Replace hardcoded CUDA device with get_device() for XPU support (#13599)

Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>

python/sglang/srt/eplb/expert_distribution.py

@@ -32,7 +32,7 @@ from sglang.srt.environ import envs
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.observability.metrics_collector import ExpertDispatchCollector
 from sglang.srt.server_args import ServerArgs
-from sglang.srt.utils import Withable, get_int_env_var
+from sglang.srt.utils import Withable, get_device, get_int_env_var
 
 if TYPE_CHECKING:
     from sglang.srt.eplb.expert_location import ExpertLocationMetadata
@@ -475,6 +475,9 @@ def _list_sum(a: List, b: List) -> List:
 class _LayerBasedGpuSinglePassGatherer(_SinglePassGatherer):
     def __init__(self, *args, enable_global_physical_experts: bool, **kwargs):
         super().__init__(*args, **kwargs)
+
+        device = get_device()
+
         self._enable_global_physical_experts = enable_global_physical_experts
         self._data = torch.zeros(
             (
@@ -486,7 +489,7 @@ class _LayerBasedGpuSinglePassGatherer(_SinglePassGatherer):
                 ),
             ),
             dtype=torch.int,
-            device="cuda",
+            device=device,
         )
 
     def reset(self):