From 2de72a2259448ac3a24dc57ce025f8bb195acb99 Mon Sep 17 00:00:00 2001
From: Bartlomiej Kocot <barkocot@amd.com>
Date: Wed, 15 Jul 2026 16:04:06 +0000
Subject: [PATCH 01/11] feat(rocke): register D256 bf16 decode dispatcher
 candidate and tune gfx942 num_segments

---
 .../attention/decode/decode_shapes.json       |   16 +
 .../attention/decode/decode_shapes_perf.csv   | 1597 ++++++++++-------
 .../attention/decode/decode_shapes.json       |   16 +
 .../rocke/library/dispatch/attention.py       |   77 +-
 .../kernels/common/attention_unified.py       |   25 +-
 .../attention/test_d256_decode_wiring.py      |  228 +++
 6 files changed, 1295 insertions(+), 664 deletions(-)
 create mode 100644 dnn-providers/hip-kernel-provider/rocke/library/tests/dispatch/attention/test_d256_decode_wiring.py

diff --git a/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx942/attention/decode/decode_shapes.json b/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx942/attention/decode/decode_shapes.json
index 1a890de7afbb..48d8abc3e7e7 100644
--- a/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx942/attention/decode/decode_shapes.json
+++ b/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx942/attention/decode/decode_shapes.json
@@ -115,6 +115,22 @@
         {"batch": 1, "seqlen_q": 1, "seqlen_k": 16384, "dtype": "fp16", "label": "qwen3_235b_fp16_b1_kv16384"},
         {"batch": 1, "seqlen_q": 1, "seqlen_k": 32768, "dtype": "fp16", "label": "qwen3_235b_fp16_b1_kv32768"}
       ]
+    },
+    {
+      "model": "Qwen3-Next-80B-A3B",
+      "num_query_heads": 16,
+      "num_kv_heads": 2,
+      "head_size": 256,
+      "block_size": 16,
+      "shapes": [
+        {"batch": 1, "seqlen_q": 1, "seqlen_k": 512,   "dtype": "bf16", "label": "qwen3next_80b_b1_kv512"},
+        {"batch": 1, "seqlen_q": 1, "seqlen_k": 1024,  "dtype": "bf16", "label": "qwen3next_80b_b1_kv1024"},
+        {"batch": 1, "seqlen_q": 1, "seqlen_k": 2048,  "dtype": "bf16", "label": "qwen3next_80b_b1_kv2048"},
+        {"batch": 1, "seqlen_q": 1, "seqlen_k": 4096,  "dtype": "bf16", "label": "qwen3next_80b_b1_kv4096"},
+        {"batch": 1, "seqlen_q": 1, "seqlen_k": 8192,  "dtype": "bf16", "label": "qwen3next_80b_b1_kv8192"},
+        {"batch": 1, "seqlen_q": 1, "seqlen_k": 16384, "dtype": "bf16", "label": "qwen3next_80b_b1_kv16384"},
+        {"batch": 1, "seqlen_q": 1, "seqlen_k": 32768, "dtype": "bf16", "label": "qwen3next_80b_b1_kv32768"}
+      ]
     }
   ]
 }
diff --git a/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx942/attention/decode/decode_shapes_perf.csv b/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx942/attention/decode/decode_shapes_perf.csv
index ed1d5f62dd1f..1f643ccb7222 100644
--- a/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx942/attention/decode/decode_shapes_perf.csv
+++ b/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx942/attention/decode/decode_shapes_perf.csv
@@ -12,30 +12,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.0456018590927124,
-    "aoTriton_ms": 0.028075098991394043,
+    "triton_ms": 0.045930638313293456,
+    "aoTriton_ms": 0.03057126045227051,
     "dsl": {
       "30": {
-        "ms": 0.05544295787811279,
+        "ms": 0.05826066017150879,
         "path": "2d"
       },
       "60": {
-        "ms": 0.058234181404113766,
+        "ms": 0.06028773784637451,
         "path": "2d"
       },
       "80": {
-        "ms": 0.054544081687927244,
+        "ms": 0.057419538497924805,
         "path": "2d"
       },
       "120": {
-        "ms": 0.05677802085876465,
+        "ms": 0.05900720119476319,
         "path": "2d"
       }
     },
     "best_sms": 80,
-    "best_ms": 0.054544081687927244,
-    "best_speedup_vs_triton": 0.8360551260835672,
-    "best_speedup_vs_aoTriton": 0.5147231032694821,
+    "best_ms": 0.057419538497924805,
+    "best_speedup_vs_triton": 0.7999130525048269,
+    "best_speedup_vs_aoTriton": 0.5324191251271617,
     "best_path": "2d"
   },
   {
@@ -51,30 +51,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.08846302032470703,
-    "aoTriton_ms": 0.04755035877227783,
+    "triton_ms": 0.08820249557495118,
+    "aoTriton_ms": 0.04751912117004395,
     "dsl": {
       "30": {
-        "ms": 0.02925057888031006,
+        "ms": 0.03284290075302124,
         "path": "3d"
       },
       "60": {
-        "ms": 0.027333381175994872,
+        "ms": 0.0328918194770813,
         "path": "3d"
       },
       "80": {
-        "ms": 0.027101640701293946,
+        "ms": 0.03156235933303833,
         "path": "3d"
       },
       "120": {
-        "ms": 0.026928460597991942,
+        "ms": 0.031061999797821045,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.026928460597991942,
-    "best_speedup_vs_triton": 3.2851124186171905,
-    "best_speedup_vs_aoTriton": 1.765803084036065,
+    "best_ms": 0.031061999797821045,
+    "best_speedup_vs_triton": 2.8395626858879335,
+    "best_speedup_vs_aoTriton": 1.5298152559184983,
     "best_path": "3d"
   },
   {
@@ -90,30 +90,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.08943408012390136,
-    "aoTriton_ms": 0.09122699737548828,
+    "triton_ms": 0.08744312286376953,
+    "aoTriton_ms": 0.09125592231750489,
     "dsl": {
       "30": {
-        "ms": 0.028017361164093018,
+        "ms": 0.03232167959213257,
         "path": "3d"
       },
       "60": {
-        "ms": 0.028060638904571535,
+        "ms": 0.03227358102798462,
         "path": "3d"
       },
       "80": {
-        "ms": 0.02763967990875244,
+        "ms": 0.031833360195159914,
         "path": "3d"
       },
       "120": {
-        "ms": 0.02915518045425415,
+        "ms": 0.03240348100662231,
         "path": "3d"
       }
     },
     "best_sms": 80,
-    "best_ms": 0.02763967990875244,
-    "best_speedup_vs_triton": 3.2357133085170413,
-    "best_speedup_vs_aoTriton": 3.3005808199175326,
+    "best_ms": 0.031833360195159914,
+    "best_speedup_vs_triton": 2.7469020652448988,
+    "best_speedup_vs_aoTriton": 2.866675769005995,
     "best_path": "3d"
   },
   {
@@ -129,30 +129,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.08883748054504395,
-    "aoTriton_ms": 0.1876980781555176,
+    "triton_ms": 0.08760432243347167,
+    "aoTriton_ms": 0.1873253631591797,
     "dsl": {
       "30": {
-        "ms": 0.04282025814056396,
+        "ms": 0.04310736179351807,
         "path": "3d"
       },
       "60": {
-        "ms": 0.029717280864715576,
+        "ms": 0.03197527885437012,
         "path": "3d"
       },
       "80": {
-        "ms": 0.02828197956085205,
+        "ms": 0.03256946086883545,
         "path": "3d"
       },
       "120": {
-        "ms": 0.027237160205841063,
+        "ms": 0.03102992057800293,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.027237160205841063,
-    "best_speedup_vs_triton": 3.2616278596471515,
-    "best_speedup_vs_aoTriton": 6.891249922422726,
+    "best_ms": 0.03102992057800293,
+    "best_speedup_vs_triton": 2.8232209687179886,
+    "best_speedup_vs_aoTriton": 6.036926929551163,
     "best_path": "3d"
   },
   {
@@ -168,30 +168,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.08854080200195312,
-    "aoTriton_ms": 0.37042232513427736,
+    "triton_ms": 0.08857695579528808,
+    "aoTriton_ms": 0.3701787948608398,
     "dsl": {
       "30": {
-        "ms": 0.07365851879119872,
+        "ms": 0.07387186050415039,
         "path": "3d"
       },
       "60": {
-        "ms": 0.04367102146148682,
+        "ms": 0.04401103973388672,
         "path": "3d"
       },
       "80": {
-        "ms": 0.032708160877227783,
+        "ms": 0.03287977933883667,
         "path": "3d"
       },
       "120": {
-        "ms": 0.03271620035171509,
+        "ms": 0.03285814046859741,
         "path": "3d"
       }
     },
-    "best_sms": 80,
-    "best_ms": 0.032708160877227783,
-    "best_speedup_vs_triton": 2.706994206561989,
-    "best_speedup_vs_aoTriton": 11.325073474007961,
+    "best_sms": 120,
+    "best_ms": 0.03285814046859741,
+    "best_speedup_vs_triton": 2.6957385455193745,
+    "best_speedup_vs_aoTriton": 11.265969089596547,
     "best_path": "3d"
   },
   {
@@ -207,30 +207,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.08822810173034668,
-    "aoTriton_ms": 0.7368000030517579,
+    "triton_ms": 0.0883963394165039,
+    "aoTriton_ms": 0.7359786224365235,
     "dsl": {
       "30": {
-        "ms": 0.13634453773498534,
+        "ms": 0.1362869167327881,
         "path": "3d"
       },
       "60": {
-        "ms": 0.07532154083251953,
+        "ms": 0.07543385982513427,
         "path": "3d"
       },
       "80": {
-        "ms": 0.05296204090118408,
+        "ms": 0.052780041694641115,
         "path": "3d"
       },
       "120": {
-        "ms": 0.05289787769317627,
+        "ms": 0.05272071838378906,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.05289787769317627,
-    "best_speedup_vs_triton": 1.6678949246715797,
-    "best_speedup_vs_aoTriton": 13.928725218909939,
+    "best_ms": 0.05272071838378906,
+    "best_speedup_vs_triton": 1.676690722857916,
+    "best_speedup_vs_aoTriton": 13.95995056590176,
     "best_path": "3d"
   },
   {
@@ -246,30 +246,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.08824432373046875,
-    "aoTriton_ms": 1.59133056640625,
+    "triton_ms": 0.0882962989807129,
+    "aoTriton_ms": 1.5905145263671876,
     "dsl": {
       "30": {
-        "ms": 0.2650442886352539,
+        "ms": 0.26536279678344726,
         "path": "3d"
       },
       "60": {
-        "ms": 0.140045862197876,
+        "ms": 0.1399962615966797,
         "path": "3d"
       },
       "80": {
-        "ms": 0.09239447593688965,
+        "ms": 0.0926046371459961,
         "path": "3d"
       },
       "120": {
-        "ms": 0.09227902412414551,
+        "ms": 0.09251482009887696,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.09227902412414551,
-    "best_speedup_vs_triton": 0.9562771666478748,
-    "best_speedup_vs_aoTriton": 17.244770211975684,
+    "best_ms": 0.09251482009887696,
+    "best_speedup_vs_triton": 0.9544016719304492,
+    "best_speedup_vs_aoTriton": 17.191997181287228,
     "best_path": "3d"
   },
   {
@@ -285,30 +285,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.044551458358764645,
-    "aoTriton_ms": 0.028409459590911866,
+    "triton_ms": 0.04450096130371094,
+    "aoTriton_ms": 0.028347740173339842,
     "dsl": {
       "30": {
-        "ms": 0.06578757762908935,
+        "ms": 0.06549816131591797,
         "path": "2d"
       },
       "60": {
-        "ms": 0.06638415813446046,
+        "ms": 0.06666163921356201,
         "path": "2d"
       },
       "80": {
-        "ms": 0.06490876197814942,
+        "ms": 0.06502828121185303,
         "path": "2d"
       },
       "120": {
-        "ms": 0.06615561962127686,
+        "ms": 0.06617331981658936,
         "path": "2d"
       }
     },
     "best_sms": 80,
-    "best_ms": 0.06490876197814942,
-    "best_speedup_vs_triton": 0.6863704837532141,
-    "best_speedup_vs_aoTriton": 0.4376829679862866,
+    "best_ms": 0.06502828121185303,
+    "best_speedup_vs_triton": 0.684332423899273,
+    "best_speedup_vs_aoTriton": 0.4359294086366342,
     "best_path": "2d"
   },
   {
@@ -324,30 +324,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.0886081600189209,
-    "aoTriton_ms": 0.04462681770324707,
+    "triton_ms": 0.08721139907836914,
+    "aoTriton_ms": 0.04458755970001221,
     "dsl": {
       "30": {
-        "ms": 0.02729327917098999,
+        "ms": 0.030773320198059083,
         "path": "3d"
       },
       "60": {
-        "ms": 0.028127200603485107,
+        "ms": 0.03075488090515137,
         "path": "3d"
       },
       "80": {
-        "ms": 0.026252501010894776,
+        "ms": 0.030913639068603515,
         "path": "3d"
       },
       "120": {
-        "ms": 0.02645133972167969,
+        "ms": 0.030938520431518554,
         "path": "3d"
       }
     },
-    "best_sms": 80,
-    "best_ms": 0.026252501010894776,
-    "best_speedup_vs_triton": 3.375227372894817,
-    "best_speedup_vs_aoTriton": 1.6999072844422314,
+    "best_sms": 60,
+    "best_ms": 0.03075488090515137,
+    "best_speedup_vs_triton": 2.8356929538218902,
+    "best_speedup_vs_aoTriton": 1.449771821179249,
     "best_path": "3d"
   },
   {
@@ -363,30 +363,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.08808615684509277,
-    "aoTriton_ms": 0.08586263656616211,
+    "triton_ms": 0.08776388168334961,
+    "aoTriton_ms": 0.08595731735229492,
     "dsl": {
       "30": {
-        "ms": 0.029641098976135254,
+        "ms": 0.033032939434051514,
         "path": "3d"
       },
       "60": {
-        "ms": 0.027211480140686035,
+        "ms": 0.0330786395072937,
         "path": "3d"
       },
       "80": {
-        "ms": 0.026828219890594484,
+        "ms": 0.03195605993270874,
         "path": "3d"
       },
       "120": {
-        "ms": 0.027005419731140137,
+        "ms": 0.031802899837493896,
         "path": "3d"
       }
     },
-    "best_sms": 80,
-    "best_ms": 0.026828219890594484,
-    "best_speedup_vs_triton": 3.2833396030116138,
-    "best_speedup_vs_aoTriton": 3.2004596993878107,
+    "best_sms": 120,
+    "best_ms": 0.031802899837493896,
+    "best_speedup_vs_triton": 2.7596188439357583,
+    "best_speedup_vs_aoTriton": 2.7028138248876257,
     "best_path": "3d"
   },
   {
@@ -402,30 +402,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.0876956558227539,
-    "aoTriton_ms": 0.18424774169921876,
+    "triton_ms": 0.087913818359375,
+    "aoTriton_ms": 0.18398164749145507,
     "dsl": {
       "30": {
-        "ms": 0.041673622131347655,
+        "ms": 0.04194387912750244,
         "path": "3d"
       },
       "60": {
-        "ms": 0.029048540592193604,
+        "ms": 0.03256063938140869,
         "path": "3d"
       },
       "80": {
-        "ms": 0.0271834397315979,
+        "ms": 0.03229363918304443,
         "path": "3d"
       },
       "120": {
-        "ms": 0.026483440399169923,
+        "ms": 0.03085592031478882,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.026483440399169923,
-    "best_speedup_vs_triton": 3.3113392558129484,
-    "best_speedup_vs_aoTriton": 6.957092391402202,
+    "best_ms": 0.03085592031478882,
+    "best_speedup_vs_triton": 2.8491718108708985,
+    "best_speedup_vs_aoTriton": 5.962604440719766,
     "best_path": "3d"
   },
   {
@@ -441,30 +441,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.08839808464050293,
-    "aoTriton_ms": 0.3638062286376953,
+    "triton_ms": 0.08863287925720215,
+    "aoTriton_ms": 0.36303592681884767,
     "dsl": {
       "30": {
-        "ms": 0.07154243946075439,
+        "ms": 0.07187925815582276,
         "path": "3d"
       },
       "60": {
-        "ms": 0.04267111778259278,
+        "ms": 0.04269440174102783,
         "path": "3d"
       },
       "80": {
-        "ms": 0.032234280109405516,
+        "ms": 0.0325734806060791,
         "path": "3d"
       },
       "120": {
-        "ms": 0.03233288049697876,
+        "ms": 0.03335206031799316,
         "path": "3d"
       }
     },
     "best_sms": 80,
-    "best_ms": 0.032234280109405516,
-    "best_speedup_vs_triton": 2.7423626133567534,
-    "best_speedup_vs_aoTriton": 11.286314675026407,
+    "best_ms": 0.0325734806060791,
+    "best_speedup_vs_triton": 2.7210134627326514,
+    "best_speedup_vs_aoTriton": 11.145137702941554,
     "best_path": "3d"
   },
   {
@@ -480,30 +480,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.08855323791503907,
-    "aoTriton_ms": 0.7220821380615234,
+    "triton_ms": 0.09064971923828125,
+    "aoTriton_ms": 0.7208148956298828,
     "dsl": {
       "30": {
-        "ms": 0.1321340560913086,
+        "ms": 0.13234583854675294,
         "path": "3d"
       },
       "60": {
-        "ms": 0.07320467948913574,
+        "ms": 0.07312935829162598,
         "path": "3d"
       },
       "80": {
-        "ms": 0.05165502071380615,
+        "ms": 0.050550899505615234,
         "path": "3d"
       },
       "120": {
-        "ms": 0.051582040786743166,
+        "ms": 0.05050759792327881,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.051582040786743166,
-    "best_speedup_vs_triton": 1.7167455293431833,
-    "best_speedup_vs_aoTriton": 13.998712091420431,
+    "best_ms": 0.05050759792327881,
+    "best_speedup_vs_triton": 1.7947739145302146,
+    "best_speedup_vs_aoTriton": 14.271415099264921,
     "best_path": "3d"
   },
   {
@@ -519,30 +519,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.0887483024597168,
-    "aoTriton_ms": 1.5994796752929688,
+    "triton_ms": 0.08860602378845214,
+    "aoTriton_ms": 1.5988369750976563,
     "dsl": {
       "30": {
-        "ms": 0.2568092918395996,
+        "ms": 0.2565561294555664,
         "path": "3d"
       },
       "60": {
-        "ms": 0.13568061828613281,
+        "ms": 0.13583308219909668,
         "path": "3d"
       },
       "80": {
-        "ms": 0.08837242126464843,
+        "ms": 0.0909199619293213,
         "path": "3d"
       },
       "120": {
-        "ms": 0.08826257705688477,
+        "ms": 0.09075475692749023,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.08826257705688477,
-    "best_speedup_vs_triton": 1.0055031862769992,
-    "best_speedup_vs_aoTriton": 18.12183292883135,
+    "best_ms": 0.09075475692749023,
+    "best_speedup_vs_triton": 0.9763237409058916,
+    "best_speedup_vs_aoTriton": 17.617114840328085,
     "best_path": "3d"
   },
   {
@@ -558,30 +558,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.045001277923583986,
-    "aoTriton_ms": 0.030519120693206787,
+    "triton_ms": 0.04499330043792724,
+    "aoTriton_ms": 0.028747859001159667,
     "dsl": {
       "30": {
-        "ms": 0.055666661262512206,
+        "ms": 0.05783810138702392,
         "path": "2d"
       },
       "60": {
-        "ms": 0.05882996082305908,
+        "ms": 0.06033505916595459,
         "path": "2d"
       },
       "80": {
-        "ms": 0.05562096118927002,
+        "ms": 0.057433180809020996,
         "path": "2d"
       },
       "120": {
-        "ms": 0.05667779922485352,
+        "ms": 0.05944180011749268,
         "path": "2d"
       }
     },
     "best_sms": 80,
-    "best_ms": 0.05562096118927002,
-    "best_speedup_vs_triton": 0.809070482806854,
-    "best_speedup_vs_aoTriton": 0.5486981893274852,
+    "best_ms": 0.057433180809020996,
+    "best_speedup_vs_triton": 0.7834025523945937,
+    "best_speedup_vs_aoTriton": 0.5005444343532556,
     "best_path": "2d"
   },
   {
@@ -597,30 +597,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.09233996391296387,
-    "aoTriton_ms": 0.04828968048095703,
+    "triton_ms": 0.09008119583129882,
+    "aoTriton_ms": 0.04822874069213867,
     "dsl": {
       "30": {
-        "ms": 0.02874703884124756,
+        "ms": 0.03165456056594849,
         "path": "3d"
       },
       "60": {
-        "ms": 0.02782572031021118,
+        "ms": 0.03115499973297119,
         "path": "3d"
       },
       "80": {
-        "ms": 0.027049520015716554,
+        "ms": 0.03141479969024658,
         "path": "3d"
       },
       "120": {
-        "ms": 0.026830620765686035,
+        "ms": 0.030816640853881836,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.026830620765686035,
-    "best_speedup_vs_triton": 3.441588799580009,
-    "best_speedup_vs_aoTriton": 1.7997973622255963,
+    "best_ms": 0.030816640853881836,
+    "best_speedup_vs_triton": 2.9231348172703804,
+    "best_speedup_vs_aoTriton": 1.5650226421762485,
     "best_path": "3d"
   },
   {
@@ -636,30 +636,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.089166259765625,
-    "aoTriton_ms": 0.09671483993530273,
+    "triton_ms": 0.08867557525634766,
+    "aoTriton_ms": 0.09666440010070801,
     "dsl": {
       "30": {
-        "ms": 0.02782249927520752,
+        "ms": 0.03254139900207519,
         "path": "3d"
       },
       "60": {
-        "ms": 0.030800580978393555,
+        "ms": 0.0334154200553894,
         "path": "3d"
       },
       "80": {
-        "ms": 0.03012701988220215,
+        "ms": 0.032741880416870116,
         "path": "3d"
       },
       "120": {
-        "ms": 0.02964751958847046,
+        "ms": 0.03257669925689697,
         "path": "3d"
       }
     },
     "best_sms": 30,
-    "best_ms": 0.02782249927520752,
-    "best_speedup_vs_triton": 3.204825665862469,
-    "best_speedup_vs_aoTriton": 3.476137746600098,
+    "best_ms": 0.03254139900207519,
+    "best_speedup_vs_triton": 2.7250080812657362,
+    "best_speedup_vs_aoTriton": 2.970505358252841,
     "best_path": "3d"
   },
   {
@@ -675,30 +675,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.08993202209472656,
-    "aoTriton_ms": 0.18871723175048827,
+    "triton_ms": 0.08681050300598145,
+    "aoTriton_ms": 0.18812557220458984,
     "dsl": {
       "30": {
-        "ms": 0.0430584192276001,
+        "ms": 0.043102540969848634,
         "path": "3d"
       },
       "60": {
-        "ms": 0.02977260112762451,
+        "ms": 0.0320755410194397,
         "path": "3d"
       },
       "80": {
-        "ms": 0.027494540214538576,
+        "ms": 0.03224152088165283,
         "path": "3d"
       },
       "120": {
-        "ms": 0.026732800006866456,
+        "ms": 0.031440460681915285,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.026732800006866456,
-    "best_speedup_vs_triton": 3.364107840242214,
-    "best_speedup_vs_aoTriton": 7.059388904342803,
+    "best_ms": 0.031440460681915285,
+    "best_speedup_vs_triton": 2.7611078566643044,
+    "best_speedup_vs_aoTriton": 5.983550117406538,
     "best_path": "3d"
   },
   {
@@ -714,30 +714,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.09003625869750977,
-    "aoTriton_ms": 0.372541618347168,
+    "triton_ms": 0.08926254272460937,
+    "aoTriton_ms": 0.3719356536865234,
     "dsl": {
       "30": {
-        "ms": 0.07413884162902833,
+        "ms": 0.07433931827545166,
         "path": "3d"
       },
       "60": {
-        "ms": 0.04388751983642578,
+        "ms": 0.04409604072570801,
         "path": "3d"
       },
       "80": {
-        "ms": 0.03298882007598877,
+        "ms": 0.03336730003356934,
         "path": "3d"
       },
       "120": {
-        "ms": 0.03285650014877319,
+        "ms": 0.033288719654083254,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.03285650014877319,
-    "best_speedup_vs_triton": 2.740287562273171,
-    "best_speedup_vs_aoTriton": 11.338444954889027,
+    "best_ms": 0.033288719654083254,
+    "best_speedup_vs_triton": 2.681465182565538,
+    "best_speedup_vs_aoTriton": 11.173023701465825,
     "best_path": "3d"
   },
   {
@@ -753,30 +753,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.0908372974395752,
-    "aoTriton_ms": 0.8037422943115234,
+    "triton_ms": 0.08850337982177735,
+    "aoTriton_ms": 0.8044700622558594,
     "dsl": {
       "30": {
-        "ms": 0.13675909996032715,
+        "ms": 0.1369837188720703,
         "path": "3d"
       },
       "60": {
-        "ms": 0.07558135986328125,
+        "ms": 0.0756126594543457,
         "path": "3d"
       },
       "80": {
-        "ms": 0.05273752212524414,
+        "ms": 0.05307432174682617,
         "path": "3d"
       },
       "120": {
-        "ms": 0.0527278995513916,
+        "ms": 0.05301018238067627,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.0527278995513916,
-    "best_speedup_vs_triton": 1.7227558505538423,
-    "best_speedup_vs_aoTriton": 15.243207128479499,
+    "best_ms": 0.05301018238067627,
+    "best_speedup_vs_triton": 1.6695543355466622,
+    "best_speedup_vs_aoTriton": 15.175764846059305,
     "best_path": "3d"
   },
   {
@@ -792,30 +792,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.08840510368347168,
-    "aoTriton_ms": 1.6082568359375,
+    "triton_ms": 0.08832216262817383,
+    "aoTriton_ms": 1.5979205322265626,
     "dsl": {
       "30": {
-        "ms": 0.26605863571166993,
+        "ms": 0.2659256935119629,
         "path": "3d"
       },
       "60": {
-        "ms": 0.14066729545593262,
+        "ms": 0.1404693603515625,
         "path": "3d"
       },
       "80": {
-        "ms": 0.09089422225952148,
+        "ms": 0.09310979843139648,
         "path": "3d"
       },
       "120": {
-        "ms": 0.09086855888366699,
+        "ms": 0.09302800178527831,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.09086855888366699,
-    "best_speedup_vs_triton": 0.9728899057005063,
-    "best_speedup_vs_aoTriton": 17.698716208281073,
+    "best_ms": 0.09302800178527831,
+    "best_speedup_vs_triton": 0.949414809876641,
+    "best_speedup_vs_aoTriton": 17.17676937654522,
     "best_path": "3d"
   },
   {
@@ -831,30 +831,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.044979639053344726,
-    "aoTriton_ms": 0.028791940212249754,
+    "triton_ms": 0.045028581619262695,
+    "aoTriton_ms": 0.029026899337768555,
     "dsl": {
       "30": {
-        "ms": 0.06576352119445801,
+        "ms": 0.06568418025970459,
         "path": "2d"
       },
       "60": {
-        "ms": 0.06689174175262451,
+        "ms": 0.0670513391494751,
         "path": "2d"
       },
       "80": {
-        "ms": 0.06504265785217285,
+        "ms": 0.06537146091461182,
         "path": "2d"
       },
       "120": {
-        "ms": 0.07906538009643554,
+        "ms": 0.06670494079589843,
         "path": "2d"
       }
     },
     "best_sms": 80,
-    "best_ms": 0.06504265785217285,
-    "best_speedup_vs_triton": 0.6915406064059252,
-    "best_speedup_vs_aoTriton": 0.44266241821924435,
+    "best_ms": 0.06537146091461182,
+    "best_speedup_vs_triton": 0.6888110039039668,
+    "best_speedup_vs_aoTriton": 0.4440301460553785,
     "best_path": "2d"
   },
   {
@@ -870,30 +870,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.0874855613708496,
-    "aoTriton_ms": 0.04548882007598877,
+    "triton_ms": 0.08888405799865723,
+    "aoTriton_ms": 0.045285181999206545,
     "dsl": {
       "30": {
-        "ms": 0.027396719455718994,
+        "ms": 0.030737240314483643,
         "path": "3d"
       },
       "60": {
-        "ms": 0.02683543920516968,
+        "ms": 0.03109325885772705,
         "path": "3d"
       },
       "80": {
-        "ms": 0.026452159881591795,
+        "ms": 0.031307361125946044,
         "path": "3d"
       },
       "120": {
-        "ms": 0.027401540279388428,
+        "ms": 0.03313397884368897,
         "path": "3d"
       }
     },
-    "best_sms": 80,
-    "best_ms": 0.026452159881591795,
-    "best_speedup_vs_triton": 3.3073125885546792,
-    "best_speedup_vs_aoTriton": 1.7196637355743751,
+    "best_sms": 30,
+    "best_ms": 0.030737240314483643,
+    "best_speedup_vs_triton": 2.8917383958108407,
+    "best_speedup_vs_aoTriton": 1.473300190123698,
     "best_path": "3d"
   },
   {
@@ -909,30 +909,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.0889762020111084,
-    "aoTriton_ms": 0.09491310119628907,
+    "triton_ms": 0.08770294189453125,
+    "aoTriton_ms": 0.09484259605407715,
     "dsl": {
       "30": {
-        "ms": 0.027806479930877686,
+        "ms": 0.03226958036422729,
         "path": "3d"
       },
       "60": {
-        "ms": 0.02758836030960083,
+        "ms": 0.03202100038528442,
         "path": "3d"
       },
       "80": {
-        "ms": 0.027181038856506346,
+        "ms": 0.03232409954071045,
         "path": "3d"
       },
       "120": {
-        "ms": 0.027177019119262694,
+        "ms": 0.03235538005828857,
         "path": "3d"
       }
     },
-    "best_sms": 120,
-    "best_ms": 0.027177019119262694,
-    "best_speedup_vs_triton": 3.2739500097729004,
-    "best_speedup_vs_aoTriton": 3.4924029298347876,
+    "best_sms": 60,
+    "best_ms": 0.03202100038528442,
+    "best_speedup_vs_triton": 2.7389194853148946,
+    "best_speedup_vs_aoTriton": 2.96188735245333,
     "best_path": "3d"
   },
   {
@@ -948,30 +948,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.08831310272216797,
-    "aoTriton_ms": 0.1852444267272949,
+    "triton_ms": 0.08773019790649414,
+    "aoTriton_ms": 0.1851034355163574,
     "dsl": {
       "30": {
-        "ms": 0.0420143985748291,
+        "ms": 0.042154760360717775,
         "path": "3d"
       },
       "60": {
-        "ms": 0.029275460243225096,
+        "ms": 0.03234416007995605,
         "path": "3d"
       },
       "80": {
-        "ms": 0.027938780784606935,
+        "ms": 0.03237060070037842,
