diff --git a/tests/unit/test_minimal_pv.cu b/tests/unit/test_minimal_pv.cu
index 55c532bd..5d0c77da 100644
--- a/tests/unit/test_minimal_pv.cu
+++ b/tests/unit/test_minimal_pv.cu
@@ -25,10 +25,16 @@ test_minimal_pv()
 
     extern __shared__ char sbuf[];
     uint32_t* sTmemBase = (uint32_t*)sbuf;
-    bf16_t* sPk = (bf16_t*)(((uintptr_t)(sbuf + 4) + 15) & ~(uintptr_t)15);
+    bf16_t* sQ0 = (bf16_t*)(((uintptr_t)(sbuf + 4) + 15) & ~(uintptr_t)15);
+    bf16_t* sK0 = sQ0 + TILE_SZ;
+    bf16_t* sPk = (bf16_t*)(((uintptr_t)(sK0 + TILE_SZ) + 127) & ~(uintptr_t)127);
     bf16_t* sV = (bf16_t*)(((uintptr_t)(sPk + TILE_SZ) + 127) & ~(uintptr_t)127);
     float* s_p_vals = (float*)(sV + 256);  // 16 P values for 1 K-tile
 
+    // Fill sQ0, sK0 (will be used by QK)
+    for (int i = tid; i < TILE_SZ; i += 128) { sQ0[i] = 0; sK0[i] = 0; }
+    __syncthreads();
+
     // Set s_p_vals = all 0.5
     for (int j = tid; j < 16; j += 128) s_p_vals[j] = 0.5f;
     __syncthreads();
@@ -59,6 +65,16 @@ test_minimal_pv()
     __syncthreads();
     uint32_t tb = *sTmemBase;
 
+    // QK GEMM (SS) — uses sQ0 and sK0
+    {
+        uint64_t dq = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sQ0), BLOCK_MN);
+        uint64_t dk = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sK0), BLOCK_MN);
+        uint32_t idesc = make_idesc(BLOCK_MN, BLOCK_MN);
+        if (tid == 0) umma_ss_f16(tb, dq, dk, idesc, false);
+        asm volatile("tcgen05.fence::after_thread_sync;" ::: "memory");
+        __syncthreads();
+    }
+
     // PV SS MMA
     uint64_t dp = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sPk), BLOCK_MN);
     uint64_t dv = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sV), 16);
@@ -83,7 +99,7 @@ test_minimal_pv()
 
 int main() {
     printf("=== Minimal PV with s_p_vals ===\n");
-    int smem = (4+16 + TILE_SZ*2 + 256*2 + 16*4 + 256 + 127) & ~127;
+    int smem = (4+16 + TILE_SZ*2 + TILE_SZ*2 + TILE_SZ*2 + 256*2 + 16*4 + 256 + 127) & ~127;
     printf("SMEM: %d bytes\n", smem);
     test_minimal_pv<<<1, 128, smem>>>();
     cudaError_t err = cudaDeviceSynchronize();