Skip test_moe_matmul_ogs on older cards

Author: jansel

test/test_examples.py

@@ -1429,6 +1429,10 @@ def _jagged_dense_add_2d_make_precompiler(x_data: torch.Tensor, x_offsets: torch
     return make_precompiler(_jagged_dense_add_2d_kernel)(x_offsets, x_data, y, out, out.size(0), out.size(1), x_offsets.size(0), y.size(0), y.size(1), out.stride(0), out.stride(1), x_data.stride(0), x_offsets.stride(0), y.stride(0), y.stride(1), _BLOCK_SIZE_1, _BLOCK_SIZE_2, num_warps=8, num_stages=4)""",
         )
 
+    @unittest.skipIf(
+        torch.cuda.get_device_capability(0) < (9, 0),
+        "Triton internal error on RTX 3090",
+    )
     def test_moe_matmul_ogs(self):
         mod = import_path(examples_dir / "moe_matmul_ogs.py")
 
@@ -1450,7 +1454,6 @@ def test_moe_matmul_ogs(self):
                 helion_kernel_args,
                 mod.moe_matmul_ogs_reference(*args),
                 block_sizes=[[16, 16], 16],
-                l2_grouping=4,
             ),
             """\
 from __future__ import annotations