Fix broadcast use cuda device lead to memory capacity unbalanced (#5416)

python/sglang/srt/utils.py

@@ -846,9 +846,12 @@ def broadcast_pyobj(
     rank: int,
     dist_group: Optional[torch.distributed.ProcessGroup] = None,
     src: int = 0,
+    force_cpu_device: bool = True,
 ):
     """Broadcast inputs from rank=0 to all other ranks with torch.dist backend."""
-    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    device = torch.device(
+        "cuda" if torch.cuda.is_available() and not force_cpu_device else "cpu"
+    )
 
     if rank == 0:
         if len(data) == 0: