diff --git a/vllm/model_executor/models/cohere2_vision.py b/vllm/model_executor/models/cohere2_vision.py
index b58df12d0d..b6def10aab 100644
--- a/vllm/model_executor/models/cohere2_vision.py
+++ b/vllm/model_executor/models/cohere2_vision.py
@@ -268,7 +268,7 @@ class Cohere2VisionMultiModalProcessor(
         num_patches = hf_inputs.get("num_patches", torch.empty(0))
         return dict(
             pixel_values=MultiModalFieldConfig.flat_from_sizes("image", num_patches),
-            num_patches=MultiModalFieldConfig.batched("image"),
+            num_patches=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
             image_embeds=MultiModalFieldConfig.batched("image"),
         )
 
diff --git a/vllm/model_executor/models/deepseek_ocr.py b/vllm/model_executor/models/deepseek_ocr.py
index ead643a7f8..4bb5598db3 100644
--- a/vllm/model_executor/models/deepseek_ocr.py
+++ b/vllm/model_executor/models/deepseek_ocr.py
@@ -320,7 +320,9 @@ class DeepseekOCRMultiModalProcessor(
         patches_per_image = torch.where(is_tiled, images_spatial_crop.prod(dim=-1), 0)
         return dict(
             pixel_values=MultiModalFieldConfig.batched("image"),
-            images_spatial_crop=MultiModalFieldConfig.batched("image"),
+            images_spatial_crop=MultiModalFieldConfig.batched(
+                "image", keep_on_cpu=True
+            ),
             images_crop=MultiModalFieldConfig.flat_from_sizes(
                 "image", patches_per_image
             ),
diff --git a/vllm/model_executor/models/deepseek_ocr2.py b/vllm/model_executor/models/deepseek_ocr2.py
index f96fb9c146..aa6e9b1aec 100644
--- a/vllm/model_executor/models/deepseek_ocr2.py
+++ b/vllm/model_executor/models/deepseek_ocr2.py
@@ -191,7 +191,9 @@ class DeepseekOCR2MultiModalProcessor(
         patches_per_image = torch.where(is_tiled, images_spatial_crop.prod(dim=-1), 0)
         return dict(
             pixel_values=MultiModalFieldConfig.batched("image"),
-            images_spatial_crop=MultiModalFieldConfig.batched("image"),
+            images_spatial_crop=MultiModalFieldConfig.batched(
+                "image", keep_on_cpu=True
+            ),
             images_crop=MultiModalFieldConfig.flat_from_sizes(
                 "image", patches_per_image
             ),
diff --git a/vllm/model_executor/models/deepseek_vl2.py b/vllm/model_executor/models/deepseek_vl2.py
index a514bef576..9ca05b6c77 100644
--- a/vllm/model_executor/models/deepseek_vl2.py
+++ b/vllm/model_executor/models/deepseek_vl2.py
@@ -270,7 +270,9 @@ class DeepseekVL2MultiModalProcessor(
 
         return dict(
             pixel_values=MultiModalFieldConfig.flat_from_sizes("image", num_patches),
-            images_spatial_crop=MultiModalFieldConfig.batched("image"),
+            images_spatial_crop=MultiModalFieldConfig.batched(
+                "image", keep_on_cpu=True
+            ),
             image_embeds=MultiModalFieldConfig.batched("image"),
         )
 
diff --git a/vllm/model_executor/models/ernie45_vl.py b/vllm/model_executor/models/ernie45_vl.py
index 0c1781b04f..9fa0447002 100644
--- a/vllm/model_executor/models/ernie45_vl.py
+++ b/vllm/model_executor/models/ernie45_vl.py
@@ -1246,11 +1246,11 @@ class Ernie4_5VLMultiModalProcessor(BaseMultiModalProcessor[Ernie4_5_VLProcessin
             pixel_values=MultiModalFieldConfig.flat_from_sizes(
                 "image", image_grid_sizes
             ),
-            image_grid_thw=MultiModalFieldConfig.batched("image"),
+            image_grid_thw=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
             pixel_values_videos=MultiModalFieldConfig.flat_from_sizes(
                 "video", video_grid_sizes
             ),
-            video_grid_thw=MultiModalFieldConfig.batched("video"),
+            video_grid_thw=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
         )
 
 
diff --git a/vllm/model_executor/models/fireredasr2.py b/vllm/model_executor/models/fireredasr2.py
index eea0c7d889..231417d833 100644
--- a/vllm/model_executor/models/fireredasr2.py
+++ b/vllm/model_executor/models/fireredasr2.py
@@ -260,7 +260,7 @@ class FireRedASR2MultiModalProcessor(
         return dict(
             input_features=MultiModalFieldConfig.batched("audio"),
             speech_lengths=MultiModalFieldConfig.batched("audio"),
-            fake_token_lengths=MultiModalFieldConfig.batched("audio"),
+            fake_token_lengths=MultiModalFieldConfig.batched("audio", keep_on_cpu=True),
         )
 
     def _get_prompt_updates(
diff --git a/vllm/model_executor/models/fireredlid.py b/vllm/model_executor/models/fireredlid.py
index 804ed2bc9f..69fe995b6b 100644
--- a/vllm/model_executor/models/fireredlid.py
+++ b/vllm/model_executor/models/fireredlid.py
@@ -482,8 +482,8 @@ class FireRedLIDMultiModalProcessor(
     ) -> Mapping[str, MultiModalFieldConfig]:
         return dict(
             input_features=MultiModalFieldConfig.batched("audio"),
-            speech_lengths=MultiModalFieldConfig.batched("audio"),
-            fake_token_lengths=MultiModalFieldConfig.batched("audio"),
+            speech_lengths=MultiModalFieldConfig.batched("audio", keep_on_cpu=True),
+            fake_token_lengths=MultiModalFieldConfig.batched("audio", keep_on_cpu=True),
         )
 
     def _get_prompt_updates(
diff --git a/vllm/model_executor/models/funasr.py b/vllm/model_executor/models/funasr.py
index 56b1b68f4f..69f63d770b 100644
--- a/vllm/model_executor/models/funasr.py
+++ b/vllm/model_executor/models/funasr.py
@@ -779,7 +779,7 @@ class FunASRMultiModalProcessor(BaseMultiModalProcessor[FunASRProcessingInfo]):
         return dict(
             input_features=MultiModalFieldConfig.batched("audio"),
             speech_lengths=MultiModalFieldConfig.batched("audio"),
-            fake_token_lengths=MultiModalFieldConfig.batched("audio"),
+            fake_token_lengths=MultiModalFieldConfig.batched("audio", keep_on_cpu=True),
         )
 
     def _get_prompt_updates(
diff --git a/vllm/model_executor/models/funaudiochat.py b/vllm/model_executor/models/funaudiochat.py
index 0a9e286783..b336b9801e 100644
--- a/vllm/model_executor/models/funaudiochat.py
+++ b/vllm/model_executor/models/funaudiochat.py
@@ -695,7 +695,9 @@ class FunAudioChatMultiModalProcessor(
     ) -> Mapping[str, MultiModalFieldConfig]:
         return {
             "speech_ids": MultiModalFieldConfig.batched("audio"),
-            "speech_attention_mask": MultiModalFieldConfig.batched("audio"),
+            "speech_attention_mask": MultiModalFieldConfig.batched(
+                "audio", keep_on_cpu=True
+            ),
             "input_features": MultiModalFieldConfig.batched("audio"),
             "feature_attention_mask": MultiModalFieldConfig.batched("audio"),
             "feature_exist_mask": MultiModalFieldConfig.batched("audio"),
diff --git a/vllm/model_executor/models/glmasr.py b/vllm/model_executor/models/glmasr.py
index 1ff7224d7e..454687f469 100644
--- a/vllm/model_executor/models/glmasr.py
+++ b/vllm/model_executor/models/glmasr.py
@@ -625,13 +625,13 @@ def _glmasr_field_config(
             feature_attention_mask=MultiModalFieldConfig.flat_from_sizes(
                 "audio", chunk_counts, dim=0
             ),
-            chunk_counts=MultiModalFieldConfig.batched("audio"),
+            chunk_counts=MultiModalFieldConfig.batched("audio", keep_on_cpu=True),
         )
     return dict(
         audio_embeds=MultiModalFieldConfig.batched("audio"),
         input_features=MultiModalFieldConfig.batched("audio"),
         feature_attention_mask=MultiModalFieldConfig.batched("audio"),
-        chunk_counts=MultiModalFieldConfig.batched("audio"),
+        chunk_counts=MultiModalFieldConfig.batched("audio", keep_on_cpu=True),
     )
 
 
diff --git a/vllm/model_executor/models/granite4_vision.py b/vllm/model_executor/models/granite4_vision.py
index 5c56717aa6..e82a42bf87 100644
--- a/vllm/model_executor/models/granite4_vision.py
+++ b/vllm/model_executor/models/granite4_vision.py
@@ -398,7 +398,7 @@ class Granite4VisionMultiModalProcessor(
     ) -> Mapping[str, MultiModalFieldConfig]:
         return dict(
             pixel_values=MultiModalFieldConfig.batched("image"),
-            image_sizes=MultiModalFieldConfig.batched("image"),
+            image_sizes=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
         )
 
 
diff --git a/vllm/model_executor/models/hyperclovax_vision.py b/vllm/model_executor/models/hyperclovax_vision.py
index 593be5d88f..647103aa44 100644
--- a/vllm/model_executor/models/hyperclovax_vision.py
+++ b/vllm/model_executor/models/hyperclovax_vision.py
@@ -324,10 +324,14 @@ class HCXVisionMultiModalProcessor(BaseMultiModalProcessor[HCXVisionProcessingIn
     ) -> Mapping[str, MultiModalFieldConfig]:
         fields = dict(
             pixel_values_images=MultiModalFieldConfig.batched("image"),
-            image_sizes_images=MultiModalFieldConfig.batched("image"),
-            vision_query_lengths_images=MultiModalFieldConfig.batched("image"),
+            image_sizes_images=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
+            vision_query_lengths_images=MultiModalFieldConfig.batched(
+                "image", keep_on_cpu=True
+            ),
             pixel_values_videos=MultiModalFieldConfig.batched("video"),
-            vision_query_lengths_videos=MultiModalFieldConfig.batched("video"),
+            vision_query_lengths_videos=MultiModalFieldConfig.batched(
+                "video", keep_on_cpu=True
+            ),
         )
 
