diff --git a/dsv4/kernels/attention/fmha_6warp_multihead.cuh b/dsv4/kernels/attention/fmha_6warp_multihead.cuh
index cdc46272..8d5094c7 100644
--- a/dsv4/kernels/attention/fmha_6warp_multihead.cuh
+++ b/dsv4/kernels/attention/fmha_6warp_multihead.cuh
@@ -229,9 +229,16 @@ fmha_6warp_multihead_kernel(FmhaParams params) {
                 *sRowMax = new_max;
             }
 
-            // Store UN-NORMALIZED P for PV (critical for multi-tile!)
-            // P_unnorm[j] = exp(s[j] - max)  (NOT divided by row_sum)
-            if (lane == 0) for (int j=0;j<SK_TILE;j++) s_p_vals[j] = s_vals[j];
+            // Store P for PV. For single KV tile, use NORMALIZED P (same as old kernel).
+            // For multi-tile, use UN-NORMALIZED P (critical for FA2 rescale).
+            // Single-tile: O = P_norm × V. Multi-tile: O = Σ(P_unnorm × V) / running_sum
+            if (lane == 0) {
+                if (n_kv_tiles == 1) {
+                    // Normalized P (backward compatible with single-tile)
+                    for (int j=0;j<SK_TILE;j++) s_vals[j] /= row_sum;
+                }
+                for (int j=0;j<SK_TILE;j++) s_p_vals[j] = s_vals[j];
+            }
 
             // Update running sum
             if (lane == 0) *sRowSum = old_sum + row_sum;
@@ -303,16 +310,25 @@ fmha_6warp_multihead_kernel(FmhaParams params) {
     } // end KV tile loop
 
     // ================================================================
-    // Epilogue: normalize sOacc by running_sum, write to GMEM
+    // Epilogue: write O to GMEM
+    // For single KV tile (normalized P): sOacc = P_norm × V (already normalized)
+    // For multi KV tile (un-normalized P): O = sOacc / running_sum
     // ================================================================
     if (wid == 0) {
         float running_max = *sRowMax;
         float running_sum = *sRowSum;
 
         if (lane == 0) {
-            float inv_sum = 1.0f / running_sum;
-            for (int d = 0; d < HD; d++) {
-                o_head[d] = f32_to_bf16(sOacc[d] * inv_sum);
+            if (n_kv_tiles > 1) {
+                float inv_sum = 1.0f / running_sum;
+                for (int d = 0; d < HD; d++) {
+                    o_head[d] = f32_to_bf16(sOacc[d] * inv_sum);
+                }
+            } else {
+                // Single tile: P was normalized, sOacc is already the correct output
+                for (int d = 0; d < HD; d++) {
+                    o_head[d] = f32_to_bf16(sOacc[d]);
+                }
             }
             if (lse_head) {
                 lse_head[0] = logf(running_sum) + running_max;