current_aspect_ratio = current_width / current_height
 
         if aspect_ratio > current_aspect_ratio:
-            new_height = (original_height * current_width) // original_width
+            new_height = int(
+                round(original_height * (current_width / original_width), 7))
             padding = (current_height - new_height) // 2
             current_height = current_height - (2 * padding)
         else:
-            new_width = (original_width * current_height) // original_height
+            new_width = int(
+                round(original_width * (current_height / original_height), 7))
             padding = (current_width - new_width) // 2
             current_width = current_width - (2 * padding)