debug: add LSE verification and merge debug prints

tests/unit/test_fmha_6warp_multirow.cu

@@ -144,6 +144,9 @@ static int test_single(int T, int n_h = 1, int batch = 1) {
                 cs /= (sqrtf(na)*sqrtf(nb)+1e-10f);
                 if(cs<min_cos) min_cos=cs;
                 if(cs<0.999f) { printf("  FAIL b=%d h=%d t=%d cos=%.6f\n",b,h,t,cs); failed++; }
+                // Check LSE
+                float lse_err = fabsf(h_lse[idx*T+t] - lse_ref[t]);
+                if(lse_err > 0.01f) { printf("  FAIL LSE b=%d h=%d t=%d kernel=%.6f ref=%.6f err=%.6f\n",b,h,t,h_lse[idx*T+t],lse_ref[t],lse_err); failed++; }
             }
         }
     }
@@ -218,6 +221,13 @@ static int test_multitile_merge(int T) {
         free(h_o_tile);
     }
 
+    // Debug: print first LSE values per tile
+    for (int tile = 0; tile < N_TILES; tile++) {
+        printf("  tile %d lse[0]=%.6f", tile, lse_per_tile[tile * T]);
+        if (T > 1) printf(" lse[1]=%.6f", lse_per_tile[tile * T + 1]);
+        printf("\n");
+    }
+
     // Python KV merge with normalized O + LSE:
     // O = Σ exp(lse_i - L) * O_i_norm / Σ exp(lse_i - L)
     // where L = max(lse_i) for numerical stability