         return fields
diff --git a/vllm/model_executor/models/isaac.py b/vllm/model_executor/models/isaac.py
index 87932b5032..4979971b02 100644
--- a/vllm/model_executor/models/isaac.py
+++ b/vllm/model_executor/models/isaac.py
@@ -440,7 +440,7 @@ class IsaacMultiModalProcessor(BaseMultiModalProcessor):
             "pixel_values": MultiModalFieldConfig.flat_from_sizes(
                 "image", image_grid_sizes
             ),
-            "image_grid_thw": MultiModalFieldConfig.batched("image"),
+            "image_grid_thw": MultiModalFieldConfig.batched("image", keep_on_cpu=True),
         }
 
     def _get_prompt_updates(
diff --git a/vllm/model_executor/models/kanana_v.py b/vllm/model_executor/models/kanana_v.py
index b1a5f78b1b..7a1421b515 100644
--- a/vllm/model_executor/models/kanana_v.py
+++ b/vllm/model_executor/models/kanana_v.py
@@ -561,8 +561,8 @@ class KananaVMultiModalProcessor(BaseMultiModalProcessor[KananaVProcessingInfo])
 
         mm_fields_config = dict(
             pixel_values=MultiModalFieldConfig.flat_from_sizes("image", pixel_sizes),
-            vision_grid_thw=MultiModalFieldConfig.batched("image"),
-            image_token_thw=MultiModalFieldConfig.batched("image"),
+            vision_grid_thw=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
+            image_token_thw=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
         )
         return mm_fields_config
 
diff --git a/vllm/model_executor/models/keye_vl1_5.py b/vllm/model_executor/models/keye_vl1_5.py
index c7ecc68db1..a4dcd729db 100644
--- a/vllm/model_executor/models/keye_vl1_5.py
+++ b/vllm/model_executor/models/keye_vl1_5.py
@@ -309,13 +309,15 @@ def _keye_field_config(
     return dict(
         pixel_values=MultiModalFieldConfig.flat_from_sizes("image", image_grid_sizes),
         image_embeds=MultiModalFieldConfig.flat_from_sizes("image", image_grid_sizes),
-        image_grid_thw=MultiModalFieldConfig.batched("image"),
+        image_grid_thw=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
         pixel_values_videos=MultiModalFieldConfig.flat_from_sizes(
             "video", video_num_patches
         ),
         video_embeds=MultiModalFieldConfig.flat_from_sizes("video", video_num_patches),
-        video_grid_thw=MultiModalFieldConfig.flat_from_sizes("video", video_num_grids),
-        num_frames=MultiModalFieldConfig.batched("video"),
+        video_grid_thw=MultiModalFieldConfig.flat_from_sizes(
+            "video", video_num_grids, keep_on_cpu=True
+        ),
+        num_frames=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
     )
 
 
diff --git a/vllm/model_executor/models/kimi_audio.py b/vllm/model_executor/models/kimi_audio.py
index cb3c83e708..900894612c 100644
--- a/vllm/model_executor/models/kimi_audio.py
+++ b/vllm/model_executor/models/kimi_audio.py
@@ -201,7 +201,7 @@ class KimiAudioDummyInputsBuilder(BaseDummyInputsBuilder[KimiAudioProcessingInfo
 # Field config for Kimi-Audio multimodal data
 _KIMIAUDIO_FIELD_CONFIG = {
     "whisper_input_features": MultiModalFieldConfig.batched("audio"),
-    "feature_attention_mask": MultiModalFieldConfig.batched("audio"),
+    "feature_attention_mask": MultiModalFieldConfig.batched("audio", keep_on_cpu=True),
 }
 
 
diff --git a/vllm/model_executor/models/kimi_vl.py b/vllm/model_executor/models/kimi_vl.py
index 2b08fc6c1f..e5fa22de7e 100644
--- a/vllm/model_executor/models/kimi_vl.py
+++ b/vllm/model_executor/models/kimi_vl.py
@@ -251,7 +251,7 @@ class KimiVLMultiModalProcessor(BaseMultiModalProcessor[KimiVLProcessingInfo]):
             pixel_values=MultiModalFieldConfig.flat_from_sizes(
                 "image", image_grid_sizes
             ),
-            image_grid_hws=MultiModalFieldConfig.batched("image"),
+            image_grid_hws=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
         )
 
