vlm: remove redundant d2h movement of mm feature tensors (#9987)

Co-authored-by: Xiang (Kevin) Li <lik@nvidia.com>

python/sglang/srt/server_args.py

@@ -381,6 +381,7 @@ class ServerArgs:
     disable_shared_experts_fusion: bool = False
     disable_chunked_prefix_cache: bool = False
     disable_fast_image_processor: bool = False
+    keep_mm_feature_on_device: bool = False
     enable_return_hidden_states: bool = False
     scheduler_recv_interval: int = 1
     numa_node: Optional[List[int]] = None
@@ -2213,6 +2214,11 @@ class ServerArgs:
             action="store_true",
             help="Adopt base image processor instead of fast image processor.",
         )
+        parser.add_argument(
+            "--keep-mm-feature-on-device",
+            action="store_true",
+            help="Keep multimodal feature tensors on device after processing to save D2H copy.",
+        )
         parser.add_argument(
             "--enable-return-hidden-states",
             action="store_true",