From d47b2bfcce860bfd44261a3a82a2cd6147d5ea46 Mon Sep 17 00:00:00 2001
From: biondizzle <biondizzle@gmail.com>
Date: Fri, 29 May 2026 19:57:54 +0000
Subject: [PATCH] fix: use un-normalized P for multi-tile PV (correct online
 softmax merge)

---
 dsv4/kernels/attention/fmha_6warp_tma_multitile.cuh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/dsv4/kernels/attention/fmha_6warp_tma_multitile.cuh b/dsv4/kernels/attention/fmha_6warp_tma_multitile.cuh
index f8f29fb4..c9b87dcc 100644
--- a/dsv4/kernels/attention/fmha_6warp_tma_multitile.cuh
+++ b/dsv4/kernels/attention/fmha_6warp_tma_multitile.cuh
@@ -201,7 +201,7 @@ fmha_6warp_tma_multitile_kernel(FmhaTmaMultiTileParams params) {
                     s_vals[j] = expf(s_vals[j] - row_max);
                     row_sum += s_vals[j];
                 }
-                for (int j=0;j<kv_len;j++) s_vals[j] /= row_sum;
+                // DO NOT normalize P — use un-normalized exp(s-max) for correct multi-tile merge
                 for (int j=0;j<kv_len;j++) s_p_vals[j] = s_vals[j];
                 *sTileMax = row_max;
                 *sTileSum = row_sum;