SLAI-AITP commited on
Commit
0860dbf
·
verified ·
1 Parent(s): ef881ce

Add files using upload-large-folder tool

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/CMakeLists.txt +20 -0
  2. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/README.md +99 -0
  3. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/docs/aclnnInterleaveRope.md +421 -0
  4. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/examples/aclnn_interleave_rope.h +58 -0
  5. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/examples/test_aclnn_interleave_rope.cpp +185 -0
  6. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/interleave_rope_collect.py +162 -0
  7. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/interleave_rope_test.py +220 -0
  8. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_graph/CMakeLists.txt +12 -0
  9. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_graph/fusion_pass/.gitkeep +0 -0
  10. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_graph/interleave_rope_proto.h +45 -0
  11. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/CMakeLists.txt +30 -0
  12. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/config/ascend910_93/interleave_rope_binary.json +99 -0
  13. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/config/ascend910_93/interleave_rope_simplified_key.ini +13 -0
  14. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/config/ascend910b/interleave_rope_binary.json +99 -0
  15. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/config/ascend910b/interleave_rope_simplified_key.ini +13 -0
  16. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/config/kirin9030/interleave_rope_binary.json +52 -0
  17. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/config/kirin9030/interleave_rope_simplified_key.ini +13 -0
  18. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/config/kirinx90/interleave_rope_binary.json +52 -0
  19. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/config/kirinx90/interleave_rope_simplified_key.ini +13 -0
  20. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/interleave_rope_def.cpp +93 -0
  21. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/interleave_rope_infershape.cpp +57 -0
  22. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/interleave_rope_tiling.cpp +359 -0
  23. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/interleave_rope_tiling.h +104 -0
  24. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/op_api/aclnn_interleave_rope.cpp +52 -0
  25. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/op_api/aclnn_interleave_rope.h +58 -0
  26. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/op_api/aclnn_rotary_position_embedding.h +58 -0
  27. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_kernel/interleave_rope.cpp +50 -0
  28. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_kernel/interleave_rope_b11d.h +204 -0
  29. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_kernel/interleave_rope_b1sd.h +205 -0
  30. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_kernel/interleave_rope_fixed_bnsd_b11d.h +188 -0
  31. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_kernel/interleave_rope_split_s.h +249 -0
  32. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/test_data/cases_manifest.json +139 -0
  33. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/CMakeLists.txt +17 -0
  34. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/ut/CMakeLists.txt +17 -0
  35. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/ut/op_host/CMakeLists.txt +21 -0
  36. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/ut/op_host/op_api/CMakeLists.txt +15 -0
  37. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/ut/op_host/op_api/test_aclnn_interleave_rope.cpp +48 -0
  38. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/ut/op_host/test_interleave_rope_infershape.cpp +92 -0
  39. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/ut/op_host/test_interleave_rope_tiling.cpp +220 -0
  40. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/ut/op_kernel/CMakeLists.txt +31 -0
  41. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/ut/op_kernel/test_interleave_rope.cpp +148 -0
  42. datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/ut/op_kernel/test_interleave_rope.h +61 -0
  43. datasets/ops_src/ops_transformer/test_kernels/kernels/qkv_rms_norm_rope_cache/CMakeLists.txt +19 -0
  44. datasets/ops_src/ops_transformer/test_kernels/kernels/qkv_rms_norm_rope_cache/README.md +302 -0
  45. datasets/ops_src/ops_transformer/test_kernels/kernels/qkv_rms_norm_rope_cache/docs/aclnnQkvRmsNormRopeCache.md +724 -0
  46. datasets/ops_src/ops_transformer/test_kernels/kernels/qkv_rms_norm_rope_cache/examples/test_aclnn_qkv_rms_norm_rope_cache.cpp +239 -0
  47. datasets/ops_src/ops_transformer/test_kernels/kernels/qkv_rms_norm_rope_cache/op_graph/CMakeLists.txt +11 -0
  48. datasets/ops_src/ops_transformer/test_kernels/kernels/qkv_rms_norm_rope_cache/op_graph/qkv_rms_norm_rope_cache_proto.h +121 -0
  49. datasets/ops_src/ops_transformer/test_kernels/kernels/qkv_rms_norm_rope_cache/op_host/CMakeLists.txt +28 -0
  50. datasets/ops_src/ops_transformer/test_kernels/kernels/qkv_rms_norm_rope_cache/op_host/config/ascend910_93/qkv_rms_norm_rope_cache_binary.json +1949 -0
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/CMakeLists.txt ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # -----------------------------------------------------------------------------------------------------------
2
+ # Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+ # This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+ # CANN Open Software License Agreement Version 2.0 (the "License").
5
+ # Please refer to the License for details. You may not use this file except in compliance with the License.
6
+ # THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+ # INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+ # See LICENSE in the root of the software repository for the full text of the License.
9
+ # -----------------------------------------------------------------------------------------------------------
10
+
11
+ message(STATUS "=== Debug: start ops.posembedding.interleave_rope.CMakeLists.txt ")
12
+ file(GLOB CURRENT_DIRS RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_SOURCE_DIR}/*)
13
+ if(NOT ENABLE_TEST AND NOT BENCHMARK)
14
+ list(REMOVE_ITEM CURRENT_DIRS tests)
15
+ endif()
16
+ foreach(SUB_DIR ${CURRENT_DIRS})
17
+ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/${SUB_DIR}/CMakeLists.txt")
18
+ add_subdirectory(${SUB_DIR})
19
+ endif()
20
+ endforeach()
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/README.md ADDED
@@ -0,0 +1,99 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # InterleaveRope
2
+
3
+ ## 产品支持情况
4
+
5
+ |产品 | 是否支持 |
6
+ |:-------------------------|:----------:|
7
+ | <term>Ascend 950PR/Ascend 950DT</term> | × |
8
+ | <term>Atlas A3 训练系列产品/Atlas A3 推理系列产品</term> | √ |
9
+ | <term>Atlas A2 训练系列产品/Atlas A2 推理系列产品</term> | √ |
10
+ | <term>Kirin X90 处理器系列产品</term> | √ |
11
+ | <term>Kirin 9030 处理器系列产品</term> | √ |
12
+
13
+ ## 功能说明
14
+
15
+ - 算子功能:针对单输入x进行旋转位置编码。
16
+ - 计算公式:
17
+ $$
18
+ q = \text{reshape}(x, [B, N, S, D//2, 2]) \cdot \text{transpose}(-1, -2) \cdot \text{reshape}([B, N, S, D])
19
+ $$
20
+
21
+ $$
22
+ q_{\text{embed}} = q \cdot \text{cos} + \text{RotateHalf}(q) \cdot \sin
23
+ $$
24
+
25
+ 其中:RotateHalf(q)表示将q的D维后半部分元素移至前半部分并乘以 -1,后半部分用前半部分的值。
26
+
27
+ $$
28
+ \text{RotateHalf}(q)_{\text{i}} =
29
+ \begin{cases}
30
+ -q_{i+D//2} & \text{if } i < D//2 \\
31
+ q_{i-D//2} & \text{otherwise}
32
+ \end{cases}
33
+ $$
34
+
35
+ ## 参数说明
36
+
37
+ <table style="undefined;table-layout: fixed; width: 1576px"><colgroup>
38
+ <col style="width: 170px">
39
+ <col style="width: 170px">
40
+ <col style="width: 312px">
41
+ <col style="width: 213px">
42
+ <col style="width: 100px">
43
+ </colgroup>
44
+ <thead>
45
+ <tr>
46
+ <th>参数名</th>
47
+ <th>输入/输出/属性</th>
48
+ <th>描述</th>
49
+ <th>数据类型</th>
50
+ <th>数据格式</th>
51
+ </tr></thead>
52
+ <tbody>
53
+ <tr>
54
+ <td>x</td>
55
+ <td>输入</td>
56
+ <td>表示待处理张量,对应公式中的x。</td>
57
+ <td>FLOAT16、BFLOAT16</td>
58
+ <td>ND</td>
59
+ </tr>
60
+ <tr>
61
+ <td>cos</td>
62
+ <td>输入</td>
63
+ <td>表示RoPE旋转位置的余弦分量,对应公式中的cos。</td>
64
+ <td>FLOAT16、BFLOAT16</td>
65
+ <td>ND</td>
66
+ </tr>
67
+ <tr>
68
+ <td>sin</td>
69
+ <td>输入</td>
70
+ <td>表示RoPE旋转位置的正弦分量,对应公式中的sin。</td>
71
+ <td>FLOAT16、BFLOAT16</td>
72
+ <td>ND</td>
73
+ </tr>
74
+ <tr>
75
+ <td>y</td>
76
+ <td>输出</td>
77
+ <td>表示旋转编码后的结果,对应公式中的q_embed。</td>
78
+ <td>FLOAT16、BFLOAT16</td>
79
+ <td>ND</td>
80
+ </tr>
81
+ </tbody>
82
+ </table>
83
+
84
+ - Kirin X90/Kirin 9030 处理器系列产品: 不支持BFLOAT16。
85
+
86
+ ## 约束说明
87
+
88
+ * 该接口支持推理场景下使用。
89
+ * x,y要求为4 维张量,shape为(B,N,S,D)。
90
+ * cos,sin要求为4 维张量,shape为(B,N,S,D),S可以为1 或与x的S相同,数据类型、数据格式与x一致。
91
+ * 输入x、cos、sin的D维度必须等于64。
92
+ * cos、sin的N维度必须等于1。
93
+ * x、cos、sin、y都不支持非连续的Tensor。
94
+
95
+ ## 调用说明
96
+
97
+ | 调用方式 | 调用样例 | 说明 |
98
+ |----------------|-----------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------|
99
+ | aclnn调用 | [test_aclnn_interleave_rope](./examples/test_aclnn_interleave_rope.cpp) | 通过[aclnnInterleaveRope](./docs/aclnnInterleaveRope.md)接口方式调用InterleaveRope算子。 |
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/docs/aclnnInterleaveRope.md ADDED
@@ -0,0 +1,421 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # aclnnInterleaveRope
2
+
3
+ [📄 查看源码](https://gitcode.com/cann/ops-transformer/tree/master/posembedding/interleave_rope)
4
+
5
+ ## 产品支持情况
6
+
7
+ | 产品 | 是否支持 |
8
+ | :----------------------------------------------------------- | :------: |
9
+ | <term>Ascend 950PR/Ascend 950DT</term> | √ |
10
+ | <term>Atlas A3 训练系列产品/Atlas A3 推理系列产品</term> | × |
11
+ | <term>Atlas A2 训练系列产品/Atlas A2 推理系列产品</term> | × |
12
+ | <term>Atlas 200I/500 A2 推理产品</term> | × |
13
+ | <term>Atlas 推理系列产品</term> | × |
14
+ | <term>Atlas 训练系列产品</term> | × |
15
+
16
+ ## 功能说明
17
+
18
+ - 接口功能:针对单输入x进行旋转位置编码。
19
+
20
+ - 计算公式:
21
+
22
+ $$
23
+ q = \text{reshape}(x, [B, N, S, D//2, 2]) \cdot \text{transpose}(-1, -2) \cdot \text{reshape}([B, N, S, D])
24
+ $$
25
+
26
+ $$
27
+ q_{\text{embed}} = q \cdot \text{cos} + \text{RotateHalf}(q) \cdot \sin
28
+ $$
29
+
30
+ 其中:RotateHalf(q)表示将q的D维后半部分元素移至前半部分并乘以 -1,后半部分用前半部分的值。
31
+
32
+ $$
33
+ \text{RotateHalf}(q)_{\text{i}} =
34
+ \begin{cases}
35
+ -q_{i+D//2} & \text{if } i < D//2 \\
36
+ q_{i-D//2} & \text{otherwise}
37
+ \end{cases}
38
+ $$
39
+
40
+ ## 函数原型
41
+
42
+ 该算子采用两段式接口,必须先调用"aclnnInterleaveRopeGetWorkspaceSize"接口获取入参并根据流程计算所需workspace大小,再调用"aclnnInterleaveRope"接口执行计算。
43
+
44
+ ```c++
45
+ aclnnStatus aclnnInterleaveRopeGetWorkspaceSize(
46
+ const aclTensor *x,
47
+ const aclTensor *cos,
48
+ const aclTensor *sin,
49
+ aclTensor *out,
50
+ uint64_t *workspaceSize,
51
+ aclOpExecutor **executor)
52
+ ```
53
+
54
+ ```c++
55
+ aclnnStatus aclnnInterleaveRope(
56
+ void *workspace,
57
+ uint64_t workspaceSize,
58
+ aclOpExecutor *executor,
59
+ aclrtStream stream)
60
+ ```
61
+
62
+ ## aclnnInterleaveRopeGetWorkspaceSize
63
+
64
+ - **参数说明**
65
+
66
+ <table style="undefined;table-layout: fixed; width: 1461px"><colgroup>
67
+ <col style="width: 162px">
68
+ <col style="width: 121px">
69
+ <col style="width: 332px">
70
+ <col style="width: 169px">
71
+ <col style="width: 275px">
72
+ <col style="width: 118px">
73
+ <col style="width: 138px">
74
+ <col style="width: 146px">
75
+ </colgroup>
76
+ <thead>
77
+ <tr>
78
+ <th>参数名</th>
79
+ <th>输入/输出</th>
80
+ <th>描述</th>
81
+ <th>使用说明</th>
82
+ <th>数据类型</th>
83
+ <th>数据格式</th>
84
+ <th>维度(shape)</th>
85
+ <th>非连续Tensor</th>
86
+ </tr>
87
+ </thead>
88
+ <tbody>
89
+ <tr>
90
+ <td>x</td>
91
+ <td>输入</td>
92
+ <td>device侧的aclTensor,待执行旋转位置编码的张量,对应公式中的x。</td>
93
+ <td>-</td>
94
+ <td>FLOAT32、FLOAT16、BFLOAT16</td>
95
+ <td>ND</td>
96
+ <td>4</td>
97
+ <td>√</td>
98
+ </tr>
99
+ <tr>
100
+ <td>cos</td>
101
+ <td>输入</td>
102
+ <td>device侧的aclTensor,位置编码张量,对应公式中的cos。</td>
103
+ <td>shape与x满足broadcast关系。</td>
104
+ <td>与x一致</td>
105
+ <td>ND</td>
106
+ <td>4</td>
107
+ <td>√</td>
108
+ </tr>
109
+ <tr>
110
+ <td>sin</td>
111
+ <td>输入</td>
112
+ <td>device侧的aclTensor,位置编码张量,对应公式中的sin。</td>
113
+ <td>shape与cos一致。</td>
114
+ <td>与x一致</td>
115
+ <td>ND</td>
116
+ <td>4</td>
117
+ <td>√</td>
118
+ </tr>
119
+ <tr>
120
+ <td>out</td>
121
+ <td>输出</td>
122
+ <td>device侧的aclTensor,旋转位置编码计算结果,对应公式中的y。</td>
123
+ <td>shape与x一致。</td>
124
+ <td>与x一致</td>
125
+ <td>ND</td>
126
+ <td>4</td>
127
+ <td>x</td>
128
+ </tr>
129
+ <tr>
130
+ <td>workspaceSize</td>
131
+ <td>输出</td>
132
+ <td>返回用户需要在npu device侧申请的workspace大小。</td>
133
+ <td>-</td>
134
+ <td>-</td>
135
+ <td>-</td>
136
+ <td>-</td>
137
+ <td>-</td>
138
+ </tr>
139
+ <tr>
140
+ <td>executor</td>
141
+ <td>输出</td>
142
+ <td>返回op执行器,包含算子计算流程。</td>
143
+ <td>-</td>
144
+ <td>-</td>
145
+ <td>-</td>
146
+ <td>-</td>
147
+ <td>-</td>
148
+ </tr>
149
+ </tbody>
150
+ </table>
151
+
152
+ - **返回值**
153
+
154
+ aclnnStatus:返回算子的执行状态码。
155
+
156
+ 第一段接口完成入参校验,出现以下场景时报错:
157
+
158
+ <table style="undefined;table-layout: fixed; width: 1155px"><colgroup>
159
+ <col style="width: 288px">
160
+ <col style="width: 125px">
161
+ <col style="width: 742px">
162
+ </colgroup>
163
+ <thead>
164
+ <tr>
165
+ <th>返回码</th>
166
+ <th>错误码</th>
167
+ <th>描述</th>
168
+ </tr>
169
+ </thead>
170
+ <tbody>
171
+ <tr>
172
+ <td>ACLNN_ERR_PARAM_NULLPTR</td>
173
+ <td>161001</td>
174
+ <td>传入的x、cos、sin或out是空指��。</td>
175
+ </tr>
176
+ <tr>
177
+ <td>ACLNN_ERR_PARAM_INVALID</td>
178
+ <td>161002</td>
179
+ <td>传入的x、cos、sin、out的数据类型和格式不在支持的范围内。</td>
180
+ </tr>
181
+ <tr>
182
+ <td>ACLNN_ERR_INNER_TILING_ERROR</td>
183
+ <td>561002</td>
184
+ <td>传入的x、cos、sin、out的shape不匹配。</td>
185
+ </tr>
186
+ </tbody>
187
+ </table>
188
+
189
+ ## aclnnInterleaveRope
190
+
191
+ - **参数说明**
192
+
193
+ <table style="undefined;table-layout: fixed; width: 1155px"><colgroup>