     def _get_prompt_updates(
diff --git a/vllm/model_executor/models/llava_onevision2.py b/vllm/model_executor/models/llava_onevision2.py
index 01f073a0d9..6c13f5b14b 100644
--- a/vllm/model_executor/models/llava_onevision2.py
+++ b/vllm/model_executor/models/llava_onevision2.py
@@ -431,20 +431,20 @@ def _create_field_factory(
             image_embeds=MultiModalFieldConfig.flat_from_sizes(
                 "image", image_embed_grid_sizes
             ),
-            image_grid_thw=MultiModalFieldConfig.batched("image"),
+            image_grid_thw=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
             # OV2 first-class MM kwarg: per-patch (t,h,w)
             # positions required by the 3-D vision RoPE.
             patch_positions=MultiModalFieldConfig.flat_from_sizes(
-                "image", image_pixel_grid_sizes
+                "image", image_pixel_grid_sizes, keep_on_cpu=True
             ),
             pixel_values_videos=MultiModalFieldConfig.flat_from_sizes(
                 "video", video_patch_sizes
             ),
             video_grid_thw=MultiModalFieldConfig.flat_from_sizes(
-                "video", video_num_frames
+                "video", video_num_frames, keep_on_cpu=True
             ),
             patch_positions_videos=MultiModalFieldConfig.flat_from_sizes(
-                "video", video_patch_sizes
+                "video", video_patch_sizes, keep_on_cpu=True
             ),
             video_num_frames=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
             frame_timestamps=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
diff --git a/vllm/model_executor/models/mimo_v2_omni.py b/vllm/model_executor/models/mimo_v2_omni.py
index 589ca4f61e..bd2bdbf901 100644
--- a/vllm/model_executor/models/mimo_v2_omni.py
+++ b/vllm/model_executor/models/mimo_v2_omni.py
@@ -893,17 +893,21 @@ class MiMoV2OmniMultiModalProcessor(BaseMultiModalProcessor[MiMoV2OmniProcessing
         merge_size = self.info.get_hf_config().vision_config.spatial_merge_size
         fields: dict[str, MultiModalFieldConfig] = dict(
             **_create_qwen2vl_field_factory(merge_size)(hf_inputs),
-            second_per_grid_ts=MultiModalFieldConfig.batched("video"),
-            video_start_times=MultiModalFieldConfig.batched("video"),
+            second_per_grid_ts=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
+            video_start_times=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
             audio_features=MultiModalFieldConfig.batched("audio"),
-            audio_token_lens=MultiModalFieldConfig.batched("audio"),
+            audio_token_lens=MultiModalFieldConfig.batched("audio", keep_on_cpu=True),
         )
         # video_audio fields: only present when video_audio content was processed
         if "video_audio_n_segs" in hf_inputs:
-            fields["video_audio_n_segs"] = MultiModalFieldConfig.batched("video")
+            fields["video_audio_n_segs"] = MultiModalFieldConfig.batched(
+                "video", keep_on_cpu=True
+            )
         # video_audio_seg_lens: list of per-video 1D tensors, batched("video")
         if "video_audio_seg_lens" in hf_inputs:
-            fields["video_audio_seg_lens"] = MultiModalFieldConfig.batched("video")
+            fields["video_audio_seg_lens"] = MultiModalFieldConfig.batched(
+                "video", keep_on_cpu=True
+            )
         if "va_audio_features" in hf_inputs:
             fields["va_audio_features"] = MultiModalFieldConfig.batched("va_audio")
         return fields
diff --git a/vllm/model_executor/models/minicpmv.py b/vllm/model_executor/models/minicpmv.py
index e85ac81ee5..5d8eecfaaa 100644
--- a/vllm/model_executor/models/minicpmv.py
+++ b/vllm/model_executor/models/minicpmv.py
@@ -472,12 +472,12 @@ def get_version_by_config(config: PretrainedConfig) -> tuple[int, ...]:
 def _minicpmv_field_config(hf_inputs: Mapping[str, torch.Tensor]):
     return dict(
         pixel_values=MultiModalFieldConfig.batched("image"),
-        image_sizes=MultiModalFieldConfig.batched("image"),
-        tgt_sizes=MultiModalFieldConfig.batched("image"),
+        image_sizes=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
+        tgt_sizes=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
         image_embeds=MultiModalFieldConfig.batched("image"),
         video_pixel_values=MultiModalFieldConfig.batched("video"),
-        video_image_sizes=MultiModalFieldConfig.batched("video"),
-        video_tgt_sizes=MultiModalFieldConfig.batched("video"),
+        video_image_sizes=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
+        video_tgt_sizes=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
         video_embeds=MultiModalFieldConfig.batched("video"),
     )
 
diff --git a/vllm/model_executor/models/minicpmv4_6.py b/vllm/model_executor/models/minicpmv4_6.py
index ac9efabaa3..a7e8d7c75f 100644
--- a/vllm/model_executor/models/minicpmv4_6.py
+++ b/vllm/model_executor/models/minicpmv4_6.py
@@ -75,15 +75,17 @@ from .vision import is_vit_use_data_parallel
 def _minicpmv4_6_field_config(hf_inputs: Mapping[str, torch.Tensor]):
     fields = dict(
         pixel_values=MultiModalFieldConfig.batched("image"),
-        tgt_sizes=MultiModalFieldConfig.batched("image"),
+        tgt_sizes=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
         image_embeds=MultiModalFieldConfig.batched("image"),
         video_pixel_values=MultiModalFieldConfig.batched("video"),
-        video_image_sizes=MultiModalFieldConfig.batched("video"),
-        video_tgt_sizes=MultiModalFieldConfig.batched("video"),
+        video_image_sizes=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
+        video_tgt_sizes=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
         video_embeds=MultiModalFieldConfig.batched("video"),
     )
     if "use_vit_merger" in hf_inputs:
-        fields["use_vit_merger"] = MultiModalFieldConfig.batched("image")
+        fields["use_vit_merger"] = MultiModalFieldConfig.batched(
+            "image", keep_on_cpu=True
+        )
     return fields
 
