diff --git a/mojo_opset/backends/ttx/kernels/npu/swa.py b/mojo_opset/backends/ttx/kernels/npu/swa.py
index a57bb7c2e..6f3337e1f 100644
--- a/mojo_opset/backends/ttx/kernels/npu/swa.py
+++ b/mojo_opset/backends/ttx/kernels/npu/swa.py
@@ -1930,6 +1930,8 @@ def _sdpa_single_block_bwd_dkdv(
 
     # Load (transposed) K block
     q = tl.load(Q_block_ptr, boundary_check=(0, 1), padding_option="zero")
+    tl.extra.cann.extension.compile_hint(q, "cv_pipeline_lazy_load", True)
+
     # q_T = tl.trans(q)
     # qkT = tl.dot(k, q_T)  # [BLOCK_N, BLOCK_M]
     k_T = tl.trans(k)
@@ -1949,6 +1951,8 @@ def _sdpa_single_block_bwd_dkdv(
     # -- Compute dV ----
     # dv = pT @ do
     do = tl.load(DO_block_ptr, boundary_check=(0, 1), padding_option="zero")
+    tl.extra.cann.extension.compile_hint(do, "cv_pipeline_lazy_load", True)
+
     dv = tl.dot(tl.trans(p_cast), do, dv_ptr)
 
     # -- Compute dS ----
@@ -1990,6 +1994,7 @@ def _sdpa_single_block_bwd_dq(
 
     # Load (transposed) K block
     k = tl.load(K_block_ptr, boundary_check=(0, 1), padding_option="zero")
+    tl.extra.cann.extension.compile_hint(k, "cv_pipeline_lazy_load", True)
     k_T = tl.trans(k)
     qk = tl.dot(q, k_T)  # [BLOCK_M, BLOCK_N]
     # tl.extra.cann.extension.compile_hint(qk, "tile_cube_loop")