diff --git a/tests/unit/test_p3_fast_decode.py b/tests/unit/test_p3_fast_decode.py
index 0df5ca3a..e1115ac7 100644
--- a/tests/unit/test_p3_fast_decode.py
+++ b/tests/unit/test_p3_fast_decode.py
@@ -88,10 +88,7 @@ def test_kernel_correctness():
             # Use the correct kernel for the KV size
             if N > 128 or hd == 512:
                 from dsv4.kernels.attention.fmha_multitile_op import fmha_multitile_decode_raw as kernel_fn
-                # Multi-tile kernel needs T in the Q shape
-                q_4d_mt = torch.randn(1, n_q, 1, hd, dtype=torch.bfloat16, device='cuda').contiguous()
-                o_4d, _ = kernel_fn(q_4d_mt, k_4d, v_4d, scale, 0, 0, False,
-                                     torch.zeros(1, n_q, dtype=torch.float32, device='cuda'))
+                o_4d, _ = kernel_fn(q_4d, k_4d, v_4d, scale)
             else:
                 sb = torch.zeros(1, n_q, dtype=torch.float32, device='cuda')
                 o_4d, _ = fmha_multihead_decode_raw(q_4d, k_4d, v_4d, scale, 0, 0, False, sb)