diff --git a/python/sglang/srt/models/deepseek_v2.py b/python/sglang/srt/models/deepseek_v2.py
index 8024ad25b..4d49cd03b 100644
--- a/python/sglang/srt/models/deepseek_v2.py
+++ b/python/sglang/srt/models/deepseek_v2.py
@@ -1753,7 +1753,7 @@ class DeepseekV2DecoderLayer(nn.Module):
             hidden_states,
             residual,
             forward_batch,
-            self._gfx95_quant_format,
+            getattr(self, "_gfx95_quant_format", ""),
         )
 
         hidden_states = self.self_attn(
diff --git a/python/sglang/srt/models/glm4_moe_lite.py b/python/sglang/srt/models/glm4_moe_lite.py
index 5b08f1f45..80a035162 100644
--- a/python/sglang/srt/models/glm4_moe_lite.py
+++ b/python/sglang/srt/models/glm4_moe_lite.py
@@ -20,7 +20,6 @@ from typing import Iterable, Optional, Tuple
 
 import torch
 import torch.nn.functional as F
-from sgl_kernel import dsv3_router_gemm
 from torch import nn
 from transformers import PretrainedConfig
 
@@ -81,6 +80,9 @@ from sglang.srt.utils.hf_transformers_utils import get_rope_config
 _is_cuda = is_cuda()
 _device_sm = get_device_sm()
 
+if _is_cuda:
+    from sgl_kernel import dsv3_router_gemm
+
 logger = logging.getLogger(__name__)