[NPU][1/N] NPU basic functions refactor and new modelslim quant type (#13359)

python/sglang/srt/eplb/expert_distribution.py

@@ -31,9 +31,7 @@ from sglang.srt.environ import envs
 from sglang.srt.metrics.collector import ExpertDispatchCollector
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.server_args import ServerArgs
-from sglang.srt.utils import Withable, get_int_env_var, is_npu
-
-_is_npu = is_npu()
+from sglang.srt.utils import Withable, get_int_env_var
 
 if TYPE_CHECKING:
     from sglang.srt.eplb.expert_location import ExpertLocationMetadata
@@ -465,10 +463,6 @@ 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)
-        if not _is_npu:
-            device = "cuda"
-        else:
-            device = "npu"
         self._enable_global_physical_experts = enable_global_physical_experts
         self._data = torch.zeros(
             (
@@ -480,7 +474,7 @@ class _LayerBasedGpuSinglePassGatherer(_SinglePassGatherer):
                 ),
             ),
             dtype=torch.int,
-            device=device,
+            device="cuda",
         )
 
     def reset(self):