diff --git a/tests/quant/test_per_channel_cast_fused.py b/tests/quant/test_per_channel_cast_fused.py
index 5acf0c1..eaff846 100644
--- a/tests/quant/test_per_channel_cast_fused.py
+++ b/tests/quant/test_per_channel_cast_fused.py
@@ -13,7 +13,7 @@
 os.environ['TILELANG_PRINT_ON_COMPILATION'] = '0'
 
 
-def generate_test_data(params):
+def generate_test_data(params, alignment=128):
     num_send_tokens = params['num_send_tokens']
     num_topk = params['num_topk']
     num_experts = params['num_experts']
@@ -28,7 +28,7 @@ def generate_test_data(params):
         topk_idx = generate_topk_idx(params)
         num_tokens = topk_idx.shape[0]
         _, pos_to_token, _, token_topk_to_pos, _, _, _, _ = (
-            tile_kernels.moe.get_fused_mapping(topk_idx, num_experts, 0, 128)
+            tile_kernels.moe.get_fused_mapping(topk_idx, num_experts, 0, alignment)
         )
         x = torch.randn((num_tokens, hidden), dtype=torch.bfloat16, device='cuda')
         x = tile_kernels.moe.expand_to_fused(x, token_topk_to_pos, pos_to_token)
@@ -74,8 +74,10 @@ def generate_test_params(is_benchmark: bool) -> list[dict]:
 
 
 @pytest.mark.parametrize('params', generate_test_params(is_benchmark=False), ids=make_param_id)
-def test_per_channel_cast_fused(params):
-    _, _, _, func, func_ref = generate_test_data(params)
+@pytest.mark.parametrize('alignment', [16, 128])
+# use alignment=16 to cover expand tails issues.
+def test_per_channel_cast_fused(params, alignment):
+    _, _, _, func, func_ref = generate_test_data(params, alignment)
 
     x_fp8, x_fp8_sf = func()
     x_fp8_ref, x_fp8_sf_ref = func_ref()
diff --git a/tile_kernels/quant/per_channel_cast_fused_kernel.py b/tile_kernels/quant/per_channel_cast_fused_kernel.py
index cce18cf..91cb31f 100644
--- a/tile_kernels/quant/per_channel_cast_fused_kernel.py
+++ b/tile_kernels/quant/per_channel_cast_fused_kernel.py
@@ -71,8 +71,12 @@ def per_channel_cast_fused_kernel(
             T.assume(num_tokens_out % 128 == 0 or (with_expand and num_tokens_out % 16 == 0))
             if with_expand:
                 tmp = T.alloc_var(T.int32)
-                if k_id < VEC_M:
-                    tmp = pos_to_token[k_id + m_offset]
+                # Invalid tail rows remain -1 and follow the existing zero-fill path,
+                # so they do not participate in the amax reduction.
+                tmp = -1
+                row = k_id + m_offset
+                if k_id < VEC_M and row < num_tokens_out:
+                    tmp = pos_to_token[row]
 
                 for i in T.serial(VEC_M):
                     pos_to_token_local[i] = T.shfl_sync(tmp, i)
@@ -128,8 +132,10 @@ def per_channel_cast_fused_kernel(
                         out_local[j] = in_local[j] * sf_invs_local[i] * amax_local[j]
                     else:
                         out_local[j] = in_local[j] * amax_local[j]
-                for j in T.vectorized(VEC_K):
-                    out[i + m_offset, j + k_offset] = out_local[j]
+                out_row = i + m_offset
+                if out_row < num_tokens_out:
+                    for j in T.vectorized(VEC_K):
+                        out[out_row, j + k_offset] = out_local[j]
 
     return per_channel_cast_fused_kernel