[VLM] feat: true on policy for vlm + fsdp (#14636)

Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com>

python/sglang/srt/multimodal/processors/base_processor.py

@@ -17,6 +17,7 @@ from sglang.srt.managers.schedule_batch import (
     MultimodalDataItem,
     MultimodalInputFormat,
 )
+from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import envs, is_npu, load_audio, load_image, load_video, logger
 from sglang.srt.utils.cuda_ipc_transport_utils import (
     MM_FEATURE_CACHE_SIZE,
@@ -316,7 +317,9 @@ class BaseMultimodalProcessor(ABC):
             and isinstance(processor.image_processor, BaseImageProcessorFast)
             and not self.server_args.disable_fast_image_processor
         ):
-            if not _is_npu:
+            if get_global_server_args().rl_on_policy_target is not None:
+                kwargs["device"] = "cpu"
+            elif not _is_npu:
                 kwargs["device"] = "cuda"
             elif processor.__class__.__name__ not in {
                 "Qwen2_5_VLProcessor",