[Bugfix] Fix HunyuanVL XD-RoPE (#29593)

Signed-off-by: Roger Wang <hey@rogerw.io>
Co-authored by: grider-transwithai <grider@transwith.ai>

vllm/model_executor/models/hunyuan_vision.py

@@ -847,7 +847,7 @@ class HunYuanVLForConditionalGeneration(
                 .expand(-1, llm_grid_w + 1)
                 .reshape(-1)
             )
-            h_index[pos : pos + token_num] = 0
+            t_index[pos : pos + token_num] = image_index
 
         if xd_num == 4:
             llm_positions = torch.stack([p_index, w_index, h_index, t_index])

vllm/transformers_utils/processors/hunyuan_vl_image.py

@@ -195,9 +195,9 @@ class HunYuanVLImageProcessor(BaseImageProcessor):
         processed_images = []
         for image in images:
             if do_resize:
-                resized_width, resized_height = smart_resize(
-                    width,
-                    height,
+                resized_height, resized_width = smart_resize(
+                    height=height,
+                    width=width,
                     factor=patch_size * merge_size,
                     min_pixels=self.min_pixels,
                     max_pixels=self.max_pixels,