194
+ <col style="width: 173px">
195
+ <col style="width: 133px">
196
+ <col style="width: 849px">
197
+ </colgroup>
198
+ <thead>
199
+ <tr>
200
+ <th>参数名</th>
201
+ <th>输入/输出</th>
202
+ <th>描述</th>
203
+ </tr></thead>
204
+ <tbody>
205
+ <tr>
206
+ <td>workspace</td>
207
+ <td>输入</td>
208
+ <td>在Device侧申请的workspace内存地址。</td>
209
+ </tr>
210
+ <tr>
211
+ <td>workspaceSize</td>
212
+ <td>输入</td>
213
+ <td>在Device侧申请的workspace大小,由第一段接口aclnnInterleaveRopeGetWorkspaceSize获取。</td>
214
+ </tr>
215
+ <tr>
216
+ <td>executor</td>
217
+ <td>输入</td>
218
+ <td>op执行器,包含了算子计算流程。</td>
219
+ </tr>
220
+ <tr>
221
+ <td>stream</td>
222
+ <td>输入</td>
223
+ <td>指定执行任务的Stream流。</td>
224
+ </tr>
225
+ </tbody>
226
+ </table>
227
+
228
+ - **返回值**
229
+
230
+ aclnnStatus:返回算子的执行状态码。
231
+
232
+ ## 约束说明
233
+
234
+ - 确定性计算:
235
+ - aclnnInterleaveRope默认确定性实现。
236
+ - 该接口支持推理场景下使用。
237
+ - x、cos、sin、out要求为4维张量,shape为(B,N,S,D)。
238
+ - cos、sin的S维度可以为1或与x的S维度相同,N维度必须等于1。
239
+ - 输入x、cos、sin的D维度必须等于64。
240
+
241
+ ## 调用示例
242
+
243
+ 以下示例代码展示接口调用流程,可供编译和运行时参考。
244
+
245
+ ```Cpp
246
+ #include "acl/acl.h"
247
+ #include "aclnnop/aclnn_interleave_rope.h"
248
+ #include <iostream>
249
+ #include <vector>
250
+
251
+ #define CHECK_RET(cond, return_expr) \
252
+ do { \
253
+ if (!(cond)) { \
254
+ return_expr; \
255
+ } \
256
+ } while (0)
257
+
258
+ #define LOG_PRINT(message, ...) \
259
+ do { \
260
+ printf(message, ##__VA_ARGS__); \
261
+ } while (0)
262
+
263
+ int64_t GetShapeSize(const std::vector<int64_t>& shape) {
264
+ int64_t shape_size = 1;
265
+ for (auto i : shape) {
266
+ shape_size *= i;
267
+ }
268
+ return shape_size;
269
+ }
270
+
271
+ std::vector<aclFloat16> ConvertToFloat16(const std::vector<float>& data) {
272
+ std::vector<aclFloat16> converted;
273
+ converted.reserve(data.size());
274
+ for (float value : data) {
275
+ converted.push_back(aclFloatToFloat16(value));
276
+ }
277
+ return converted;
278
+ }
279
+
280
+ int Init(int32_t deviceId, aclrtStream* stream) {
281
+ // 固定写法,资源初始化
282
+ auto ret = aclInit(nullptr);
283
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclInit failed. ERROR: %d\n", ret); return ret);
284
+ ret = aclrtSetDevice(deviceId);
285
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret); return ret);
286
+ ret = aclrtCreateStream(stream);
287
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateStream failed. ERROR: %d\n", ret); return ret);
288
+ return 0;
289
+ }
290
+
291
+ template <typename T>
292
+ int CreateAclTensor(const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr,
293
+ aclDataType dataType, aclTensor** tensor) {
294
+ auto size = GetShapeSize(shape) * sizeof(T);
295
+ // 调用aclrtMalloc申请device侧内存
296
+ auto ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
297
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMalloc failed. ERROR: %d\n", ret); return ret);
298
+
299
+ // 调用aclrtMemcpy将host侧数据拷贝到device侧内存上
300
+ ret = aclrtMemcpy(*deviceAddr, size, hostData.data(), size, ACL_MEMCPY_HOST_TO_DEVICE);
301
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMemcpy failed. ERROR: %d\n", ret); return ret);
302
+
303
+ // 计算连续tensor的strides
304
+ std::vector<int64_t> strides(shape.size(), 1);
305
+ for (int64_t i = shape.size() - 2; i >= 0; i--) {
306
+ strides[i] = shape[i + 1] * strides[i + 1];
307
+ }
308
+
309
+ // 调用aclCreateTensor接口创建aclTensor
310
+ *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND,
311
+ shape.data(), shape.size(), *deviceAddr);
312
+ return 0;
313
+ }
314
+
315
+ int main() {
316
+ // 1. 固定写法,device/stream初始化,参考acl API手册
317
+ // 根据自己的实际device填写deviceId
318
+ int32_t deviceId = 0;
319
+ aclrtStream stream;
320
+ auto ret = Init(deviceId, &stream);
321
+ // check根据自己的需要处理
322
+ CHECK_RET(ret == 0, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);
323
+ // 2. 构造输入与输出,需要根据API的接口定义构造
324
+ // interleave_rope约束:D=64, N=1
325
+ std::vector<int64_t> xShape = {1, 1, 1, 64};
326
+ std::vector<int64_t> cosShape = {1, 1, 1, 64};
327
+ std::vector<int64_t> sinShape = {1, 1, 1, 64};
328
+ std::vector<int64_t> outShape = {1, 1, 1, 64};
329
+
330
+ void* xDeviceAddr = nullptr;
331
+ void* cosDeviceAddr = nullptr;
332
+ void* sinDeviceAddr = nullptr;
333
+ void* outDeviceAddr = nullptr;
334
+ aclTensor* x = nullptr;
335
+ aclTensor* cos = nullptr;
336
+ aclTensor* sin = nullptr;
337
+ aclTensor* out = nullptr;
338
+
339
+ std::vector<aclFloat16> xHostData = ConvertToFloat16({
340
+ 74, 54, 84, 125, 23, 78, 37, 72, 27, 98, 34, 107, 29, 23, 54, 60,
341
+ 70, 49, 119, 54, 29, 54, 41, 99, 27, 62, 5, 46, 108, 39, 24, 123,
342
+ 33, 82, 6, 40, 88, 24, 6, 116, 38, 119, 110, 5, 30, 79, 87, 18,
343
+ 29, 100, 90, 24, 21, 93, 63, 68, 34, 112, 119, 48, 74, 43, 85, 64
344
+ });
345
+ std::vector<aclFloat16> cosHostData = ConvertToFloat16({
346
+ 41, 37, 17, 25, 49, 25, 22, 24, 110, 120, 107, 3, 82, 66, 75, 86,
347
+ 85, 115, 110, 56, 52, 39, 86, 23, 36, 71, 20, 73, 113, 25, 114, 56,
348
+ 125, 80, 95, 82, 31, 63, 99, 62, 23, 55, 30, 99, 42, 121, 15, 24,
349
+ 97, 87, 81, 67, 43, 21, 13, 9, 33, 29, 117, 10, 114, 61, 98, 15
350
+ });
351
+ std::vector<aclFloat16> sinHostData = ConvertToFloat16({
352
+ 46, 56, 56, 101, 66, 10, 96, 16, 86, 57, 102, 66, 12, 105, 76, 58,
353
+ 90, 6, 79, 128, 126, 82, 41, 3, 45, 7, 66, 4, 46, 22, 31, 26,
354
+ 37, 63, 97, 84, 91, 90, 47, 77, 90, 34, 41, 83, 91, 108, 120, 13,
355
+ 90, 32, 85, 37, 119, 31, 51, 82, 122, 125, 7, 116, 121, 108, 38, 56
356
+ });
357
+ std::vector<aclFloat16> outHostData(64, aclFloatToFloat16(0.0f));
358
+
359
+ // 创建x aclTensor
360
+ ret = CreateAclTensor(xHostData, xShape, &xDeviceAddr, aclDataType::ACL_FLOAT16, &x);
361
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
362
+ // 创建cos aclTensor
363
+ ret = CreateAclTensor(cosHostData, cosShape, &cosDeviceAddr, aclDataType::ACL_FLOAT16, &cos);
364
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
365
+ // 创建sin aclTensor
366
+ ret = CreateAclTensor(sinHostData, sinShape, &sinDeviceAddr, aclDataType::ACL_FLOAT16, &sin);
367
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
368
+ // 创建out aclTensor
369
+ ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, aclDataType::ACL_FLOAT16, &out);
370
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
371
+
372
+ // 3. 调用CANN算子库API,需要修改为具体的API
373
+ uint64_t workspaceSize = 0;
374
+ aclOpExecutor* executor;
375
+ // 调用aclnnInterleaveRope第一段接口
376
+ ret = aclnnInterleaveRopeGetWorkspaceSize(x, cos, sin, out, &workspaceSize, &executor);
377
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnInterleaveRopeGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
378
+ // 根据第一段接口计算出的workspaceSize申请device内存
379
+ void* workspaceAddr = nullptr;
380
+ if (workspaceSize > 0) {
381
+ ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
382
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("allocate workspace failed. ERROR: %d\n", ret); return ret);
383
+ }
384
+ // 调用aclnnInterleaveRope第二段接口
385
+ ret = aclnnInterleaveRope(workspaceAddr, workspaceSize, executor, stream);
386
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnInterleaveRope failed. ERROR: %d\n", ret); return ret);
387
+ // 4. 固定写法,同步等待任务执行结束
388
+ ret = aclrtSynchronizeStream(stream);
389
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);
390
+ // 5. 获取输出的值,将device侧内存上的结果拷贝至host侧,需要根据具体API的接口定义修改
391
+ auto size = GetShapeSize(outShape);
392
+ std::vector<aclFloat16> resultData(size, aclFloatToFloat16(0.0f));
393
+ ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]), outDeviceAddr,
394
+ size * sizeof(resultData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
395
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return ret);
396
+
397
+ for (int64_t i = 0; i < size; i++) {
398
+ LOG_PRINT("result[%ld] is: %f\n", i, aclFloat16ToFloat(resultData[i]));
399
+ }
400
+
401
+ // 6. 释放aclTensor和aclScalar,需要根据具体API的接口定义修改
402
+ aclDestroyTensor(x);
403
+ aclDestroyTensor(cos);
404
+ aclDestroyTensor(sin);
405
+ aclDestroyTensor(out);
406
+
407
+ // 7. 释放device资源
408
+ aclrtFree(xDeviceAddr);
409
+ aclrtFree(cosDeviceAddr);
410
+ aclrtFree(sinDeviceAddr);
411
+ aclrtFree(outDeviceAddr);
412
+ if (workspaceSize > 0) {
413
+ aclrtFree(workspaceAddr);
414
+ }
415
+ aclrtDestroyStream(stream);
416
+ aclrtResetDevice(deviceId);
417
+ aclFinalize();
418
+
419
+ return 0;
420
+ }
421
+ ```
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/examples/aclnn_interleave_rope.h ADDED
@@ -0,0 +1,58 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ #ifndef OP_API_INC_INTERLEAVE_ROPE_H_
12
+ #define OP_API_INC_INTERLEAVE_ROPE_H_
13
+
14
+ #include "aclnn/aclnn_base.h"
15
+ #include "aclnn_util.h"
16
+
17
+ #ifdef __cplusplus
18
+ extern "C" {
19
+ #endif
20
+
21
+ /**
22
+ * @brief aclnnInterleaveRope的第一段接口,根据具体的计算流程,计算workspace大小。
23
+ * @domain aclnn_ops_infer
24
+ *
25
+ * @param [in] x: npu
26
+ * device侧的aclTensor, 数据类型支持FLOAT32,FLOAT16,BFLOAT16,支持非连续的Tensor,数据格式支持ND。
27
+ * @param [in] sin: npu
28
+ * device侧的aclTensor, 数据类型支持FLOAT32,FLOAT16,BFLOAT16,数据类型与x一致,shape与x满足broadcast关系
29
+ * 支持非连续的Tensor,数据格式支持ND。
30
+ * @param [in] cos: npu
31
+ * device侧的aclTensor, 数据类型支持FLOAT32,FLOAT16,BFLOAT16,数据类型与x一致,shape与sin一致
32
+ * 支持非连续的Tensor,数据格式支持ND。
33
+ * @param [in] out: npu
34
+ * device侧的aclTensor, 数据类型支持FLOAT32,FLOAT16,BFLOAT16,且数据类型与x一致,
35
+ * shape与x相同, 数据格式支持ND, 且数据格式需要与x一致。
36
+ * @param [out] workspaceSize: 返回用户需要在npu device侧申请的workspace大小.
37
+ * @param [out] executor: 返回op执行器,包含算子计算流程。
38
+ * @return aclnnStatus: 返回状态码。
39
+ */
40
+ ACLNN_API aclnnStatus aclnnInterleaveRopeGetWorkspaceSize(
41
+ const aclTensor* x, const aclTensor* cos, const aclTensor* sin, aclTensor* out, uint64_t* workspaceSize,
42
+ aclOpExecutor** executor);
43
+
44
+ /* @brief aclnnInterleaveRope的第二段接口,用于执行计算.
45
+ * @param [in] workspace: 在npu device侧申请的workspace内存起址
46
+ * @param [in] workspace_size: 在npu device侧申请的workspace大小,由第一段接口aclnnBitwiseNotGetWorkspaceSize获取.
47
+ * @param [in] executor: op执行器,包含了算子计算流程.
48
+ * @param [in] stream: acl stream流.
49
+ * @return aclnnStatus: 返回状态码。
50
+ */
51
+ ACLNN_API aclnnStatus
52
+ aclnnInterleaveRope(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, aclrtStream stream);
53
+
54
+ #ifdef __cplusplus
55
+ }
56
+ #endif
57
+
58
+ #endif
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/examples/test_aclnn_interleave_rope.cpp ADDED
@@ -0,0 +1,185 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3
+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+ * CANN Open Software License Agreement Version 2.0 (the "License").
5
+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+ * See LICENSE in the root of the software repository for the full text of the License.
9
+ */
10
+
11
+ #include "acl/acl.h"
12
+ #include "aclnnop/aclnn_interleave_rope.h"
13
+ #include <iostream>
14
+ #include <vector>
15
+
16
+ #define CHECK_RET(cond, return_expr) \
17
+ do { \
18
+ if (!(cond)) { \
19
+ return_expr; \
20
+ } \
21
+ } while (0)
22
+
23
+ #define LOG_PRINT(message, ...) \
24
+ do { \
25
+ printf(message, ##__VA_ARGS__); \
26
+ } while (0)
27
+
28
+ int64_t GetShapeSize(const std::vector<int64_t>& shape) {
29
+ int64_t shape_size = 1;
30
+ for (auto i : shape) {
31
+ shape_size *= i;
32
+ }
33
+ return shape_size;
34
+ }
35
+
36
+ std::vector<aclFloat16> ConvertToFloat16(const std::vector<float>& data) {
37
+ std::vector<aclFloat16> converted;
38
+ converted.reserve(data.size());
39
+ for (float value : data) {
40
+ converted.push_back(aclFloatToFloat16(value));
41
+ }
42
+ return converted;
43
+ }
44
+
45
+ int Init(int32_t deviceId, aclrtStream* stream) {
46
+ // 固定写法,资源初始化
47
+ auto ret = aclInit(nullptr);
48
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclInit failed. ERROR: %d\n", ret); return ret);
49
+ ret = aclrtSetDevice(deviceId);
50
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret); return ret);
51
+ ret = aclrtCreateStream(stream);
52
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateStream failed. ERROR: %d\n", ret); return ret);
53
+ return 0;
54
+ }
55
+
56
+ template <typename T>
57
+ int CreateAclTensor(const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr,
58
+ aclDataType dataType, aclTensor** tensor) {
59
+ auto size = GetShapeSize(shape) * sizeof(T);
60
+ // 调用aclrtMalloc申请device侧内存
61
+ auto ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
62
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMalloc failed. ERROR: %d\n", ret); return ret);
63
+
64
+ // 调用aclrtMemcpy将host侧数据拷贝到device侧内存上
65
+ ret = aclrtMemcpy(*deviceAddr, size, hostData.data(), size, ACL_MEMCPY_HOST_TO_DEVICE);
66
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMemcpy failed. ERROR: %d\n", ret); return ret);
67
+
68
+ // 计算连续tensor的strides
69
+ std::vector<int64_t> strides(shape.size(), 1);
70
+ for (int64_t i = shape.size() - 2; i >= 0; i--) {
71
+ strides[i] = shape[i + 1] * strides[i + 1];
72
+ }
73
+
74
+ // 调用aclCreateTensor接口创建aclTensor
75
+ *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND,
76
+ shape.data(), shape.size(), *deviceAddr);
77
+ return 0;
78
+ }
79
+
80
+ int main() {
81
+ // 1. 固定写法,device/stream初始化, 参考AscendCL对外接口列表
82
+ // 根据自己的实际device填写deviceId
83
+ int32_t deviceId = 0;
84
+ aclrtStream stream;
85
+ auto ret = Init(deviceId, &stream);
86
+ // check根据自己的需要处理
87
+ CHECK_RET(ret == 0, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);
88
+ // 2. 构造输入与输出,需要根据API的接口定义构造
89
+ // interleave_rope约束:D=64, N=1
90
+ std::vector<int64_t> xShape = {1, 1, 1, 64};
91
+ std::vector<int64_t> cosShape = {1, 1, 1, 64};
92
+ std::vector<int64_t> sinShape = {1, 1, 1, 64};
93
+ std::vector<int64_t> outShape = {1, 1, 1, 64};
94
+
95
+ void* xDeviceAddr = nullptr;
96
+ void* cosDeviceAddr = nullptr;
97
+ void* sinDeviceAddr = nullptr;
98
+ void* outDeviceAddr = nullptr;
99
+ aclTensor* x = nullptr;
100
+ aclTensor* cos = nullptr;
101
+ aclTensor* sin = nullptr;
102
+ aclTensor* out = nullptr;
103
+
104
+ std::vector<aclFloat16> xHostData = ConvertToFloat16({
105
+ 74, 54, 84, 125, 23, 78, 37, 72, 27, 98, 34, 107, 29, 23, 54, 60,
106
+ 70, 49, 119, 54, 29, 54, 41, 99, 27, 62, 5, 46, 108, 39, 24, 123,
107
+ 33, 82, 6, 40, 88, 24, 6, 116, 38, 119, 110, 5, 30, 79, 87, 18,
108
+ 29, 100, 90, 24, 21, 93, 63, 68, 34, 112, 119, 48, 74, 43, 85, 64
109
+ });
110
+ std::vector<aclFloat16> cosHostData = ConvertToFloat16({
111
+ 41, 37, 17, 25, 49, 25, 22, 24, 110, 120, 107, 3, 82, 66, 75, 86,
112
+ 85, 115, 110, 56, 52, 39, 86, 23, 36, 71, 20, 73, 113, 25, 114, 56,
113
+ 125, 80, 95, 82, 31, 63, 99, 62, 23, 55, 30, 99, 42, 121, 15, 24,
114
+ 97, 87, 81, 67, 43, 21, 13, 9, 33, 29, 117, 10, 114, 61, 98, 15
115
+ });
116
+ std::vector<aclFloat16> sinHostData = ConvertToFloat16({
117
+ 46, 56, 56, 101, 66, 10, 96, 16, 86, 57, 102, 66, 12, 105, 76, 58,
118
+ 90, 6, 79, 128, 126, 82, 41, 3, 45, 7, 66, 4, 46, 22, 31, 26,
119
+ 37, 63, 97, 84, 91, 90, 47, 77, 90, 34, 41, 83, 91, 108, 120, 13,
120
+ 90, 32, 85, 37, 119, 31, 51, 82, 122, 125, 7, 116, 121, 108, 38, 56
121
+ });
122
+ std::vector<aclFloat16> outHostData(64, aclFloatToFloat16(0.0f));
123
+
124
+ // 创建x aclTensor
125
+ ret = CreateAclTensor(xHostData, xShape, &xDeviceAddr, aclDataType::ACL_FLOAT16, &x);
126
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
127
+ // 创建cos aclTensor
128
+ ret = CreateAclTensor(cosHostData, cosShape, &cosDeviceAddr, aclDataType::ACL_FLOAT16, &cos);
129
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
130
+ // 创建sin aclTensor
131
+ ret = CreateAclTensor(sinHostData, sinShape, &sinDeviceAddr, aclDataType::ACL_FLOAT16, &sin);
132
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
133
+ // 创建out aclTensor
134
+ ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, aclDataType::ACL_FLOAT16, &out);
135
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
136
+
137
+ // 3. 调用CANN算子库API,需要修改为具体的API
138
+ uint64_t workspaceSize = 0;
139
+ aclOpExecutor* executor;
140
+ // 调用aclnnInterleaveRope第一段接口
141
+ ret = aclnnInterleaveRopeGetWorkspaceSize(x, cos, sin, out, &workspaceSize, &executor);
142
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnInterleaveRopeGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
143
+ // 根据第一段接口计算出的workspaceSize申请device内存
144
+ void* workspaceAddr = nullptr;
145
+ if (workspaceSize > 0) {
146
+ ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
147
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("allocate workspace failed. ERROR: %d\n", ret); return ret);
148
+ }
149
+ // 调用aclnnInterleaveRope第二段接口
150
+ ret = aclnnInterleaveRope(workspaceAddr, workspaceSize, executor, stream);
151
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnInterleaveRope failed. ERROR: %d\n", ret); return ret);
152
+ // 4. 固定写法,同步等待任务执行结束
153
+ ret = aclrtSynchronizeStream(stream);
154
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);
155
+ // 5. 获取输出的值,将device侧内存上的结果拷贝至host侧,需要根据具体API的接口定义修改
156
+ auto size = GetShapeSize(outShape);
157
+ std::vector<aclFloat16> resultData(size, aclFloatToFloat16(0.0f));
158
+ ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]), outDeviceAddr,
159
+ size * sizeof(resultData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
160
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return ret);
161
+
162
+ for (int64_t i = 0; i < size; i++) {
163
+ LOG_PRINT("result[%ld] is: %f\n", i, aclFloat16ToFloat(resultData[i]));
164
+ }
165
+
166
+ // 6. 释放aclTensor和aclScalar,需要根据具体API的接口定义修改
167
+ aclDestroyTensor(x);
168
+ aclDestroyTensor(cos);
169
+ aclDestroyTensor(sin);
170
+ aclDestroyTensor(out);
171
+
172
+ // 7. 释放device 资源
173
+ aclrtFree(xDeviceAddr);
174
+ aclrtFree(cosDeviceAddr);
175
+ aclrtFree(sinDeviceAddr);
176
+ aclrtFree(outDeviceAddr);
177
+ if (workspaceSize > 0) {
178
+ aclrtFree(workspaceAddr);
179
+ }
180
+ aclrtDestroyStream(stream);
181
+ aclrtResetDevice(deviceId);
182
+ aclFinalize();
183
+
184
+ return 0;
185
+ }
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/interleave_rope_collect.py ADDED
@@ -0,0 +1,162 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ """Collect golden data for interleave_rope by running the NPU operator on every case.
3
+
4
+ Re-run ONLY after an intentional kernel change — this overwrites the regression baseline.
5
+
6
+ Shape constraints from tiling code:
7
+ - x: (B, N, S, 64) — N>=1, S>=1, B>=1
8
+ - cos/sin: (B, 1, cosS, 64) — cosB MUST equal x's B, cosN=1, cosS=1 or S
9
+ - D must be exactly 64
10
+
11
+ Tiling branches (from DoOpTiling):
12
+ 1. BN1D: seq >= 512 AND cosS == seq (seq-per-core split)
13
+ 2. FIXED_BNSD: B==32, N==32, S==1, cosN==1, cosS==1 (batch-per-block=4)
14
+ 3. B11D: cosS==1 and (B >= coreNum OR B >= N*S) (batch split)
15
+ 4. B11D: cosS==1 and B < coreNum AND B < N*S (NS split)
16
+ 5. B1SD: cosS != 1 and cosS != seq (batch split) — NOT reachable due to shape check
17
+
18
+ All shapes are multiples of 32 for alignment.
19
+ """
20
+ import json, os, sys
21
+ import torch
22
+ import torch_npu # noqa: F401
23
+ import gtsrusops as ops
24
+
25
+ OP_NAME = "interleave_rope"
26
+ OP_CAMEL = "InterleaveRope"
27
+ DEVICE = "npu:0"
28
+ SEED = 2026
29
+
30
+ # Positional order expected by ops.npu_interleave_rope(...), excluding output buffers.
31
+ PARAM_NAMES = ["x", "cos", "sin"]
32
+ OUTPUT_KEYS = ["y"]
33
+
34
+ # 10 cases spanning dtypes / sizes / cos/sin broadcast patterns
35
+ # cos/sin batch MUST equal x batch (enforced in tiling code: cosB_ != batchSize_)
36
+ # cos/sin N must be 1 (enforced in tiling code: cosN_ != 1)
37
+ # cosS must be 1 or equal to seq (enforced in tiling code)
38
+ # D must be 64
39
+ CASES = [
40
+ {"id": 0, "dtype": torch.float16, "batch": 32, "seq": 32, "cos_s": 1}, # B11D, batch split
41
+ {"id": 1, "dtype": torch.float16, "batch": 32, "seq": 64, "cos_s": 1}, # B11D, batch split
42
+ {"id": 2, "dtype": torch.float16, "batch": 32, "seq": 128, "cos_s": 1}, # B11D, batch split
43
+ {"id": 3, "dtype": torch.bfloat16, "batch": 64, "seq": 32, "cos_s": 1}, # B11D, batch split
44
+ {"id": 4, "dtype": torch.bfloat16, "batch": 64, "seq": 64, "cos_s": 1}, # B11D, batch split
45
+ {"id": 5, "dtype": torch.float16, "batch": 128,"seq": 32, "cos_s": 1}, # B11D, batch split
46
+ {"id": 6, "dtype": torch.float16, "batch": 32, "seq": 512, "cos_s": 512}, # BN1D (seq >= 512)
47
+ {"id": 7, "dtype": torch.bfloat16, "batch": 128,"seq": 64, "cos_s": 1}, # B11D, batch split
48
+ {"id": 8, "dtype": torch.float16, "batch": 64, "seq": 1024, "cos_s": 1024},# BN1D (seq >= 512)
49
+ {"id": 9, "dtype": torch.bfloat16, "batch": 32, "seq": 1, "cos_s": 1}, # B11D, batch split
50
+ ]
51
+
52
+ DATA_ROOT = "test_data"
53
+ INPUT_DIR = os.path.join(DATA_ROOT, "golden_input", OP_CAMEL)
54
+ RESULT_DIR = os.path.join(DATA_ROOT, "golden_result", OP_CAMEL)
55
+
56
+
57
+ def make_inputs(case):
58
+ """Build every entry of PARAM_NAMES for this case.
59
+
60
+ Key constraint from tiling code:
61
+ cosB_ must equal batchSize_ (cos/sin batch == x batch)
62
+ cosN_ must be 1
63
+ cosS_ must be 1 or seqLength_
64
+ cosD_ must equal hiddenDim_ (64)
65
+ """
66
+ torch.manual_seed(SEED + case["id"])
67
+ dtype = case["dtype"]
68
+ batch, seq = case["batch"], case["seq"]
69
+ cos_s = case["cos_s"]
70
+ # Shape is always (B, N, S, D) with N=1, D=64
71
+ # cos/sin batch MUST equal x batch (not 1!)
72
+ x_shape = (batch, 1, seq, 64)
73
+ cos_shape = (batch, 1, cos_s, 64)
74
+ sin_shape = (batch, 1, cos_s, 64)
75
+ return {
76
+ "x": torch.randn(x_shape, dtype=dtype, device=DEVICE),
77
+ "cos": torch.randn(cos_shape, dtype=dtype, device=DEVICE),
78
+ "sin": torch.randn(sin_shape, dtype=dtype, device=DEVICE),
79
+ }
80
+
81
+
82
+ def make_output_bufs(case):
83
+ """(shape, dtype) per output, in OUTPUT_KEYS order."""
84
+ batch, seq = case["batch"], case["seq"]
85
+ dtype = case["dtype"]
86
+ return [torch.empty((batch, 1, seq, 64), dtype=dtype, device=DEVICE)]
87
+
88
+
89
+ def build_call_args(case):
90
+ inputs = make_inputs(case)
91
+ positional = [inputs[p] for p in PARAM_NAMES]
92
+ return positional, inputs
93
+
94
+
95
+ def collect_case(case):
96
+ call_args, inputs = build_call_args(case)
97
+
98
+ out_bufs = make_output_bufs(case)
99
+ out_arg = out_bufs[0] if len(out_bufs) == 1 else out_bufs
100
+
101
+ outputs = ops.npu_interleave_rope(*call_args, out_arg)
102
+ torch.npu.synchronize()
103
+
104
+ outputs = [outputs] if isinstance(outputs, torch.Tensor) else list(outputs)
105
+ if len(outputs) != len(OUTPUT_KEYS):
106
+ raise RuntimeError(f"case_{case['id']}: got {len(outputs)} outputs, expected {OUTPUT_KEYS}")
107
+
108
+ saved_out = {k: outputs[i].cpu() for i, k in enumerate(OUTPUT_KEYS)}
109
+ for k, t in saved_out.items():
110
+ if not torch.isfinite(t.float()).all():
111
+ raise RuntimeError(f"case_{case['id']}: output '{k}' has NaN/Inf — bad golden data")
112
+
113
+ saved_in = {k: (v.cpu() if isinstance(v, torch.Tensor) else v) for k, v in inputs.items()}
114
+ return saved_in, saved_out
115
+
116
+
117
+ def describe_call(call_args):
118
+ """Compact argument summary."""
119
+ parts = []
120
+ for name, val in zip(PARAM_NAMES, call_args):
121
+ if isinstance(val, torch.Tensor):
122
+ parts.append(f"{name}=Tensor{tuple(val.shape)}:{str(val.dtype).replace('torch.', '')}")
123
+ else:
124
+ parts.append(f"{name}={val!r}")
125
+ return ", ".join(parts)
126
+
127
+
128
+ def main():
129
+ os.makedirs(INPUT_DIR, exist_ok=True)
130
+ os.makedirs(RESULT_DIR, exist_ok=True)
131
+ manifest = {"op_name": OP_NAME, "op_camel": OP_CAMEL, "output_keys": OUTPUT_KEYS, "cases": []}
132
+
133
+ for case in CASES:
134
+ name = f"case_{case['id']}"
135
+ try:
136
+ saved_in, saved_out = collect_case(case)
137
+ except TypeError as exc:
138
+ print(f"{name}: signature mismatch")
139
+ print(f" passed: {describe_call(build_call_args(case)[0])}")
140
+ print(f" binding expects: {ops.npu_interleave_rope.__doc__.splitlines()[0] if ops.npu_interleave_rope.__doc__ else '?'}")
141
+ print(f" {type(exc).__name__}: {str(exc)[:400]}")
142
+ sys.exit(1)
143
+ except Exception as exc:
144
+ print(f"{name}: COLLECT FAILED — {type(exc).__name__}: {str(exc)[:400]}")
145
+ sys.exit(1)
146
+ in_path, out_path = os.path.join(INPUT_DIR, f"{name}.pt"), os.path.join(RESULT_DIR, f"{name}.pt")
147
+ torch.save(saved_in, in_path)
148
+ torch.save(saved_out, out_path)
149
+ manifest["cases"].append({
150
+ "id": case["id"], "dtype": str(case["dtype"]).replace("torch.", ""),
151
+ "shape": [case["batch"], 1, case["seq"], 64],
152
+ "attrs": {k: v for k, v in case.items() if k not in ("id", "dtype", "batch", "seq")},
153
+ })
154
+ print(f"{name}: OK input={os.path.getsize(in_path)}B result={os.path.getsize(out_path)}B")
155
+
156
+ with open(os.path.join(DATA_ROOT, "cases_manifest.json"), "w") as fh:
157
+ json.dump(manifest, fh, indent=2)
158
+ print(f"\nCollected {len(manifest['cases'])}/{len(CASES)} cases into {DATA_ROOT}/")
159
+
160
+
161
+ if __name__ == "__main__":
162
+ main()
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/interleave_rope_test.py ADDED
@@ -0,0 +1,220 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ """Reference regression test: NPU operator vs golden data collected from the same operator.
3
+
4
+ This is the canonical test script for the sandbox workflow. To adapt it to another operator,
5
+ change only OP_NAME/OP_CAMEL, PARAM_NAMES, OUTPUT_KEYS, INTERMEDIATE_OUTPUT_KEYS and the
6
+ ops.npu_<op_name>(...) call in run_case(). Leave path resolution, the comparator and the exit code alone.
7
+
8
+ Inputs are LOADED from golden_input/, never regenerated from a seed: the comparison is only
9
+ meaningful if the inputs are byte-identical to the ones used at collection time.
10
+
11
+ Tolerance: golden is frozen while the sources keep changing, so the check carries slack —
12
+ |ref - cur| / max(|ref|, 1) <= 1e-4 for every element. Note this is stricter than one ulp for
13
+ fp16/bf16 outputs of order 1 (spacing 2**-10 and 2**-8 there); the slack is real for float32 and
14
+ for small-magnitude outputs. A legitimate rounding shift in bf16 needs a per-case override.
15
+
16
+ Naming: the operator is called as ops.npu_{op_name} verbatim — the npu_ prefix is part of the
17
+ registered name; no _custom/_v2/_kernel suffixes, no wrapper function, no alias.
18
+ """
19
+ import json
20
+ import os
21
+ import sys
22
+
23
+ import numpy as np
24
+ import torch
25
+ import torch_npu # noqa: F401
26
+ import gtsrusops as ops
27
+
28
+ OP_NAME = "interleave_rope"
29
+ OP_CAMEL = "InterleaveRope"
30
+ DEVICE = "npu:0"
31
+
32
+ # Actual binding signature (from op_desc inputs + attrs):
33
+ # npu_interleave_rope(x, cos, sin, out)
34
+ PARAM_NAMES = ["x", "cos", "sin"]
35
+ # Order in which the operator returns its outputs.
36
+ OUTPUT_KEYS = ["y"]
37
+
38
+ # Outputs that are intermediate statistics: stored as float32 but derived from the input tensor,
39
+ # so their real precision floor is the INPUT dtype, not their storage dtype.
40
+ INTERMEDIATE_OUTPUT_KEYS = set()
41
+
42
+ # Default tolerance. atol applies per element to diff / max(|ref|, 1); err_rate is the allowed
43
+ # fraction of elements exceeding atol. Do not widen these globally.
44
+ ATOL_DEFAULT = 1e-4
45
+ ERR_RATE_DEFAULT = 0.0
46
+
47
+ # Dtype floor, a couple of ulp of headroom — used only when an override sets "atol": None.
48
+ ATOL_BY_DTYPE = {torch.float16: 2 ** -8, torch.bfloat16: 2 ** -7, torch.float32: 2 ** -11}
49
+
50
+ # Per-case tolerance overrides, for cases where an intentional source change shifted the rounding
51
+ # (or where the kernel's reduction order is nondeterministic). Each entry needs a comment naming
52
+ # the case, the output tensor and the cause. {case_id: {"atol": float or None, "err_rate": float}}
53
+ TOLERANCE_OVERRIDES = {}
54
+
55
+
56
+ def resolve_data_root():
57
+ """Find test_data/ in both layouts: sandbox (CWD/<op>/test_data) and local (script dir/test_data).
58
+
59
+ Do not branch on __file__.startswith("tmp") — that breaks when the client's temp naming changes.
60
+ Absolute host paths never resolve inside the sandbox, so everything is relative to here or CWD.
61
+ """
62
+ here = os.path.dirname(os.path.abspath(__file__))
63
+ candidates = [
64
+ os.path.join(here, "test_data"),
65
+ os.path.join(here, OP_NAME, "test_data"),
66
+ os.path.join(os.getcwd(), OP_NAME, "test_data"),
67
+ os.path.join(os.getcwd(), "test_data"),
68
+ ]
69
+ for path in candidates:
70
+ if os.path.isdir(os.path.join(path, "golden_result", OP_CAMEL)):
71
+ return path
72
+ raise FileNotFoundError(
73
+ "golden data not found. Tried:\n " + "\n ".join(candidates)
74
+ + f"\nRun datamining/{OP_NAME}_collect.sh first."
75
+ )
76
+
77
+
78
+ DATA_ROOT = resolve_data_root()
79
+ INPUT_DIR = os.path.join(DATA_ROOT, "golden_input", OP_CAMEL)
80
+ RESULT_DIR = os.path.join(DATA_ROOT, "golden_result", OP_CAMEL)
81
+
82
+
83
+ def compare_tensor(ref_t, cur_t, atol=ATOL_DEFAULT, err_rate=ERR_RATE_DEFAULT):
84
+ """Two independent knobs:
85
+ atol — per-element threshold on diff / max(|ref|, 1)
86
+ err_rate — allowed fraction of elements exceeding atol
87
+ Never compare error_rate against atol; they are different quantities.
88
+ """
89
+ if tuple(ref_t.shape) != tuple(cur_t.shape):
90
+ return float("inf"), 1.0, f"SHAPE {tuple(cur_t.shape)}!={tuple(ref_t.shape)}"
91
+ if ref_t.dtype != cur_t.dtype:
92
+ return float("inf"), 1.0, f"DTYPE {cur_t.dtype}!={ref_t.dtype}"
93
+
94
+ ref_np = ref_t.float().cpu().detach().numpy().flatten().astype("float64")
95
+ cur_np = cur_t.float().cpu().detach().numpy().flatten().astype("float64")
96
+ diff = np.abs(ref_np - cur_np)
97
+ rel = diff / np.maximum(np.abs(ref_np), 1.0)
98
+
99
+ max_diff = float(diff.max()) if diff.size else 0.0
100
+ error_rate = float((rel > atol).sum()) / max(ref_np.size, 1)
101
+ status = "PASS" if error_rate <= err_rate else "FAIL"
102
+ return max_diff, error_rate, status
103
+
104
+
105
+ def tolerance_for(case_id, output_key, input_dtype, ref_dtype):
106
+ override = TOLERANCE_OVERRIDES.get(case_id)
107
+ if override is None:
108
+ return ATOL_DEFAULT, ERR_RATE_DEFAULT
109
+ atol = override.get("atol", ATOL_DEFAULT)
110
+ if atol is None:
111
+ dtype = input_dtype if output_key in INTERMEDIATE_OUTPUT_KEYS else ref_dtype
112
+ atol = max(ATOL_DEFAULT, ATOL_BY_DTYPE.get(dtype, 0.0))
113
+ return atol, override.get("err_rate", ERR_RATE_DEFAULT)
114
+
115
+
116
+ def run_case(case_id):
117
+ name = f"case_{case_id}"
118
+ input_data = torch.load(os.path.join(INPUT_DIR, f"{name}.pt"), weights_only=False)
119
+ golden = torch.load(os.path.join(RESULT_DIR, f"{name}.pt"), weights_only=False)
120
+
121
+ # Build kwargs dict — the wrapper (binding.py) calls the raw func by keyword,
122
+ # so we do the same here.
123
+ call_kwargs = {}
124
+ for pname in PARAM_NAMES:
125
+ if pname not in input_data:
126
+ raise KeyError(f"{name}: missing param '{pname}' in golden_input")
127
+ val = input_data[pname]
128
+ if isinstance(val, torch.Tensor):
129
+ call_kwargs[pname] = val.to(DEVICE)
130
+ elif isinstance(val, list):
131
+ call_kwargs[pname] = [v.to(DEVICE) if isinstance(v, torch.Tensor) else v for v in val]
132
+ else:
133
+ call_kwargs[pname] = val
134
+
135
+ # Size the output buffers from the golden tensors — this also checks shape and dtype.
136
+ out_bufs = []
137
+ for k in OUTPUT_KEYS:
138
+ if k in golden:
139
+ out_bufs.append(torch.empty_like(golden[k]).to(DEVICE))
140
+ # Call with positional args: (x, cos, sin, out)
141
+ positional = [call_kwargs[p] for p in PARAM_NAMES] + [out_bufs[0]]
142
+ outputs = ops.npu_interleave_rope(*positional)
143
+ torch.npu.synchronize()
144
+
145
+ outputs = [outputs] if isinstance(outputs, torch.Tensor) else list(outputs)
146
+ input_dtype = input_data[PARAM_NAMES[0]].dtype
147
+
148
+ results = []
149
+ for i, okey in enumerate(OUTPUT_KEYS):
150
+ if okey not in golden:
151
+ continue
152
+ if i >= len(outputs):
153
+ results.append((okey, float("nan"), 1.0, "MISSING"))
154
+ continue
155
+ atol, err_rate = tolerance_for(case_id, okey, input_dtype, golden[okey].dtype)
156
+ results.append((okey,) + compare_tensor(
157
+ golden[okey].to(DEVICE), outputs[i], atol=atol, err_rate=err_rate
158
+ ))
159
+ return results
160
+
161
+
162
+ def main():
163
+ manifest_path = os.path.join(DATA_ROOT, "cases_manifest.json")
164
+ if os.path.isfile(manifest_path):
165
+ with open(manifest_path) as fh:
166
+ collected = [c["id"] for c in json.load(fh)["cases"]]
167
+ else:
168
+ collected = sorted(
169
+ int(f[len("case_"):-len(".pt")])
170
+ for f in os.listdir(RESULT_DIR)
171
+ if f.startswith("case_") and f.endswith(".pt")
172
+ )
173
+
174
+ if not collected:
175
+ print(f"No golden cases in {RESULT_DIR} — run the collect script first.")
176
+ sys.exit(1)
177
+
178
+ print(f"Golden data: {DATA_ROOT} (atol={ATOL_DEFAULT:.0e})")
179
+ print(f"{'case':<12}{'output':<16}{'max_diff':<18}{'err_rate':<14}{'status':<8}")
180
+ print("-" * 68)
181
+
182
+ all_pass = True
183
+ for case_id in collected:
184
+ for okey, max_diff, error_rate, status in run_case(case_id):
185
+ if status != "PASS":
186
+ all_pass = False
187
+ print(f"{'case_' + str(case_id):<12}{okey:<16}{max_diff:<18.10f}{error_rate:<14.6f}{status:<8}")
188
+
189
+ print("-" * 68)
190
+ if all_pass:
191
+ print(f"ALL CHECKS PASSED ({len(collected)} cases)")
192
+ else:
193
+ print("REGRESSION DETECTED — do NOT re-collect golden data to make this pass.")
