[BUG] Fix async rlhf tests (#35811)

Signed-off-by: ahao-anyscale <ahao@anyscale.com>

vllm/v1/worker/gpu_worker.py

@@ -1006,6 +1006,10 @@ class Worker(WorkerBase):
                 load_weights=load_weights_direct,
             )
 
+        # NCCL broadcast/packed path are asynchronous.
+        # Sync here so the next step uses the new weights.
+        torch.accelerator.synchronize()
+
     def shutdown(self) -> None:
         # has_kv_transfer_group can be None during interpreter shutdown.
         if ensure_kv_transfer_shutdown is not None: