From 79f837cdc0245ea59d974e2db8762da6e287d1fe Mon Sep 17 00:00:00 2001
From: vorapolsiloai <115975949+vorapolsiloai@users.noreply.github.com>
Date: Wed, 16 Sep 2026 19:37:53 +0000
Subject: [PATCH] [Triton/Gluon] Enable unified-attention skip-mask for gfx950
 hd256 FP8 prefill

Long FP8 prefill at head size 256 matches `D_GEQ_256` (`BLOCK_M=16`, no `SPLIT_UNMASKED_LOOP`) because D outranks Q in the lookup, so it never reaches the hd128 skip key. Add the same `Q_GEQ_256` family under `D_GEQ_256`, with `SW` and `SHUF` siblings left skip-off, so the existing kernel fast path fires only for non-windowed, non-shuffled prefill.
---
 .../attention/unified_attention/DEFAULT.json  | 25 +++++++++++++++++++
 1 file changed, 25 insertions(+)

diff --git a/aiter/ops/triton/configs/gfx950/triton/attention/unified_attention/DEFAULT.json b/aiter/ops/triton/configs/gfx950/triton/attention/unified_attention/DEFAULT.json
index 658f90ac639..f7285bc9383 100644
--- a/aiter/ops/triton/configs/gfx950/triton/attention/unified_attention/DEFAULT.json
+++ b/aiter/ops/triton/configs/gfx950/triton/attention/unified_attention/DEFAULT.json
@@ -132,6 +132,31 @@
       "TILE_SIZE_MAX": 64,
       "SPLIT_UNMASKED_LOOP": true
     },
+    "D_GEQ_256.Q_GEQ_256.SW.DT_fp8_fp8": {
+      "BLOCK_M": 128,
+      "num_warps": 4,
+      "num_stages": 1,
+      "waves_per_eu": 2,
+      "TILE_SIZE_MIN": 64,
+      "TILE_SIZE_MAX": 64
+    },
+    "D_GEQ_256.Q_GEQ_256.SHUF.DT_fp8_fp8": {
+      "BLOCK_M": 128,
+      "num_warps": 4,
+      "num_stages": 1,
+      "waves_per_eu": 2,
+      "TILE_SIZE_MIN": 64,
+      "TILE_SIZE_MAX": 64
+    },
+    "D_GEQ_256.Q_GEQ_256.DT_fp8_fp8": {
+      "BLOCK_M": 128,
+      "num_warps": 4,
+      "num_stages": 1,
+      "waves_per_eu": 2,
+      "TILE_SIZE_MIN": 64,
+      "TILE_SIZE_MAX": 64,
+      "SPLIT_UNMASKED_LOOP": true
+    },
     "Q_GEQ_256": {
       "BLOCK_M": 128,
       "num_warps": 4,