         "path": "3d"
       },
       "120": {
-        "ms": 0.028197760581970214,
+        "ms": 0.03205627918243408,
         "path": "3d"
       }
     },
-    "best_sms": 80,
-    "best_ms": 0.027938780784606935,
-    "best_speedup_vs_triton": 3.1609504868166862,
-    "best_speedup_vs_aoTriton": 6.6303690256003085,
+    "best_sms": 120,
+    "best_ms": 0.03205627918243408,
+    "best_speedup_vs_triton": 2.7367554857884993,
+    "best_speedup_vs_aoTriton": 5.774326909961178,
     "best_path": "3d"
   },
   {
@@ -987,30 +987,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.08881423950195312,
-    "aoTriton_ms": 0.36550777435302734,
+    "triton_ms": 0.08720999717712402,
+    "aoTriton_ms": 0.36465408325195314,
     "dsl": {
       "30": {
-        "ms": 0.07217269897460937,
+        "ms": 0.0722777795791626,
         "path": "3d"
       },
       "60": {
-        "ms": 0.04299026012420654,
+        "ms": 0.043315038681030274,
         "path": "3d"
       },
       "80": {
-        "ms": 0.03272742033004761,
+        "ms": 0.03283087968826294,
         "path": "3d"
       },
       "120": {
-        "ms": 0.03280839920043945,
+        "ms": 0.03363271951675415,
         "path": "3d"
       }
     },
     "best_sms": 80,
-    "best_ms": 0.03272742033004761,
-    "best_speedup_vs_triton": 2.7137561899558347,
-    "best_speedup_vs_aoTriton": 11.16824273550972,
+    "best_ms": 0.03283087968826294,
+    "best_speedup_vs_triton": 2.6563405551481964,
+    "best_speedup_vs_aoTriton": 11.107045766498825,
     "best_path": "3d"
   },
   {
@@ -1026,30 +1026,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.08911312103271485,
-    "aoTriton_ms": 0.8088990020751953,
+    "triton_ms": 0.08944835662841796,
+    "aoTriton_ms": 0.8090968322753906,
     "dsl": {
       "30": {
-        "ms": 0.13337692260742187,
+        "ms": 0.1332615566253662,
         "path": "3d"
       },
       "60": {
-        "ms": 0.07382612228393555,
+        "ms": 0.07380290031433105,
         "path": "3d"
       },
       "80": {
-        "ms": 0.05120997905731201,
+        "ms": 0.052009482383728024,
         "path": "3d"
       },
       "120": {
-        "ms": 0.051262102127075194,
+        "ms": 0.051995038986206055,
         "path": "3d"
       }
     },
-    "best_sms": 80,
-    "best_ms": 0.05120997905731201,
-    "best_speedup_vs_triton": 1.7401514836196919,
-    "best_speedup_vs_aoTriton": 15.795729991803164,
+    "best_sms": 120,
+    "best_ms": 0.051995038986206055,
+    "best_speedup_vs_triton": 1.7203248304544598,
+    "best_speedup_vs_aoTriton": 15.56103905393818,
     "best_path": "3d"
   },
   {
@@ -1065,30 +1065,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.08814408302307129,
-    "aoTriton_ms": 1.6142866516113281,
+    "triton_ms": 0.08780757904052734,
+    "aoTriton_ms": 1.6059767150878905,
     "dsl": {
       "30": {
-        "ms": 0.25890134811401366,
+        "ms": 0.259506893157959,
         "path": "3d"
       },
       "60": {
-        "ms": 0.13674548149108887,
+        "ms": 0.13713447570800782,
         "path": "3d"
       },
       "80": {
-        "ms": 0.09165038108825684,
+        "ms": 0.09189417839050293,
         "path": "3d"
       },
       "120": {
-        "ms": 0.09168326377868652,
+        "ms": 0.09187013626098633,
         "path": "3d"
       }
     },
-    "best_sms": 80,
-    "best_ms": 0.09165038108825684,
-    "best_speedup_vs_triton": 0.9617426788241167,
-    "best_speedup_vs_aoTriton": 17.61352907040085,
+    "best_sms": 120,
+    "best_ms": 0.09187013626098633,
+    "best_speedup_vs_triton": 0.955779349135632,
+    "best_speedup_vs_aoTriton": 17.48094408530758,
     "best_path": "3d"
   },
   {
@@ -1104,30 +1104,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.04507905960083008,
-    "aoTriton_ms": 0.02874382019042969,
+    "triton_ms": 0.04444643974304199,
+    "aoTriton_ms": 0.029012479782104493,
     "dsl": {
       "30": {
-        "ms": 0.05536595821380615,
+        "ms": 0.05781003952026367,
         "path": "2d"
       },
       "60": {
-        "ms": 0.056996941566467285,
+        "ms": 0.06902068138122558,
         "path": "2d"
       },
       "80": {
-        "ms": 0.054440641403198244,
+        "ms": 0.057149319648742675,
         "path": "2d"
       },
       "120": {
-        "ms": 0.05614215850830078,
+        "ms": 0.05867763996124267,
         "path": "2d"
       }
     },
     "best_sms": 80,
-    "best_ms": 0.054440641403198244,
-    "best_speedup_vs_triton": 0.8280405674680718,
-    "best_speedup_vs_aoTriton": 0.527984598446356,
+    "best_ms": 0.057149319648742675,
+    "best_speedup_vs_triton": 0.7777247396158607,
+    "best_speedup_vs_aoTriton": 0.507660982850262,
     "best_path": "2d"
   },
   {
@@ -1143,30 +1143,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.09279301643371582,
-    "aoTriton_ms": 0.05153153896331787,
+    "triton_ms": 0.09051502227783204,
+    "aoTriton_ms": 0.05142652034759521,
     "dsl": {
       "30": {
-        "ms": 0.027747139930725098,
+        "ms": 0.03153507947921753,
         "path": "3d"
       },
       "60": {
-        "ms": 0.027781620025634765,
+        "ms": 0.031368300914764405,
         "path": "3d"
       },
       "80": {
-        "ms": 0.026323859691619874,
+        "ms": 0.03163450002670288,
         "path": "3d"
       },
       "120": {
-        "ms": 0.026744821071624757,
+        "ms": 0.03077091932296753,
         "path": "3d"
       }
     },
-    "best_sms": 80,
-    "best_ms": 0.026323859691619874,
-    "best_speedup_vs_triton": 3.525053602350578,
-    "best_speedup_vs_aoTriton": 1.9575981473462567,
+    "best_sms": 120,
+    "best_ms": 0.03077091932296753,
+    "best_speedup_vs_triton": 2.941576796188578,
+    "best_speedup_vs_aoTriton": 1.6712701953370066,
     "best_path": "3d"
   },
   {
@@ -1182,30 +1182,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.08860896110534668,
-    "aoTriton_ms": 0.09785987854003907,
+    "triton_ms": 0.08857695579528808,
+    "aoTriton_ms": 0.09773006439208984,
     "dsl": {
       "30": {
-        "ms": 0.027875421047210695,
+        "ms": 0.03311553955078125,
         "path": "3d"
       },
       "60": {
-        "ms": 0.027792840003967284,
+        "ms": 0.0324179196357727,
         "path": "3d"
       },
       "80": {
-        "ms": 0.027780001163482667,
+        "ms": 0.03251976013183594,
         "path": "3d"
       },
       "120": {
-        "ms": 0.027802460193634033,
+        "ms": 0.03172754049301148,
         "path": "3d"
       }
     },
-    "best_sms": 80,
-    "best_ms": 0.027780001163482667,
-    "best_speedup_vs_triton": 3.189667292808642,
-    "best_speedup_vs_aoTriton": 3.522673665999615,
+    "best_sms": 120,
+    "best_ms": 0.03172754049301148,
+    "best_speedup_vs_triton": 2.791800259928709,
+    "best_speedup_vs_aoTriton": 3.080291219346691,
     "best_path": "3d"
   },
   {
@@ -1221,30 +1221,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.08992881774902343,
-    "aoTriton_ms": 0.1905278205871582,
+    "triton_ms": 0.08880707740783692,
+    "aoTriton_ms": 0.19016790390014648,
     "dsl": {
       "30": {
-        "ms": 0.04296140193939209,
+        "ms": 0.04312498092651367,
         "path": "3d"
       },
       "60": {
-        "ms": 0.0428964376449585,
+        "ms": 0.043127398490905765,
         "path": "3d"
       },
       "80": {
-        "ms": 0.03393739938735962,
+        "ms": 0.0346069598197937,
         "path": "3d"
       },
       "120": {
-        "ms": 0.03301448106765747,
+        "ms": 0.03356215953826904,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.03301448106765747,
-    "best_speedup_vs_triton": 2.7239203779920054,
-    "best_speedup_vs_aoTriton": 5.771037872644564,
+    "best_ms": 0.03356215953826904,
+    "best_speedup_vs_triton": 2.64604777015541,
+    "best_speedup_vs_aoTriton": 5.666140275726558,
     "best_path": "3d"
   },
   {
@@ -1260,30 +1260,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.0890413761138916,
-    "aoTriton_ms": 0.41119384765625,
+    "triton_ms": 0.09006077766418458,
+    "aoTriton_ms": 0.41084617614746094,
     "dsl": {
       "30": {
-        "ms": 0.07379161834716796,
+        "ms": 0.0735727596282959,
         "path": "3d"
       },
       "60": {
-        "ms": 0.07339230060577392,
+        "ms": 0.07360243797302246,
         "path": "3d"
       },
       "80": {
-        "ms": 0.05137998104095459,
+        "ms": 0.05093500137329102,
         "path": "3d"
       },
       "120": {
-        "ms": 0.0514176607131958,
+        "ms": 0.050983099937438964,
         "path": "3d"
       }
     },
     "best_sms": 80,
-    "best_ms": 0.05137998104095459,
-    "best_speedup_vs_triton": 1.732997449783358,
-    "best_speedup_vs_aoTriton": 8.002997263243257,
+    "best_ms": 0.05093500137329102,
+    "best_speedup_vs_triton": 1.76815107953271,
+    "best_speedup_vs_aoTriton": 8.066087465797104,
     "best_path": "3d"
   },
   {
@@ -1299,30 +1299,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.08870557785034179,
-    "aoTriton_ms": 0.8168212890625,
+    "triton_ms": 0.08959148406982421,
+    "aoTriton_ms": 0.8131292724609375,
     "dsl": {
       "30": {
-        "ms": 0.1353382396697998,
+        "ms": 0.13511542320251466,
         "path": "3d"
       },
       "60": {
-        "ms": 0.13534946441650392,
+        "ms": 0.13516994476318359,
         "path": "3d"
       },
       "80": {
-        "ms": 0.08831950187683106,
+        "ms": 0.08824578285217285,
         "path": "3d"
       },
       "120": {
-        "ms": 0.08822648048400879,
+        "ms": 0.08824418067932129,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.08822648048400879,
-    "best_speedup_vs_triton": 1.0054303125740105,
-    "best_speedup_vs_aoTriton": 9.258232727650858,
+    "best_ms": 0.08824418067932129,
+    "best_speedup_vs_triton": 1.0152679007287633,
+    "best_speedup_vs_aoTriton": 9.21453705163679,
     "best_path": "3d"
   },
   {
@@ -1338,30 +1338,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.08800919532775879,
-    "aoTriton_ms": 1.634215087890625,
+    "triton_ms": 0.08674391746520996,
+    "aoTriton_ms": 1.6234425354003905,
     "dsl": {
       "30": {
-        "ms": 0.2626940727233887,
+        "ms": 0.26254512786865236,
         "path": "3d"
       },
       "60": {
-        "ms": 0.26278228759765626,
+        "ms": 0.26253389358520507,
         "path": "3d"
       },
       "80": {
-        "ms": 0.16474920272827148,
+        "ms": 0.1641303062438965,
         "path": "3d"
       },
       "120": {
-        "ms": 0.16466421127319336,
+        "ms": 0.164119873046875,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.16466421127319336,
-    "best_speedup_vs_triton": 0.5344767672784907,
-    "best_speedup_vs_aoTriton": 9.924531112466868,
+    "best_ms": 0.164119873046875,
+    "best_speedup_vs_triton": 0.528539998568209,
+    "best_speedup_vs_aoTriton": 9.89180959783409,
     "best_path": "3d"
   },
   {
@@ -1377,30 +1377,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.04533805847167969,
-    "aoTriton_ms": 0.028132820129394533,
+    "triton_ms": 0.045794358253479,
+    "aoTriton_ms": 0.02811038017272949,
     "dsl": {
       "30": {
-        "ms": 0.06560555934906005,
+        "ms": 0.06548051834106446,
         "path": "2d"
       },
       "60": {
-        "ms": 0.06834867954254151,
+        "ms": 0.0675613021850586,
         "path": "2d"
       },
       "80": {
-        "ms": 0.06513807773590088,
+        "ms": 0.0655671215057373,
         "path": "2d"
       },
       "120": {
-        "ms": 0.06627992153167725,
+        "ms": 0.06702888011932373,
         "path": "2d"
       }
     },
-    "best_sms": 80,
-    "best_ms": 0.06513807773590088,
-    "best_speedup_vs_triton": 0.6960300341606733,
-    "best_speedup_vs_aoTriton": 0.4318951542208178,
+    "best_sms": 30,
+    "best_ms": 0.06548051834106446,
+    "best_speedup_vs_triton": 0.6993585178259077,
+    "best_speedup_vs_aoTriton": 0.42929379432081827,
     "best_path": "2d"
   },
   {
@@ -1416,30 +1416,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.09251155853271484,
-    "aoTriton_ms": 0.0507794189453125,
+    "triton_ms": 0.0889939022064209,
+    "aoTriton_ms": 0.05071847915649414,
     "dsl": {
       "30": {
-        "ms": 0.028225839138031006,
+        "ms": 0.031199100017547606,
         "path": "3d"
       },
       "60": {
-        "ms": 0.027027881145477294,
+        "ms": 0.030775740146636962,
         "path": "3d"
       },
       "80": {
-        "ms": 0.026645400524139405,
+        "ms": 0.03259432077407837,
         "path": "3d"
       },
       "120": {
-        "ms": 0.026113760471343995,
+        "ms": 0.03157517910003662,
         "path": "3d"
       }
     },
-    "best_sms": 120,
-    "best_ms": 0.026113760471343995,
-    "best_speedup_vs_triton": 3.542636405592854,
-    "best_speedup_vs_aoTriton": 1.9445464011602402,
+    "best_sms": 60,
+    "best_ms": 0.030775740146636962,
+    "best_speedup_vs_triton": 2.8916900708932505,
+    "best_speedup_vs_aoTriton": 1.6480019299238993,
     "best_path": "3d"
   },
   {
@@ -1455,30 +1455,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.09079400062561035,
-    "aoTriton_ms": 0.09634679794311524,
+    "triton_ms": 0.08832356452941895,
+    "aoTriton_ms": 0.09612316131591797,
     "dsl": {
       "30": {
-        "ms": 0.028393421173095703,
+        "ms": 0.031762020587921144,
         "path": "3d"
       },
       "60": {
-        "ms": 0.02823146104812622,
+        "ms": 0.032356960773468016,
         "path": "3d"
       },
       "80": {
-        "ms": 0.02762605905532837,
+        "ms": 0.03253098011016846,
         "path": "3d"
       },
       "120": {
-        "ms": 0.027657320499420168,
+        "ms": 0.03221585988998413,
         "path": "3d"
       }
     },
-    "best_sms": 80,
-    "best_ms": 0.02762605905532837,
-    "best_speedup_vs_triton": 3.2865346607625705,
-    "best_speedup_vs_aoTriton": 3.487533192850841,
+    "best_sms": 30,
+    "best_ms": 0.031762020587921144,
+    "best_speedup_vs_triton": 2.780791741033243,
+    "best_speedup_vs_aoTriton": 3.0263553620537884,
     "best_path": "3d"
   },
   {
@@ -1494,30 +1494,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.08829383850097657,
-    "aoTriton_ms": 0.18741264343261718,
+    "triton_ms": 0.08817522048950195,
+    "aoTriton_ms": 0.18696212768554688,
     "dsl": {
       "30": {
-        "ms": 0.04261260032653809,
+        "ms": 0.04256289958953857,
         "path": "3d"
       },
       "60": {
-        "ms": 0.04255805969238281,
+        "ms": 0.04265992164611816,
         "path": "3d"
       },
       "80": {
-        "ms": 0.033874039649963376,
+        "ms": 0.03397831916809082,
         "path": "3d"
       },
       "120": {
-        "ms": 0.033635900020599366,
+        "ms": 0.033762619495391846,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.033635900020599366,
-    "best_speedup_vs_triton": 2.624988136095763,
-    "best_speedup_vs_aoTriton": 5.571804034315762,
+    "best_ms": 0.033762619495391846,
+    "best_speedup_vs_triton": 2.6116226112590795,
+    "best_speedup_vs_aoTriton": 5.537548048102866,
     "best_path": "3d"
   },
   {
@@ -1533,30 +1533,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.089725341796875,
-    "aoTriton_ms": 0.41206703186035154,
+    "triton_ms": 0.08959349632263183,
+    "aoTriton_ms": 0.4116327667236328,
     "dsl": {
       "30": {
-        "ms": 0.07289196014404296,
+        "ms": 0.07278696060180664,
         "path": "3d"
       },
       "60": {
-        "ms": 0.07292404174804687,
+        "ms": 0.07282626152038574,
         "path": "3d"
       },
       "80": {
-        "ms": 0.050614237785339355,
+        "ms": 0.050463500022888186,
         "path": "3d"
       },
       "120": {
-        "ms": 0.050541257858276366,
+        "ms": 0.050438637733459475,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.050541257858276366,
-    "best_speedup_vs_triton": 1.775289052925343,
-    "best_speedup_vs_aoTriton": 8.153082240569397,
+    "best_ms": 0.050438637733459475,
+    "best_speedup_vs_triton": 1.776286996411051,
+    "best_speedup_vs_aoTriton": 8.16106035414529,
     "best_path": "3d"
   },
   {
@@ -1572,30 +1572,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.0897864818572998,
-    "aoTriton_ms": 0.8200230407714844,
+    "triton_ms": 0.08964061737060547,
+    "aoTriton_ms": 0.8160665130615234,
     "dsl": {
       "30": {
-        "ms": 0.13409618377685548,
+        "ms": 0.13393989562988282,
         "path": "3d"
       },
       "60": {
-        "ms": 0.133944616317749,
+        "ms": 0.13398480415344238,
         "path": "3d"
       },
       "80": {
-        "ms": 0.08971792221069336,
+        "ms": 0.08989277839660645,
         "path": "3d"
       },
       "120": {
-        "ms": 0.08969867706298829,
+        "ms": 0.08994649887084961,
         "path": "3d"
       }
     },
-    "best_sms": 120,
-    "best_ms": 0.08969867706298829,
-    "best_speedup_vs_triton": 1.0009788861685203,
-    "best_speedup_vs_aoTriton": 9.141974749478713,
+    "best_sms": 80,
+    "best_ms": 0.08989277839660645,
+    "best_speedup_vs_triton": 0.9971948689260839,
+    "best_speedup_vs_aoTriton": 9.078221049760442,
     "best_path": "3d"
   },
   {
@@ -1611,30 +1611,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.08800357818603516,
-    "aoTriton_ms": 1.6411614990234376,
+    "triton_ms": 0.08979875564575196,
+    "aoTriton_ms": 1.6301219177246093,
     "dsl": {
       "30": {
-        "ms": 0.2601089286804199,
+        "ms": 0.26009708404541015,
         "path": "3d"
       },
       "60": {
-        "ms": 0.26014421463012694,
+        "ms": 0.26022857666015625,
         "path": "3d"
       },
       "80": {
-        "ms": 0.16899820327758788,
+        "ms": 0.16287057876586913,
         "path": "3d"
       },
       "120": {
-        "ms": 0.1689837646484375,
+        "ms": 0.1626877784729004,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.1689837646484375,
-    "best_speedup_vs_triton": 0.5207812618515294,
-    "best_speedup_vs_aoTriton": 9.711947786450338,
+    "best_ms": 0.1626877784729004,
+    "best_speedup_vs_triton": 0.551969892813492,
+    "best_speedup_vs_aoTriton": 10.019940852509372,
     "best_path": "3d"
   },
   {
@@ -1650,30 +1650,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.04420825958251953,
-    "aoTriton_ms": 0.028420701026916503,
+    "triton_ms": 0.045893778800964354,
+    "aoTriton_ms": 0.028690121173858642,
     "dsl": {
       "30": {
-        "ms": 0.055137438774108885,
+        "ms": 0.05837212085723877,
         "path": "2d"
       },
       "60": {
-        "ms": 0.05861266136169434,
+        "ms": 0.060725560188293455,
         "path": "2d"
       },
       "80": {
-        "ms": 0.05526254177093506,
+        "ms": 0.05784853935241699,
         "path": "2d"
       },
       "120": {
-        "ms": 0.05782201766967773,
+        "ms": 0.05872496128082275,
         "path": "2d"
       }
     },
-    "best_sms": 30,
-    "best_ms": 0.055137438774108885,
-    "best_speedup_vs_triton": 0.8017829729747727,
-    "best_speedup_vs_aoTriton": 0.5154519625649012,
+    "best_sms": 80,
+    "best_ms": 0.05784853935241699,
+    "best_speedup_vs_triton": 0.7933437786799858,
+    "best_speedup_vs_aoTriton": 0.49595238695789007,
     "best_path": "2d"
   },
   {
@@ -1689,30 +1689,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.09269118309020996,
-    "aoTriton_ms": 0.04760168075561524,
+    "triton_ms": 0.09034903526306153,
+    "aoTriton_ms": 0.04758887767791748,
     "dsl": {
       "30": {
-        "ms": 0.028255500793457032,
+        "ms": 0.031321799755096434,
         "path": "3d"
       },
       "60": {
-        "ms": 0.026503479480743407,
+        "ms": 0.032119619846344,
         "path": "3d"
       },
       "80": {
-        "ms": 0.02620919942855835,
+        "ms": 0.031534280776977536,
         "path": "3d"
       },
       "120": {
-        "ms": 0.026571619510650634,
+        "ms": 0.03145410060882568,
         "path": "3d"
       }
     },
-    "best_sms": 80,
-    "best_ms": 0.02620919942855835,
-    "best_speedup_vs_triton": 3.536589636889511,
-    "best_speedup_vs_aoTriton": 1.8162203269645458,
+    "best_sms": 30,
+    "best_ms": 0.031321799755096434,
+    "best_speedup_vs_triton": 2.8845416281789698,
+    "best_speedup_vs_aoTriton": 1.5193532316154406,
     "best_path": "3d"
   },
   {
@@ -1728,30 +1728,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.0907218360900879,
-    "aoTriton_ms": 0.0911404037475586,
+    "triton_ms": 0.08836928367614746,
+    "aoTriton_ms": 0.09123027801513672,
     "dsl": {
       "30": {
-        "ms": 0.027978079319000246,
+        "ms": 0.03224149942398071,
         "path": "3d"
       },
       "60": {
-        "ms": 0.02634310007095337,
+        "ms": 0.03194082021713257,
         "path": "3d"
       },
       "80": {
-        "ms": 0.02670151948928833,
+        "ms": 0.031952059268951415,
         "path": "3d"
       },
       "120": {
-        "ms": 0.026656620502471924,
+        "ms": 0.03147655963897705,
         "path": "3d"
       }
     },
-    "best_sms": 60,
-    "best_ms": 0.02634310007095337,
-    "best_speedup_vs_triton": 3.4438557286627134,
-    "best_speedup_vs_aoTriton": 3.4597448099152355,
+    "best_sms": 120,
+    "best_ms": 0.03147655963897705,
+    "best_speedup_vs_triton": 2.8074632262771444,
+    "best_speedup_vs_aoTriton": 2.898356080254951,
     "best_path": "3d"
   },
   {
@@ -1767,30 +1767,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.08673266410827636,
-    "aoTriton_ms": 0.18755374908447264,
+    "triton_ms": 0.08864912033081054,
+    "aoTriton_ms": 0.19174753189086913,
     "dsl": {
       "30": {
-        "ms": 0.0287758994102478,
+        "ms": 0.032793979644775394,
         "path": "3d"
       },
       "60": {
-        "ms": 0.027432820796966552,
+        "ms": 0.031122119426727296,
         "path": "3d"
       },
       "80": {
-        "ms": 0.02771346092224121,
+        "ms": 0.03246201992034912,
         "path": "3d"
       },
       "120": {
-        "ms": 0.027459259033203124,
+        "ms": 0.0322038197517395,
         "path": "3d"
       }
     },
     "best_sms": 60,
-    "best_ms": 0.027432820796966552,
-    "best_speedup_vs_triton": 3.161638562442949,
-    "best_speedup_vs_aoTriton": 6.8368379056816435,
+    "best_ms": 0.031122119426727296,
+    "best_speedup_vs_triton": 2.8484281264816356,
+    "best_speedup_vs_aoTriton": 6.161133477503421,
     "best_path": "3d"
   },
   {
@@ -1806,30 +1806,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.08940600395202637,
-    "aoTriton_ms": 0.370882568359375,
+    "triton_ms": 0.08741988182067871,
+    "aoTriton_ms": 0.37002723693847656,
     "dsl": {
       "30": {
-        "ms": 0.043399200439453126,
+        "ms": 0.04355078220367432,
         "path": "3d"
       },
       "60": {
-        "ms": 0.02877671957015991,
+        "ms": 0.03151263952255249,
         "path": "3d"
       },
       "80": {
-        "ms": 0.028844060897827147,
+        "ms": 0.03135385990142822,
         "path": "3d"
       },
       "120": {
-        "ms": 0.028807981014251707,
+        "ms": 0.03205548048019409,
         "path": "3d"
       }
     },
-    "best_sms": 60,
-    "best_ms": 0.02877671957015991,
-    "best_speedup_vs_triton": 3.106886583581825,
-    "best_speedup_vs_aoTriton": 12.888285179800778,
+    "best_sms": 80,
+    "best_ms": 0.03135385990142822,
+    "best_speedup_vs_triton": 2.78816968933055,
+    "best_speedup_vs_aoTriton": 11.801648604088493,
     "best_path": "3d"
   },
   {
@@ -1845,30 +1845,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.08896657943725586,
-    "aoTriton_ms": 0.7369083404541016,
+    "triton_ms": 0.08916872024536132,
+    "aoTriton_ms": 0.7358119201660156,
     "dsl": {
       "30": {
-        "ms": 0.07429518222808838,
+        "ms": 0.07421905994415283,
         "path": "3d"
       },
       "60": {
-        "ms": 0.04415373802185059,
+        "ms": 0.04428445816040039,
         "path": "3d"
       },
       "80": {
-        "ms": 0.038428540229797366,
+        "ms": 0.03841494083404541,
         "path": "3d"
       },
       "120": {
-        "ms": 0.038398880958557126,
+        "ms": 0.03847830057144165,
         "path": "3d"
       }
     },
-    "best_sms": 120,
-    "best_ms": 0.038398880958557126,
-    "best_speedup_vs_triton": 2.316905524754096,
-    "best_speedup_vs_aoTriton": 19.19088062095942,
+    "best_sms": 80,
+    "best_ms": 0.03841494083404541,
+    "best_speedup_vs_triton": 2.321198947841022,
+    "best_speedup_vs_aoTriton": 19.154316112180474,
     "best_path": "3d"
   },
   {
@@ -1884,30 +1884,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.08905318260192871,
-    "aoTriton_ms": 1.591803741455078,
+    "triton_ms": 0.08786731719970703,
+    "aoTriton_ms": 1.5880705261230468,
     "dsl": {
       "30": {
-        "ms": 0.1366468620300293,
+        "ms": 0.13689149856567384,
         "path": "3d"
       },
       "60": {
-        "ms": 0.07566314220428466,
+        "ms": 0.07597829818725586,
         "path": "3d"
       },
       "80": {
-        "ms": 0.05785170078277588,
+        "ms": 0.0574267578125,
         "path": "3d"
       },
       "120": {
-        "ms": 0.05782926082611084,
+        "ms": 0.05730408191680908,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.05782926082611084,
-    "best_speedup_vs_triton": 1.539932922015143,
-    "best_speedup_vs_aoTriton": 27.525922322291784,
+    "best_ms": 0.05730408191680908,
+    "best_speedup_vs_triton": 1.53335180078913,
+    "best_speedup_vs_aoTriton": 27.713043696058516,
     "best_path": "3d"
   },
   {
@@ -1923,30 +1923,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.044889822006225585,
-    "aoTriton_ms": 0.028620340824127198,
+    "triton_ms": 0.045061440467834474,
+    "aoTriton_ms": 0.028816800117492675,
     "dsl": {
       "30": {
-        "ms": 0.0658156394958496,
+        "ms": 0.06606266021728516,
         "path": "2d"
       },
       "60": {
-        "ms": 0.07647702217102051,
+        "ms": 0.06671857833862305,
         "path": "2d"
       },
       "80": {
-        "ms": 0.06483580112457275,
+        "ms": 0.06513092041015625,
         "path": "2d"
       },
       "120": {
-        "ms": 0.06569697856903076,
+        "ms": 0.07696378231048584,
         "path": "2d"
       }
     },
     "best_sms": 80,
-    "best_ms": 0.06483580112457275,
-    "best_speedup_vs_triton": 0.6923616463067401,
-    "best_speedup_vs_aoTriton": 0.4414280432679052,
+    "best_ms": 0.06513092041015625,
+    "best_speedup_vs_triton": 0.6918594146077471,
+    "best_speedup_vs_aoTriton": 0.44244423287773926,
     "best_path": "2d"
   },
   {
@@ -1962,30 +1962,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.08907485961914062,
-    "aoTriton_ms": 0.044630842208862306,
+    "triton_ms": 0.08891533851623536,
+    "aoTriton_ms": 0.04459238052368164,
     "dsl": {
       "30": {
-        "ms": 0.027693419456481932,
+        "ms": 0.030452580451965333,
         "path": "3d"
       },
       "60": {
-        "ms": 0.026848258972167968,
+        "ms": 0.030843899250030518,
         "path": "3d"
       },
       "80": {
-        "ms": 0.026192340850830078,
+        "ms": 0.031624081134796145,
         "path": "3d"
       },
       "120": {
-        "ms": 0.02604882001876831,
+        "ms": 0.03139797925949097,
         "path": "3d"
       }
     },
-    "best_sms": 120,
-    "best_ms": 0.02604882001876831,
-    "best_speedup_vs_triton": 3.419535301597605,
-    "best_speedup_vs_aoTriton": 1.71335370188383,
+    "best_sms": 30,
+    "best_ms": 0.030452580451965333,
+    "best_speedup_vs_triton": 2.9197965228755183,
+    "best_speedup_vs_aoTriton": 1.4643219018506446,
     "best_path": "3d"
   },
   {
@@ -2001,30 +2001,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.09032974243164063,
-    "aoTriton_ms": 0.08602060317993164,
+    "triton_ms": 0.08794427871704101,
+    "aoTriton_ms": 0.08593485832214355,
     "dsl": {
       "30": {
-        "ms": 0.028189759254455566,
+        "ms": 0.03236419916152954,
         "path": "3d"
       },
       "60": {
-        "ms": 0.026867499351501466,
+        "ms": 0.03296878099441528,
         "path": "3d"
       },
       "80": {
-        "ms": 0.027014238834381102,
