From 143ce80a45c221e3b02a3415f784afb6f946af7e Mon Sep 17 00:00:00 2001
From: "Lv, Tao A" <tao.a.lv@intel.com>
Date: Thu, 2 Jan 2025 00:39:59 -0800
Subject: [PATCH] benchdnn: inputs: graph: rm case
 sdpa-compressed-kv-int8-gs128.json

rewrite sdpa-compressed-kv-int4-gs32.json for it.
---
 .../graph/complex_fusion/harness_mha_all      |   1 -
 .../graph/complex_fusion/harness_mha_ci       |   3 +-
 .../mha/sdpa-compressed-kv-int8-gs128.json    | 548 ------------------
 3 files changed, 2 insertions(+), 550 deletions(-)
 delete mode 100644 tests/benchdnn/inputs/graph/complex_fusion/mha/sdpa-compressed-kv-int8-gs128.json

diff --git a/tests/benchdnn/inputs/graph/complex_fusion/harness_mha_all b/tests/benchdnn/inputs/graph/complex_fusion/harness_mha_all
index 80199e56b07..0b0691e2cf5 100644
--- a/tests/benchdnn/inputs/graph/complex_fusion/harness_mha_all
+++ b/tests/benchdnn/inputs/graph/complex_fusion/harness_mha_all
@@ -24,7 +24,6 @@
 --reset --expected-n-partitions=0 --case=complex_fusion/mha/dynamic_quantized_mha-Bert_large-inf-int8-bs1-fake.json
 --reset --case=complex_fusion/mha/sdpa-plain-wo-scale-int8-bs1.json
 --reset --case=complex_fusion/mha/sdpa-compressed-kv-int4-gs32.json
---reset --case=complex_fusion/mha/sdpa-compressed-kv-int8-gs128.json
 --reset --case=complex_fusion/mha/sdpa-compressed-k-int8-gs32.json
 --reset --case=complex_fusion/mha/sdpa-compressed-v-int8-gs32.json
 
diff --git a/tests/benchdnn/inputs/graph/complex_fusion/harness_mha_ci b/tests/benchdnn/inputs/graph/complex_fusion/harness_mha_ci
index 8b86b687abc..63c0fe01ddf 100644
--- a/tests/benchdnn/inputs/graph/complex_fusion/harness_mha_ci
+++ b/tests/benchdnn/inputs/graph/complex_fusion/harness_mha_ci
@@ -21,5 +21,6 @@
 --reset --expected-n-partitions=0 --case=complex_fusion/mha/MHA-starcoder-inf-int8-bs1.json
 --reset --expected-n-partitions=0 --case=complex_fusion/mha/dynamic_quantized_mha-Bert_large-inf-int8-bs1-fake.json
 --reset --case=complex_fusion/mha/sdpa-plain-wo-scale-int8-bs1.json
---reset --case=complex_fusion/mha/sdpa-compressed-kv-int8-gs128.json
 --reset --case=complex_fusion/mha/sdpa-compressed-v-int8-gs32.json
+--reset --case=complex_fusion/mha/sdpa-compressed-kv-int4-gs32.json
+--reset --dt=0:s8+2:s8+6:s8+8:s8 --case=complex_fusion/mha/sdpa-compressed-kv-int4-gs32.json
diff --git a/tests/benchdnn/inputs/graph/complex_fusion/mha/sdpa-compressed-kv-int8-gs128.json b/tests/benchdnn/inputs/graph/complex_fusion/mha/sdpa-compressed-kv-int8-gs128.json
deleted file mode 100644
index ec71b67c8c9..00000000000
--- a/tests/benchdnn/inputs/graph/complex_fusion/mha/sdpa-compressed-kv-int8-gs128.json
+++ /dev/null
@@ -1,548 +0,0 @@
-{
-  "version": "3.2.0",
-  "engine_kind": "cpu",
-  "fpmath_mode": "f16",
-  "fpmath_mode_apply_to_int": "true",
-  "input_ports": [
-    0,
-    1,
-    2,
-    3,
-    4,
-    5,
-    6,
-    7,
-    8
-  ],
-  "output_ports": [
-    50
-  ],
-  "graph": [
-    {
-      "id": 34107656704,
-      "name": "aten::dequantize",
-      "kind": "DynamicDequantize",
-      "attrs": {
-        "qtype": {
-          "type": "string",
-          "value": "per_group"
-        },
-        "group_shape": {
-          "type": "s64[]",
-          "value": [
-            1,
-            1,
-            128,
-            1
-          ]
-        },
-        "axis": {
-          "type": "s64",
-          "value": 2
-        }
-      },
-      "inputs": [
-        {
-          "id": 0,
-          "dtype": "s8",
-          "shape": [
-            1,
-            32,
-            128,
-            32
-          ],
-          "stride": [
-            131072,
-            4096,
-            1,
-            128
-          ],
-          "layout_type": "strided",
-          "property_type": "variable"
-        },
-        {
-          "id": 1,
-          "dtype": "f16",
-          "shape": [
-            1,
-            32,
-            1,
-            32
-          ],
-          "stride": [
-            4096,
-            128,
-            32,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "undef"
-        },
-        {
-          "id": 2,
-          "dtype": "s8",
