[Model] Pass mm_features directly into get_mrope_input_positions (#28399)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

vllm/model_executor/models/qwen3_vl.py

@@ -34,7 +34,7 @@ import numpy as np
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from transformers import BatchFeature, PretrainedConfig
+from transformers import BatchFeature
 from transformers.models.qwen2_vl import Qwen2VLImageProcessorFast
 from transformers.models.qwen2_vl.image_processing_qwen2_vl import (
     smart_resize as image_smart_resize,
@@ -70,6 +70,7 @@ from vllm.model_executor.models.module_mapping import MultiModelKeys
 from vllm.multimodal import MULTIMODAL_REGISTRY
 from vllm.multimodal.inputs import (
     MultiModalDataDict,
+    MultiModalFeatureSpec,
     MultiModalFieldConfig,
     MultiModalKwargsItem,
     MultiModalKwargsItems,
@@ -1416,17 +1417,18 @@ class Qwen3VLForConditionalGeneration(
     def get_mrope_input_positions(
         self,
         input_tokens: list[int],
-        hf_config: PretrainedConfig,
-        image_grid_thw: list[list[int]] | torch.Tensor,
-        video_grid_thw: list[list[int]] | torch.Tensor,
-        second_per_grid_ts: list[float] | None = None,
-        audio_feature_lengths: torch.Tensor | None = None,
-        use_audio_in_video: bool = False,
+        mm_features: list[MultiModalFeatureSpec],
     ) -> tuple[torch.Tensor, int]:
-        """Get mrope input positions and delta value."""
+        kwargs = MultiModalFeatureSpec.gather_kwargs(
+            mm_features,
+            {"image_grid_thw", "video_grid_thw"},
+        )
+        image_grid_thw = [item.tolist() for item in kwargs.get("image_grid_thw", [])]
+        video_grid_thw = [item.tolist() for item in kwargs.get("video_grid_thw", [])]
 
         video_grid_thw = [[1, h, w] for t, h, w in video_grid_thw for _ in range(t)]
 
+        hf_config = self.config
         image_token_id = hf_config.image_token_id
         video_token_id = hf_config.video_token_id
         vision_start_token_id = hf_config.vision_start_token_id
@@ -1455,20 +1457,12 @@ class Qwen3VLForConditionalGeneration(
             else:
                 ed_video = len(input_tokens) + 1
             if ed_image < ed_video:
-                t, h, w = (
-                    image_grid_thw[image_index][0],
-                    image_grid_thw[image_index][1],
-                    image_grid_thw[image_index][2],
-                )
+                t, h, w = image_grid_thw[image_index]
                 image_index += 1
                 remain_images -= 1
                 ed = ed_image
             else:
-                t, h, w = (
-                    video_grid_thw[video_index][0],
-                    video_grid_thw[video_index][1],
-                    video_grid_thw[video_index][2],
-                )
+                t, h, w = video_grid_thw[video_index]
                 video_index += 1
                 remain_videos -= 1
                 ed = ed_video