+        "ms": 0.03165293931961059,
         "path": "3d"
       },
       "120": {
-        "ms": 0.026614940166473387,
+        "ms": 0.03163290023803711,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.026614940166473387,
-    "best_speedup_vs_triton": 3.3939487320519413,
-    "best_speedup_vs_aoTriton": 3.2320419524478607,
+    "best_ms": 0.03163290023803711,
+    "best_speedup_vs_triton": 2.7801522483003964,
+    "best_speedup_vs_aoTriton": 2.716629132184688,
     "best_path": "3d"
   },
   {
@@ -2040,30 +2040,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.09064805984497071,
-    "aoTriton_ms": 0.184149112701416,
+    "triton_ms": 0.0895071029663086,
+    "aoTriton_ms": 0.18381805419921876,
     "dsl": {
       "30": {
-        "ms": 0.027656519412994386,
+        "ms": 0.032332119941711424,
         "path": "3d"
       },
       "60": {
-        "ms": 0.02733897924423218,
+        "ms": 0.031495800018310545,
         "path": "3d"
       },
       "80": {
-        "ms": 0.02742640018463135,
+        "ms": 0.03189752101898193,
         "path": "3d"
       },
       "120": {
-        "ms": 0.026449759006500245,
+        "ms": 0.03159362077713013,
         "path": "3d"
       }
     },
-    "best_sms": 120,
-    "best_ms": 0.026449759006500245,
-    "best_speedup_vs_triton": 3.4271790462322627,
-    "best_speedup_vs_aoTriton": 6.962222705173225,
+    "best_sms": 60,
+    "best_ms": 0.031495800018310545,
+    "best_speedup_vs_triton": 2.84187424717811,
+    "best_speedup_vs_aoTriton": 5.836271950303007,
     "best_path": "3d"
   },
   {
@@ -2079,30 +2079,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.08993441581726075,
-    "aoTriton_ms": 0.363888053894043,
+    "triton_ms": 0.08876057624816895,
+    "aoTriton_ms": 0.3630471420288086,
     "dsl": {
       "30": {
-        "ms": 0.04249392032623291,
+        "ms": 0.042635078430175784,
         "path": "3d"
       },
       "60": {
-        "ms": 0.028272340297698973,
+        "ms": 0.03238503932952881,
         "path": "3d"
       },
       "80": {
-        "ms": 0.028530540466308593,
+        "ms": 0.031031520366668702,
         "path": "3d"
       },
       "120": {
-        "ms": 0.028481619358062746,
+        "ms": 0.03154150009155274,
         "path": "3d"
       }
     },
-    "best_sms": 60,
-    "best_ms": 0.028272340297698973,
-    "best_speedup_vs_triton": 3.1810035840782636,
-    "best_speedup_vs_aoTriton": 12.870814727836983,
+    "best_sms": 80,
+    "best_ms": 0.031031520366668702,
+    "best_speedup_vs_triton": 2.8603360453942717,
+    "best_speedup_vs_aoTriton": 11.699302442775622,
     "best_path": "3d"
   },
   {
@@ -2118,30 +2118,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.08970749855041504,
-    "aoTriton_ms": 0.7218479919433594,
+    "triton_ms": 0.08905323982238769,
+    "aoTriton_ms": 0.7208982849121094,
     "dsl": {
       "30": {
-        "ms": 0.07234990119934082,
+        "ms": 0.07276050090789794,
         "path": "3d"
       },
       "60": {
-        "ms": 0.04325967788696289,
+        "ms": 0.04328455924987793,
         "path": "3d"
       },
       "80": {
-        "ms": 0.03797068119049072,
+        "ms": 0.0383187198638916,
         "path": "3d"
       },
       "120": {
-        "ms": 0.03797068119049072,
+        "ms": 0.03828263998031616,
         "path": "3d"
       }
     },
-    "best_sms": 80,
-    "best_ms": 0.03797068119049072,
-    "best_speedup_vs_triton": 2.3625464631611917,
-    "best_speedup_vs_aoTriton": 19.01066742316272,
+    "best_sms": 120,
+    "best_ms": 0.03828263998031616,
+    "best_speedup_vs_triton": 2.3262042499727373,
+    "best_speedup_vs_aoTriton": 18.83094492132138,
     "best_path": "3d"
   },
   {
@@ -2157,30 +2157,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.08845179557800292,
-    "aoTriton_ms": 1.599663543701172,
+    "triton_ms": 0.08829509735107421,
+    "aoTriton_ms": 1.5961700439453126,
     "dsl": {
       "30": {
-        "ms": 0.1330770206451416,
+        "ms": 0.13313644409179687,
         "path": "3d"
       },
       "60": {
-        "ms": 0.07412280082702637,
+        "ms": 0.07399213790893555,
         "path": "3d"
       },
       "80": {
-        "ms": 0.05636106014251709,
+        "ms": 0.057519779205322266,
         "path": "3d"
       },
       "120": {
-        "ms": 0.056324977874755856,
+        "ms": 0.05747568130493164,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.056324977874755856,
-    "best_speedup_vs_triton": 1.5703831393362322,
-    "best_speedup_vs_aoTriton": 28.400606694567756,
+    "best_ms": 0.05747568130493164,
+    "best_speedup_vs_triton": 1.5362166284316519,
+    "best_speedup_vs_aoTriton": 27.771224415365996,
     "best_path": "3d"
   },
   {
@@ -2196,30 +2196,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.04488984107971192,
-    "aoTriton_ms": 0.0287951397895813,
+    "triton_ms": 0.045275559425354,
+    "aoTriton_ms": 0.0288360595703125,
     "dsl": {
       "30": {
-        "ms": 0.05560492038726807,
+        "ms": 0.05801209926605225,
         "path": "2d"
       },
       "60": {
-        "ms": 0.05834564208984375,
+        "ms": 0.060641360282897946,
         "path": "2d"
       },
       "80": {
-        "ms": 0.05495222091674805,
+        "ms": 0.05824304103851318,
         "path": "2d"
       },
       "120": {
-        "ms": 0.05656633853912354,
+        "ms": 0.05985156059265137,
         "path": "2d"
       }
     },
-    "best_sms": 80,
-    "best_ms": 0.05495222091674805,
-    "best_speedup_vs_triton": 0.8168885684842381,
-    "best_speedup_vs_aoTriton": 0.5240032033137585,
+    "best_sms": 30,
+    "best_ms": 0.05801209926605225,
+    "best_speedup_vs_triton": 0.780450285339847,
+    "best_speedup_vs_aoTriton": 0.4970697481238522,
     "best_path": "2d"
   },
   {
@@ -2235,30 +2235,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.09206893920898437,
-    "aoTriton_ms": 0.048240737915039064,
+    "triton_ms": 0.09175308227539063,
+    "aoTriton_ms": 0.04829771995544434,
     "dsl": {
       "30": {
-        "ms": 0.028501660823822023,
+        "ms": 0.031134960651397706,
         "path": "3d"
       },
       "60": {
-        "ms": 0.02725882053375244,
+        "ms": 0.030717999935150148,
         "path": "3d"
       },
       "80": {
-        "ms": 0.02656280040740967,
+        "ms": 0.03125123977661133,
         "path": "3d"
       },
       "120": {
-        "ms": 0.026729578971862792,
+        "ms": 0.03115580081939697,
         "path": "3d"
       }
     },
-    "best_sms": 80,
-    "best_ms": 0.02656280040740967,
-    "best_speedup_vs_triton": 3.4660855706803346,
-    "best_speedup_vs_aoTriton": 1.8161013588605799,
+    "best_sms": 60,
+    "best_ms": 0.030717999935150148,
+    "best_speedup_vs_triton": 2.986948449413822,
+    "best_speedup_vs_aoTriton": 1.572293770994445,
     "best_path": "3d"
   },
   {
@@ -2274,30 +2274,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.09045561790466308,
-    "aoTriton_ms": 0.09681026458740234,
+    "triton_ms": 0.08873412132263184,
+    "aoTriton_ms": 0.09668044090270995,
     "dsl": {
       "30": {
-        "ms": 0.02790911912918091,
+        "ms": 0.032699360847473144,
         "path": "3d"
       },
       "60": {
-        "ms": 0.027597200870513917,
+        "ms": 0.03262399911880493,
         "path": "3d"
       },
       "80": {
-        "ms": 0.026666240692138674,
+        "ms": 0.03210520029067993,
         "path": "3d"
       },
       "120": {
-        "ms": 0.026412060260772707,
+        "ms": 0.03207953929901123,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.026412060260772707,
-    "best_speedup_vs_triton": 3.4247846253405725,
-    "best_speedup_vs_aoTriton": 3.665381027892978,
+    "best_ms": 0.03207953929901123,
+    "best_speedup_vs_triton": 2.7660659492503012,
+    "best_speedup_vs_aoTriton": 3.013772735373724,
     "best_path": "3d"
   },
   {
@@ -2313,30 +2313,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.0873661231994629,
-    "aoTriton_ms": 0.18880943298339845,
+    "triton_ms": 0.08981579780578614,
+    "aoTriton_ms": 0.18818254470825196,
     "dsl": {
       "30": {
-        "ms": 0.04285714149475098,
+        "ms": 0.042782621383666994,
         "path": "3d"
       },
       "60": {
-        "ms": 0.029591379165649415,
+        "ms": 0.032862141132354736,
         "path": "3d"
       },
       "80": {
-        "ms": 0.02852734088897705,
+        "ms": 0.032752280235290525,
         "path": "3d"
       },
       "120": {
-        "ms": 0.027249979972839355,
+        "ms": 0.031102879047393797,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.027249979972839355,
-    "best_speedup_vs_triton": 3.206098620495964,
-    "best_speedup_vs_aoTriton": 6.928791623758582,
+    "best_ms": 0.031102879047393797,
+    "best_speedup_vs_triton": 2.8877004494962364,
+    "best_speedup_vs_aoTriton": 6.050325579876515,
     "best_path": "3d"
   },
   {
@@ -2352,30 +2352,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.08877094268798828,
-    "aoTriton_ms": 0.37275646209716795,
+    "triton_ms": 0.0889049243927002,
+    "aoTriton_ms": 0.3715660095214844,
     "dsl": {
       "30": {
-        "ms": 0.07356872081756592,
+        "ms": 0.07368261814117431,
         "path": "3d"
       },
       "60": {
-        "ms": 0.04355234146118164,
+        "ms": 0.04371915817260742,
         "path": "3d"
       },
       "80": {
-        "ms": 0.0325325608253479,
+        "ms": 0.032833280563354494,
         "path": "3d"
       },
       "120": {
-        "ms": 0.03252213954925537,
+        "ms": 0.03291425943374634,
         "path": "3d"
       }
     },
-    "best_sms": 120,
-    "best_ms": 0.03252213954925537,
-    "best_speedup_vs_triton": 2.7295542027160633,
-    "best_speedup_vs_aoTriton": 11.461621752548647,
+    "best_sms": 80,
+    "best_ms": 0.032833280563354494,
+    "best_speedup_vs_triton": 2.707768546647384,
+    "best_speedup_vs_aoTriton": 11.316749442825778,
     "best_path": "3d"
   },
   {
@@ -2391,30 +2391,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.08894094467163086,
-    "aoTriton_ms": 0.8054093933105468,
+    "triton_ms": 0.09038031578063965,
+    "aoTriton_ms": 0.8055245971679688,
     "dsl": {
       "30": {
-        "ms": 0.13500867843627928,
+        "ms": 0.13514588356018067,
         "path": "3d"
       },
       "60": {
-        "ms": 0.07434892177581787,
+        "ms": 0.07460073947906494,
         "path": "3d"
       },
       "80": {
-        "ms": 0.05206155776977539,
+        "ms": 0.05247213840484619,
         "path": "3d"
       },
       "120": {
-        "ms": 0.05210244178771973,
+        "ms": 0.052292518615722657,
         "path": "3d"
       }
     },
-    "best_sms": 80,
-    "best_ms": 0.05206155776977539,
-    "best_speedup_vs_triton": 1.7083803958564219,
-    "best_speedup_vs_aoTriton": 15.470328353834457,
+    "best_sms": 120,
+    "best_ms": 0.052292518615722657,
+    "best_speedup_vs_triton": 1.7283603500686089,
+    "best_speedup_vs_aoTriton": 15.404203478655429,
     "best_path": "3d"
   },
   {
@@ -2430,30 +2430,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "bf16",
-    "triton_ms": 0.0886867618560791,
-    "aoTriton_ms": 1.6079008483886719,
+    "triton_ms": 0.08971216201782227,
+    "aoTriton_ms": 1.5985075378417968,
     "dsl": {
       "30": {
-        "ms": 0.25982667922973635,
+        "ms": 0.2601179122924805,
         "path": "3d"
       },
       "60": {
-        "ms": 0.13682565689086915,
+        "ms": 0.13718338012695314,
         "path": "3d"
       },
       "80": {
-        "ms": 0.0913095760345459,
+        "ms": 0.09123027801513672,
         "path": "3d"
       },
       "120": {
-        "ms": 0.09127511978149414,
+        "ms": 0.09123669624328613,
         "path": "3d"
       }
     },
-    "best_sms": 120,
-    "best_ms": 0.09127511978149414,
-    "best_speedup_vs_triton": 0.9716422401678423,
-    "best_speedup_vs_aoTriton": 17.615981794796514,
+    "best_sms": 80,
+    "best_ms": 0.09123027801513672,
+    "best_speedup_vs_triton": 0.9833595158280394,
+    "best_speedup_vs_aoTriton": 17.521677809384467,
     "best_path": "3d"
   },
   {
@@ -2469,30 +2469,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.044984440803527835,
-    "aoTriton_ms": 0.02837739944458008,
+    "triton_ms": 0.04485377788543701,
+    "aoTriton_ms": 0.028953120708465577,
     "dsl": {
       "30": {
-        "ms": 0.06576432228088379,
+        "ms": 0.06578281879425049,
         "path": "2d"
       },
       "60": {
-        "ms": 0.0668596601486206,
+        "ms": 0.06734481811523438,
         "path": "2d"
       },
       "80": {
-        "ms": 0.06541711807250977,
+        "ms": 0.06542277812957764,
         "path": "2d"
       },
       "120": {
-        "ms": 0.06596960067749023,
+        "ms": 0.06649003982543945,
         "path": "2d"
       }
     },
     "best_sms": 80,
-    "best_ms": 0.06541711807250977,
-    "best_speedup_vs_triton": 0.6876554964354451,
-    "best_speedup_vs_aoTriton": 0.43379164782413593,
+    "best_ms": 0.06542277812957764,
+    "best_speedup_vs_triton": 0.6855987955234604,
+    "best_speedup_vs_aoTriton": 0.4425541307817968,
     "best_path": "2d"
   },
   {
@@ -2508,30 +2508,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.0881214427947998,
-    "aoTriton_ms": 0.04545032024383545,
+    "triton_ms": 0.08859780311584473,
+    "aoTriton_ms": 0.04542789936065674,
     "dsl": {
       "30": {
-        "ms": 0.027949199676513672,
+        "ms": 0.03128809928894043,
         "path": "3d"
       },
       "60": {
-        "ms": 0.026525120735168457,
+        "ms": 0.031096460819244384,
         "path": "3d"
       },
       "80": {
-        "ms": 0.02577860116958618,
+        "ms": 0.03141881942749023,
         "path": "3d"
       },
       "120": {
-        "ms": 0.02625730037689209,
+        "ms": 0.030839879512786866,
         "path": "3d"
       }
     },
-    "best_sms": 80,
-    "best_ms": 0.02577860116958618,
-    "best_speedup_vs_triton": 3.4183950562362653,
-    "best_speedup_vs_aoTriton": 1.7631026580859683,
+    "best_sms": 120,
+    "best_ms": 0.030839879512786866,
+    "best_speedup_vs_triton": 2.8728323364269372,
+    "best_speedup_vs_aoTriton": 1.4730245408974887,
     "best_path": "3d"
   },
   {
@@ -2547,30 +2547,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.08767642021179199,
-    "aoTriton_ms": 0.09485535621643067,
+    "triton_ms": 0.09333592414855957,
+    "aoTriton_ms": 0.09492839813232422,
     "dsl": {
       "30": {
-        "ms": 0.028471999168395996,
+        "ms": 0.032620799541473386,
         "path": "3d"
       },
       "60": {
-        "ms": 0.027374279499053956,
+        "ms": 0.03265768051147461,
         "path": "3d"
       },
       "80": {
-        "ms": 0.026719160079956054,
+        "ms": 0.034404900074005124,
         "path": "3d"
       },
       "120": {
-        "ms": 0.026230840682983397,
+        "ms": 0.03271460056304932,
         "path": "3d"
       }
     },
-    "best_sms": 120,
-    "best_ms": 0.026230840682983397,
-    "best_speedup_vs_triton": 3.342493718421685,
-    "best_speedup_vs_aoTriton": 3.6161767502162334,
+    "best_sms": 30,
+    "best_ms": 0.032620799541473386,
+    "best_speedup_vs_triton": 2.8612396219747547,
+    "best_speedup_vs_aoTriton": 2.910057370348458,
     "best_path": "3d"
   },
   {
@@ -2586,30 +2586,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.09001540184020997,
-    "aoTriton_ms": 0.1853005599975586,
+    "triton_ms": 0.09680312156677245,
+    "aoTriton_ms": 0.1851491355895996,
     "dsl": {
       "30": {
-        "ms": 0.04154612064361572,
+        "ms": 0.041851658821105954,
         "path": "3d"
       },
       "60": {
-        "ms": 0.02888495922088623,
+        "ms": 0.03154069900512695,
         "path": "3d"
       },
       "80": {
-        "ms": 0.028716559410095214,
+        "ms": 0.032591118812561035,
         "path": "3d"
       },
       "120": {
-        "ms": 0.02776236057281494,
+        "ms": 0.03318449974060059,
         "path": "3d"
       }
     },
-    "best_sms": 120,
-    "best_ms": 0.02776236057281494,
-    "best_speedup_vs_triton": 3.2423540355697833,
-    "best_speedup_vs_aoTriton": 6.674524650436459,
+    "best_sms": 60,
+    "best_ms": 0.03154069900512695,
+    "best_speedup_vs_triton": 3.0691495312465036,
+    "best_speedup_vs_aoTriton": 5.870165894532127,
     "best_path": "3d"
   },
   {
@@ -2625,30 +2625,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.08826335906982422,
-    "aoTriton_ms": 0.36562889099121093,
+    "triton_ms": 0.0888864803314209,
+    "aoTriton_ms": 0.36516647338867186,
     "dsl": {
       "30": {
-        "ms": 0.07130750179290772,
+        "ms": 0.07135647773742676,
         "path": "3d"
       },
       "60": {
-        "ms": 0.042390480041503906,
+        "ms": 0.04260700225830078,
         "path": "3d"
       },
       "80": {
-        "ms": 0.032393839359283444,
+        "ms": 0.03258147954940796,
         "path": "3d"
       },
       "120": {
-        "ms": 0.03242671966552734,
+        "ms": 0.032712979316711424,
         "path": "3d"
       }
     },
     "best_sms": 80,
-    "best_ms": 0.032393839359283444,
-    "best_speedup_vs_triton": 2.724695831540254,
-    "best_speedup_vs_aoTriton": 11.286988459008604,
+    "best_ms": 0.03258147954940796,
+    "best_speedup_vs_triton": 2.728129034061501,
+    "best_speedup_vs_aoTriton": 11.2077928454697,
     "best_path": "3d"
   },
   {
@@ -2664,30 +2664,30 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.08976041793823242,
-    "aoTriton_ms": 0.8092534637451172,
+    "triton_ms": 0.0895792579650879,
+    "aoTriton_ms": 0.8084449768066406,
     "dsl": {
       "30": {
-        "ms": 0.13074206352233886,
+        "ms": 0.13076457977294922,
         "path": "3d"
       },
       "60": {
-        "ms": 0.07218070030212402,
+        "ms": 0.07233871936798096,
         "path": "3d"
       },
       "80": {
-        "ms": 0.05109131813049316,
+        "ms": 0.05040257930755615,
         "path": "3d"
       },
       "120": {
-        "ms": 0.051117777824401855,
+        "ms": 0.050364079475402834,
         "path": "3d"
       }
     },
-    "best_sms": 80,
-    "best_ms": 0.05109131813049316,
-    "best_speedup_vs_triton": 1.7568624420488406,
-    "best_speedup_vs_aoTriton": 15.839353795456791,
+    "best_sms": 120,
+    "best_ms": 0.050364079475402834,
+    "best_speedup_vs_triton": 1.7786338775205301,
+    "best_speedup_vs_aoTriton": 16.052015349580145,
     "best_path": "3d"
   },
   {
@@ -2703,30 +2703,303 @@
     "head_size": 128,
     "block_size": 16,
     "dtype": "fp16",
-    "triton_ms": 0.0879610824584961,
-    "aoTriton_ms": 1.6157373046875,
+    "triton_ms": 0.08756823539733886,
+    "aoTriton_ms": 1.6062213134765626,
+    "dsl": {
+      "30": {
+        "ms": 0.25116289138793946,
+        "path": "3d"
+      },
+      "60": {
+        "ms": 0.1324973773956299,
+        "path": "3d"
+      },
+      "80": {
+        "ms": 0.08718576431274414,
+        "path": "3d"
+      },
+      "120": {
+        "ms": 0.08711518287658691,
+        "path": "3d"
+      }
+    },
+    "best_sms": 120,
+    "best_ms": 0.08711518287658691,
+    "best_speedup_vs_triton": 1.0052006149306232,
+    "best_speedup_vs_aoTriton": 18.437903249908125,
+    "best_path": "3d"
+  },
+  {
+    "label": "qwen3next_80b_b1_kv512",
+    "signature": "b1_sq1_sk512_nhq16_nhk2_hd256_bs16_bf16",
+    "bias": "none",
+    "arch": "gfx942",
+    "batch": 1,
+    "seqlen_q": 1,
+    "seqlen_k": 512,
+    "num_query_heads": 16,
+    "num_kv_heads": 2,
+    "head_size": 256,
+    "block_size": 16,
+    "dtype": "bf16",
+    "triton_ms": 0.045553002357482913,
+    "aoTriton_ms": 0.03935631990432739,
+    "dsl": {
+      "30": {
+        "ms": 0.19017671585083007,
+        "path": "2d"
+      },
+      "60": {
+        "ms": 0.19033388137817384,
+        "path": "2d"
+      },
+      "80": {
+        "ms": 0.18995941162109375,
+        "path": "2d"
+      },
+      "120": {
+        "ms": 0.18989126205444337,
+        "path": "2d"
+      }
+    },
+    "best_sms": 120,
+    "best_ms": 0.18989126205444337,
+    "best_speedup_vs_triton": 0.23988993418992863,
+    "best_speedup_vs_aoTriton": 0.20725714010497026,
+    "best_path": "2d"
+  },
+  {
+    "label": "qwen3next_80b_b1_kv1024",
+    "signature": "b1_sq1_sk1024_nhq16_nhk2_hd256_bs16_bf16",
+    "bias": "none",
+    "arch": "gfx942",
+    "batch": 1,
+    "seqlen_q": 1,
+    "seqlen_k": 1024,
+    "num_query_heads": 16,
+    "num_kv_heads": 2,
+    "head_size": 256,
+    "block_size": 16,
+    "dtype": "bf16",
+    "triton_ms": 0.09127595901489258,
+    "aoTriton_ms": 0.07233071804046631,
+    "dsl": {
+      "30": {
+        "ms": 0.032355360984802246,
+        "path": "3d"
+      },
+      "60": {
+        "ms": 0.034133880138397216,
+        "path": "3d"
+      },
+      "80": {
+        "ms": 0.03435358047485351,
+        "path": "3d"
+      },
+      "120": {
+        "ms": 0.03628361940383911,
+        "path": "3d"
+      }
+    },
+    "best_sms": 30,
+    "best_ms": 0.032355360984802246,
+    "best_speedup_vs_triton": 2.8210459174838487,
+    "best_speedup_vs_aoTriton": 2.2355095365630766,
+    "best_path": "3d"
+  },
+  {
+    "label": "qwen3next_80b_b1_kv2048",
+    "signature": "b1_sq1_sk2048_nhq16_nhk2_hd256_bs16_bf16",
+    "bias": "none",
+    "arch": "gfx942",
+    "batch": 1,
+    "seqlen_q": 1,
+    "seqlen_k": 2048,
+    "num_query_heads": 16,
+    "num_kv_heads": 2,
+    "head_size": 256,
+    "block_size": 16,
+    "dtype": "bf16",
+    "triton_ms": 0.09660985946655273,
+    "aoTriton_ms": 0.1393355369567871,
+    "dsl": {
+      "30": {
+        "ms": 0.03554994106292725,
+        "path": "3d"
+      },
+      "60": {
+        "ms": 0.03410582065582275,
+        "path": "3d"
+      },
+      "80": {
+        "ms": 0.03171710014343262,
+        "path": "3d"
+      },
+      "120": {
+        "ms": 0.03234813928604126,
+        "path": "3d"
+      }
+    },
+    "best_sms": 80,
+    "best_ms": 0.03171710014343262,
+    "best_speedup_vs_triton": 3.045986519248573,
+    "best_speedup_vs_aoTriton": 4.393073021388372,
+    "best_path": "3d"
+  },
+  {
+    "label": "qwen3next_80b_b1_kv4096",
+    "signature": "b1_sq1_sk4096_nhq16_nhk2_hd256_bs16_bf16",
+    "bias": "none",
+    "arch": "gfx942",
+    "batch": 1,
+    "seqlen_q": 1,
+    "seqlen_k": 4096,
+    "num_query_heads": 16,
+    "num_kv_heads": 2,
+    "head_size": 256,
+    "block_size": 16,
+    "dtype": "bf16",
+    "triton_ms": 0.08857053756713867,
+    "aoTriton_ms": 0.2826033592224121,
+    "dsl": {
+      "30": {
+        "ms": 0.03523721933364868,
+        "path": "3d"
+      },
+      "60": {
+        "ms": 0.03387006044387817,
+        "path": "3d"
+      },
+      "80": {
+        "ms": 0.03377546072006225,
+        "path": "3d"
+      },
+      "120": {
+        "ms": 0.03397269964218139,
+        "path": "3d"
+      }
+    },
+    "best_sms": 80,
+    "best_ms": 0.03377546072006225,
+    "best_speedup_vs_triton": 2.6223339572250084,
+    "best_speedup_vs_aoTriton": 8.367120779333996,
+    "best_path": "3d"
+  },
+  {
+    "label": "qwen3next_80b_b1_kv8192",
+    "signature": "b1_sq1_sk8192_nhq16_nhk2_hd256_bs16_bf16",
+    "bias": "none",
+    "arch": "gfx942",
+    "batch": 1,
+    "seqlen_q": 1,
+    "seqlen_k": 8192,
+    "num_query_heads": 16,
+    "num_kv_heads": 2,
+    "head_size": 256,
+    "block_size": 16,
+    "dtype": "bf16",
+    "triton_ms": 0.0891510772705078,
+    "aoTriton_ms": 0.5559293365478516,
+    "dsl": {
+      "30": {
+        "ms": 0.04642940044403076,
+        "path": "3d"
+      },
+      "60": {
+        "ms": 0.03900991916656494,
+        "path": "3d"
+      },
+      "80": {
+        "ms": 0.03926569938659668,
+        "path": "3d"
+      },
+      "120": {
+        "ms": 0.03922800064086914,
+        "path": "3d"
+      }
+    },
+    "best_sms": 60,
+    "best_ms": 0.03900991916656494,
+    "best_speedup_vs_triton": 2.285343809348839,
+    "best_speedup_vs_aoTriton": 14.250973814483926,
+    "best_path": "3d"
+  },
+  {
+    "label": "qwen3next_80b_b1_kv16384",
+    "signature": "b1_sq1_sk16384_nhq16_nhk2_hd256_bs16_bf16",
+    "bias": "none",
+    "arch": "gfx942",
+    "batch": 1,
+    "seqlen_q": 1,
+    "seqlen_k": 16384,
+    "num_query_heads": 16,
+    "num_kv_heads": 2,
+    "head_size": 256,
+    "block_size": 16,
+    "dtype": "bf16",
+    "triton_ms": 0.08883835792541504,
+    "aoTriton_ms": 1.1062545776367188,
+    "dsl": {
+      "30": {
+        "ms": 0.06837520122528076,
+        "path": "3d"
+      },
+      "60": {
+        "ms": 0.05022213935852051,
+        "path": "3d"
+      },
+      "80": {
+        "ms": 0.050214138031005856,
+        "path": "3d"
+      },
+      "120": {
+        "ms": 0.05013636112213135,
+        "path": "3d"
+      }
+    },
+    "best_sms": 120,
+    "best_ms": 0.05013636112213135,
+    "best_speedup_vs_triton": 1.771934698431868,
+    "best_speedup_vs_aoTriton": 22.06491561966176,
+    "best_path": "3d"
+  },
+  {
+    "label": "qwen3next_80b_b1_kv32768",
+    "signature": "b1_sq1_sk32768_nhq16_nhk2_hd256_bs16_bf16",
+    "bias": "none",
+    "arch": "gfx942",
+    "batch": 1,
+    "seqlen_q": 1,
+    "seqlen_k": 32768,
+    "num_query_heads": 16,
+    "num_kv_heads": 2,
+    "head_size": 256,
+    "block_size": 16,
+    "dtype": "bf16",
+    "triton_ms": 0.09160473823547363,
+    "aoTriton_ms": 2.3580221557617187,
     "dsl": {
       "30": {
-        "ms": 0.25100074768066405,
+        "ms": 0.11201016426086426,
         "path": "3d"
       },
       "60": {
-        "ms": 0.13238825798034667,
+        "ms": 0.07200193881988526,
         "path": "3d"
       },
       "80": {
-        "ms": 0.08947735786437988,
+        "ms": 0.07207491874694824,
         "path": "3d"
       },
       "120": {
-        "ms": 0.08942846298217773,
+        "ms": 0.07196588039398194,
         "path": "3d"
       }
     },
     "best_sms": 120,
-    "best_ms": 0.08942846298217773,
-    "best_speedup_vs_triton": 0.9835915716903905,
-    "best_speedup_vs_aoTriton": 18.067371961983756,
+    "best_ms": 0.07196588039398194,
+    "best_speedup_vs_triton": 1.2728912330951483,
+    "best_speedup_vs_aoTriton": 32.76583490471556,
     "best_path": "3d"
   }
 ]
diff --git a/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx950/attention/decode/decode_shapes.json b/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx950/attention/decode/decode_shapes.json
index 1a890de7afbb..48d8abc3e7e7 100644
--- a/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx950/attention/decode/decode_shapes.json
+++ b/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx950/attention/decode/decode_shapes.json
@@ -115,6 +115,22 @@
         {"batch": 1, "seqlen_q": 1, "seqlen_k": 16384, "dtype": "fp16", "label": "qwen3_235b_fp16_b1_kv16384"},
         {"batch": 1, "seqlen_q": 1, "seqlen_k": 32768, "dtype": "fp16", "label": "qwen3_235b_fp16_b1_kv32768"}
       ]
+    },
+    {
+      "model": "Qwen3-Next-80B-A3B",
+      "num_query_heads": 16,
+      "num_kv_heads": 2,
+      "head_size": 256,
+      "block_size": 16,
+      "shapes": [
+        {"batch": 1, "seqlen_q": 1, "seqlen_k": 512,   "dtype": "bf16", "label": "qwen3next_80b_b1_kv512"},
+        {"batch": 1, "seqlen_q": 1, "seqlen_k": 1024,  "dtype": "bf16", "label": "qwen3next_80b_b1_kv1024"},
+        {"batch": 1, "seqlen_q": 1, "seqlen_k": 2048,  "dtype": "bf16", "label": "qwen3next_80b_b1_kv2048"},
+        {"batch": 1, "seqlen_q": 1, "seqlen_k": 4096,  "dtype": "bf16", "label": "qwen3next_80b_b1_kv4096"},
+        {"batch": 1, "seqlen_q": 1, "seqlen_k": 8192,  "dtype": "bf16", "label": "qwen3next_80b_b1_kv8192"},
+        {"batch": 1, "seqlen_q": 1, "seqlen_k": 16384, "dtype": "bf16", "label": "qwen3next_80b_b1_kv16384"},
+        {"batch": 1, "seqlen_q": 1, "seqlen_k": 32768, "dtype": "bf16", "label": "qwen3next_80b_b1_kv32768"}
+      ]
     }
   ]
 }
diff --git a/dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention.py b/dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention.py
index a0bd12d5a44f..ee77ff5add2b 100644
--- a/dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention.py
+++ b/dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention.py
@@ -40,7 +40,7 @@
 from __future__ import annotations
 
 from dataclasses import asdict, dataclass
-from typing import Sequence, Tuple
+from typing import Any, Optional, Sequence, Tuple
 
 from rocke.core.arch import ArchTarget
 from kernels.common.attention_unified import (
@@ -150,6 +150,9 @@ class AttentionSpec:
     num_query_heads: int
     num_kv_heads: int
     name: str = "rocke_attention_unified"
+    # Pinned codegen knobs from a specialized candidate; None for generic paths.
+    # Stored as a sorted tuple of (key, value) pairs for hashability.
+    # tiled_overrides: Optional[Tuple[Tuple[str, Any], ...]] = None
 
     def kernel_name(self) -> str:
         from rocke.helpers.spec import kernel_name_join
@@ -301,6 +304,78 @@ def select(req: OperatorRequest) -> AttentionSpec:
 ATTENTION_REGISTRY.register(_make_gfx942_dense_pipe_candidate())
 
 
+def _make_d256_decode_candidate() -> KernelCandidate:
+    """D256 bf16 decode candidate for gfx950 and gfx942 — 3D split-KV path.
+
+    Registered at priority 5 so it outranks the generic unified_3d candidate
+    (priority 10) for eligible D256 bf16 decode shapes. Callers can also force
+    this path explicitly via algorithm="d256_decode".
+    """
+    from kernels.common.attention_unified import _d256_decode_cohort
+
+    spec_id = "d256_decode"
+    name = "attention_d256_decode"
+
+    def support(req: OperatorRequest) -> Tuple[bool, str]:
+        errors = _request_errors(req)
+        if errors:
+            return False, "; ".join(errors)
+        assert isinstance(req, AttentionRequest)
+        if req.arch not in ("gfx942", "gfx950"):
+            return False, f"d256_decode requires gfx942 or gfx950 (got {req.arch!r})"
+        if req.dtype != "bf16":
+            return False, f"d256_decode is bf16-only (got {req.dtype!r})"
+        ok, why = _selector_matches(req, candidate)
+        if not ok:
+            return False, why
+        problem = _problem(req)
+        ok, why = supports_native_unified_attention(problem)
+        if not ok:
+            return False, why
+        if not _d256_decode_cohort(problem):
+            return False, "problem is not in the D256 bf16 decode cohort"
+        if problem.select_path() != "3d":
+            return False, (
+                f"problem routes to {problem.select_path()!r} path, not '3d'"
+            )
+        return True, "ok"
+
+    def select(req: OperatorRequest) -> AttentionSpec:
+        ok, why = support(req)
+        if not ok:
+            raise ValueError(f"{name} does not support request: {why}")
+        assert isinstance(req, AttentionRequest)
+        problem = _problem(req)
+        return AttentionSpec(
+            path="3d",
+            head_size=problem.head_size,
+            block_size=problem.block_size,
+            dtype=problem.dtype,
+            num_query_heads=problem.num_query_heads,
+            num_kv_heads=problem.num_kv_heads,
+            name="rocke_attention_d256_decode",
+        )
+
+    candidate = KernelCandidate(
+        name=name,
+        family=_FAMILY,
+        algorithm="d256_decode",
+        spec_id=spec_id,
+        abi_version=ATTENTION_ABI_VERSION,
+        priority=5,
+        supports=support,
+        select_spec=select,
+        signature=lambda _spec: (),
+        grid=lambda spec, req: (0, 0, 0),
+        block=lambda spec: (0, 0, 0),
+        sweep_space=lambda req: (select(req),) if support(req)[0] else (),
+    )
+    return candidate
+
+
+ATTENTION_REGISTRY.register(_make_d256_decode_candidate())
+
+
 def attention_candidates() -> Tuple[KernelCandidate, ...]:
     return ATTENTION_REGISTRY.candidates()
 
diff --git a/dnn-providers/hip-kernel-provider/rocke/library/kernels/common/attention_unified.py b/dnn-providers/hip-kernel-provider/rocke/library/kernels/common/attention_unified.py
index cc4926e62da4..190982ac0295 100644
--- a/dnn-providers/hip-kernel-provider/rocke/library/kernels/common/attention_unified.py
+++ b/dnn-providers/hip-kernel-provider/rocke/library/kernels/common/attention_unified.py
@@ -2176,7 +2176,10 @@ def _num_segments(problem: UnifiedAttentionProblem) -> int:
             return min(segments, 32)
         if problem.head_size == 128:
             return min(segments, 16)
-        return min(segments, 64)
+        # D256: sweep (decode_shapes_perf gfx942) shows seg128 wins or ties at
+        # every kv_len — do not cap below the formula value.
+        if problem.head_size != 256:
+            return min(segments, 64)
     return segments
 
 
@@ -2223,6 +2226,26 @@ def _enable_gfx942_3d_wide_kv_load(problem: UnifiedAttentionProblem) -> bool:
     return True
 
 
+def _d256_decode_cohort(problem: UnifiedAttentionProblem) -> bool:
+    """Predicate: true when this problem belongs to the D256 bf16 decode cohort.
+
+    Covers gfx950 and gfx942, bf16, head_size=256, decode (all_decode=True).
+    No sliding window, softcap, sinks, alibi, or qq_bias — these features
+    are not validated for the D256 decode fast path.
+    """
+    return (
+        problem.head_size == 256
+        and problem.dtype == "bf16"
+        and problem.all_decode
+        and not problem.use_fp8
+        and problem.sliding_window == 0
+        and problem.softcap == 0.0
+        and not problem.use_sinks
+        and not problem.use_alibi
+        and not problem.use_qq_bias
+    )
+
+
 def _env_enabled_true(var: str) -> bool:
     env = __import__("os").environ.get(var, "").strip().lower()
     return env in ("1", "on", "enable", "enabled", "yes", "true")
diff --git a/dnn-providers/hip-kernel-provider/rocke/library/tests/dispatch/attention/test_d256_decode_wiring.py b/dnn-providers/hip-kernel-provider/rocke/library/tests/dispatch/attention/test_d256_decode_wiring.py
new file mode 100644
index 000000000000..ed316716e32c
--- /dev/null
+++ b/dnn-providers/hip-kernel-provider/rocke/library/tests/dispatch/attention/test_d256_decode_wiring.py
@@ -0,0 +1,228 @@
+# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+# SPDX-License-Identifier: MIT
+"""Unit tests for the D256 bf16 decode dispatcher wiring (gfx950 + gfx942).
+
+Covers:
+  - d256_decode candidate is registered and discoverable
+  - d256_decode priority outranks unified_3d (priority 10) for eligible shapes
+  - gfx950 decode GQA (16/2 heads, hd256, bf16) routes to d256_decode
+  - gfx942 decode routes to d256_decode
+  - explicit algorithm="d256_decode" forces the candidate
+  - spec path is "3d" and name is "rocke_attention_d256_decode"
+  - gates: rejects non-gfx942/gfx950 arch, non-bf16, prefill (2D), hd!=256,
+    sliding_window, sinks, softcap
+"""
+
+from __future__ import annotations
+
+import os
+import unittest
+
+import kernels.common.attention_unified as au
+from dispatch.attention import (
+    AttentionRequest,
+    attention_candidates,
+    dispatch_attention,
+)
+
+
+def _gfx950_d256_decode(**kw) -> AttentionRequest:
+    base = dict(
+        batch=1,
+        nhead_q=16,
+        nhead_k=2,
+        seqlen_q=1,
+        seqlen_k=8192,
+        hdim_q=256,
+        hdim_v=256,
+        arch="gfx950",
+        dtype="bf16",
+    )
+    base.update(kw)
+    return AttentionRequest(**base)
+
+
+class _PinnedArch:
+    """Context-manager that pins _RESOLVED_ATTENTION_ARCH to a given arch."""
+
+    def __init__(self, arch: str):
+        self._arch = arch
+
+    def __enter__(self):
+        self._old = au._RESOLVED_ATTENTION_ARCH
+        au._RESOLVED_ATTENTION_ARCH = self._arch
+        return self
+
+    def __exit__(self, *_):
+        au._RESOLVED_ATTENTION_ARCH = self._old
+
+
+class TestD256DecodeRegistration(unittest.TestCase):
+    def test_candidate_is_registered(self):
+        names = [c.name for c in attention_candidates()]
+        self.assertIn("attention_d256_decode", names)
+
+    def test_spec_id_is_d256_decode(self):
+        candidate = next(
+            c for c in attention_candidates() if c.name == "attention_d256_decode"
+        )
+        self.assertEqual(candidate.spec_id, "d256_decode")
+
+    def test_algorithm_is_d256_decode(self):
+        candidate = next(
+            c for c in attention_candidates() if c.name == "attention_d256_decode"
+        )
+        self.assertEqual(candidate.algorithm, "d256_decode")
+
+    def test_priority_beats_unified_3d(self):
+        # The registry sorts ascending; d256_decode must have a lower priority
+        # number than unified_3d so it is selected first for eligible requests.
+        d256 = next(
+            c for c in attention_candidates() if c.name == "attention_d256_decode"
+        )
+        unified_3d = next(
+            c for c in attention_candidates() if c.name == "attention_unified_3d"
+        )
+        self.assertLess(d256.priority, unified_3d.priority)
+
+
+class TestD256DecodeSupportGates(unittest.TestCase):
+    def _candidate(self):
+        return next(
+            c for c in attention_candidates() if c.name == "attention_d256_decode"
+        )
+
+    def test_rejects_non_gfx942_gfx950_arch(self):
+        with _PinnedArch("gfx950"):
+            ok, why = self._candidate().supports(_gfx950_d256_decode(arch="gfx1250"))
+            self.assertFalse(ok)
+            self.assertIn("gfx942", why)
+
+    def test_rejects_non_bf16_dtype(self):
+        with _PinnedArch("gfx950"):
+            ok, why = self._candidate().supports(_gfx950_d256_decode(dtype="fp16"))
+            self.assertFalse(ok)
+            self.assertIn("bf16", why)
+
+    def test_rejects_prefill(self):
+        # Prefill (seqlen_q>1) routes to 2d, not 3d — d256_decode must refuse.
+        with _PinnedArch("gfx950"):
+            req = _gfx950_d256_decode(
+                batch=2, nhead_q=16, nhead_k=2, seqlen_q=512, seqlen_k=512
+            )
+            ok, why = self._candidate().supports(req)
+            self.assertFalse(ok)
+            # Either the cohort predicate (not all_decode) or the path check fires.
+            self.assertTrue("cohort" in why or "3d" in why or "2d" in why)
+
+    def test_rejects_non_d256_head_size(self):
+        with _PinnedArch("gfx950"):
+            ok, why = self._candidate().supports(
+                _gfx950_d256_decode(hdim_q=128, hdim_v=128)
+            )
+            self.assertFalse(ok)
+            self.assertIn("cohort", why)
+
+    def test_rejects_sliding_window(self):
+        with _PinnedArch("gfx950"):
+            ok, why = self._candidate().supports(_gfx950_d256_decode(sliding_window=64))
+            self.assertFalse(ok)
+            self.assertIn("cohort", why)
+
+    def test_rejects_sinks(self):
+        with _PinnedArch("gfx950"):
+            ok, why = self._candidate().supports(_gfx950_d256_decode(use_sinks=True))
+            self.assertFalse(ok)
+            self.assertIn("cohort", why)
+
+
+class TestD256DecodeRouting(unittest.TestCase):
+    """Verify that dispatch picks d256_decode for eligible shapes."""
+
+    def test_gfx950_gqa_decode_routes_d256_decode(self):
+        with _PinnedArch("gfx950"):
+            r = dispatch_attention(_gfx950_d256_decode())
+        self.assertEqual(r.candidate.spec_id, "d256_decode")
+        self.assertEqual(r.spec.path, "3d")
+        self.assertEqual(r.spec.name, "rocke_attention_d256_decode")
+
+    def test_gfx942_gqa_decode_routes_d256_decode(self):
+        with _PinnedArch("gfx942"):
+            r = dispatch_attention(_gfx950_d256_decode(arch="gfx942"))
+        self.assertEqual(r.candidate.spec_id, "d256_decode")
+        self.assertEqual(r.spec.path, "3d")
+
+    def test_spec_records_correct_dims(self):
+        with _PinnedArch("gfx950"):
+            r = dispatch_attention(
+                _gfx950_d256_decode(
+                    batch=2,
+                    nhead_q=16,
+                    nhead_k=2,
+                    seqlen_q=1,
+                    seqlen_k=4096,
+                    kv_block_size=32,
+                )
+            )
+        self.assertEqual(r.spec.head_size, 256)
+        self.assertEqual(r.spec.block_size, 32)
+        self.assertEqual(r.spec.dtype, "bf16")
+
+    def test_explicit_algorithm_forces_d256_decode(self):
+        with _PinnedArch("gfx950"):
+            r = dispatch_attention(_gfx950_d256_decode(algorithm="d256_decode"))
+        self.assertEqual(r.candidate.algorithm, "d256_decode")
+
+    def test_explicit_algorithm_unified_3d_bypasses_d256_decode(self):
+        with _PinnedArch("gfx950"):
+            r = dispatch_attention(
+                _gfx950_d256_decode(algorithm="auto", spec_id="unified_3d")
+            )
+        self.assertEqual(r.candidate.spec_id, "unified_3d")
+
+    def test_d256_decode_cohort_predicate_all_decode(self):
+        # The cohort helper itself: all_decode=True required.
+        p_decode = au.UnifiedAttentionProblem(
+            total_q=1,
+            num_seqs=1,
+            num_query_heads=16,
+            num_kv_heads=2,
+            head_size=256,
+            block_size=16,
+            max_seqlen_q=1,
+            max_seqlen_k=8192,
+            dtype="bf16",
+        )
+        self.assertTrue(au._d256_decode_cohort(p_decode))
+
+    def test_d256_decode_cohort_predicate_rejects_prefill(self):
+        p_prefill = au.UnifiedAttentionProblem(
+            total_q=512,
+            num_seqs=1,
+            num_query_heads=16,
+            num_kv_heads=2,
+            head_size=256,
+            block_size=16,
+            max_seqlen_q=512,
+            max_seqlen_k=8192,
+            dtype="bf16",
+        )
+        self.assertFalse(au._d256_decode_cohort(p_prefill))
+
+    def test_d256_decode_cohort_predicate_rejects_fp16(self):
+        p = au.UnifiedAttentionProblem(
+            total_q=1,
+            num_seqs=1,
+            num_query_heads=16,
+            num_kv_heads=2,
+            head_size=256,
+            block_size=16,
+            max_seqlen_q=1,
+            max_seqlen_k=8192,
+            dtype="fp16",
+        )
+        self.assertFalse(au._d256_decode_cohort(p))
+
+
+if __name__ == "__main__":
+    unittest.main()

From 115dc66ccf83cbe0c2f0db66711b97d2430742f0 Mon Sep 17 00:00:00 2001
From: Bartlomiej Kocot <barkocot@amd.com>
Date: Wed, 15 Jul 2026 16:37:00 +0000
Subject: [PATCH 02/11] extend tests

---
 .../prefill/parity_unified_attention.py       | 366 ++++++++++++++++--
 .../prefill/parity_unified_attention.py       |   2 +-
 2 files changed, 324 insertions(+), 44 deletions(-)

diff --git a/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/prefill/parity_unified_attention.py b/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/prefill/parity_unified_attention.py
index 29dcaf8935c7..34a457ea881a 100644
--- a/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/prefill/parity_unified_attention.py
+++ b/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/prefill/parity_unified_attention.py
@@ -55,8 +55,6 @@
 from pathlib import Path
 from typing import List, Optional
 
-SHAPES_JSON = Path(__file__).resolve().parent / "shapes.json"
-
 # Block size for the paged-KV mapping. The dense SDPA problems are mapped onto
 # one contiguous run of paged-KV blocks per sequence; block_size=64 matches the
 # gfx942 flash-regime tile (T=64) so the default wide4 / L4 geometries apply.
@@ -86,57 +84,339 @@ def num_queries_per_kv(self) -> int:
         return self.heads // self.kv_heads
 
 
-def load_shapes(path: Path = SHAPES_JSON) -> List[Shape]:
-    """Load the shipped canonical shape set from ``shapes.json``."""
-    data = json.loads(path.read_text())
+# ---------------------------------------------------------------------------
+# In-code scenario functions (adapted from gfx950's parity_unified_attention).
+#
+# gfx942 constraints vs. gfx950:
+#   - causal-only (no sliding window / softcap / ALiBi / QQ-bias / sinks)
+#   - BLOCK_SIZE = 64 (tile size T = BLOCK_SIZE)
+#   - head_size ∈ {64, 128} for the shipped paths
+#   - Scenarios with non-uniform seq_lens are split into one Shape per unique
+#     (seqlen_q, seqlen_k) pair (batch=1) and given a disambiguating suffix.
+# ---------------------------------------------------------------------------
+
+
+def _shape(
+    name: str,
+    dtype: str,
+    seqlen_q: int,
+    seqlen_k: int,
+    heads: int,
+    kv_heads: int,
+    head_size: int,
+    batch: int,
+    group: str,
+) -> Shape:
+    return Shape(
+        name=name,
+        dtype=dtype,
+        seqlen_q=seqlen_q,
+        seqlen_k=seqlen_k,
+        head_size=head_size,
+        heads=heads,
+        kv_heads=kv_heads,
+        batch=batch,
+        causal=True,
+        group=group,
+    )
+
+
+def default_scenarios() -> List[Shape]:
+    """Core production scenarios adapted from the gfx950 default set.
+
+    Sliding-window / softcap / ALiBi / QQ-bias / sinks scenarios are dropped
+    (unsupported on the gfx942 SDPA-fwd net). Decode shapes (seqlen_q == 1)
+    are included; they exercise the T=64 single-query path.
+    Non-uniform seq_lens from the gfx950 mixed/prefill scenarios are split
+    into one Shape per unique (q, k) pair.
+    """
+    g = "default"
     shapes: List[Shape] = []
-    for group, items in data["scenarios"].items():
-        for it in items:
-            shapes.append(
-                Shape(
-                    name=it["name"],
-                    dtype=it["dtype"],
-                    seqlen_q=int(it["seqlen_q"]),
-                    seqlen_k=int(it["seqlen_k"]),
-                    head_size=int(it["head_size"]),
-                    heads=int(it["heads"]),
-                    kv_heads=int(it["kv_heads"]),
-                    batch=int(it["batch"]),
-                    causal=bool(it["causal"]),
-                    group=group,
-                )
+
+    # decode_d128_b16 / b64 equivalents (block_size pinned to 64 on gfx942)
+    for sq, sk in [(1, 1024), (1, 2048), (1, 4096), (1, 512)]:
+        shapes.append(
+            _shape(f"decode_d128_q{sq}k{sk}", "fp16", sq, sk, 16, 2, 128, 1, g)
+        )
+
+    # decode_d256 (head_size=256, fp16)
+    for sq, sk in [(1, 1024), (1, 2048)]:
+        shapes.append(
+            _shape(f"decode_d256_q{sq}k{sk}", "fp16", sq, sk, 16, 2, 256, 1, g)
+        )
+
+    # bf16 D256 decode — Qwen3-Next-80B-A3B gated_attn GQA 16/2 (d256_decode cohort)
+    for sq, sk in [(1, 1024), (1, 2048), (1, 4096), (1, 8192)]:
+        shapes.append(
+            _shape(f"bf16_decode_d256_q{sq}k{sk}", "bf16", sq, sk, 16, 2, 256, 1, g)
+        )
+
+    # prefill_d128 (uniform batch from gfx950 prefill_d128_b16)
+    for sq, sk in [(64, 64), (128, 256), (32, 256)]:
+        shapes.append(
+            _shape(f"prefill_d128_q{sq}k{sk}", "fp16", sq, sk, 16, 2, 128, 1, g)
+        )
+
+    # mixed_d128 split into individual shapes
+    for sq, sk in [(1, 1328), (5, 18), (129, 463)]:
+        shapes.append(
+            _shape(f"mixed_d128_q{sq}k{sk}", "fp16", sq, sk, 16, 2, 128, 1, g)
+        )
+
+    # bf16 decode
+    for sq, sk in [(1, 1024), (1, 2048), (1, 4096)]:
+        shapes.append(
+            _shape(f"bf16_decode_d128_q{sq}k{sk}", "bf16", sq, sk, 16, 2, 128, 1, g)
+        )
+
+    # GQA non-power-of-2 NQK ratios (from gfx950 gqa_nqk* scenarios)
+    for sq, sk in [(1024, 1024), (512, 512)]:
+        shapes.append(
+            _shape(f"gqa_nqk5_d128_q{sq}k{sk}", "bf16", sq, sk, 40, 8, 128, 1, g)
+        )
+        shapes.append(
+            _shape(f"gqa_nqk7_d128_q{sq}k{sk}", "bf16", sq, sk, 28, 4, 128, 1, g)
+        )
+        shapes.append(
+            _shape(f"gqa_nqk3_d128_q{sq}k{sk}", "bf16", sq, sk, 24, 8, 128, 1, g)
+        )
+        shapes.append(
+            _shape(f"gqa_nqk8_d128_q{sq}k{sk}", "bf16", sq, sk, 32, 4, 128, 1, g)
+        )
+
+    # combo bf16 d64/b32/h64kv8 prefill (from gfx950 combo_bf16_d64_b32_gqa8_64x8)
+    shapes.append(
+        _shape("combo_bf16_d64_h64kv8_q512k1024", "bf16", 512, 1024, 64, 8, 64, 2, g)
+    )
+
+    return shapes
+
+
+def fmha_scenarios() -> List[Shape]:
+    """FMHA matrix shapes adapted from gfx950 (causal-compatible subset only).
+
+    Drops sliding-window / softcap / ALiBi / QQ-bias variants. Non-uniform
+    seq_lens are split per (q, k) pair. Batch counts are kept from the
+    gfx950 ``rep(n, q, k)`` helper where the layout is uniform.
+    """
+    g = "fmha"
+    shapes: List[Shape] = []
+
+    # GQA_4to1_Prefill_Basic (32:8 → 16:4 to stay ≤ 16 NQK)
+    shapes.append(
+        _shape("fmha_gqa_4to1_prefill_2k_b1", "fp16", 2048, 2048, 16, 4, 128, 1, g)
+    )
+    shapes.append(
+        _shape("fmha_gqa_4to1_prefill_2k_b4", "fp16", 2048, 2048, 16, 4, 128, 4, g)
+    )
+    shapes.append(
+        _shape("fmha_gqa_4to1_prefill_2k_bf16", "bf16", 2048, 2048, 16, 4, 128, 2, g)
+    )
+
+    # GQA_16to1
+    shapes.append(
+        _shape("fmha_gqa_16to1_prefill_2k_b1", "fp16", 2048, 2048, 16, 1, 128, 1, g)
+    )
+
+    # MQA decode (seqlen_q = 1)
+    shapes.append(
+        _shape("fmha_mqa_16to1_decode_1k_b8", "fp16", 1, 1024, 16, 1, 128, 8, g)
+    )
+    shapes.append(
+        _shape("fmha_mqa_16to1_decode_4k_b8", "bf16", 1, 4096, 16, 1, 128, 8, g)
+    )
+
+    # Tiny sequences (split per pair)
+    for sq, sk in [(1, 10), (3, 99), (33, 33), (1, 33), (3, 10)]:
+        shapes.append(
+            _shape(f"fmha_tiny_seqs_q{sq}k{sk}", "fp16", sq, sk, 16, 1, 128, 1, g)
+        )
+
+    # Asymmetric seqlen (split per pair)
+    for sq, sk in [(100, 256), (51, 99), (256, 1024)]:
+        shapes.append(
+            _shape(f"fmha_asym_b1_q{sq}k{sk}", "fp16", sq, sk, 16, 1, 128, 1, g)
+        )
+    for sq, sk in [(100, 256), (51, 99), (256, 1024)]:
+        shapes.append(
+            _shape(f"fmha_asym_b2_d64_q{sq}k{sk}", "fp16", sq, sk, 16, 2, 128, 2, g)
+        )
+
+    # Cross-attention (sq << sk)
+    for sq, sk in [
+        (1, 1024),
+        (1, 4096),
+        (32, 1024),
+        (32, 4096),
+        (128, 1024),
+        (128, 4096),
+    ]:
+        shapes.append(
+            _shape(f"fmha_cross_attn_q{sq}k{sk}", "fp16", sq, sk, 16, 4, 128, 1, g)
+        )
+
+    # Paged decode (many sequences, short Q)
+    shapes.append(_shape("fmha_paged_decode_b80", "fp16", 1, 4096, 16, 4, 128, 80, g))
+    shapes.append(
+        _shape("fmha_paged_decode_q4_b16", "fp16", 4, 4096, 16, 4, 128, 16, g)
+    )
+
+    # Padding / boundary stress (odd lengths)
+    for sq in [259, 500, 987, 1023]:
+        shapes.append(
+            _shape(f"fmha_pad_boundary_q{sq}", "fp16", sq, sq, 16, 4, 128, 1, g)
+        )
+
+    # Prefill odd lengths
+    for sq, sk in [(113, 203), (339, 339), (799, 799), (1023, 1024), (3131, 3131)]:
+        shapes.append(
+            _shape(f"fmha_prefill_odd_q{sq}k{sk}", "fp16", sq, sk, 16, 4, 128, 1, g)
+        )
+
+    # H256 high-LDS shapes
+    shapes.append(_shape("fmha_h256_2k", "bf16", 2048, 2048, 16, 2, 256, 1, g))
+    shapes.append(_shape("fmha_h256_4k_b2", "bf16", 4096, 4096, 8, 4, 256, 2, g))
+
+    # Long sequence stress (8K / 16K)
+    shapes.append(_shape("fmha_long_8k", "bf16", 8192, 8192, 16, 4, 128, 1, g))
+    shapes.append(_shape("fmha_long_16k", "bf16", 16384, 16384, 16, 4, 128, 1, g))
+
+    # Extreme batch (many short sequences)
+    shapes.append(_shape("fmha_batch_64_s128", "fp16", 128, 128, 8, 8, 128, 64, g))
+    shapes.append(_shape("fmha_batch_128_s128", "fp16", 128, 128, 8, 8, 128, 128, g))
+
+    # CK benchmark standard
+    shapes.append(_shape("fmha_bench_1k_b4", "fp16", 1024, 1024, 16, 16, 128, 4, g))
+    shapes.append(_shape("fmha_bench_4k_b2_d256", "fp16", 4096, 4096, 8, 8, 256, 2, g))
+
+    # Vision transformer shapes
+    shapes.append(_shape("fmha_vit_1k_b4", "fp16", 1024, 1024, 16, 8, 128, 4, g))
+    shapes.append(_shape("fmha_vit_256_b4", "fp16", 256, 256, 16, 8, 128, 4, g))
+
+    return shapes
+
+
+def creative_scenarios() -> List[Shape]:
+    """Exploratory sweep adapted from gfx950 creative scenarios (causal-only).
+
+    Long-context decode, GQA/MQA variants, head_size=256, bf16, chunked
+    prefill, and batch sweeps. Sliding-window / softcap / ALiBi / QQ-bias
+    shapes are omitted entirely.
+    """
+    g = "creative"
+    shapes: List[Shape] = []
+
+    # Long-context decode
+    shapes.append(_shape("creative_decode_8k", "fp16", 1, 8192, 16, 2, 128, 1, g))
+    shapes.append(_shape("creative_decode_32k", "fp16", 1, 32768, 16, 2, 128, 1, g))
+    shapes.append(
+        _shape("creative_decode_64k_bf16", "bf16", 1, 65536, 16, 2, 128, 1, g)
+    )
+
+    # Large decode batch (varied kv per seq → one shape each)
+    for sk in [512, 1024, 2048, 4096, 768, 1536, 3072, 6144, 384, 896, 1280, 2560]:
+        shapes.append(
+            _shape(f"creative_decode_b16_k{sk}", "fp16", 1, sk, 16, 2, 128, 1, g)
+        )
+
+    # Tiny + large prefill
+    shapes.append(_shape("creative_prefill_tiny", "fp16", 4, 4, 16, 2, 128, 1, g))
+    shapes.append(_shape("creative_prefill_512", "fp16", 512, 512, 16, 2, 128, 1, g))
+    shapes.append(_shape("creative_prefill_2k", "fp16", 2048, 2048, 16, 2, 128, 1, g))
+
+    # Chunked prefill (sq < sk)
+    for sq, sk in [(128, 2048), (256, 4096), (512, 8192)]:
+        shapes.append(
+            _shape(
+                f"creative_chunk_prefill_q{sq}k{sk}", "fp16", sq, sk, 16, 2, 128, 1, g
+            )
+        )
+
+    # GQA / MQA variants
+    for sq, sk in [(1, 4096), (1, 8192)]:
+        shapes.append(
+            _shape(f"creative_gqa_h32k4_q{sq}k{sk}", "fp16", sq, sk, 32, 4, 128, 1, g)
+        )
+        shapes.append(
+            _shape(f"creative_mqa_h16k1_q{sq}k{sk}", "fp16", sq, sk, 16, 1, 128, 1, g)
+        )
+    for sq, sk in [(1, 2048), (1, 4096)]:
+        shapes.append(
+            _shape(f"creative_gqa_h64k8_q{sq}k{sk}", "fp16", sq, sk, 64, 8, 128, 1, g)
+        )
+
+    # D64 bf16 (from gfx950 creative_r1r4_bf16_d64_sinks -- sinks dropped)
+    for sq, sk in [(640, 704), (640, 768)]:
+        shapes.append(
+            _shape(f"creative_d64_bf16_q{sq}k{sk}", "bf16", sq, sk, 64, 8, 64, 1, g)
+        )
+
+    # head_size=256 with bf16
+    for sq, sk in [(1, 4096), (1, 8192)]:
+        shapes.append(
+            _shape(
+                f"creative_d256_bf16_decode_q{sq}k{sk}",
+                "bf16",
+                sq,
+                sk,
+                16,
+                2,
+                256,
+                1,
+                g,
+            )
+        )
+    for sq, sk in [(128, 1024), (256, 2048)]:
+        shapes.append(
+            _shape(
+                f"creative_d256_prefill_q{sq}k{sk}", "fp16", sq, sk, 16, 2, 256, 1, g
             )
+        )
+
+    # bf16 chunked prefill (from gfx950 creative_bf16_b64_chunk)
+    for sq, sk in [(64, 2048), (128, 4096)]:
+        shapes.append(
+            _shape(f"creative_bf16_chunk_q{sq}k{sk}", "bf16", sq, sk, 16, 2, 128, 1, g)
+        )
+
     return shapes
 
 
-def select_shapes(shapes: List[Shape], selectors: Optional[List[str]]) -> List[Shape]:
+def all_inline_scenarios() -> List[Shape]:
+    return default_scenarios() + fmha_scenarios() + creative_scenarios()
+
+
+def select_shapes(selectors: Optional[List[str]]) -> List[Shape]:
     """Resolve ``--scenario`` selectors to a concrete shape list.
 
-    A selector is either a group name (``correctness`` / ``perf`` / ``all``)
-    or an exact shape name. Multiple ``--scenario`` flags union together.
+    A selector is either a group name (``default`` / ``fmha`` / ``creative`` /
+    ``all``) or an exact shape name. Multiple ``--scenario`` flags union.
+    No selector → all inline scenarios.
     """
+    all_shapes = all_inline_scenarios()
     if not selectors:
-        return shapes
+        return all_shapes
+
     wanted: List[Shape] = []
-    seen = set()
+    seen: set = set()
+
+    def _add(s: Shape) -> None:
+        if s.name not in seen:
+            seen.add(s.name)
+            wanted.append(s)
+
     for sel in selectors:
-        if sel == "all":
-            picked = shapes
-        elif sel in (
-            "correctness",
-            "perf",
-            "decode",
-            "gqa_nqk_nonpow2",
-            "short_prefill",
-            "long_prefill",
-        ):
-            picked = [s for s in shapes if s.group == sel]
+        if sel in ("all", "inline"):
+            for s in all_shapes:
+                _add(s)
+        elif sel in ("default", "fmha", "creative"):
+            for s in all_shapes:
+                if s.group == sel:
+                    _add(s)
         else:
-            picked = [s for s in shapes if s.name == sel]
-        for s in picked:
-            if s.name not in seen:
-                seen.add(s.name)
-                wanted.append(s)
+            for s in all_shapes:
+                if s.name == sel:
+                    _add(s)
     return wanted
 
 
@@ -663,7 +943,7 @@ def main() -> int:
         "--scenario",
         action="append",
         default=None,
-        help="group (correctness|perf|decode|short_prefill|long_prefill|gqa_nqk_nonpow2|all) or exact shape name; repeatable",
+        help="group (default|fmha|creative|all) or exact shape name; repeatable",
     )
     parser.add_argument("--attempts", type=int, default=30)
     parser.add_argument("--warmup", type=int, default=10)
@@ -717,7 +997,7 @@ def main() -> int:
     if _q_major_grid_enabled():
         print("HIPDNN_GFX942_Q_MAJOR_GRID -> enabled")
 
-    shapes = select_shapes(load_shapes(), args.scenario)
+    shapes = select_shapes(args.scenario)
     if not shapes:
         print(f"no shapes matched {args.scenario!r}", file=sys.stderr)
         return 2
diff --git a/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx950/attention/prefill/parity_unified_attention.py b/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx950/attention/prefill/parity_unified_attention.py
index a4ec68b820a6..006122c73c7c 100644
--- a/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx950/attention/prefill/parity_unified_attention.py
+++ b/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx950/attention/prefill/parity_unified_attention.py
@@ -891,7 +891,7 @@ def creative_scenarios() -> List[Scenario]:
         # --- head_size=256 with bf16 ---
         Scenario(
             name="creative_d256_bf16_decode",
-            seq_lens=[(1, 4096), (1, 8192)],
+            seq_lens=[(1, 1024), (1, 2048), (1, 4096), (1, 8192)],
             num_query_heads=16,
             num_kv_heads=2,
             head_size=256,

From df94b362823a010d18f47220e8fa3ad66572c4d7 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Bart=C5=82omiej=20Kocot?= <barkocot@amd.com>
Date: Thu, 16 Jul 2026 00:34:46 +0200
Subject: [PATCH 03/11] Potential fix for pull request finding

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
---
 .../hip-kernel-provider/rocke/library/dispatch/attention.py     | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention.py b/dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention.py
index ee77ff5add2b..cc19290e1e38 100644
--- a/dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention.py
+++ b/dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention.py
@@ -40,7 +40,7 @@
 from __future__ import annotations
 
 from dataclasses import asdict, dataclass
-from typing import Any, Optional, Sequence, Tuple
+from typing import Sequence, Tuple
 
 from rocke.core.arch import ArchTarget
 from kernels.common.attention_unified import (

From 0e65dafc22f697a5aca76d21c31c7b5e7e6d78e1 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Bart=C5=82omiej=20Kocot?= <barkocot@amd.com>
Date: Thu, 16 Jul 2026 00:34:55 +0200
Subject: [PATCH 04/11] Potential fix for pull request finding

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
---
 .../library/tests/dispatch/attention/test_d256_decode_wiring.py  | 1 -
 1 file changed, 1 deletion(-)

diff --git a/dnn-providers/hip-kernel-provider/rocke/library/tests/dispatch/attention/test_d256_decode_wiring.py b/dnn-providers/hip-kernel-provider/rocke/library/tests/dispatch/attention/test_d256_decode_wiring.py
index ed316716e32c..75324add11e5 100644
--- a/dnn-providers/hip-kernel-provider/rocke/library/tests/dispatch/attention/test_d256_decode_wiring.py
+++ b/dnn-providers/hip-kernel-provider/rocke/library/tests/dispatch/attention/test_d256_decode_wiring.py
@@ -15,7 +15,6 @@
 
 from __future__ import annotations
 
-import os
 import unittest
 
 import kernels.common.attention_unified as au

From 034eb15bfcbffbf5c58dece176624277d0df4cf7 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Bart=C5=82omiej=20Kocot?= <barkocot@amd.com>
Date: Thu, 16 Jul 2026 00:35:06 +0200
Subject: [PATCH 05/11] Potential fix for pull request finding

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
---
 .../gfx942/attention/prefill/parity_unified_attention.py        | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/prefill/parity_unified_attention.py b/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/prefill/parity_unified_attention.py
index 34a457ea881a..01481f0f9d95 100644
--- a/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/prefill/parity_unified_attention.py
+++ b/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/prefill/parity_unified_attention.py
@@ -90,7 +90,7 @@ def num_queries_per_kv(self) -> int:
 # gfx942 constraints vs. gfx950:
 #   - causal-only (no sliding window / softcap / ALiBi / QQ-bias / sinks)
 #   - BLOCK_SIZE = 64 (tile size T = BLOCK_SIZE)
-#   - head_size ∈ {64, 128} for the shipped paths
+#   - head_size ∈ {64, 128} for the shipped paths (+ select 256 decode shapes)
 #   - Scenarios with non-uniform seq_lens are split into one Shape per unique
 #     (seqlen_q, seqlen_k) pair (batch=1) and given a disambiguating suffix.
 # ---------------------------------------------------------------------------

From 1d79977018d22504cb433cf22ccb229f504c31f8 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Bart=C5=82omiej=20Kocot?= <barkocot@amd.com>
Date: Thu, 16 Jul 2026 00:37:12 +0200
Subject: [PATCH 06/11] Potential fix for pull request finding

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
---
 .../rocke/library/kernels/common/attention_unified.py       | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/dnn-providers/hip-kernel-provider/rocke/library/kernels/common/attention_unified.py b/dnn-providers/hip-kernel-provider/rocke/library/kernels/common/attention_unified.py
index 190982ac0295..d18bbf311e2a 100644
--- a/dnn-providers/hip-kernel-provider/rocke/library/kernels/common/attention_unified.py
+++ b/dnn-providers/hip-kernel-provider/rocke/library/kernels/common/attention_unified.py
@@ -2229,9 +2229,9 @@ def _enable_gfx942_3d_wide_kv_load(problem: UnifiedAttentionProblem) -> bool:
 def _d256_decode_cohort(problem: UnifiedAttentionProblem) -> bool:
     """Predicate: true when this problem belongs to the D256 bf16 decode cohort.
 
-    Covers gfx950 and gfx942, bf16, head_size=256, decode (all_decode=True).
-    No sliding window, softcap, sinks, alibi, or qq_bias — these features
-    are not validated for the D256 decode fast path.
+    Intended for bf16, head_size=256, decode-only (all_decode=True) shapes; the
+    architecture gate (gfx942/gfx950) is enforced by the dispatcher/caller.
+    Excludes sliding window, softcap, sinks, ALiBi, and QQ-bias.
     """
     return (
         problem.head_size == 256

From 3a4b0ca0333a574ac533b6d64be9037449cb9f76 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Bart=C5=82omiej=20Kocot?= <barkocot@amd.com>
Date: Thu, 16 Jul 2026 00:37:30 +0200
Subject: [PATCH 07/11] Potential fix for pull request finding

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
---
 .../hip-kernel-provider/rocke/library/dispatch/attention.py      | 1 -
 1 file changed, 1 deletion(-)

diff --git a/dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention.py b/dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention.py
index cc19290e1e38..0c47f213d0e9 100644
--- a/dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention.py
+++ b/dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention.py
@@ -152,7 +152,6 @@ class AttentionSpec:
     name: str = "rocke_attention_unified"
     # Pinned codegen knobs from a specialized candidate; None for generic paths.
     # Stored as a sorted tuple of (key, value) pairs for hashability.
-    # tiled_overrides: Optional[Tuple[Tuple[str, Any], ...]] = None
 
     def kernel_name(self) -> str:
         from rocke.helpers.spec import kernel_name_join

From 4eae68b6bb31b3236a1961c24a5e55a105fcc316 Mon Sep 17 00:00:00 2001
From: Bartlomiej Kocot <barkocot@amd.com>
Date: Wed, 15 Jul 2026 23:28:52 +0000
Subject: [PATCH 08/11] readme

---
 .../gfx942/attention/decode/TUNING_D256.md    | 108 ++++++++++++++++++
 .../rocke/library/dispatch/AGENTS.md          |  67 +++++++++++
 2 files changed, 175 insertions(+)
 create mode 100644 dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx942/attention/decode/TUNING_D256.md
 create mode 100644 dnn-providers/hip-kernel-provider/rocke/library/dispatch/AGENTS.md

