diff --git a/sgl-kernel/python/sgl_kernel/attention.py b/sgl-kernel/python/sgl_kernel/attention.py
index d80a6fbbd..52b2d62af 100644
--- a/sgl-kernel/python/sgl_kernel/attention.py
+++ b/sgl-kernel/python/sgl_kernel/attention.py
@@ -74,9 +74,11 @@ def cutlass_mla_decode(
         f"but got D_q = {D_q}, D_ckv = {D_ckv}, D_latent = {D_latent}, D_rope = {D_rope}"
     )
     assert H == 128, f"H must be 128, but got {H}"
-    # TODO: There is currently an illegal memory access issue with page size !=
-    # 128. Change this when it is fixed.
-    assert PAGE_SIZE == 128, f"PAGE_SIZE must be 128, but got {PAGE_SIZE}"
+
+    assert len(page_table.shape) == 2
+    B_block_table, block_num = page_table.shape
+    assert B_block_table == B_q
+    assert block_num % (128 / PAGE_SIZE) == 0
 
     # TODO(kaixih@nvidia): support fp8
     assert q_nope_and_q_pe.dtype in (
diff --git a/sgl-kernel/tests/test_cutlass_mla.py b/sgl-kernel/tests/test_cutlass_mla.py
index 26a59ad7c..a8aab8d15 100644
--- a/sgl-kernel/tests/test_cutlass_mla.py
+++ b/sgl-kernel/tests/test_cutlass_mla.py
@@ -39,7 +39,7 @@ def ref_mla(
 @pytest.mark.parametrize("mean_seq_len", [128, 1024, 4096])
 @pytest.mark.parametrize("bs", [1, 2, 4])
 @pytest.mark.parametrize("varlen", [False, True])
-@pytest.mark.parametrize("block_size", [128])
+@pytest.mark.parametrize("block_size", [1, 16, 64, 128])
 def test_cutlass_mla_decode(
     dtype: torch.dtype, mean_seq_len: int, bs: int, varlen: bool, block_size: int
 ):
@@ -62,6 +62,11 @@ def test_cutlass_mla_decode(
     max_seq_len = seq_lens.max().item()
     block_num = (max_seq_len + block_size - 1) // block_size
 
+    # Pad block_num so that small blocks can be packed into full 128-sized CUTLASS tiles.
+    # One 128-wide tile can hold (128 // block_size) small blocks.
+    pack_factor = 128 // block_size
+    block_num = ((block_num + pack_factor - 1) // pack_factor) * pack_factor
+
     q = torch.randn(bs, h_q, d)
     block_table = torch.randint(0, bs * block_num, (bs, block_num), dtype=torch.int32)