diff --git a/src/transform/loop_vectorize.cc b/src/transform/loop_vectorize.cc
index e8a18b004..5525a6e42 100644
--- a/src/transform/loop_vectorize.cc
+++ b/src/transform/loop_vectorize.cc
@@ -85,6 +85,8 @@ class VectorizePlanner : public arith::IRMutatorWithAnalyzer {
         !disable_vectorize_256 &&
         VectorizeFindGlobalAccess().HasGlobalAccess(node)) {
       vector_load_bits_max_ = vector_size_ = 256;
+    } else if (tvm::tl::TargetIsSunmmio(Target::Current(false))) {
+      vector_load_bits_max_ = vector_size_ = 8192;
     } else {
       vector_load_bits_max_ = vector_size_ = 128;
     }
diff --git a/testing/python/transform/test_tilelang_mesh_transform_legalize_vectorized_loop.py b/testing/python/transform/test_tilelang_mesh_transform_legalize_vectorized_loop.py
new file mode 100644
index 000000000..7dcd39c16
--- /dev/null
+++ b/testing/python/transform/test_tilelang_mesh_transform_legalize_vectorized_loop.py
@@ -0,0 +1,52 @@
+from tilelang import tvm as tvm
+import tilelang as tl
+import tilelang.language as T
+import tilelang.testing
+from tilelang.utils.target import determine_target
+
+
+def vectorize_access_legalize(M: int = 64, N: int = 1024):
+    dtype = "float32"
+    vec_len = 512
+
+    @T.prim_func
+    def main(A: T.Tensor((M, N, vec_len), dtype="float32"),):
+        with T.Kernel(1, 1, threads=M) as (bx, by):
+            A_shared = T.alloc_shared((M, N, vec_len), dtype=dtype)
+            tid = T.get_thread_binding()
+            for j in T.serial(N):
+                for v in T.vectorized(vec_len):
+                    A_shared[tid, j, v] = A[tid, j, v]
+
+    @T.prim_func
+    def expected(A: T.Tensor((M, N, vec_len), dtype="float32"),):
+        with T.Kernel(1, 1, threads=M) as (bx, by):
+            A_shared = T.alloc_shared((M, N, vec_len), dtype=dtype)
+            tid = T.get_thread_binding()
+            for j, v_2 in T.grid(M, vec_len // 256):
+                for vec in T.vectorized(256):
+                    A_shared[tid, j, v_2 * 256 + vec] = A[tid, j, v_2 * 256 + vec]
+
+    return main, expected
+
+
+def assert_vectorize_access(M: int = 64, N: int = 1024):
+    func, expected = vectorize_access_legalize(M, N)
+    mod = tvm.IRModule({func.attrs["global_symbol"]: func})
+    mod.show()
+    target = determine_target("Sunmmio", return_object=True)
+    # mod = tvm.tir.transform.BindTarget(target)(mod)
+    # transformed = tl.transform.LegalizeVectorizedLoop()(mod)
+    with tvm.target.Target(target):
+        transformed = tl.transform.LegalizeVectorizedLoop()(mod)
+    transformed.show()
+    print("Expected:{}".format(expected.body))
+    tvm.ir.assert_structural_equal(transformed["main"].body, expected.body)
+
+
+def test_vectorize_access():
+    assert_vectorize_access(1024, 1024)
+
+
+if __name__ == "__main__":
+    tilelang.testing.main()