 
diff --git a/vllm/model_executor/models/mllama4.py b/vllm/model_executor/models/mllama4.py
index 178dae506c..fe5ff17a3b 100644
--- a/vllm/model_executor/models/mllama4.py
+++ b/vllm/model_executor/models/mllama4.py
@@ -653,8 +653,8 @@ class Mllama4MultiModalProcessor(BaseMultiModalProcessor[Mllama4ProcessingInfo])
             pixel_values=MultiModalFieldConfig.flat_from_sizes(
                 "image", patches_per_image
             ),
-            patches_per_image=MultiModalFieldConfig.batched("image"),
-            aspect_ratios=MultiModalFieldConfig.batched("image"),
+            patches_per_image=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
+            aspect_ratios=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
         )
 
     def _get_prompt_updates(
diff --git a/vllm/model_executor/models/molmo.py b/vllm/model_executor/models/molmo.py
index bdc46d83b9..3fccb7a86c 100644
--- a/vllm/model_executor/models/molmo.py
+++ b/vllm/model_executor/models/molmo.py
@@ -1208,8 +1208,10 @@ class MolmoMultiModalProcessor(BaseMultiModalProcessor[MolmoProcessingInfo]):
             images=MultiModalFieldConfig.flat_from_sizes("image", num_crops),
             image_masks=MultiModalFieldConfig.flat_from_sizes("image", num_crops),
             image_input_idx=MultiModalFieldConfig.flat_from_sizes("image", num_crops),
-            num_crops=MultiModalFieldConfig.batched("image"),
-            img_patch_id=MultiModalFieldConfig.shared("image", num_images),
+            num_crops=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
+            img_patch_id=MultiModalFieldConfig.shared(
+                "image", num_images, keep_on_cpu=True
+            ),
         )
 
     def _get_prompt_updates(
diff --git a/vllm/model_executor/models/molmo2.py b/vllm/model_executor/models/molmo2.py
index d8de4f5ed1..42f5ebfdc5 100644
--- a/vllm/model_executor/models/molmo2.py
+++ b/vllm/model_executor/models/molmo2.py
@@ -2127,26 +2127,30 @@ class Molmo2MultiModalProcessor(BaseMultiModalProcessor[Molmo2ProcessingInfo]):
             image_token_pooling=MultiModalFieldConfig.flat_from_sizes(
                 "image", image_num_pooled_patches
             ),
-            image_num_crops=MultiModalFieldConfig.batched("image"),
-            image_num_pooled_patches=MultiModalFieldConfig.batched("image"),
-            image_num_patches=MultiModalFieldConfig.batched("image"),
+            image_num_crops=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
+            image_num_pooled_patches=MultiModalFieldConfig.batched(
+                "image", keep_on_cpu=True
+            ),
+            image_num_patches=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
             image_tokens=MultiModalFieldConfig.flat_from_sizes(
                 "image", num_image_tokens
             ),
-            num_image_tokens=MultiModalFieldConfig.batched("image"),
+            num_image_tokens=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
             pixel_values_videos=MultiModalFieldConfig.flat_from_sizes(
                 "video", video_num_crops
             ),
             video_token_pooling=MultiModalFieldConfig.flat_from_sizes(
                 "video", video_num_pooled_patches
             ),
-            video_num_crops=MultiModalFieldConfig.batched("video"),
-            video_num_pooled_patches=MultiModalFieldConfig.batched("video"),
-            video_num_patches=MultiModalFieldConfig.batched("video"),
+            video_num_crops=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
+            video_num_pooled_patches=MultiModalFieldConfig.batched(
+                "video", keep_on_cpu=True
+            ),
+            video_num_patches=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
             video_tokens=MultiModalFieldConfig.flat_from_sizes(
                 "video", num_video_tokens
             ),
-            num_video_tokens=MultiModalFieldConfig.batched("video"),
+            num_video_tokens=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
         )
 
     def _get_prompt_updates(
diff --git a/vllm/model_executor/models/moss_transcribe_diarize.py b/vllm/model_executor/models/moss_transcribe_diarize.py
index 819881e0f2..347c61cb54 100644
--- a/vllm/model_executor/models/moss_transcribe_diarize.py
+++ b/vllm/model_executor/models/moss_transcribe_diarize.py
@@ -352,9 +352,12 @@ def _mtd_field_config(
             audio_feature_lengths=MultiModalFieldConfig.flat_from_sizes(
                 "audio",
                 audio_chunk_counts,
+                keep_on_cpu=True,
+            ),
+            audio_chunk_counts=MultiModalFieldConfig.batched("audio", keep_on_cpu=True),
+            audio_token_lengths=MultiModalFieldConfig.batched(
+                "audio", keep_on_cpu=True
             ),
-            audio_chunk_counts=MultiModalFieldConfig.batched("audio"),
-            audio_token_lengths=MultiModalFieldConfig.batched("audio"),
         )
     return fields
 