diff --git a/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx942/attention/decode/TUNING_D256.md b/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx942/attention/decode/TUNING_D256.md
new file mode 100644
index 000000000000..6506409f9024
--- /dev/null
+++ b/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx942/attention/decode/TUNING_D256.md
@@ -0,0 +1,108 @@
+# D256 bf16 decode tuning — gfx942
+
+**Workload:** Qwen3-Next-80B-A3B gated attention decode, GQA 16/2,
+`head_size=256`, `block_size=16`, `dtype=bf16`, `batch=1`,
+`kv_len` sweep 512–32768.
+
+**Dispatcher candidate:** `attention_d256_decode` (priority 5, algorithm
+`d256_decode`) registered in `dispatch/attention.py`. Routes D256 bf16 decode
+to the 3D split-KV path on both gfx942 and gfx950.
+
+**Cohort predicate:** `_d256_decode_cohort()` in
+`kernels/common/attention_unified.py` — `head_size==256`, `dtype=="bf16"`,
+`all_decode==True`, no sliding_window / softcap / sinks / alibi / qq_bias / fp8.
+
+---
+
+## Problem found
+
+`_num_segments()` in `kernels/common/attention_unified.py` had a gfx942
+short-KV cap block:
+
+```python
+if arch == "gfx942" and max_seqlen_q == 1 and max_seqlen_k <= 2048:
+    if head_size == 64:  return min(segments, 32)
+    if head_size == 128: return min(segments, 16)
+    return min(segments, 64)   # ← D256 fell here — wrong
+```
+
+D256 inherited `min(segments, 64)` via the `else` branch, which was intended
+for other head sizes. The formula produces 128 segments for typical D256 decode
+shapes; capping to 64 at short kv_len was a regression.
+
+---
+
+## Tuning workflow
+
+```bash
+# Sweep num_segments for D256 bf16 decode shapes on gfx942
+python -m benchmarks.gfx942.attention.decode.benchmark_decode_live \
+    --shapes library/benchmarks/gfx942/attention/decode/decode_shapes.json \
+    --segments-sweep 8 16 32 64 128 \
+    --output-json /tmp/seg_sweep_gfx942.json
+```
+
+The benchmark prints a `segments_sweep:` line per D256 shape:
+```
+qwen3next_80b_b1_kv8192   ...  segments_sweep: seg8=187.8us(0.52x)  seg16=102.2us(0.95x)  seg32=59.7us(1.63x)  seg64=48.0us(2.03x)  seg128=40.8us(2.39x)
+```
+
+---
+
+## Sweep results (gfx942, 2026-07)
+
+| kv_len | seg8 | seg16 | seg32 | seg64 | seg128 | winner |
+|---|---|---|---|---|---|---|
+| 512 | 0.26x | 0.26x | 0.26x | 0.26x | 0.26x | — (2D path) |
+| 1024 | 3.23x | 3.21x | 3.18x | 3.23x | **3.24x** | seg128 |
+| 2048 | 3.21x | 3.19x | **3.23x** | 3.19x | 3.19x | seg32 ≈ seg128 |
+| 4096 | 0.87x | 1.47x | 2.37x | 2.40x | **2.54x** | seg128 |
+| 8192 | 0.52x | 0.95x | 1.63x | 2.03x | **2.39x** | seg128 |
+| 16384 | 0.25x | 0.48x | 0.88x | 1.30x | **1.76x** | seg128 |
+| 32768 | 0.13x | 0.25x | 0.47x | 0.79x | **1.22x** | seg128 |
+
+kv=512 routes to 2D (`max_seqlen_k <= 512` triggers `use_2d_kernel`); segment
+count has no effect there.
+
+**Conclusion:** seg128 wins or ties at every kv_len. Do not cap below the
+formula value for D256.
+
+---
+
+## Fix applied
+
+```python
+# kernels/common/attention_unified.py — _num_segments()
+if arch == "gfx942" and max_seqlen_q == 1 and max_seqlen_k <= 2048:
+    if head_size == 64:  return min(segments, 32)
+    if head_size == 128: return min(segments, 16)
+    # D256: seg128 wins sweep — do not cap below formula value.
+    if head_size != 256:
+        return min(segments, 64)
+```
+
+---
+
+## Before / after speedup vs Triton (gfx942, batch=1)
+
+| kv_len | before | after | delta |
+|---|---|---|---|
+| 512 | 0.25x | 0.56x | +0.32x |
+| 1024 | 3.09x | 3.16x | +0.07x |
+| 2048 | 3.04x | 2.89x | −0.15x† |
+| 4096 | 2.53x | 2.74x | +0.21x |
+| 8192 | 2.26x | 2.62x | +0.37x |
+| 16384 | 1.90x | 2.19x | +0.28x |
+| 32768 | 1.40x | 1.66x | +0.26x |
+
+† kv=2048 delta is within run-to-run noise; the sweep showed seg128 ≈ seg32
+at this point, so no regression is expected structurally.
+
+Full post-tuning results saved in `decode_shapes_perf.csv`.
+
+---
+
+## gfx950
+
+No tuning needed. `_num_segments()` has no cap for gfx950 — it already returns
+the raw formula value (128) for all D256 decode shapes.
diff --git a/dnn-providers/hip-kernel-provider/rocke/library/dispatch/AGENTS.md b/dnn-providers/hip-kernel-provider/rocke/library/dispatch/AGENTS.md
new file mode 100644
index 000000000000..1ade20903ca2
--- /dev/null
+++ b/dnn-providers/hip-kernel-provider/rocke/library/dispatch/AGENTS.md
@@ -0,0 +1,67 @@
+# Attention dispatcher — agent guide
+
+This folder owns the **path-level dispatch** for unified attention: which kernel
+family (2D tiled prefill vs 3D split-KV decode) and which specialized candidate
+handles a given `AttentionRequest`.
+
+## What the dispatcher decides (and what it does NOT)
+
+**Decides:** kernel path (`"2d"` or `"3d"`), candidate name, algorithm tag,
+and spec identity `(path, head_size, block_size)`.
+
+**Does NOT decide:** CTA geometry (`num_warps`, `tile_size`), `num_segments`,
+`waves_per_eu`, or any other performance knob. Those live in
+`builders/common/attention_spec_builder.py` and
+`kernels/common/attention_unified.py`. The parity identity with C++ is
+`(path, head_size, block_size)` only — C++ reads `num_segments` as a parameter
+passed from Python, it does not recompute it.
+
+## Candidate registry — priority table
+
+| priority | candidate | scope |
+|---|---|---|
+| 5 | `attention_gfx942_dense_pipe` | gfx942 fp16 2D prefill flash |
+| 5 | `attention_d256_decode` | gfx942/gfx950 bf16 D256 3D decode |
+| 10 | `attention_unified_2d` | generic 2D prefill fallback |
+| 10 | `attention_unified_3d` | generic 3D decode fallback |
+
+Lower priority number = higher precedence. Generic candidates (10) remain the
+fallback for everything a specialized candidate (5) does not claim.
+
+## How to add a new specialized candidate
+
+Follow the `_make_d256_decode_candidate()` pattern in `attention.py`:
+
+1. **Add a cohort predicate** in `kernels/common/attention_unified.py` —
+   a pure function of `UnifiedAttentionProblem` that returns `True` for the
+   target shape family. This is the single source of truth for membership;
+   import it lazily inside the factory to keep `dispatch/` arch-neutral.
+
+2. **Add a factory function** `_make_<name>_candidate()` in `attention.py`.
+   The `support()` closure must check in order: request errors → arch/dtype
+   gates → `_selector_matches` → `supports_native_unified_attention` →
+   cohort predicate → path check (`select_path() == "2d"` or `"3d"`).
+
+3. **Register** with `ATTENTION_REGISTRY.register(_make_<name>_candidate())`.
+
+4. **Add CPU-only dispatch tests** in
+   `tests/dispatch/attention/test_<name>_wiring.py`. Cover: registration,
+   spec_id, algorithm, priority ordering, support gates (wrong arch/dtype/
+   cohort/path), and routing for each target arch. Use `_PinnedArch` context
+   manager to avoid GPU dependency.
+
+5. **No C++ changes needed.** The dispatcher is Python-only.
+
+## How to tune `num_segments` for a new cohort
+
+See the worked example:
+`benchmarks/gfx942/attention/decode/TUNING_D256.md`
+
+## Testing (CPU-only, no GPU)
+
+```bash
+PYTHONPATH=library:platform/python python -m unittest discover \
+    -s library/tests/dispatch/attention -p "test_*.py" -v
+```
+
+All dispatch tests complete in < 1 s.

From eaddb05d079e0bc61b7b5d2fbca84365e9548186 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Bart=C5=82omiej=20Kocot?= <barkocot@amd.com>
Date: Mon, 20 Jul 2026 00:27:10 +0200
Subject: [PATCH 09/11] Update
 dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/prefill/parity_unified_attention.py

Co-authored-by: Anagha Rao <127366241+AnaghaRaoAMD@users.noreply.github.com>
---
 .../gfx942/attention/prefill/parity_unified_attention.py        | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/prefill/parity_unified_attention.py b/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/prefill/parity_unified_attention.py
index 01481f0f9d95..43d76046d5c7 100644
--- a/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/prefill/parity_unified_attention.py
+++ b/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/prefill/parity_unified_attention.py
@@ -125,7 +125,7 @@ def default_scenarios() -> List[Shape]:
     """Core production scenarios adapted from the gfx950 default set.
 
     Sliding-window / softcap / ALiBi / QQ-bias / sinks scenarios are dropped
-    (unsupported on the gfx942 SDPA-fwd net). Decode shapes (seqlen_q == 1)
+    (the harness's torch reference is causal-only, not a kernel limitation).
     are included; they exercise the T=64 single-query path.
     Non-uniform seq_lens from the gfx950 mixed/prefill scenarios are split
     into one Shape per unique (q, k) pair.

From f632ed5558a3432bbb5379aec52f1d82b321855b Mon Sep 17 00:00:00 2001
From: Bartlomiej Kocot <barkocot@amd.com>
Date: Sun, 19 Jul 2026 19:53:15 -0400
Subject: [PATCH 10/11] fixes

---
 .../gfx942/attention/decode/TUNING_D256.md    | 108 ------------------
 .../builders/gfx942/attention/README.md       |  40 +++----
 .../builders/gfx942/attention/__init__.py     |   8 +-
 .../prefill/parity_unified_attention.py       |  10 +-
 .../rocke/library/dispatch/attention.py       |   2 -
 5 files changed, 27 insertions(+), 141 deletions(-)
 delete mode 100644 dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx942/attention/decode/TUNING_D256.md

diff --git a/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx942/attention/decode/TUNING_D256.md b/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx942/attention/decode/TUNING_D256.md
deleted file mode 100644
index 6506409f9024..000000000000
--- a/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx942/attention/decode/TUNING_D256.md
+++ /dev/null
@@ -1,108 +0,0 @@
-# D256 bf16 decode tuning — gfx942
-
-**Workload:** Qwen3-Next-80B-A3B gated attention decode, GQA 16/2,
-`head_size=256`, `block_size=16`, `dtype=bf16`, `batch=1`,
-`kv_len` sweep 512–32768.
-
-**Dispatcher candidate:** `attention_d256_decode` (priority 5, algorithm
-`d256_decode`) registered in `dispatch/attention.py`. Routes D256 bf16 decode
-to the 3D split-KV path on both gfx942 and gfx950.
-
-**Cohort predicate:** `_d256_decode_cohort()` in
-`kernels/common/attention_unified.py` — `head_size==256`, `dtype=="bf16"`,
-`all_decode==True`, no sliding_window / softcap / sinks / alibi / qq_bias / fp8.
-
----
-
-## Problem found
-
-`_num_segments()` in `kernels/common/attention_unified.py` had a gfx942
-short-KV cap block:
-
-```python
-if arch == "gfx942" and max_seqlen_q == 1 and max_seqlen_k <= 2048:
-    if head_size == 64:  return min(segments, 32)
-    if head_size == 128: return min(segments, 16)
-    return min(segments, 64)   # ← D256 fell here — wrong
-```
-
-D256 inherited `min(segments, 64)` via the `else` branch, which was intended
-for other head sizes. The formula produces 128 segments for typical D256 decode
-shapes; capping to 64 at short kv_len was a regression.
-
----
-
-## Tuning workflow
-
-```bash
-# Sweep num_segments for D256 bf16 decode shapes on gfx942
-python -m benchmarks.gfx942.attention.decode.benchmark_decode_live \
-    --shapes library/benchmarks/gfx942/attention/decode/decode_shapes.json \
-    --segments-sweep 8 16 32 64 128 \
-    --output-json /tmp/seg_sweep_gfx942.json
-```
-
-The benchmark prints a `segments_sweep:` line per D256 shape:
-```
-qwen3next_80b_b1_kv8192   ...  segments_sweep: seg8=187.8us(0.52x)  seg16=102.2us(0.95x)  seg32=59.7us(1.63x)  seg64=48.0us(2.03x)  seg128=40.8us(2.39x)
-```
-
----
-
-## Sweep results (gfx942, 2026-07)
-
-| kv_len | seg8 | seg16 | seg32 | seg64 | seg128 | winner |
-|---|---|---|---|---|---|---|
-| 512 | 0.26x | 0.26x | 0.26x | 0.26x | 0.26x | — (2D path) |
-| 1024 | 3.23x | 3.21x | 3.18x | 3.23x | **3.24x** | seg128 |
-| 2048 | 3.21x | 3.19x | **3.23x** | 3.19x | 3.19x | seg32 ≈ seg128 |
-| 4096 | 0.87x | 1.47x | 2.37x | 2.40x | **2.54x** | seg128 |
-| 8192 | 0.52x | 0.95x | 1.63x | 2.03x | **2.39x** | seg128 |
-| 16384 | 0.25x | 0.48x | 0.88x | 1.30x | **1.76x** | seg128 |
-| 32768 | 0.13x | 0.25x | 0.47x | 0.79x | **1.22x** | seg128 |
-
-kv=512 routes to 2D (`max_seqlen_k <= 512` triggers `use_2d_kernel`); segment
-count has no effect there.
-
-**Conclusion:** seg128 wins or ties at every kv_len. Do not cap below the
-formula value for D256.
-
----
-
-## Fix applied
-
-```python
-# kernels/common/attention_unified.py — _num_segments()
-if arch == "gfx942" and max_seqlen_q == 1 and max_seqlen_k <= 2048:
-    if head_size == 64:  return min(segments, 32)
-    if head_size == 128: return min(segments, 16)
-    # D256: seg128 wins sweep — do not cap below formula value.
-    if head_size != 256:
-        return min(segments, 64)
-```
-
----
-
-## Before / after speedup vs Triton (gfx942, batch=1)
-
-| kv_len | before | after | delta |
-|---|---|---|---|
-| 512 | 0.25x | 0.56x | +0.32x |
-| 1024 | 3.09x | 3.16x | +0.07x |
-| 2048 | 3.04x | 2.89x | −0.15x† |
-| 4096 | 2.53x | 2.74x | +0.21x |
-| 8192 | 2.26x | 2.62x | +0.37x |
-| 16384 | 1.90x | 2.19x | +0.28x |
-| 32768 | 1.40x | 1.66x | +0.26x |
-
-† kv=2048 delta is within run-to-run noise; the sweep showed seg128 ≈ seg32
-at this point, so no regression is expected structurally.
-
-Full post-tuning results saved in `decode_shapes_perf.csv`.
-
----
-
-## gfx950
-
-No tuning needed. `_num_segments()` has no cap for gfx950 — it already returns
-the raw formula value (128) for all D256 decode shapes.
diff --git a/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/README.md b/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/README.md
index 18e9bd0879b1..4bef9f5ef095 100644
--- a/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/README.md
+++ b/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/README.md
@@ -15,9 +15,9 @@ torch reference — so it runs on any box with torch + a gfx942 GPU.
 
 ## What it shows
 
-- A correct gfx942 tiled SDPA-forward over a canonical shape set
-  (`shapes.json`): fp16/bf16, `head_size` 64 / 128, MHA + GQA, and the
-  decode / short-prefill / long-prefill regimes — each gated against an fp32
+- A correct gfx942 tiled SDPA-forward over a canonical inline shape set
+  (`default` / `fmha` / `creative` groups): fp16/bf16, `head_size` 64 / 128,
+  MHA + GQA, and decode / prefill regimes — each gated against an fp32
   paged reference.
 - The two shipped D128-fp16 flash geometries side by side: the provider's
   analytic-default **wide4** (workgroup 256, `num_warps=4`) and the **L4**
@@ -38,8 +38,7 @@ torch reference — so it runs on any box with torch + a gfx942 GPU.
 
 | file | role |
 |------|------|
-| `shapes.json` | the canonical shape set, grouped by regime: `decode`, `short_prefill`, `long_prefill`, and `d256_disabled` (off by default) |
-| `parity_unified_attention.py` | per-shape parity + latency harness; builds the gfx942 spec **explicitly** (wide4 vs L4) and exposes the `HIPDNN_GFX942_*` lever knobs via `--scenario` |
+| `parity_unified_attention.py` | per-shape parity + latency harness; defines inline scenario groups (`default` / `fmha` / `creative`), builds the gfx942 spec **explicitly** (wide4 vs L4), and exposes the `HIPDNN_GFX942_*` lever knobs via `--scenario` |
 | `final_shapes_check.py` | the definitive correctness + perf check over every shape via the **production dispatcher** (`run_unified_attention_torch`), timed eager + graph against PyTorch's flash SDPA |
 | `__init__.py` | package marker + module docstring |
 
@@ -54,7 +53,7 @@ directory (so `python/` is the package root):
 PYTHONPATH=python python python/rocke/examples/gfx942/attention/final_shapes_check.py
 
 # 2) The explicit parity + latency harness (builds the spec by hand; default
-#    runs the whole shapes.json set):
+#    runs all inline scenario groups):
 PYTHONPATH=python python python/rocke/examples/gfx942/attention/parity_unified_attention.py
 ```
 
@@ -62,7 +61,7 @@ PYTHONPATH=python python python/rocke/examples/gfx942/attention/parity_unified_a
 
 | flag | default | meaning |
 |------|---------|---------|
-| `--groups G [G ...]` | all non-`*_disabled` groups | which `shapes.json` groups to run (e.g. `--groups decode`, or `--groups d256_disabled` to include the off-by-default d256 set) |
+| `--groups G [G ...]` | all non-`*_disabled` groups | which shape groups to run (e.g. `--groups decode`, or `--groups d256_disabled` to include the off-by-default d256 set) |
 | `--warmup N` | 10 | warm-up launches before timing |
 | `--iters N` | 50 | inner timed launches per measurement |
 | `--reps N` | 10 | timed measurements per cell, reduced per `--reduce` |
@@ -78,20 +77,17 @@ correctness failure.
 
 | flag | default | meaning |
 |------|---------|---------|
-| `--scenario S` (repeatable) | all shapes | a group (`correctness` / `perf` / `decode` / `all`) **or** an exact shape name; multiple flags union together |
+| `--scenario S` (repeatable) | all shapes | a group (`default` / `fmha` / `creative` / `all`) **or** an exact shape name; multiple flags union together |
 | `--attempts N` | 30 | timed launches |
 | `--warmup N` | 10 | warm-up launches |
 | `--tol F` | `2e-2` fp16 / `4e-2` bf16 | absolute-tolerance override |
 | `--report PATH` | none | write a JSON results report |
 | `--debug-mismatch N` | 0 | on failure, print the `N` worst mismatch samples |
 
-> **Note on `--scenario`:** the selector special-cases the group names
-> `correctness`, `perf`, `decode`, and `all`. The current `shapes.json` groups
-> are `decode` / `short_prefill` / `long_prefill` / `d256_disabled`, so
-> `--scenario decode` selects the decode group and `--scenario all` (or no flag)
-> runs everything; any other value is matched as an **exact shape name**, e.g.
-> `--scenario fp16_h32kv8_b1_s2048x2048_d128`. To target the other regimes by
-> group, prefer `final_shapes_check.py --groups short_prefill long_prefill`.
+> **Note on `--scenario`:** the selector accepts the group names
+> `default`, `fmha`, `creative`, and `all`. `--scenario all` (or no flag)
+> runs all three groups; any other value is matched as an **exact shape name**,
+> e.g. `--scenario decode_d128_q1k1024`.
 
