diff --git a/python/sglang/srt/layers/attention/linear/kernels/gdn_flashinfer.py b/python/sglang/srt/layers/attention/linear/kernels/gdn_flashinfer.py
index c60a0e1b1..24e716cf9 100644
--- a/python/sglang/srt/layers/attention/linear/kernels/gdn_flashinfer.py
+++ b/python/sglang/srt/layers/attention/linear/kernels/gdn_flashinfer.py
@@ -17,6 +17,7 @@ import torch
 from sglang.srt.layers.attention.linear.kernels.kernel_backend import (
     LinearAttnKernelBase,
 )
+from sglang.srt.utils import is_cuda
 
 logger = logging.getLogger(__name__)
 
@@ -54,7 +55,7 @@ def _get_flashinfer_gdn_kernels():
             _flashinfer_gated_delta_rule_mtp_bf16 = gated_delta_rule_mtp_bf16
             _flashinfer_gated_delta_rule_decode = gated_delta_rule_decode_pretranspose
             _flashinfer_gdn_available = (
-                torch.cuda.is_available() and torch.cuda.get_device_capability()[0] >= 9
+                is_cuda() and torch.cuda.get_device_capability()[0] >= 9
             )
             if _flashinfer_gdn_available:
                 logger.info("FlashInfer GDN kernels loaded successfully")
diff --git a/python/sglang/srt/layers/moe/utils.py b/python/sglang/srt/layers/moe/utils.py
index e29f7efec..f7eda3c36 100644
--- a/python/sglang/srt/layers/moe/utils.py
+++ b/python/sglang/srt/layers/moe/utils.py
@@ -14,7 +14,7 @@ from sglang.srt.layers.dp_attention import (
     get_attention_dp_size,
     is_dp_attention_enabled,
 )
-from sglang.srt.utils import is_npu
+from sglang.srt.utils import is_cuda, is_npu
 
 _is_npu = is_npu()
 
@@ -288,7 +288,7 @@ def initialize_moe_config(server_args: ServerArgs):
     DEEPEP_CONFIG = server_args.deepep_config or ""
     IS_TBO_ENABLED = server_args.enable_two_batch_overlap
     IS_SBO_ENABLED = server_args.enable_single_batch_overlap
-    if IS_SBO_ENABLED and torch.cuda.is_available():
+    if IS_SBO_ENABLED and is_cuda():
         if torch.cuda.get_device_capability()[0] == 9:
             raise ValueError(
                 "SBO (single batch overlap) is not supported on SM90 GPUs with latest sgl-deep-gemm wheel. Please try removing --enable-single-batch-overlap argument."
diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py
index f2f30f4c1..cfe1237f2 100644
--- a/python/sglang/srt/server_args.py
+++ b/python/sglang/srt/server_args.py
@@ -3399,7 +3399,7 @@ class ServerArgs:
         if (
             decode == "flashinfer"
             and self.mamba_ssm_dtype != "bfloat16"
-            and torch.cuda.is_available()
+            and is_cuda()
             and torch.cuda.get_device_capability()[0] >= 10
         ):
             raise ValueError(
@@ -3415,7 +3415,7 @@ class ServerArgs:
         cuda_major = int(cuda_version.split(".")[0]) if cuda_version is not None else 0
         if (
             prefill == "flashinfer"
-            and torch.cuda.is_available()
+            and is_cuda()
             and torch.cuda.get_device_capability()[0] >= 10
             and cuda_major < 13
         ):