diff --git a/vllm/model_executor/models/muse_glimmer.py b/vllm/model_executor/models/muse_glimmer.py
index 74e0001318..83ec17a75d 100644
--- a/vllm/model_executor/models/muse_glimmer.py
+++ b/vllm/model_executor/models/muse_glimmer.py
@@ -378,12 +378,16 @@ class MuseGlimmerMultiModalProcessor(
         if "image_pixel_values" in hf_inputs:
             fields.update(
                 image_pixel_values=MultiModalFieldConfig.batched("image"),
-                image_feature_sizes=MultiModalFieldConfig.batched("image"),
+                image_feature_sizes=MultiModalFieldConfig.batched(
+                    "image", keep_on_cpu=True
+                ),
             )
         if "video_pixel_values" in hf_inputs:
             fields.update(
                 video_pixel_values=MultiModalFieldConfig.batched("video"),
-                video_feature_sizes=MultiModalFieldConfig.batched("video"),
+                video_feature_sizes=MultiModalFieldConfig.batched(
+                    "video", keep_on_cpu=True
+                ),
             )
         return fields
 
diff --git a/vllm/model_executor/models/nano_nemotron_vl.py b/vllm/model_executor/models/nano_nemotron_vl.py
index c339b0e435..7ecf8677bb 100644
--- a/vllm/model_executor/models/nano_nemotron_vl.py
+++ b/vllm/model_executor/models/nano_nemotron_vl.py
@@ -386,10 +386,12 @@ class NanoNemotronVLMultiModalProcessor(
 
         return dict(
             pixel_values_flat=pixel_values_flat,
-            image_num_patches=MultiModalFieldConfig.batched("image"),
+            image_num_patches=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
             image_embeds=MultiModalFieldConfig.batched("image"),
-            num_tokens_per_image=MultiModalFieldConfig.batched("image"),
-            imgs_sizes=MultiModalFieldConfig.batched("image"),
+            num_tokens_per_image=MultiModalFieldConfig.batched(
+                "image", keep_on_cpu=True
+            ),
+            imgs_sizes=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
         )
 
     def _get_video_fields_config(self, hf_inputs: BatchFeature):
@@ -399,9 +401,9 @@ class NanoNemotronVLMultiModalProcessor(
             pixel_values_flat_video=MultiModalFieldConfig.flat_from_sizes(
                 "video", video_num_patches
             ),
-            video_num_patches=MultiModalFieldConfig.batched("video"),
-            frames_indices=MultiModalFieldConfig.batched("video"),
-            frame_duration_ms=MultiModalFieldConfig.batched("video"),
+            video_num_patches=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
+            frames_indices=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
+            frame_duration_ms=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
         )
 
     def _get_audio_fields_config(self, hf_inputs: BatchFeature):
diff --git a/vllm/model_executor/models/ovis.py b/vllm/model_executor/models/ovis.py
index f25585fd76..5ec5f8cd18 100644
--- a/vllm/model_executor/models/ovis.py
+++ b/vllm/model_executor/models/ovis.py
@@ -386,7 +386,7 @@ class OvisMultiModalProcessor(BaseMultiModalProcessor[OvisProcessingInfo]):
     ) -> Mapping[str, MultiModalFieldConfig]:
         return dict(
             pixel_values=MultiModalFieldConfig.batched("image"),
-            grids=MultiModalFieldConfig.batched("image"),
+            grids=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
             indicator_tokens=MultiModalFieldConfig.batched("image"),
         )
 
diff --git a/vllm/model_executor/models/paddleocr_vl.py b/vllm/model_executor/models/paddleocr_vl.py
index bd635aaada..5f5302e7b7 100644
--- a/vllm/model_executor/models/paddleocr_vl.py
+++ b/vllm/model_executor/models/paddleocr_vl.py
@@ -279,7 +279,7 @@ class PaddleOCRVLMultiModalProcessor(
     ) -> Mapping[str, MultiModalFieldConfig]:
         return dict(
             pixel_values=MultiModalFieldConfig.batched("image"),
-            image_grid_thw=MultiModalFieldConfig.batched("image"),
+            image_grid_thw=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
         )
 
     def _get_prompt_updates(
diff --git a/vllm/model_executor/models/phi4mm.py b/vllm/model_executor/models/phi4mm.py
index 9a8da5ed32..359cc1613b 100644
--- a/vllm/model_executor/models/phi4mm.py
+++ b/vllm/model_executor/models/phi4mm.py
@@ -918,7 +918,7 @@ class Phi4MMMultiModalProcessor(BaseMultiModalProcessor[Phi4MMProcessingInfo]):
             input_image_embeds=MultiModalFieldConfig.batched("image"),
             image_attention_mask=MultiModalFieldConfig.batched("image"),
             image_sizes=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
-            num_img_tokens=MultiModalFieldConfig.batched("image"),
+            num_img_tokens=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
             input_audio_embeds=MultiModalFieldConfig.batched("audio"),
         )
 