 ```bash
 # Force the L4 (WG=64) flash geometry instead of the default wide4, on one shape:
@@ -100,17 +96,17 @@ HIPDNN_GFX942_FLASH_WIDE=0 PYTHONPATH=python python \
     --scenario fp16_h32kv8_b1_s2048x2048_d128
 ```
 
-## The shape set (`shapes.json`)
+## The shape set (inline scenarios)
 
-Grouped by regime; every shape is causal (the kernel always applies a causal
-mask), so non-causal *prefill* shapes are not included.
+Shapes are defined inline in `parity_unified_attention.py` across three groups;
+every shape is causal (the kernel always applies a causal mask), so non-causal
+prefill shapes are not included.
 
 | group | what it covers |
 |-------|----------------|
-| `decode` | `seqlen_q == 1`, long KV (a single query attends to all keys) — fp16, D64/D128, MHA + GQA, batch 1–64 |
-| `short_prefill` | 1–2 KV tiles (`seqlen_q` 64 / 512 / 528) — fp16/bf16, D64/D128, MHA + GQA |
-| `long_prefill` | square prefill `seqlen_q == seqlen_k` up to 8192 — fp16/bf16, D64/D128, MHA + GQA, batch 1–51 |
-| `d256_disabled` | `head_size = 256` — **off by default** (no tiled d256 path on gfx942 → scalar fallback; see below) |
+| `default` | core production shapes — decode (`seqlen_q == 1`) and prefill, fp16/bf16, D64/D128, MHA + GQA, varied batch sizes |
+| `fmha` | integration-test net shapes — GQA ratios, asymmetric seq-lens, tiny seqs, cross-attention, large-batch decode, and select D256 shapes |
+| `creative` | exploratory sweep — long-KV decode (up to 64 k), chunked prefill, non-standard GQA ratios, D64/D256 bf16, large-batch variants |
 
 ## Arch notes
 
diff --git a/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/__init__.py b/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/__init__.py
index 94143eaac1bd..086d7f6392c9 100644
--- a/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/__init__.py
+++ b/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/__init__.py
@@ -8,8 +8,8 @@
 harness it has **no** Triton/AITER dependency: the oracle is a fp32 torch
 reference, so the example runs on any box with torch + a gfx942 GPU.
 
-Ships the canonical shapes we validate (``shapes.json``) and the definitive
-correctness + perf check (``final_shapes_check.py``); see ``README.md``. The
-exhaustive sweep / probe / regression harnesses that found & gated the levers
-are archived under ``~/attention-archive/``.
+Ships inline scenario groups (``default`` / ``fmha`` / ``creative``) and the
+definitive correctness + perf check (``final_shapes_check.py``); see
+``README.md``. The exhaustive sweep / probe / regression harnesses that found
+& gated the levers are archived under ``~/attention-archive/``.
 """
diff --git a/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/prefill/parity_unified_attention.py b/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/prefill/parity_unified_attention.py
index 6a49c4adbc68..b77662a14773 100644
--- a/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/prefill/parity_unified_attention.py
+++ b/dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/prefill/parity_unified_attention.py
@@ -12,9 +12,9 @@
 
 The harness:
 
-  1. Loads the canonical shapes we validate from the shipped ``shapes.json``
-     (``--scenario`` selects subsets); these mirror the rocke-provider
-     integration-test net plus the case-study perf shapes.
+  1. Defines shapes inline across three scenario groups (``default`` /
+     ``fmha`` / ``creative``); ``--scenario`` selects subsets by group name
+     or exact shape name, defaulting to all groups.
   2. For each shape, builds the gfx942 tiled-2D kernel via
      ``UnifiedAttention2DTiledSpec`` / ``build_unified_attention_2d_tiled``
      (auto-routes to ``instances/gfx942/attention_tiled_2d.py``), gated by
@@ -37,7 +37,7 @@
 
     PYTHONPATH=python .venv/bin/python \\
         python/rocke/library/builders/gfx942/attention/parity_unified_attention.py \\
-        --scenario correctness
+        --scenario default
 
     # force the L4 (WG=64) fallback instead of the default wide4:
     HIPDNN_GFX942_FLASH_WIDE=0 PYTHONPATH=python .venv/bin/python \\
@@ -77,7 +77,7 @@ class Shape:
     kv_heads: int  # Hkv
     batch: int  # B
     causal: bool
-    group: str  # "correctness" | "perf" | "decode"
+    group: str  # "default" | "fmha" | "creative"
 
     @property
     def num_queries_per_kv(self) -> int:
diff --git a/dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention.py b/dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention.py
index 0c47f213d0e9..de7717159937 100644
--- a/dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention.py
+++ b/dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention.py
@@ -150,8 +150,6 @@ class AttentionSpec:
     num_query_heads: int
     num_kv_heads: int
     name: str = "rocke_attention_unified"
-    # Pinned codegen knobs from a specialized candidate; None for generic paths.
-    # Stored as a sorted tuple of (key, value) pairs for hashability.
 
     def kernel_name(self) -> str:
         from rocke.helpers.spec import kernel_name_join

From ad5d2bdcddd075e77bca20bd49aa11767950060e Mon Sep 17 00:00:00 2001
From: Bartlomiej Kocot <barkocot@amd.com>
Date: Mon, 20 Jul 2026 09:28:50 +0000
Subject: [PATCH 11/11] gfx950 perf

---
 .../attention/prefill/prefill_bench_perf.json | 956 +++++++++---------
 1 file changed, 478 insertions(+), 478 deletions(-)

diff --git a/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx950/attention/prefill/prefill_bench_perf.json b/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx950/attention/prefill/prefill_bench_perf.json
index 0d4e8bdab152..52ecffe245d6 100644
--- a/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx950/attention/prefill/prefill_bench_perf.json
+++ b/dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx950/attention/prefill/prefill_bench_perf.json
@@ -6,34 +6,34 @@
     "num_seqs": 1,
     "total_q": 512,
     "max_seqlen_k": 512,
-    "triton_ms": 0.04494441986083984,
-    "aoTriton_ms": 0.038186779022216795,
-    "flydsl_ms": 0.21248931884765626,
+    "triton_ms": 0.04530444145202637,
+    "aoTriton_ms": 0.03762197971343994,
+    "flydsl_ms": 0.21236127853393555,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.054742140769958494,
-        "speedup": 0.8210205013667374,
+        "ms": 0.05295730113983154,
+        "speedup": 0.8554899981100224,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.029582679271697998,
-        "speedup": 1.5192815852835397,
+        "ms": 0.02966507911682129,
+        "speedup": 1.527197728804874,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.21151378386723826,
+    "speedup_flydsl_vs_triton": 0.21333663916883355,
     "best_variant": "fallback",
-    "best_speedup_vs_triton": 1.5192815852835397,
-    "best_speedup_vs_aoTriton": 1.2908492388906239,
-    "best_speedup_vs_flydsl": 7.182896346070541
+    "best_speedup_vs_triton": 1.527197728804874,
+    "best_speedup_vs_aoTriton": 1.2682244859447136,
+    "best_speedup_vs_flydsl": 7.158628422922971
   },
   {
     "signature": "d128_b64_h32kv8_q1024_k1024_ns1_tq1024_sw0_sc0_sinks0_bfloat16",
@@ -42,34 +42,34 @@
     "num_seqs": 1,
     "total_q": 1024,
     "max_seqlen_k": 1024,
-    "triton_ms": 0.04407966136932373,
-    "aoTriton_ms": 0.06277822017669678,
-    "flydsl_ms": 0.22260065078735353,
+    "triton_ms": 0.04453643798828125,
+    "aoTriton_ms": 0.0625582218170166,
+    "flydsl_ms": 0.2167757797241211,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.058950982093811034,
-        "speedup": 0.7477341310307268,
+        "ms": 0.05854057788848877,
+        "speedup": 0.7607789262537968,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.08023520469665528,
-        "speedup": 0.5493805560286837,
+        "ms": 0.08046321868896485,
+        "speedup": 0.5535005772070762,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.19802126010598348,
+    "speedup_flydsl_vs_triton": 0.2054493266958162,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.7477341310307268,
-    "best_speedup_vs_aoTriton": 1.0649223803734993,
-    "best_speedup_vs_flydsl": 3.7760295567785502
+    "best_speedup_vs_triton": 0.7607789262537968,
+    "best_speedup_vs_aoTriton": 1.068630069490958,
+    "best_speedup_vs_flydsl": 3.7030003382789833
   },
   {
     "signature": "d128_b64_h32kv8_q2048_k2048_ns1_tq2048_sw0_sc0_sinks0_bfloat16",
@@ -78,34 +78,34 @@
     "num_seqs": 1,
     "total_q": 2048,
     "max_seqlen_k": 2048,
-    "triton_ms": 0.08524806022644044,
-    "aoTriton_ms": 0.14309745788574219,
-    "flydsl_ms": 0.24806970596313477,
+    "triton_ms": 0.08698328018188477,
+    "aoTriton_ms": 0.14490785598754882,
+    "flydsl_ms": 0.24641443252563477,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.16571046829223632,
-        "speedup": 0.5144398003637431,
+        "ms": 0.16628883361816407,
+        "speedup": 0.5230855150600047,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.25876741409301757,
-        "speedup": 0.32943893080678555,
+        "ms": 0.2592129707336426,
+        "speedup": 0.3355668504384585,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.3436455890309678,
+    "speedup_flydsl_vs_triton": 0.35299588295355144,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.5144398003637431,
-    "best_speedup_vs_aoTriton": 0.8635390350438495,
-    "best_speedup_vs_flydsl": 1.4970068488712192
+    "best_speedup_vs_triton": 0.5230855150600047,
+    "best_speedup_vs_aoTriton": 0.871422649582649,
+    "best_speedup_vs_flydsl": 1.4818459373613553
   },
   {
     "signature": "d128_b64_h32kv8_q4096_k4096_ns1_tq4096_sw0_sc0_sinks0_bfloat16",
@@ -114,34 +114,34 @@
     "num_seqs": 1,
     "total_q": 4096,
     "max_seqlen_k": 4096,
-    "triton_ms": 0.263809871673584,
-    "aoTriton_ms": 0.438018798828125,
-    "flydsl_ms": 0.3631324768066406,
+    "triton_ms": 0.26673065185546874,
+    "aoTriton_ms": 0.43830917358398436,
+    "flydsl_ms": 0.3602244186401367,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.542783088684082,
-        "speedup": 0.48603185540132077,
+        "ms": 0.5445582580566406,
+        "speedup": 0.4898110494317865,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.9393407440185547,
-        "speedup": 0.28084576694181274,
+        "ms": 0.9365765380859375,
+        "speedup": 0.28479322405468405,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.726483827592365,
+    "speedup_flydsl_vs_triton": 0.7404568875768858,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.48603185540132077,
-    "best_speedup_vs_aoTriton": 0.8069868202601034,
-    "best_speedup_vs_flydsl": 0.669019511435066
+    "best_speedup_vs_triton": 0.4898110494317865,
+    "best_speedup_vs_aoTriton": 0.8048894073302162,
+    "best_speedup_vs_flydsl": 0.661498404093009
   },
   {
     "signature": "d128_b64_h32kv8_q8192_k8192_ns1_tq8192_sw0_sc0_sinks0_bfloat16",
@@ -150,34 +150,34 @@
     "num_seqs": 1,
     "total_q": 8192,
     "max_seqlen_k": 8192,
-    "triton_ms": 0.8871473693847656,
-    "aoTriton_ms": 1.3932284545898437,
-    "flydsl_ms": 0.7676717376708985,
+    "triton_ms": 0.886360092163086,
+    "aoTriton_ms": 1.3923817443847657,
+    "flydsl_ms": 0.7604667663574218,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 1.9165232849121094,
-        "speedup": 0.46289412519475315,
+        "ms": 1.9168678283691407,
+        "speedup": 0.4624002130168754,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 3.483743896484375,
-        "speedup": 0.2546534405930446,
+        "ms": 3.4786669921875,
+        "speedup": 0.2547987761270916,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 1.155633750535553,
+    "speedup_flydsl_vs_triton": 1.1655474392506113,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.46289412519475315,
-    "best_speedup_vs_aoTriton": 0.7269561844398548,
-    "best_speedup_vs_flydsl": 0.400554349490256
+    "best_speedup_vs_triton": 0.4624002130168754,
+    "best_speedup_vs_aoTriton": 0.726383803712432,
+    "best_speedup_vs_flydsl": 0.396723631698917
   },
   {
     "signature": "d128_b64_h40kv8_q512_k512_ns1_tq512_sw0_sc0_sinks0_bfloat16",
@@ -186,34 +186,34 @@
     "num_seqs": 1,
     "total_q": 512,
     "max_seqlen_k": 512,
-    "triton_ms": 0.044002838134765625,
-    "aoTriton_ms": 0.04145962238311768,
-    "flydsl_ms": 0.20985809326171875,
+    "triton_ms": 0.044145240783691406,
+    "aoTriton_ms": 0.04022439956665039,
+    "flydsl_ms": 0.21304689407348631,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.0542109203338623,
-        "speedup": 0.8116969397267306,
+        "ms": 0.05329651832580566,
+        "speedup": 0.8282950213338177,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.045933260917663574,
-        "speedup": 0.9579733129255013,
+        "ms": 0.04588764190673828,
+        "speedup": 0.9620289679171548,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.20967901428461325,
+    "speedup_flydsl_vs_triton": 0.2072090324323826,
     "best_variant": "fallback",
-    "best_speedup_vs_triton": 0.9579733129255013,
-    "best_speedup_vs_aoTriton": 0.9026056838732831,
-    "best_speedup_vs_flydsl": 4.568761047422568
+    "best_speedup_vs_triton": 0.9620289679171548,
+    "best_speedup_vs_aoTriton": 0.8765845856364154,
+    "best_speedup_vs_flydsl": 4.6427945568979405
   },
   {
     "signature": "d128_b64_h40kv8_q1024_k1024_ns1_tq1024_sw0_sc0_sinks0_bfloat16",
@@ -222,34 +222,34 @@
     "num_seqs": 1,
     "total_q": 1024,
     "max_seqlen_k": 1024,
-    "triton_ms": 0.04420683860778809,
-    "aoTriton_ms": 0.08162883758544921,
-    "flydsl_ms": 0.22243425369262695,
+    "triton_ms": 0.043613219261169435,
+    "aoTriton_ms": 0.08163359642028808,
+    "flydsl_ms": 0.22130861282348632,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.08035200119018554,
-        "speedup": 0.5501647495145107,
+        "ms": 0.07992318153381348,
+        "speedup": 0.5456892283838549,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.12070201873779297,
-        "speedup": 0.36624771540748474,
+        "ms": 0.12036441802978516,
+        "speedup": 0.3623431241147777,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.19874114653616146,
+    "speedup_flydsl_vs_triton": 0.19706968791112947,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.5501647495145107,
-    "best_speedup_vs_aoTriton": 1.0158905363444717,
-    "best_speedup_vs_flydsl": 2.768247839480018
+    "best_speedup_vs_triton": 0.5456892283838549,
+    "best_speedup_vs_aoTriton": 1.0214007357271053,
+    "best_speedup_vs_flydsl": 2.7690165553514188
   },
   {
     "signature": "d128_b64_h40kv8_q2048_k2048_ns1_tq2048_sw0_sc0_sinks0_bfloat16",
@@ -258,34 +258,34 @@
     "num_seqs": 1,
     "total_q": 2048,
     "max_seqlen_k": 2048,
-    "triton_ms": 0.10507227897644043,
-    "aoTriton_ms": 0.18317228317260742,
-    "flydsl_ms": 0.2665970993041992,
+    "triton_ms": 0.10779308319091797,
+    "aoTriton_ms": 0.18331941604614257,
+    "flydsl_ms": 0.2638418388366699,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.22542627334594725,
-        "speedup": 0.46610484845833716,
+        "ms": 0.22508943557739258,
+        "speedup": 0.4788900150485092,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.3947103881835938,
-        "speedup": 0.26620094662309113,
+        "ms": 0.3917671203613281,
+        "speedup": 0.27514581389959436,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.39412386425310747,
+    "speedup_flydsl_vs_triton": 0.4085518948253192,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.46610484845833716,
-    "best_speedup_vs_aoTriton": 0.8125596029860488,
-    "best_speedup_vs_flydsl": 1.1826354370640273
+    "best_speedup_vs_triton": 0.4788900150485092,
+    "best_speedup_vs_aoTriton": 0.8144292315447732,
+    "best_speedup_vs_flydsl": 1.1721644694690838
   },
   {
     "signature": "d128_b64_h40kv8_q4096_k4096_ns1_tq4096_sw0_sc0_sinks0_bfloat16",
@@ -294,34 +294,34 @@
     "num_seqs": 1,
     "total_q": 4096,
     "max_seqlen_k": 4096,
-    "triton_ms": 0.33019214630126953,
-    "aoTriton_ms": 0.5222876739501953,
-    "flydsl_ms": 0.4055417251586914,
+    "triton_ms": 0.32686485290527345,
+    "aoTriton_ms": 0.5218148040771484,
+    "flydsl_ms": 0.4034936141967773,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.7208232879638672,
-        "speedup": 0.45807641319965386,
+        "ms": 0.7182536315917969,
+        "speedup": 0.4550827709438435,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 1.3519033813476562,
-        "speedup": 0.2442424146998691,
+        "ms": 1.3523646545410157,
+        "speedup": 0.2416987547017852,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.8142001816756659,
+    "speedup_flydsl_vs_triton": 0.8100868053536697,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.45807641319965386,
-    "best_speedup_vs_aoTriton": 0.7245710324169994,
-    "best_speedup_vs_flydsl": 0.5626090776065771
+    "best_speedup_vs_triton": 0.4550827709438435,
+    "best_speedup_vs_aoTriton": 0.7265049296314732,
+    "best_speedup_vs_flydsl": 0.5617703781080131
   },
   {
     "signature": "d128_b64_h40kv8_q8192_k8192_ns1_tq8192_sw0_sc0_sinks0_bfloat16",
@@ -330,34 +330,34 @@
     "num_seqs": 1,
     "total_q": 8192,
     "max_seqlen_k": 8192,
-    "triton_ms": 1.1060104370117188,
-    "aoTriton_ms": 1.7166990661621093,
-    "flydsl_ms": 0.9068356323242187,
+    "triton_ms": 1.102583770751953,
+    "aoTriton_ms": 1.7152835083007814,
+    "flydsl_ms": 0.8979777526855469,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 2.540103302001953,
-        "speedup": 0.4354194713813527,
+        "ms": 2.5480255126953124,
+        "speedup": 0.43272085199242577,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 4.979765625,
-        "speedup": 0.2221009019900849,
+        "ms": 4.9808740234375,
+        "speedup": 0.22136351282199584,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 1.2196371620036757,
+    "speedup_flydsl_vs_triton": 1.2278519901573273,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.4354194713813527,
-    "best_speedup_vs_aoTriton": 0.6758382876826753,
-    "best_speedup_vs_flydsl": 0.3570073829711991
+    "best_speedup_vs_triton": 0.43272085199242577,
+    "best_speedup_vs_aoTriton": 0.6731814496183545,
+    "best_speedup_vs_flydsl": 0.35242102098721223
   },
   {
     "signature": "d128_b64_h28kv4_q512_k512_ns1_tq512_sw0_sc0_sinks0_bfloat16",
@@ -366,34 +366,34 @@
     "num_seqs": 1,
     "total_q": 512,
     "max_seqlen_k": 512,
-    "triton_ms": 0.043571639060974124,
-    "aoTriton_ms": 0.040881199836730955,
-    "flydsl_ms": 0.23406875610351563,
+    "triton_ms": 0.043930039405822754,
+    "aoTriton_ms": 0.04031879901885986,
+    "flydsl_ms": 0.21134849548339843,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.05508453845977783,
-        "speedup": 0.7909958089744129,
+        "ms": 0.05293409824371338,
+        "speedup": 0.8299005907981066,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h28kv4_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.030811500549316407,
-        "speedup": 1.4141355754885758,
+        "ms": 0.030913100242614747,
+        "speedup": 1.4210816469732053,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h28kv4_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.1861488896950638,
+    "speedup_flydsl_vs_triton": 0.2078559362598987,
     "best_variant": "fallback",
-    "best_speedup_vs_triton": 1.4141355754885758,
-    "best_speedup_vs_aoTriton": 1.3268162571731015,
-    "best_speedup_vs_flydsl": 7.596798336026148
+    "best_speedup_vs_triton": 1.4210816469732053,
+    "best_speedup_vs_aoTriton": 1.3042625522004112,
+    "best_speedup_vs_flydsl": 6.8368586076671605
   },
   {
     "signature": "d128_b64_h28kv4_q1024_k1024_ns1_tq1024_sw0_sc0_sinks0_bfloat16",
@@ -402,34 +402,34 @@
     "num_seqs": 1,
     "total_q": 1024,
     "max_seqlen_k": 1024,
-    "triton_ms": 0.04372446060180664,
-    "aoTriton_ms": 0.06314464092254639,
-    "flydsl_ms": 0.2212246322631836,
+    "triton_ms": 0.04390123844146729,
+    "aoTriton_ms": 0.06060860157012939,
+    "flydsl_ms": 0.21886457443237306,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.058229379653930664,
-        "speedup": 0.7509003335029536,
+        "ms": 0.05814698219299316,
+        "speedup": 0.7550045898470975,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h28kv4_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.08373605728149414,
-        "speedup": 0.5221700426474444,
+        "ms": 0.08384324073791503,
+        "speedup": 0.5236109441272415,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h28kv4_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.19764734222629016,
+    "speedup_flydsl_vs_triton": 0.200586314872224,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.7509003335029536,
-    "best_speedup_vs_aoTriton": 1.084412049343959,
-    "best_speedup_vs_flydsl": 3.7991926683397224
+    "best_speedup_vs_triton": 0.7550045898470975,
+    "best_speedup_vs_aoTriton": 1.0423344305120767,
+    "best_speedup_vs_flydsl": 3.76398853694503
   },
   {
     "signature": "d128_b64_h28kv4_q2048_k2048_ns1_tq2048_sw0_sc0_sinks0_bfloat16",
@@ -438,34 +438,34 @@
     "num_seqs": 1,
     "total_q": 2048,
     "max_seqlen_k": 2048,
-    "triton_ms": 0.0806496238708496,
-    "aoTriton_ms": 0.1389558219909668,
-    "flydsl_ms": 0.24630250930786132,
+    "triton_ms": 0.08260961532592774,
+    "aoTriton_ms": 0.13859098434448242,
+    "flydsl_ms": 0.24371044158935548,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.16475366592407226,
-        "speedup": 0.48951641481542196,
+        "ms": 0.16432323455810546,
+        "speedup": 0.502726321984105,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h28kv4_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.2672114944458008,
-        "speedup": 0.3018194409567511,
+        "ms": 0.2667170524597168,
+        "speedup": 0.30972753546908877,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h28kv4_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.32744134072155584,
+    "speedup_flydsl_vs_triton": 0.33896625350637366,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.48951641481542196,
-    "best_speedup_vs_aoTriton": 0.843415660656713,
-    "best_speedup_vs_flydsl": 1.4949743784236726
+    "best_speedup_vs_triton": 0.502726321984105,
+    "best_speedup_vs_aoTriton": 0.8434046756514886,
+    "best_speedup_vs_flydsl": 1.4831161414558105
   },
   {
     "signature": "d128_b64_h28kv4_q4096_k4096_ns1_tq4096_sw0_sc0_sinks0_bfloat16",
@@ -474,34 +474,34 @@
     "num_seqs": 1,
     "total_q": 4096,
     "max_seqlen_k": 4096,
-    "triton_ms": 0.24569210052490234,
-    "aoTriton_ms": 0.39886962890625,
-    "flydsl_ms": 0.34984195709228516,
+    "triton_ms": 0.24535205841064453,
+    "aoTriton_ms": 0.39729434967041016,
+    "flydsl_ms": 0.3501715087890625,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.5172180557250976,
-        "speedup": 0.4750261476863216,
+        "ms": 0.5166155624389649,
+        "speedup": 0.4749219269592395,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h28kv4_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.9295486450195313,
-        "speedup": 0.26431333297219745,
+        "ms": 0.9285749053955078,
+        "speedup": 0.2642243043453148,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h28kv4_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.7022945519942051,
+    "speedup_flydsl_vs_triton": 0.7006625389344298,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.4750261476863216,
-    "best_speedup_vs_aoTriton": 0.7711827235943403,
-    "best_speedup_vs_flydsl": 0.676391617074998
+    "best_speedup_vs_triton": 0.4749219269592395,
+    "best_speedup_vs_aoTriton": 0.7690328719382087,
+    "best_speedup_vs_flydsl": 0.6778183513014733
   },
   {
     "signature": "d128_b64_h28kv4_q8192_k8192_ns1_tq8192_sw0_sc0_sinks0_bfloat16",
@@ -510,34 +510,34 @@
     "num_seqs": 1,
     "total_q": 8192,
     "max_seqlen_k": 8192,
-    "triton_ms": 0.7988545227050782,
-    "aoTriton_ms": 1.2386493682861328,
-    "flydsl_ms": 0.7192121124267579,
+    "triton_ms": 0.7977047729492187,
+    "aoTriton_ms": 1.2378363800048828,
+    "flydsl_ms": 0.7142520141601563,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 1.7971022033691406,
-        "speedup": 0.4445237011047092,
+        "ms": 1.7844523620605468,
+        "speedup": 0.4470305791901843,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h28kv4_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 3.3761279296875,
-        "speedup": 0.23661855810630417,
+        "ms": 3.368626708984375,
+        "speedup": 0.23680414657453183,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h28kv4_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 1.1107356354297089,
+    "speedup_flydsl_vs_triton": 1.116839374806929,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.4445237011047092,
-    "best_speedup_vs_aoTriton": 0.6892481495843469,
-    "best_speedup_vs_flydsl": 0.4002065720460448
+    "best_speedup_vs_triton": 0.4470305791901843,
+    "best_speedup_vs_aoTriton": 0.6936785796710906,
+    "best_speedup_vs_flydsl": 0.4002639853805868
   },
   {
     "signature": "d128_b64_h64kv8_q512_k512_ns1_tq512_sw0_sc0_sinks0_bfloat16",
@@ -546,34 +546,34 @@
     "num_seqs": 1,
     "total_q": 512,
     "max_seqlen_k": 512,
-    "triton_ms": 0.04379563808441162,
-    "aoTriton_ms": 0.04252284049987793,
-    "flydsl_ms": 0.2295951271057129,
+    "triton_ms": 0.04382123947143555,
+    "aoTriton_ms": 0.04216201782226563,
+    "flydsl_ms": 0.20882768630981446,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.05434214115142822,
-        "speedup": 0.8059240426756829,
+        "ms": 0.05297011852264404,
+        "speedup": 0.8272822620304792,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.045822858810424805,
-        "speedup": 0.9557596191368142,
+        "ms": 0.04572524070739746,
+        "speedup": 0.9583599516042814,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.19075160103135255,
+    "speedup_flydsl_vs_triton": 0.20984401180609183,
     "best_variant": "fallback",
-    "best_speedup_vs_triton": 0.9557596191368142,
-    "best_speedup_vs_aoTriton": 0.9279831421212831,
-    "best_speedup_vs_flydsl": 5.01049330107443
+    "best_speedup_vs_triton": 0.9583599516042814,
+    "best_speedup_vs_aoTriton": 0.9220731738093316,
+    "best_speedup_vs_flydsl": 4.567011197297648
   },
   {
     "signature": "d128_b64_h64kv8_q1024_k1024_ns1_tq1024_sw0_sc0_sinks0_bfloat16",
@@ -582,34 +582,34 @@
     "num_seqs": 1,
     "total_q": 1024,
     "max_seqlen_k": 1024,
-    "triton_ms": 0.04899809837341309,
-    "aoTriton_ms": 0.11047634124755859,
-    "flydsl_ms": 0.22387027740478516,
+    "triton_ms": 0.04941810131072998,
+    "aoTriton_ms": 0.11047069549560547,
+    "flydsl_ms": 0.22591028213500977,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.09331294059753419,
-        "speedup": 0.5250943551842998,
+        "ms": 0.09379371643066406,
+        "speedup": 0.5268807249711845,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.13965262413024904,
-        "speedup": 0.3508569830217748,
+        "ms": 0.13936539649963378,
+        "speedup": 0.35459376970136086,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.21886826130482015,
+    "speedup_flydsl_vs_triton": 0.21875100523842672,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.5250943551842998,
-    "best_speedup_vs_aoTriton": 1.18393376674358,
-    "best_speedup_vs_flydsl": 2.3991343105385
+    "best_speedup_vs_triton": 0.5268807249711845,
+    "best_speedup_vs_aoTriton": 1.177804864756262,
+    "best_speedup_vs_flydsl": 2.4085865315083383
   },
   {
     "signature": "d128_b64_h64kv8_q2048_k2048_ns1_tq2048_sw0_sc0_sinks0_bfloat16",
@@ -618,34 +618,34 @@
     "num_seqs": 1,
     "total_q": 2048,
     "max_seqlen_k": 2048,
-    "triton_ms": 0.1500327205657959,
-    "aoTriton_ms": 0.24794811248779297,
-    "flydsl_ms": 0.28379167556762697,
+    "triton_ms": 0.15025431632995606,
+    "aoTriton_ms": 0.2465560531616211,
+    "flydsl_ms": 0.2817244148254395,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.2901189422607422,
-        "speedup": 0.5171421052230196,
+        "ms": 0.2902532958984375,
+        "speedup": 0.5176661848571447,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.48578174591064455,
-        "speedup": 0.30884799980399663,
+        "ms": 0.4846656799316406,
+        "speedup": 0.3100164145130899,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.5286720276967511,
+    "speedup_flydsl_vs_triton": 0.5333379303425151,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.5171421052230196,
-    "best_speedup_vs_aoTriton": 0.8546429631780178,
-    "best_speedup_vs_flydsl": 0.978190783946025
+    "best_speedup_vs_triton": 0.5176661848571447,
+    "best_speedup_vs_aoTriton": 0.8494513469638377,
+    "best_speedup_vs_flydsl": 0.9706157304894744
   },
   {
     "signature": "d128_b64_h64kv8_q4096_k4096_ns1_tq4096_sw0_sc0_sinks0_bfloat16",
@@ -654,34 +654,34 @@
     "num_seqs": 1,
     "total_q": 4096,
     "max_seqlen_k": 4096,
-    "triton_ms": 0.47394485473632814,
-    "aoTriton_ms": 0.7778615570068359,
-    "flydsl_ms": 0.5009763336181641,
+    "triton_ms": 0.4728183364868164,
+    "aoTriton_ms": 0.7772989654541016,
+    "flydsl_ms": 0.49553455352783204,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.9863083648681641,
-        "speedup": 0.48052401421099017,
+        "ms": 0.9807850646972657,
+        "speedup": 0.4820815013458214,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 1.781826934814453,
-        "speedup": 0.2659881526516948,
+        "ms": 1.7736672973632812,
+        "speedup": 0.2665766782697658,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.9460424034672287,
+    "speedup_flydsl_vs_triton": 0.9541581573287002,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.48052401421099017,
-    "best_speedup_vs_aoTriton": 0.788659596444576,
-    "best_speedup_vs_flydsl": 0.5079307359267176
+    "best_speedup_vs_triton": 0.4820815013458214,
+    "best_speedup_vs_aoTriton": 0.7925273267635115,
+    "best_speedup_vs_flydsl": 0.5052427604826817
   },
   {
     "signature": "d128_b64_h64kv8_q8192_k8192_ns1_tq8192_sw0_sc0_sinks0_bfloat16",
@@ -690,34 +690,34 @@
     "num_seqs": 1,
     "total_q": 8192,
     "max_seqlen_k": 8192,
-    "triton_ms": 1.7028677368164062,
-    "aoTriton_ms": 2.63884521484375,
-    "flydsl_ms": 1.3335743713378907,
+    "triton_ms": 1.6922450256347656,
+    "aoTriton_ms": 2.63490966796875,
+    "flydsl_ms": 1.3214260864257812,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 3.6593307495117187,
-        "speedup": 0.4653495006002471,
+        "ms": 3.65565185546875,
+        "speedup": 0.46291197645181004,
         "max_abs": 0.0078125,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 6.8523486328125,
-        "speedup": 0.24850862500810553,
+        "ms": 6.8215673828125,
+        "speedup": 0.24807275669496665,
         "max_abs": 0.0078125,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 1.2769199629324213,
+    "speedup_flydsl_vs_triton": 1.2806202654981502,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.4653495006002471,
-    "best_speedup_vs_aoTriton": 0.7211278224019141,
-    "best_speedup_vs_flydsl": 0.36443122052190435
+    "best_speedup_vs_triton": 0.46291197645181004,
+    "best_speedup_vs_aoTriton": 0.7207769700572009,
+    "best_speedup_vs_flydsl": 0.36147481726110375
   },
   {
     "signature": "d128_b64_h128kv8_q512_k512_ns1_tq512_sw0_sc0_sinks0_bfloat16",
@@ -726,34 +726,34 @@
     "num_seqs": 1,
     "total_q": 512,
     "max_seqlen_k": 512,
-    "triton_ms": 0.0437308406829834,
-    "aoTriton_ms": 0.07787360191345215,
-    "flydsl_ms": 0.23373918533325194,
+    "triton_ms": 0.04345004081726074,
+    "aoTriton_ms": 0.07788477897644043,
+    "flydsl_ms": 0.21163089752197264,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.05625016212463379,
-        "speedup": 0.7774349269623213,
+        "ms": 0.05568294048309326,
+        "speedup": 0.7803115359982339,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h128kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.08354244232177735,
-        "speedup": 0.5234565744983481,
+        "ms": 0.08225842475891114,
+        "speedup": 0.5282138692128766,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h128kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.18709246641993071,
