Skip some tests on Blackwell (#9777)

Signed-off-by: Hao Lu <14827759+hlu1@users.noreply.github.com>

sgl-kernel/tests/test_cutlass_w4a8_moe_mm.py

 import pytest
 import torch
 from sgl_kernel import cutlass_w4a8_moe_mm
+from utils import is_hopper
 
 
 def pack_int4_values_to_int8(int4_values_interleaved: torch.Tensor) -> torch.Tensor:
@@ -38,6 +39,10 @@ def pack_interleave(num_experts, ref_weight, ref_scale):
     return w_q, w_scale
 
 
+@pytest.mark.skipif(
+    not is_hopper(),
+    reason="cutlass_w4a8_moe_mm is only supported on sm90",
+)
 @pytest.mark.parametrize("batch_size", [1, 2, 4, 8, 16])
 def test_int4_fp8_grouped_gemm_single_expert(batch_size):
     # Test parameters
@@ -127,6 +132,10 @@ def test_int4_fp8_grouped_gemm_single_expert(batch_size):
         raise
 
 
+@pytest.mark.skipif(
+    not is_hopper(),
+    reason="cutlass_w4a8_moe_mm is only supported on sm90",
+)
 @pytest.mark.parametrize("batch_size", [2, 4, 8, 16])
 @pytest.mark.parametrize("k", [512, 1024])
 @pytest.mark.parametrize("n", [1024, 2048])

sgl-kernel/tests/test_int8_gemm.py

 import pytest
 import torch
 from sgl_kernel import int8_scaled_mm
+from utils import is_sm10x
 
 
 def to_int8(tensor: torch.Tensor) -> torch.Tensor:
@@ -30,6 +31,10 @@ def _test_accuracy_once(M, N, K, with_bias, out_dtype, device):
     torch.testing.assert_close(o, o1)
 
 
+@pytest.mark.skipif(
+    is_sm10x(),
+    reason="int8_scaled_mm is only supported on sm90 and lower",
+)
 @pytest.mark.parametrize("M", [1, 16, 32, 64, 128, 512, 1024, 4096, 8192])
 @pytest.mark.parametrize("N", [16, 128, 512, 1024, 4096, 8192, 16384])
 @pytest.mark.parametrize("K", [512, 1024, 4096, 8192, 16384])

sgl-kernel/tests/utils.py

+import torch
+
+
+def is_sm10x():
+    return torch.cuda.get_device_capability() >= (10, 0)
+
+
+def is_hopper():
+    return torch.cuda.get_device_capability() == (9, 0)