diff --git a/vllm/model_executor/models/qwen2_5_omni_thinker.py b/vllm/model_executor/models/qwen2_5_omni_thinker.py
index a512be850d..149ab7bc15 100644
--- a/vllm/model_executor/models/qwen2_5_omni_thinker.py
+++ b/vllm/model_executor/models/qwen2_5_omni_thinker.py
@@ -301,8 +301,12 @@ def create_qwen2_5_omni_thinker_field_factory(
             input_audio_features=MultiModalFieldConfig.flat_from_sizes(
                 "audio", audio_feature_lengths, dim=1
             ),
-            feature_attention_mask=MultiModalFieldConfig.batched("audio"),
-            audio_feature_lengths=MultiModalFieldConfig.batched("audio"),
+            feature_attention_mask=MultiModalFieldConfig.batched(
+                "audio", keep_on_cpu=True
+            ),
+            audio_feature_lengths=MultiModalFieldConfig.batched(
+                "audio", keep_on_cpu=True
+            ),
             pixel_values=MultiModalFieldConfig.flat_from_sizes(
                 "image", image_pixel_grid_sizes
             ),
@@ -318,7 +322,9 @@ def create_qwen2_5_omni_thinker_field_factory(
             ),
             video_grid_thw=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
             second_per_grid_ts=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
-            use_audio_in_video=MultiModalFieldConfig.shared("video", num_videos),
+            use_audio_in_video=MultiModalFieldConfig.shared(
+                "video", num_videos, keep_on_cpu=True
+            ),
         )
 
     return _qwen2_5_omni_thinker_field_config
@@ -1031,7 +1037,11 @@ class Qwen2_5OmniConditionalGenerationMixin:
         audio_input: Qwen2_5OmniAudioFeatureInputs,
     ) -> torch.Tensor:
         input_features = audio_input["input_features"]
-        audio_feature_lengths = audio_input["audio_feature_lengths"]
+        # audio_feature_lengths is keep_on_cpu; the audio tower derives
+        # device placement from feature_lens, so move it explicitly.
+        audio_feature_lengths = audio_input["audio_feature_lengths"].to(
+            input_features.device, non_blocking=True
+        )
 
         audio_feat_lengths, audio_output_lengths = (
             self.audio_tower._get_feat_extract_output_lengths(audio_feature_lengths)
diff --git a/vllm/model_executor/models/qwen2_audio.py b/vllm/model_executor/models/qwen2_audio.py
index 115a7f7f79..c989ec85b4 100644
--- a/vllm/model_executor/models/qwen2_audio.py
+++ b/vllm/model_executor/models/qwen2_audio.py
@@ -130,7 +130,7 @@ def _qwen2audio_field_config(hf_inputs: Mapping[str, torch.Tensor]):
     return dict(
         audio_embeds=MultiModalFieldConfig.batched("audio"),
         input_features=MultiModalFieldConfig.batched("audio"),
-        feature_attention_mask=MultiModalFieldConfig.batched("audio"),
+        feature_attention_mask=MultiModalFieldConfig.batched("audio", keep_on_cpu=True),
     )
 
 
diff --git a/vllm/model_executor/models/qwen3_asr.py b/vllm/model_executor/models/qwen3_asr.py
index f5ec1a4526..328fcdbdac 100644
--- a/vllm/model_executor/models/qwen3_asr.py
+++ b/vllm/model_executor/models/qwen3_asr.py
@@ -254,8 +254,8 @@ def _qwen3asr_field_config(hf_inputs: Mapping[str, torch.Tensor]):
         input_audio_features=MultiModalFieldConfig.flat_from_sizes(
             "audio", audio_feature_lengths, dim=1
         ),
-        feature_attention_mask=MultiModalFieldConfig.batched("audio"),
-        audio_feature_lengths=MultiModalFieldConfig.batched("audio"),
+        feature_attention_mask=MultiModalFieldConfig.batched("audio", keep_on_cpu=True),
+        audio_feature_lengths=MultiModalFieldConfig.batched("audio", keep_on_cpu=True),
     )
 
 
@@ -459,7 +459,11 @@ class Qwen3ASRForConditionalGeneration(
         audio_input: Qwen2_5OmniAudioFeatureInputs,
     ) -> torch.Tensor:
         input_features = audio_input["input_features"]
-        audio_feature_lengths = audio_input["audio_feature_lengths"]
+        # audio_feature_lengths is keep_on_cpu; the audio tower derives
+        # device placement from feature_lens, so move it explicitly.
+        audio_feature_lengths = audio_input["audio_feature_lengths"].to(
+            input_features.device, non_blocking=True
+        )
 
         audio_output_lengths = _get_feat_extract_output_lengths(audio_feature_lengths)
 
diff --git a/vllm/model_executor/models/qwen3_omni_moe_thinker.py b/vllm/model_executor/models/qwen3_omni_moe_thinker.py
index 48a16e97d1..0df29a7e71 100755
--- a/vllm/model_executor/models/qwen3_omni_moe_thinker.py
+++ b/vllm/model_executor/models/qwen3_omni_moe_thinker.py
@@ -1553,7 +1553,11 @@ class Qwen3OmniMoeConditionalGenerationMixin(Qwen2_5OmniConditionalGenerationMix
         audio_input: Qwen2_5OmniAudioFeatureInputs,
     ) -> tuple[torch.Tensor, ...]:
         input_features = audio_input["input_features"]
-        audio_feature_lengths = audio_input["audio_feature_lengths"]
+        # audio_feature_lengths is keep_on_cpu; the audio tower derives
+        # device placement from feature_lens, so move it explicitly.
+        audio_feature_lengths = audio_input["audio_feature_lengths"].to(
+            input_features.device, non_blocking=True
+        )
 
