Support sanity checking weight consistency especially for RL (#13854)

python/sglang/srt/model_executor/model_runner.py

@@ -170,6 +170,7 @@ from sglang.srt.utils.offloader import (
 )
 from sglang.srt.utils.patch_torch import monkey_patch_torch_reductions
 from sglang.srt.utils.torch_memory_saver_adapter import TorchMemorySaverAdapter
+from sglang.srt.utils.weight_checker import WeightChecker
 from sglang.srt.weight_sync.tensor_bucket import (
     FlattenedTensorBucket,
     FlattenedTensorMetadata,
@@ -328,6 +329,8 @@ class ModelRunner:
         # CPU offload
         set_offloader(create_offloader_from_server_args(server_args, dp_rank=dp_rank))
 
+        self._weight_checker = WeightChecker(model_runner=self)
+
         if get_bool_env_var("SGLANG_DETECT_SLOW_RANK"):
             slow_rank_detector.execute()
         # Init mindspore running environment when model impl is "mindspore"
@@ -2508,6 +2511,9 @@ class ModelRunner:
         )
         ShardedStateLoader.save_model(self.model, path, pattern, max_size)
 
+    def check_weights(self, action: str):
+        self._weight_checker.handle(action=action)
+
     def update_weights_from_ipc(self, recv_req):
         """Update weights from IPC for checkpoint-engine integration."""
         try: