diff --git a/python/sglang/srt/managers/schedule_batch.py b/python/sglang/srt/managers/schedule_batch.py
index d423e61d7..50288f3e2 100644
--- a/python/sglang/srt/managers/schedule_batch.py
+++ b/python/sglang/srt/managers/schedule_batch.py
@@ -1574,6 +1574,9 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
                     mm_item.feature = pixel_values.reconstruct_on_target_device(
                         torch.cuda.current_device()
                     )
+                    # The reference by CudaIpcTensorTransportProxy was cut off,
+                    # proactively delete to avoid slow gc.
+                    del pixel_values
         self.multimodal_inputs = multimodal_inputs
         self.token_type_ids = token_type_ids_tensor
         self.seq_lens_sum = sum(seq_lens)
diff --git a/python/sglang/srt/multimodal/processors/base_processor.py b/python/sglang/srt/multimodal/processors/base_processor.py
index 0fa304431..306ef5d19 100644
--- a/python/sglang/srt/multimodal/processors/base_processor.py
+++ b/python/sglang/srt/multimodal/processors/base_processor.py
@@ -880,6 +880,8 @@ class BaseMultimodalProcessor(ABC):
                             info_data=item.feature,
                             sync_buffer_meta=sync_flag,
                         )
+                    elif not self.server_args.keep_mm_feature_on_device:
+                        item.feature = item.feature.cpu()
                 elif (
                     isinstance(item.precomputed_embeddings, torch.Tensor)
                     and item.precomputed_embeddings.is_cuda
@@ -900,5 +902,7 @@ class BaseMultimodalProcessor(ABC):
                             info_data=item.precomputed_embeddings,
                             sync_buffer_meta=sync_flag,
                         )
+                    elif not self.server_args.keep_mm_feature_on_device:
+                        item.precomputed_embeddings = item.precomputed_embeddings.cpu()
 
         return all_collected_items, input_ids, ret