         audio_output_lengths = _get_feat_extract_output_lengths(audio_feature_lengths)
 
diff --git a/vllm/model_executor/models/step3_vl.py b/vllm/model_executor/models/step3_vl.py
index f9b22705ed..a06c221968 100644
--- a/vllm/model_executor/models/step3_vl.py
+++ b/vllm/model_executor/models/step3_vl.py
@@ -202,7 +202,7 @@ class Step3VLMultiModalProcessor(BaseMultiModalProcessor[Step3VLProcessingInfo])
             ),
             num_patches=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
             patch_newline_mask=MultiModalFieldConfig.flat_from_sizes(
-                "image", num_patches
+                "image", num_patches, keep_on_cpu=True
             ),
         )
 
diff --git a/vllm/model_executor/models/transformers/multimodal.py b/vllm/model_executor/models/transformers/multimodal.py
index f328a18a3c..e94490232c 100644
--- a/vllm/model_executor/models/transformers/multimodal.py
+++ b/vllm/model_executor/models/transformers/multimodal.py
@@ -300,11 +300,17 @@ class MultiModalProcessor(BaseMultiModalProcessor[MultiModalProcessingInfo]):
 
         # Keep these as batched, as they always have batch size as first dim
         if "audio" in modalities:
-            mm_fields["num_audio_tokens"] = MultiModalFieldConfig.batched("audio")
+            mm_fields["num_audio_tokens"] = MultiModalFieldConfig.batched(
+                "audio", keep_on_cpu=True
+            )
         if "image" in modalities:
-            mm_fields["image_grid_thw"] = MultiModalFieldConfig.batched("image")
+            mm_fields["image_grid_thw"] = MultiModalFieldConfig.batched(
+                "image", keep_on_cpu=True
+            )
             # TODO: route to "video" once the video modality is supported
-            mm_fields["video_grid_thw"] = MultiModalFieldConfig.batched("image")
+            mm_fields["video_grid_thw"] = MultiModalFieldConfig.batched(
+                "image", keep_on_cpu=True
+            )
             mm_fields["num_image_patches"] = MultiModalFieldConfig.batched(
                 "image", keep_on_cpu=True
             )
@@ -755,6 +761,12 @@ class MultiModalMixin(SupportsMultiModal, SupportsMRoPE):
 
         num_image_patches = kwargs.pop("num_image_patches")
 
+        # grid_thw fields are registered keep_on_cpu; restore the on-device
+        # placement that HF get_image_features implementations expect.
+        for key, value in kwargs.items():
+            if isinstance(value, torch.Tensor):
+                kwargs[key] = value.to(pixel_values.device, non_blocking=True)
+
         # The underlying HuggingFace `get_image_features` implementations
         # contain model-internal syncs (e.g. Idefics3 filters all-zero
         # padding images via boolean-mask indexing, LlavaOnevision
diff --git a/vllm/model_executor/models/ultravox.py b/vllm/model_executor/models/ultravox.py
index 551c2c5694..6bd41f3003 100644
--- a/vllm/model_executor/models/ultravox.py
+++ b/vllm/model_executor/models/ultravox.py
@@ -251,7 +251,7 @@ class UltravoxMultiModalProcessor(BaseMultiModalProcessor[UltravoxProcessingInfo
                 "audio", num_chunks, keep_on_cpu=True
             ),
             # num_chunks can convert audio_chunked to audio batch dimension
-            audio_num_chunks=MultiModalFieldConfig.batched("audio"),
+            audio_num_chunks=MultiModalFieldConfig.batched("audio", keep_on_cpu=True),
             audio_embeds=MultiModalFieldConfig.batched("audio"),
         )
 
diff --git a/vllm/models/inkling/common/mm_preprocess.py b/vllm/models/inkling/common/mm_preprocess.py
index eb3c7b591f..6688d37cd8 100644
--- a/vllm/models/inkling/common/mm_preprocess.py
+++ b/vllm/models/inkling/common/mm_preprocess.py
@@ -307,12 +307,12 @@ class InklingMultiModalProcessor(BaseMultiModalProcessor[InklingProcessingInfo])
         return dict(
             # Ragged per-image patches, grouped by num_patches.
             pixel_values=MultiModalFieldConfig.flat_from_sizes("image", num_patches),
-            num_patches=MultiModalFieldConfig.batched("image"),
+            num_patches=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
             # Ragged per-audio frames, grouped by num_audio_tokens.
             input_audio_features=MultiModalFieldConfig.flat_from_sizes(
                 "audio", num_audio_tokens
             ),
-            num_audio_tokens=MultiModalFieldConfig.batched("audio"),
+            num_audio_tokens=MultiModalFieldConfig.batched("audio", keep_on_cpu=True),
         )
 
     def _get_prompt_updates(