+    "speedup_flydsl_vs_triton": 0.2053104784132455,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.7774349269623213,
-    "best_speedup_vs_aoTriton": 1.3844155993880904,
-    "best_speedup_vs_flydsl": 4.15535131819451
+    "best_speedup_vs_triton": 0.7803115359982339,
+    "best_speedup_vs_aoTriton": 1.3987188589670154,
+    "best_speedup_vs_flydsl": 3.8006415553113455
   },
   {
     "signature": "d128_b64_h128kv8_q1024_k1024_ns1_tq1024_sw0_sc0_sinks0_bfloat16",
@@ -762,34 +762,34 @@
     "num_seqs": 1,
     "total_q": 1024,
     "max_seqlen_k": 1024,
-    "triton_ms": 0.08492566108703613,
-    "aoTriton_ms": 0.1791330337524414,
-    "flydsl_ms": 0.24456247329711914,
+    "triton_ms": 0.08457045555114746,
+    "aoTriton_ms": 0.18007461547851564,
+    "flydsl_ms": 0.24390565872192382,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.15975842475891114,
-        "speedup": 0.5315879974104406,
+        "ms": 0.15820158004760743,
+        "speedup": 0.5345740259085767,
         "max_abs": 0.0078125,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h128kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.26060985565185546,
-        "speedup": 0.3258727912442688,
+        "ms": 0.2635706520080566,
+        "speedup": 0.3208644623626851,
         "max_abs": 0.0078125,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h128kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.3472554883097697,
+    "speedup_flydsl_vs_triton": 0.3467342906035895,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.5315879974104406,
-    "best_speedup_vs_aoTriton": 1.1212744118049998,
-    "best_speedup_vs_flydsl": 1.5308267696441324
+    "best_speedup_vs_triton": 0.5345740259085767,
+    "best_speedup_vs_aoTriton": 1.1382605371218542,
+    "best_speedup_vs_flydsl": 1.5417397136521998
   },
   {
     "signature": "d128_b64_h128kv8_q2048_k2048_ns1_tq2048_sw0_sc0_sinks0_bfloat16",
@@ -798,34 +798,34 @@
     "num_seqs": 1,
     "total_q": 2048,
     "max_seqlen_k": 2048,
-    "triton_ms": 0.25996583938598633,
-    "aoTriton_ms": 0.4551030349731445,
-    "flydsl_ms": 0.38356952667236327,
+    "triton_ms": 0.25761377334594726,
+    "aoTriton_ms": 0.4580598068237305,
+    "flydsl_ms": 0.36473087310791014,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.5086603546142578,
-        "speedup": 0.5110794207327819,
+        "ms": 0.5112283325195313,
+        "speedup": 0.5039113776740949,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h128kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.9293822479248047,
-        "speedup": 0.2797189638240431,
+        "ms": 0.9221965026855469,
+        "speedup": 0.27934802679878423,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h128kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.6777541522688362,
+    "speedup_flydsl_vs_triton": 0.7063119476308468,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.5110794207327819,
-    "best_speedup_vs_aoTriton": 0.8947090742274804,
-    "best_speedup_vs_flydsl": 0.7540778894852989
+    "best_speedup_vs_triton": 0.5039113776740949,
+    "best_speedup_vs_aoTriton": 0.8959984759964972,
+    "best_speedup_vs_flydsl": 0.7134402573315433
   },
   {
     "signature": "d128_b64_h128kv8_q4096_k4096_ns1_tq4096_sw0_sc0_sinks0_bfloat16",
@@ -834,34 +834,34 @@
     "num_seqs": 1,
     "total_q": 4096,
     "max_seqlen_k": 4096,
-    "triton_ms": 0.8741985321044922,
-    "aoTriton_ms": 1.6216500854492188,
-    "flydsl_ms": 0.7871688079833984,
+    "triton_ms": 0.8720351409912109,
+    "aoTriton_ms": 1.6252706909179688,
+    "flydsl_ms": 0.7837214660644531,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 1.8524493408203124,
-        "speedup": 0.4719149467900561,
+        "ms": 1.8211102294921875,
+        "speedup": 0.4788480822681316,
         "max_abs": 0.0078125,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h128kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 3.448607177734375,
-        "speedup": 0.2534932182907572,
+        "ms": 3.433183288574219,
+        "speedup": 0.2540019182469463,
         "max_abs": 0.0078125,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h128kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 1.1105604328302212,
+    "speedup_flydsl_vs_triton": 1.1126850274629263,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.4719149467900561,
-    "best_speedup_vs_aoTriton": 0.8754086007723754,
-    "best_speedup_vs_flydsl": 0.42493405386989946
+    "best_speedup_vs_triton": 0.4788480822681316,
+    "best_speedup_vs_aoTriton": 0.8924614581793722,
+    "best_speedup_vs_flydsl": 0.4303536674345035
   },
   {
     "signature": "d128_b64_h128kv8_q8192_k8192_ns1_tq8192_sw0_sc0_sinks0_bfloat16",
@@ -870,34 +870,34 @@
     "num_seqs": 1,
     "total_q": 8192,
     "max_seqlen_k": 8192,
-    "triton_ms": 3.2899957275390626,
-    "aoTriton_ms": 5.07345458984375,
-    "flydsl_ms": 2.4707867431640627,
+    "triton_ms": 3.2916650390625,
+    "aoTriton_ms": 5.080816650390625,
+    "flydsl_ms": 2.4647311401367187,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 6.960784912109375,
-        "speedup": 0.47264723290265725,
+        "ms": 6.964030151367187,
+        "speedup": 0.4726666840200679,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h128kv8_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 13.448505859375,
-        "speedup": 0.2446365240823831,
+        "ms": 13.340902099609375,
+        "speedup": 0.24673481706749656,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h128kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 1.3315579487551927,
+    "speedup_flydsl_vs_triton": 1.3355067355865482,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.47264723290265725,
-    "best_speedup_vs_aoTriton": 0.728862427715254,
-    "best_speedup_vs_flydsl": 0.35495806498283583
+    "best_speedup_vs_triton": 0.4726666840200679,
+    "best_speedup_vs_aoTriton": 0.7295799328773946,
+    "best_speedup_vs_flydsl": 0.353923100067687
   },
   {
     "signature": "d128_b64_h64kv4_q512_k512_ns1_tq512_sw0_sc0_sinks0_bfloat16",
@@ -906,34 +906,34 @@
     "num_seqs": 1,
     "total_q": 512,
     "max_seqlen_k": 512,
-    "triton_ms": 0.043572459220886234,
-    "aoTriton_ms": 0.042516441345214845,
-    "flydsl_ms": 0.23091592788696289,
+    "triton_ms": 0.04352283954620361,
+    "aoTriton_ms": 0.04244524002075195,
+    "flydsl_ms": 0.21005647659301757,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.054987740516662595,
-        "speedup": 0.7924031577126313,
+        "ms": 0.053204498291015624,
+        "speedup": 0.8180293197793972,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv4_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.04578125953674316,
-        "speedup": 0.9517531772125188,
+        "ms": 0.04603643894195557,
+        "speedup": 0.9453997864838939,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv4_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.18869403951301125,
+    "speedup_flydsl_vs_triton": 0.20719589441904568,
     "best_variant": "fallback",
-    "best_speedup_vs_triton": 0.9517531772125188,
-    "best_speedup_vs_aoTriton": 0.9286865799551007,
-    "best_speedup_vs_flydsl": 5.043896350244235
+    "best_speedup_vs_triton": 0.9453997864838939,
+    "best_speedup_vs_aoTriton": 0.9219922521433178,
+    "best_speedup_vs_flydsl": 4.562830692831487
   },
   {
     "signature": "d128_b64_h64kv4_q1024_k1024_ns1_tq1024_sw0_sc0_sinks0_bfloat16",
@@ -942,34 +942,34 @@
     "num_seqs": 1,
     "total_q": 1024,
     "max_seqlen_k": 1024,
-    "triton_ms": 0.04681325912475586,
-    "aoTriton_ms": 0.11070431709289551,
-    "flydsl_ms": 0.22523910522460938,
+    "triton_ms": 0.04762685775756836,
+    "aoTriton_ms": 0.11055312156677247,
+    "flydsl_ms": 0.2218622398376465,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.09327374458312988,
-        "speedup": 0.5018910662800046,
+        "ms": 0.09276412010192871,
+        "speedup": 0.5134189566530274,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv4_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.1385646152496338,
-        "speedup": 0.33784425439653926,
+        "ms": 0.137773380279541,
+        "speedup": 0.345689839800213,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv4_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.20783806203666758,
+    "speedup_flydsl_vs_triton": 0.2146686060341794,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.5018910662800046,
-    "best_speedup_vs_aoTriton": 1.1868754448282142,
-    "best_speedup_vs_flydsl": 2.4148178700370053
+    "best_speedup_vs_triton": 0.5134189566530274,
+    "best_speedup_vs_aoTriton": 1.1917659699170036,
+    "best_speedup_vs_flydsl": 2.3916816070034996
   },
   {
     "signature": "d128_b64_h64kv4_q2048_k2048_ns1_tq2048_sw0_sc0_sinks0_bfloat16",
@@ -978,34 +978,34 @@
     "num_seqs": 1,
     "total_q": 2048,
     "max_seqlen_k": 2048,
-    "triton_ms": 0.1409382152557373,
-    "aoTriton_ms": 0.24619770050048828,
-    "flydsl_ms": 0.28387807846069335,
+    "triton_ms": 0.14402706146240235,
+    "aoTriton_ms": 0.24974409103393555,
+    "flydsl_ms": 0.28221883773803713,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.28835172653198243,
-        "speedup": 0.48877187922821463,
+        "ms": 0.28485004425048827,
+        "speedup": 0.5056241498623375,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv4_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.48250972747802734,
-        "speedup": 0.2920940392070238,
+        "ms": 0.48078319549560544,
+        "speedup": 0.29956758641269693,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv4_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.49647445840116905,
+    "speedup_flydsl_vs_triton": 0.5103382276561284,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.48877187922821463,
-    "best_speedup_vs_aoTriton": 0.8538103914324278,
-    "best_speedup_vs_flydsl": 0.9844854472518898
+    "best_speedup_vs_triton": 0.5056241498623375,
+    "best_speedup_vs_aoTriton": 0.8767563708514589,
+    "best_speedup_vs_flydsl": 0.9907628362165978
   },
   {
     "signature": "d128_b64_h64kv4_q4096_k4096_ns1_tq4096_sw0_sc0_sinks0_bfloat16",
@@ -1014,34 +1014,34 @@
     "num_seqs": 1,
     "total_q": 4096,
     "max_seqlen_k": 4096,
-    "triton_ms": 0.4675344085693359,
-    "aoTriton_ms": 0.7772415161132813,
-    "flydsl_ms": 0.5011907196044922,
+    "triton_ms": 0.4639966583251953,
+    "aoTriton_ms": 0.7759101867675782,
+    "flydsl_ms": 0.49666576385498046,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.979204330444336,
-        "speedup": 0.4774635834761696,
+        "ms": 0.9803058624267578,
+        "speedup": 0.47331825311802844,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv4_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 1.7834877014160155,
-        "speedup": 0.2621461354615078,
+        "ms": 1.7833779907226563,
+        "speedup": 0.2601785267839801,
         "max_abs": 0.0078125,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv4_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.9328472980072022,
+    "speedup_flydsl_vs_triton": 0.9342231578915029,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.4774635834761696,
-    "best_speedup_vs_aoTriton": 0.7937480380224529,
-    "best_speedup_vs_flydsl": 0.5118346641472323
+    "best_speedup_vs_triton": 0.47331825311802844,
+    "best_speedup_vs_aoTriton": 0.7914980584190368,
+    "best_speedup_vs_flydsl": 0.5066436740727828
   },
   {
     "signature": "d128_b64_h64kv4_q8192_k8192_ns1_tq8192_sw0_sc0_sinks0_bfloat16",
@@ -1050,34 +1050,34 @@
     "num_seqs": 1,
     "total_q": 8192,
     "max_seqlen_k": 8192,
-    "triton_ms": 1.6826394653320313,
-    "aoTriton_ms": 2.6346820068359373,
-    "flydsl_ms": 1.3292518615722657,
+    "triton_ms": 1.677855224609375,
+    "aoTriton_ms": 2.6302423095703125,
+    "flydsl_ms": 1.320194854736328,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 3.640749816894531,
-        "speedup": 0.46216838562317936,
+        "ms": 3.6115850830078124,
+        "speedup": 0.4645758541044859,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv4_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 6.788587646484375,
-        "speedup": 0.24786296545842265,
+        "ms": 6.752616577148437,
+        "speedup": 0.24847482534213672,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h64kv4_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 1.2658545110794668,
+    "speedup_flydsl_vs_triton": 1.2709148339655358,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.46216838562317936,
-    "best_speedup_vs_aoTriton": 0.7236646678137459,
-    "best_speedup_vs_flydsl": 0.365103873769081
+    "best_speedup_vs_triton": 0.4645758541044859,
+    "best_speedup_vs_aoTriton": 0.7282792040385174,
+    "best_speedup_vs_flydsl": 0.3655444422305673
   },
   {
     "signature": "d128_b64_h32kv32_q512_k512_ns1_tq512_sw0_sc0_sinks0_bfloat16",
@@ -1086,34 +1086,34 @@
     "num_seqs": 1,
     "total_q": 512,
     "max_seqlen_k": 512,
-    "triton_ms": 0.04412444114685059,
-    "aoTriton_ms": 0.04076759815216065,
-    "flydsl_ms": 0.2098573112487793,
+    "triton_ms": 0.04379404067993164,
+    "aoTriton_ms": 0.04099720001220703,
+    "flydsl_ms": 0.21101810455322265,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.05419174194335938,
-        "speedup": 0.8142281381722141,
+        "ms": 0.05316609859466553,
+        "speedup": 0.8237211651322062,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv32_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.036842761039733884,
-        "speedup": 1.197642084947532,
+        "ms": 0.036573159694671634,
+        "speedup": 1.19743661869915,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv32_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.21025925131835144,
+    "speedup_flydsl_vs_triton": 0.2075368877597229,
     "best_variant": "fallback",
-    "best_speedup_vs_triton": 1.197642084947532,
-    "best_speedup_vs_aoTriton": 1.1065293968656131,
-    "best_speedup_vs_flydsl": 5.696025632347534
+    "best_speedup_vs_triton": 1.19743661869915,
+    "best_speedup_vs_aoTriton": 1.1209641265471504,
+    "best_speedup_vs_flydsl": 5.769753182795579
   },
   {
     "signature": "d128_b64_h32kv32_q1024_k1024_ns1_tq1024_sw0_sc0_sinks0_bfloat16",
@@ -1122,34 +1122,34 @@
     "num_seqs": 1,
     "total_q": 1024,
     "max_seqlen_k": 1024,
-    "triton_ms": 0.04409403800964355,
-    "aoTriton_ms": 0.062225418090820314,
-    "flydsl_ms": 0.22093664169311522,
+    "triton_ms": 0.044235639572143555,
+    "aoTriton_ms": 0.06207181930541992,
+    "flydsl_ms": 0.2188134002685547,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.05861658096313477,
-        "speedup": 0.7522451375554511,
+        "ms": 0.0583909797668457,
+        "speedup": 0.7575765940009193,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv32_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.10173303604125977,
-        "speedup": 0.4334289010283776,
+        "ms": 0.10131540298461914,
+        "speedup": 0.4366131730123903,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv32_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.19957775076028778,
+    "speedup_flydsl_vs_triton": 0.20216147419605993,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.7522451375554511,
-    "best_speedup_vs_aoTriton": 1.0615668308930406,
-    "best_speedup_vs_flydsl": 3.7691833618215815
+    "best_speedup_vs_triton": 0.7575765940009193,
+    "best_speedup_vs_aoTriton": 1.063037810861742,
+    "best_speedup_vs_flydsl": 3.747383605177945
   },
   {
     "signature": "d128_b64_h32kv32_q2048_k2048_ns1_tq2048_sw0_sc0_sinks0_bfloat16",
@@ -1158,34 +1158,34 @@
     "num_seqs": 1,
     "total_q": 2048,
     "max_seqlen_k": 2048,
-    "triton_ms": 0.09208494186401367,
-    "aoTriton_ms": 0.14362706184387208,
-    "flydsl_ms": 0.2510353660583496,
+    "triton_ms": 0.0914377212524414,
+    "aoTriton_ms": 0.14257424354553222,
+    "flydsl_ms": 0.2466640853881836,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.17913301467895507,
-        "speedup": 0.5140590193775821,
+        "ms": 0.1812649917602539,
+        "speedup": 0.5044422553108295,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv32_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.46751754760742187,
-        "speedup": 0.19696574456995167,
+        "ms": 0.46714229583740235,
+        "speedup": 0.1957384764069148,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv32_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.3668205930896997,
+    "speedup_flydsl_vs_triton": 0.3706973437521023,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.5140590193775821,
-    "best_speedup_vs_aoTriton": 0.801790011189633,
-    "best_speedup_vs_flydsl": 1.4013908408132305
+    "best_speedup_vs_triton": 0.5044422553108295,
+    "best_speedup_vs_aoTriton": 0.786551457956674,
+    "best_speedup_vs_flydsl": 1.3607927432255003
   },
   {
     "signature": "d128_b64_h32kv32_q4096_k4096_ns1_tq4096_sw0_sc0_sinks0_bfloat16",
@@ -1194,34 +1194,34 @@
     "num_seqs": 1,
     "total_q": 4096,
     "max_seqlen_k": 4096,
-    "triton_ms": 0.2745643997192383,
-    "aoTriton_ms": 0.43816768646240234,
-    "flydsl_ms": 0.36884136199951173,
+    "triton_ms": 0.2736651420593262,
+    "aoTriton_ms": 0.4387900161743164,
+    "flydsl_ms": 0.36355484008789063,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.5957004928588867,
-        "speedup": 0.4609101436219205,
+        "ms": 0.5921171188354493,
+        "speedup": 0.46218076349077547,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv32_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 1.740249786376953,
-        "speedup": 0.15777298286061406,
+        "ms": 1.7290013122558594,
+        "speedup": 0.15827931426047925,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv32_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.7443969901607774,
+    "speedup_flydsl_vs_triton": 0.7527478990326926,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.4609101436219205,
-    "best_speedup_vs_aoTriton": 0.7355503171728922,
-    "best_speedup_vs_flydsl": 0.6191724976243812
+    "best_speedup_vs_triton": 0.46218076349077547,
+    "best_speedup_vs_aoTriton": 0.741052744830803,
+    "best_speedup_vs_flydsl": 0.6139914360235266
   },
   {
     "signature": "d128_b64_h32kv32_q8192_k8192_ns1_tq8192_sw0_sc0_sinks0_bfloat16",
@@ -1230,34 +1230,34 @@
     "num_seqs": 1,
     "total_q": 8192,
     "max_seqlen_k": 8192,
-    "triton_ms": 0.9736331176757812,
-    "aoTriton_ms": 1.3939646911621093,
-    "flydsl_ms": 0.8177291107177734,
+    "triton_ms": 0.9738137817382813,
+    "aoTriton_ms": 1.3924674987792969,
+    "flydsl_ms": 0.8154657745361328,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 2.2362603759765625,
-        "speedup": 0.43538450537120527,
+        "ms": 2.2351527404785156,
+        "speedup": 0.435681089753983,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv32_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 6.579153442382813,
-        "speedup": 0.14798759843533218,
+        "ms": 6.578958740234375,
+        "speedup": 0.14801943897030567,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h32kv32_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 1.1906548328958972,
+    "speedup_flydsl_vs_triton": 1.194181058416857,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.43538450537120527,
-    "best_speedup_vs_aoTriton": 0.6233463268128483,
-    "best_speedup_vs_flydsl": 0.3656681124892157
+    "best_speedup_vs_triton": 0.435681089753983,
+    "best_speedup_vs_aoTriton": 0.6229853886769227,
+    "best_speedup_vs_flydsl": 0.36483671105248616
   },
   {
     "signature": "d128_b64_h40kv40_q512_k512_ns1_tq512_sw0_sc0_sinks0_bfloat16",
@@ -1266,34 +1266,34 @@
     "num_seqs": 1,
     "total_q": 512,
     "max_seqlen_k": 512,
-    "triton_ms": 0.043777241706848144,
-    "aoTriton_ms": 0.04000842094421387,
-    "flydsl_ms": 0.2411896324157715,
+    "triton_ms": 0.04303324222564697,
+    "aoTriton_ms": 0.04191162109375,
+    "flydsl_ms": 0.21084770202636718,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.054942140579223635,
-        "speedup": 0.796788061865258,
+        "ms": 0.05351652145385742,
+        "speedup": 0.804111348357175,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv40_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.03990520000457764,
-        "speedup": 1.0970310060299493,
+        "ms": 0.03968597888946533,
+        "speedup": 1.0843437261684925,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv40_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.18150548706580943,
+    "speedup_flydsl_vs_triton": 0.20409633025199167,
     "best_variant": "fallback",
-    "best_speedup_vs_triton": 1.0970310060299493,
-    "best_speedup_vs_aoTriton": 1.0025866538602586,
-    "best_speedup_vs_flydsl": 6.044065244331665
+    "best_speedup_vs_triton": 1.0843437261684925,
+    "best_speedup_vs_aoTriton": 1.0560813230910493,
+    "best_speedup_vs_flydsl": 5.312901632428597
   },
   {
     "signature": "d128_b64_h40kv40_q1024_k1024_ns1_tq1024_sw0_sc0_sinks0_bfloat16",
@@ -1302,34 +1302,34 @@
     "num_seqs": 1,
     "total_q": 1024,
     "max_seqlen_k": 1024,
-    "triton_ms": 0.04385245800018311,
-    "aoTriton_ms": 0.08130962371826173,
-    "flydsl_ms": 0.22213586807250976,
+    "triton_ms": 0.04425084114074707,
+    "aoTriton_ms": 0.08166241645812988,
+    "flydsl_ms": 0.2209870147705078,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.0785359811782837,
-        "speedup": 0.5583741024465475,
+        "ms": 0.07835837841033935,
+        "speedup": 0.5647237990176197,
         "max_abs": 0.015625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv40_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.11019071578979492,
-        "speedup": 0.3979687189239986,
+        "ms": 0.10902667999267578,
+        "speedup": 0.40587167419680914,
         "max_abs": 0.015625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv40_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.1974127743560475,
+    "speedup_flydsl_vs_triton": 0.20024181595783358,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.5583741024465475,
-    "best_speedup_vs_aoTriton": 1.0353168381977886,
-    "best_speedup_vs_flydsl": 2.828459831274553
+    "best_speedup_vs_triton": 0.5647237990176197,
+    "best_speedup_vs_aoTriton": 1.0421657277092728,
+    "best_speedup_vs_flydsl": 2.8202091372190607
   },
   {
     "signature": "d128_b64_h40kv40_q2048_k2048_ns1_tq2048_sw0_sc0_sinks0_bfloat16",
@@ -1338,34 +1338,34 @@
     "num_seqs": 1,
     "total_q": 2048,
     "max_seqlen_k": 2048,
-    "triton_ms": 0.11193553924560547,
-    "aoTriton_ms": 0.1822602653503418,
-    "flydsl_ms": 0.2665699005126953,
+    "triton_ms": 0.11295233726501465,
+    "aoTriton_ms": 0.1828402328491211,
+    "flydsl_ms": 0.26466745376586914,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.24078643798828125,
-        "speedup": 0.4648747669544961,
+        "ms": 0.23846317291259767,
+        "speedup": 0.4736678451662401,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv40_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 0.3662621307373047,
-        "speedup": 0.305615923274059,
+        "ms": 0.36528209686279295,
+        "speedup": 0.30921947238887465,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv40_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.4199106464395239,
+    "speedup_flydsl_vs_triton": 0.4267707859725543,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.4648747669544961,
-    "best_speedup_vs_aoTriton": 0.7569374208659216,
-    "best_speedup_vs_flydsl": 1.107080210745378
+    "best_speedup_vs_triton": 0.4736678451662401,
+    "best_speedup_vs_aoTriton": 0.7667441081820057,
+    "best_speedup_vs_flydsl": 1.1098881665173346
   },
   {
     "signature": "d128_b64_h40kv40_q4096_k4096_ns1_tq4096_sw0_sc0_sinks0_bfloat16",
@@ -1374,34 +1374,34 @@
     "num_seqs": 1,
     "total_q": 4096,
     "max_seqlen_k": 4096,
-    "triton_ms": 0.3363090133666992,
-    "aoTriton_ms": 0.5225525283813477,
-    "flydsl_ms": 0.4169034576416016,
+    "triton_ms": 0.3392945861816406,
+    "aoTriton_ms": 0.5222820663452148,
+    "flydsl_ms": 0.4198842239379883,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.7758271026611329,
-        "speedup": 0.4334844867021781,
+        "ms": 0.7759901428222656,
+        "speedup": 0.4372408455442885,
         "max_abs": 0.0078125,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv40_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 1.28006103515625,
-        "speedup": 0.2627288888030623,
+        "ms": 1.2757288360595702,
+        "speedup": 0.26596136780104673,
         "max_abs": 0.0078125,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv40_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.8066831953593755,
+    "speedup_flydsl_vs_triton": 0.8080670023738501,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.4334844867021781,
-    "best_speedup_vs_aoTriton": 0.6735425026903051,
-    "best_speedup_vs_flydsl": 0.5373664521535766
+    "best_speedup_vs_triton": 0.4372408455442885,
+    "best_speedup_vs_aoTriton": 0.6730524494108675,
+    "best_speedup_vs_flydsl": 0.5410947907287521
   },
   {
     "signature": "d128_b64_h40kv40_q8192_k8192_ns1_tq8192_sw0_sc0_sinks0_bfloat16",
@@ -1410,34 +1410,34 @@
     "num_seqs": 1,
     "total_q": 8192,
     "max_seqlen_k": 8192,
-    "triton_ms": 1.235236587524414,
-    "aoTriton_ms": 1.7191558837890626,
-    "flydsl_ms": 0.9735499572753906,
+    "triton_ms": 1.2324427032470704,
+    "aoTriton_ms": 1.7143052673339845,
+    "flydsl_ms": 0.9693681335449219,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 2.8946142578125,
-        "speedup": 0.4267361649969552,
+        "ms": 2.936159973144531,
+        "speedup": 0.4197464424689247,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv40_bf16_w2_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_ksb"
       },
       "fallback": {
-        "ms": 4.756542663574219,
-        "speedup": 0.25969210724922154,
+        "ms": 4.775755310058594,
+        "speedup": 0.25806236359122625,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d128_b64_t128_h40kv40_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 1.2687963039733352,
+    "speedup_flydsl_vs_triton": 1.2713876809010631,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.4267361649969552,
-    "best_speedup_vs_aoTriton": 0.593915365112674,
-    "best_speedup_vs_flydsl": 0.336331500699204
+    "best_speedup_vs_triton": 0.4197464424689247,
+    "best_speedup_vs_aoTriton": 0.5838596272048555,
+    "best_speedup_vs_flydsl": 0.3301482693079425
   },
   {
     "signature": "d64_b32_h64kv8_q2048_k2048_ns1_tq2048_sw0_sc0_sinks0_bfloat16",
@@ -1446,9 +1446,9 @@
     "num_seqs": 1,
     "total_q": 2048,
     "max_seqlen_k": 2048,
-    "triton_ms": 0.09678019523620605,
-    "aoTriton_ms": 0.1319981575012207,
-    "flydsl_ms": 0.17625856399536133,
+    "triton_ms": 0.09536895751953126,
+    "aoTriton_ms": 0.1331269359588623,
+    "flydsl_ms": 0.14998629570007324,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
@@ -1458,18 +1458,18 @@
         "error": "ValueError('use_fast_paged_kv_desc is restricted to bf16 h64kv8 HD=64 BS=32 T=64 num_warps=4')"
       },
       "fallback": {
-        "ms": 0.2011308479309082,
-        "speedup": 0.48118026763080946,
+        "ms": 0.20134199142456055,
+        "speedup": 0.4736665056542087,
         "max_abs": 0.001953125,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d64_b32_t64_h64kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.5490808108407887,
+    "speedup_flydsl_vs_triton": 0.6358511427620029,
     "best_variant": "fallback",
-    "best_speedup_vs_triton": 0.48118026763080946,
-    "best_speedup_vs_aoTriton": 0.6562800229757112,
-    "best_speedup_vs_flydsl": 0.8763377960595536
+    "best_speedup_vs_triton": 0.4736665056542087,
+    "best_speedup_vs_aoTriton": 0.661198069100964,
+    "best_speedup_vs_flydsl": 0.7449330099442797
   },
   {
     "signature": "d64_b32_h64kv8_q4096_k4096_ns1_tq4096_sw0_sc0_sinks0_bfloat16",
@@ -1478,9 +1478,9 @@
     "num_seqs": 1,
     "total_q": 4096,
     "max_seqlen_k": 4096,
-    "triton_ms": 0.3134895706176758,
-    "aoTriton_ms": 0.3490867614746094,
-    "flydsl_ms": 0.23817441940307618,
+    "triton_ms": 0.30757747650146483,
+    "aoTriton_ms": 0.34462265014648436,
+    "flydsl_ms": 0.23331274032592775,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
@@ -1490,18 +1490,18 @@
         "error": "ValueError('use_fast_paged_kv_desc is restricted to bf16 h64kv8 HD=64 BS=32 T=64 num_warps=4')"
       },
       "fallback": {
-        "ms": 0.6903694152832032,
-        "speedup": 0.4540895985217946,
+        "ms": 0.688012466430664,
+        "speedup": 0.4470521851110406,
         "max_abs": 0.001953125,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d64_b32_t64_h64kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 1.316218473013844,
+    "speedup_flydsl_vs_triton": 1.318305533044584,
     "best_variant": "fallback",
-    "best_speedup_vs_triton": 0.4540895985217946,
-    "best_speedup_vs_aoTriton": 0.5056521244229903,
-    "best_speedup_vs_flydsl": 0.3449956126827726
+    "best_speedup_vs_triton": 0.4470521851110406,
+    "best_speedup_vs_aoTriton": 0.500895938607552,
+    "best_speedup_vs_flydsl": 0.3391112104935098
   },
   {
     "signature": "d64_b32_h64kv8_q8192_k8192_ns1_tq8192_sw0_sc0_sinks0_bfloat16",
@@ -1510,9 +1510,9 @@
     "num_seqs": 1,
     "total_q": 8192,
     "max_seqlen_k": 8192,
-    "triton_ms": 1.0929367065429687,
-    "aoTriton_ms": 1.5019953918457032,
-    "flydsl_ms": 0.786251220703125,
+    "triton_ms": 1.0870700836181642,
+    "aoTriton_ms": 1.5029534912109375,
+    "flydsl_ms": 0.7879878997802734,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
@@ -1522,18 +1522,18 @@
         "error": "ValueError('use_fast_paged_kv_desc is restricted to bf16 h64kv8 HD=64 BS=32 T=64 num_warps=4')"
       },
       "fallback": {
-        "ms": 2.6061056518554686,
-        "speedup": 0.4193754408094817,
+        "ms": 2.5890548706054686,
+        "speedup": 0.4198713962998958,
         "max_abs": 0.001953125,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d64_b32_t64_h64kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 1.390060425681225,
+    "speedup_flydsl_vs_triton": 1.3795517468241434,
     "best_variant": "fallback",
-    "best_speedup_vs_triton": 0.4193754408094817,
-    "best_speedup_vs_aoTriton": 0.5763371069689088,
-    "best_speedup_vs_flydsl": 0.3016958349878632
+    "best_speedup_vs_triton": 0.4198713962998958,
+    "best_speedup_vs_aoTriton": 0.5805027573090644,
+    "best_speedup_vs_flydsl": 0.3043534954498654
   },
   {
     "signature": "d64_b32_h64kv8_q1024_k1024_ns2_tq2048_sw0_sc0_sinks0_bfloat16",
@@ -1542,33 +1542,33 @@
     "num_seqs": 2,
     "total_q": 2048,
     "max_seqlen_k": 1024,
-    "triton_ms": 0.055994181632995604,
-    "aoTriton_ms": 0.09300733566284179,
-    "flydsl_ms": 0.14775670051574707,
+    "triton_ms": 0.05494696140289307,
+    "aoTriton_ms": 0.09310653686523437,
+    "flydsl_ms": 0.15140392303466796,
     "flydsl_ok": true,
     "flydsl_max_abs": 0.015625,
     "flydsl_skip_reason": "",
     "flydsl_error": null,
     "variants": {
       "prod": {
-        "ms": 0.07463913917541504,
-        "speedup": 0.7501986524978468,
+        "ms": 0.0730239200592041,
+        "speedup": 0.7524515440741177,
         "max_abs": 0.00390625,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d64_b32_t64_h64kv8_bf16_w4_mw32_mfma32_stqk_s1_mask1_hlpv_skipqreg_mlim_fastkvdesc"
       },
       "fallback": {
-        "ms": 0.11481156349182128,
-        "speedup": 0.4877050702038772,
+        "ms": 0.1151971435546875,
+        "speedup": 0.4769819780887893,
         "max_abs": 0.001953125,
         "ok": true,
         "kernel": "rocke_uattn2d_tiled_d64_b32_t64_h64kv8_bf16_w4"
       }
     },
-    "speedup_flydsl_vs_triton": 0.3789620466452421,
+    "speedup_flydsl_vs_triton": 0.36291636505555736,
     "best_variant": "prod",
-    "best_speedup_vs_triton": 0.7501986524978468,
-    "best_speedup_vs_aoTriton": 1.2460933591993644,
-    "best_speedup_vs_flydsl": 1.9796142097578722
+    "best_speedup_vs_triton": 0.7524515440741177,
+    "best_speedup_vs_aoTriton": 1.275014225335319,
+    "best_speedup_vs_flydsl": 2.0733469651029046
   }
 ]