-          "shape": [
-            1,
-            32,
-            1,
-            32
-          ],
-          "stride": [
-            4096,
-            128,
-            32,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "undef"
-        }
-      ],
-      "outputs": [
-        {
-          "id": 10,
-          "dtype": "f16",
-          "shape": [
-            1,
-            32,
-            128,
-            32
-          ],
-          "stride": [
-            131072,
-            4096,
-            32,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "variable"
-        }
-      ]
-    },
-    {
-      "id": 34107654464,
-      "name": "aten::matmul",
-      "kind": "MatMul",
-      "attrs": {
-        "transpose_a": {
-          "type": "bool",
-          "value": 0
-        },
-        "transpose_b": {
-          "type": "bool",
-          "value": 0
-        }
-      },
-      "inputs": [
-        {
-          "id": 3,
-          "dtype": "f16",
-          "shape": [
-            1,
-            32,
-            32,
-            128
-          ],
-          "stride": [
-            131072,
-            4096,
-            128,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "variable"
-        },
-        {
-          "id": 10,
-          "dtype": "f16",
-          "shape": [
-            1,
-            32,
-            128,
-            32
-          ],
-          "stride": [
-            131072,
-            4096,
-            32,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "variable"
-        }
-      ],
-      "outputs": [
-        {
-          "id": 15,
-          "dtype": "f16",
-          "shape": [
-            1,
-            32,
-            32,
-            32
-          ],
-          "stride": [
-            32768,
-            1024,
-            32,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "variable"
-        }
-      ]
-    },
-    {
-      "id": 34107661824,
-      "name": "aten::div",
-      "kind": "Divide",
-      "attrs": {
-        "auto_broadcast": {
-          "type": "string",
-          "value": "numpy"
-        }
-      },
-      "inputs": [
-        {
-          "id": 15,
-          "dtype": "f16",
-          "shape": [
-            1,
-            32,
-            32,
-            32
-          ],
-          "stride": [
-            32768,
-            1024,
-            32,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "variable"
-        },
-        {
-          "id": 4,
-          "dtype": "f16",
-          "shape": [],
-          "stride": [],
-          "layout_type": "strided",
-          "property_type": "undef"
-        }
-      ],
-      "outputs": [
-        {
-          "id": 16,
-          "dtype": "f16",
-          "shape": [
-            1,
-            32,
-            32,
-            32
-          ],
-          "stride": [
-            32768,
-            1024,
-            32,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "variable"
-        }
-      ]
-    },
-    {
-      "id": 34106997632,
-      "name": "aten::add",
-      "kind": "Add",
-      "attrs": {
-        "auto_broadcast": {
-          "type": "string",
-          "value": "numpy"
-        }
-      },
-      "inputs": [
-        {
-          "id": 16,
-          "dtype": "f16",
-          "shape": [
-            1,
-            32,
-            32,
-            32
-          ],
-          "stride": [
-            32768,
-            1024,
-            32,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "variable"
-        },
-        {
-          "id": 5,
-          "dtype": "f16",
-          "shape": [
-            1,
-            1,
-            1,
-            32
-          ],
-          "stride": [
-            1024,
-            32,
-            32,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "variable"
-        }
-      ],
-      "outputs": [
-        {
-          "id": 18,
-          "dtype": "f16",
-          "shape": [
-            1,
-            32,
-            32,
-            32
-          ],
-          "stride": [
-            32768,
-            1024,
