diff --git a/tests/kernels/quantization/test_per_token_group_quant.py b/tests/kernels/quantization/test_per_token_group_quant.py
index a2788d0583..d88a39d084 100644
--- a/tests/kernels/quantization/test_per_token_group_quant.py
+++ b/tests/kernels/quantization/test_per_token_group_quant.py
@@ -413,9 +413,12 @@ def test_per_token_group_quant_fp8_packed_large_mn():
 
 @pytest.mark.parametrize("shape", [(32, 128), (64, 256), (16, 512)])
 @pytest.mark.parametrize("group_size", [64, 128])
-@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
+@pytest.mark.skipif(
+    not (current_platform.is_cuda_alike() or current_platform.is_xpu()),
+    reason="Only test on CUDA/ROCm/XPU.",
+)
 def test_per_token_group_quant_int8(shape, group_size: int):
-    device = "cuda"
+    device = current_platform.device_type
 
     torch.manual_seed(42)
     num_tokens, hidden_dim = shape