-            32,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "variable"
-        }
-      ]
-    },
-    {
-      "id": 34426356992,
-      "name": "aten::softmax",
-      "kind": "SoftMax",
-      "attrs": {
-        "axis": {
-          "type": "s64",
-          "value": 3
-        }
-      },
-      "inputs": [
-        {
-          "id": 18,
-          "dtype": "f16",
-          "shape": [
-            1,
-            32,
-            32,
-            32
-          ],
-          "stride": [
-            32768,
-            1024,
-            32,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "variable"
-        }
-      ],
-      "outputs": [
-        {
-          "id": 27,
-          "dtype": "f16",
-          "shape": [
-            1,
-            32,
-            32,
-            32
-          ],
-          "stride": [
-            32768,
-            1024,
-            32,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "variable"
-        }
-      ]
-    },
-    {
-      "id": 34107752448,
-      "name": "aten::dequantize",
-      "kind": "DynamicDequantize",
-      "attrs": {
-        "qtype": {
-          "type": "string",
-          "value": "per_group"
-        },
-        "group_shape": {
-          "type": "s64[]",
-          "value": [
-            1,
-            1,
-            1,
-            128
-          ]
-        }, 
-        "axis": {
-          "type": "s64",
-          "value": 3
-        }
-      },
-      "inputs": [
-        {
-          "id": 6,
-          "dtype": "s8",
-          "shape": [
-            1,
-            32,
-            32,
-            128
-          ],
-          "stride": [
-            131072,
-            4096,
-            128,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "variable"
-        },
-        {
-          "id": 7,
-          "dtype": "f16",
-          "shape": [
-            1,
-            32,
-            32,
-            1
-          ],
-          "stride": [
-            4096,
-            128,
-            1,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "undef"
-        },
-        {
-          "id": 8,
-          "dtype": "s8",
-          "shape": [
-            1,
-            32,
-            32,
-            1
-          ],
-          "stride": [
-            4096,
-            128,
-            1,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "undef"
-        }
-      ],
-      "outputs": [
-        {
-          "id": 45,
-          "dtype": "f16",
-          "shape": [
-            1,
-            32,
-            32,
-            128
-          ],
-          "stride": [
-            131072,
-            4096,
-            128,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "variable"
-        }
-      ]
-    },
-    {
-      "id": 34105676800,
-      "name": "aten::matmul",
-      "kind": "MatMul",
-      "attrs": {
-        "transpose_a": {
-          "type": "bool",
-          "value": 0
-        },
-        "transpose_b": {
-          "type": "bool",
-          "value": 0
-        }
-      },
-      "inputs": [
-        {
-          "id": 27,
-          "dtype": "f16",
-          "shape": [
-            1,
-            32,
-            32,
-            32
-          ],
-          "stride": [
-            32768,
-            1024,
-            32,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "variable"
-        },
-        {
-          "id": 45,
-          "dtype": "f16",
-          "shape": [
-            1,
-            32,
-            32,
-            128
-          ],
-          "stride": [
-            131072,
-            4096,
-            128,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "variable"
-        }
-      ],
-      "outputs": [
-        {
-          "id": 50,
-          "dtype": "f16",
-          "shape": [
-            1,
-            32,
-            32,
-            128
-          ],
-          "stride": [
-            131072,
-            4096,
-            128,
-            1
-          ],
-          "layout_type": "strided",
-          "property_type": "variable"
-        }
-      ]
-    }
-  ]
-}
-