194
+ sys.exit(1)
195
+
196
+
197
+
198
+
199
+ def _run_suite(runs):
200
+ """重复整套测试 runs 次(warmup + 测量);单次 FAIL 不中断后续轮次。"""
201
+ all_ok = True
202
+ for run_idx in range(1, runs + 1):
203
+ print(f"\n{'=' * 98}")
204
+ print(f"Run {run_idx}/{runs}")
205
+ print(f"{'=' * 98}")
206
+ try:
207
+ main()
208
+ except SystemExit as exc:
209
+ if exc.code not in (0, None):
210
+ all_ok = False
211
+ return all_ok
212
+
213
+
214
+ if __name__ == "__main__":
215
+ import argparse
216
+ parser = argparse.ArgumentParser(description="repeated suite runs for warmup + measurement")
217
+ parser.add_argument("--runs", type=int, default=1,
218
+ help="Number of repeated runs (each run iterates all cases)")
219
+ args = parser.parse_args()
220
+ raise SystemExit(0 if _run_suite(args.runs) else 1)
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_graph/CMakeLists.txt ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # -----------------------------------------------------------------------------------------------------------
2
+ # Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+ # This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+ # CANN Open Software License Agreement Version 2.0 (the "License").
5
+ # Please refer to the License for details. You may not use this file except in compliance with the License.
6
+ # THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+ # INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+ # See LICENSE in the root of the software repository for the full text of the License.
9
+ # -----------------------------------------------------------------------------------------------------------
10
+
11
+ message(STATUS "=== Debug: start ops.posembedding.interleave_rope.op_graph.CMakeLists.txt ")
12
+ add_graph_plugin_sources()
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_graph/fusion_pass/.gitkeep ADDED
File without changes
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_graph/interleave_rope_proto.h ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ /*!
12
+ * \file interleave_rope_proto.h
13
+ * \brief
14
+ */
15
+ #ifndef OPS_OP_PROTO_INC_INTERLEAVEROPE_OPS_H_
16
+ #define OPS_OP_PROTO_INC_INTERLEAVEROPE_OPS_H_
17
+
18
+ #include "graph/operator_reg.h"
19
+ namespace ge {
20
+
21
+ /**
22
+ * @brief The fusion operator of Interleave RotaryPositionEmbedding.
23
+ *
24
+ * @par Inputs:
25
+ * @li x: A tensor. The type support float16, bfloat16. Format: ND.
26
+ * @li cos: A tensor. The type support float16, bfloat16. Format: ND.
27
+ * @li sin: A tensor. The type support float16, bfloat16. Format: ND.
28
+ *
29
+ * @par Outputs:
30
+ * y: A tensor. The type support float16, bfloat16. Format: ND.
31
+ *
32
+ * @par Restrictions:
33
+ * Warning: THIS FUNCTION IS EXPERIMENTAL. Please do not use.
34
+ */
35
+
36
+ REG_OP(InterleaveRope)
37
+ .INPUT(x, TensorType({DT_FLOAT16, DT_BF16}))
38
+ .INPUT(cos, TensorType({DT_FLOAT16, DT_BF16}))
39
+ .INPUT(sin, TensorType({DT_FLOAT16, DT_BF16}))
40
+ .OUTPUT(y, TensorType({DT_FLOAT16, DT_BF16}))
41
+ .OP_END_FACTORY_REG(InterleaveRope)
42
+
43
+ } // namespace ge
44
+
45
+ #endif
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/CMakeLists.txt ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # -----------------------------------------------------------------------------------------------------------
2
+ # Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+ # This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+ # CANN Open Software License Agreement Version 2.0 (the "License").
5
+ # Please refer to the License for details. You may not use this file except in compliance with the License.
6
+ # THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+ # INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+ # See LICENSE in the root of the software repository for the full text of the License.
9
+ # -----------------------------------------------------------------------------------------------------------
10
+
11
+ message(STATUS "=== Debug: start ops.posembedding.interleave_rope.op_host.CMakeLists.txt ")
12
+
13
+ if (BUILD_OPEN_PROJECT)
14
+ message(STATUS "=== Debug: BUILD is BUILD_OPEN_PROJECT")
15
+ set(interleave_rope_depends posembedding/rotary_position_embedding CACHE INTERNAL "Dependencies for rotary_position_embedding")
16
+ target_sources(op_host_aclnnInner PRIVATE
17
+ interleave_rope_def.cpp
18
+ )
19
+ add_ops_compile_options(
20
+ OP_NAME InterleaveRope
21
+ OPTIONS --cce-auto-sync=on
22
+ -Wno-deprecated-declarations
23
+ -Werror
24
+ )
25
+ endif()
26
+
27
+ if(NOT BUILD_OPS_RTY_KERNEL)
28
+ message(STATUS "=== Debug: BUILD is NOT BUILD_OPS_RTY_KERNEL")
29
+ add_modules_sources(OPTYPE interleave_rope ACLNNTYPE aclnn_inner)
30
+ endif()
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/config/ascend910_93/interleave_rope_binary.json ADDED
@@ -0,0 +1,99 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "op_type": "InterleaveRope",
3
+ "op_list": [
4
+ {
5
+ "bin_filename": "InterleaveRope_B8B1918CF4B1F919BC0F2A58138E37AB",
6
+ "inputs": [
7
+ {
8
+ "name": "x",
9
+ "index": 0,
10
+ "dtype": "float16",
11
+ "format": "ND",
12
+ "paramType": "required",
13
+ "shape": [
14
+ -2
15
+ ]
16
+ },
17
+ {
18
+ "name": "cos",
19
+ "index": 1,
20
+ "dtype": "float16",
21
+ "format": "ND",
22
+ "paramType": "required",
23
+ "shape": [
24
+ -2
25
+ ]
26
+ },
27
+ {
28
+ "name": "sin",
29
+ "index": 2,
30
+ "dtype": "float16",
31
+ "format": "ND",
32
+ "paramType": "required",
33
+ "shape": [
34
+ -2
35
+ ]
36
+ }
37
+ ],
38
+ "outputs": [
39
+ {
40
+ "name": "y",
41
+ "index": 0,
42
+ "dtype": "float16",
43
+ "format": "ND",
44
+ "paramType": "required",
45
+ "shape": [
46
+ -2
47
+ ]
48
+ }
49
+ ]
50
+ },
51
+ {
52
+ "bin_filename": "InterleaveRope_5A99F786F3FB24A70D3CBC0A8D16CA01",
53
+ "inputs": [
54
+ {
55
+ "name": "x",
56
+ "index": 0,
57
+ "dtype": "bfloat16",
58
+ "format": "ND",
59
+ "paramType": "required",
60
+ "shape": [
61
+ -2
62
+ ]
63
+ },
64
+ {
65
+ "name": "cos",
66
+ "index": 1,
67
+ "dtype": "bfloat16",
68
+ "format": "ND",
69
+ "paramType": "required",
70
+ "shape": [
71
+ -2
72
+ ]
73
+ },
74
+ {
75
+ "name": "sin",
76
+ "index": 2,
77
+ "dtype": "bfloat16",
78
+ "format": "ND",
79
+ "paramType": "required",
80
+ "shape": [
81
+ -2
82
+ ]
83
+ }
84
+ ],
85
+ "outputs": [
86
+ {
87
+ "name": "y",
88
+ "index": 0,
89
+ "dtype": "bfloat16",
90
+ "format": "ND",
91
+ "paramType": "required",
92
+ "shape": [
93
+ -2
94
+ ]
95
+ }
96
+ ]
97
+ }
98
+ ]
99
+ }
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/config/ascend910_93/interleave_rope_simplified_key.ini ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ; 该文件主要影响 opc 工具 编译二进制kernel时, --simplified_key_mode 选项中填写的值,格式如下所示:
2
+ ; [某算子]
3
+ ; default=xx
4
+ ; ascendxx=xx
5
+ ; 其中,default为默认mode,ascnedxx为可选mode,如果不同芯片有差异化要求时,需要配置;
6
+ ; 1)如果没有配置:非ascendC算子继续按空处理,即opc编译命令中不添加 --simplified_key_mode 选项,AscendC算子按照 simplified_key_mode=0 处理
7
+ ; 2)如果仅有default配置:各个版本按default配置
8
+ ; 3)如果仅有某些平台的配置,没有default配置:对应平台的按照配置的值传递,非对应平台的:非AscendC算子继续按空处理,AscendC算子按照 simplified_key_mode=0 处理
9
+ ; 4)如果default配置和平台配置都有:对应平台的使用平台的配置,非对应的平台的以default值配置。
10
+ ; 5)对于自定义simplified key的情况,需要在binary_simplified_key_mode.ini 文件中显式配置为None,不传入 --simplified_key_mode 选项,由opc工具和FE框架自行判断使用何种模式
11
+ ; 6)是否是AscendC算子,由 ops/build-in/tbe/op_info_cfg/parser/ascendc_config.json 中配置的算子名字和对于的平台决定
12
+ [InterleaveRope]
13
+ default=0
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/config/ascend910b/interleave_rope_binary.json ADDED
@@ -0,0 +1,99 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "op_type": "InterleaveRope",
3
+ "op_list": [
4
+ {
5
+ "bin_filename": "InterleaveRope_B8B1918CF4B1F919BC0F2A58138E37AB",
6
+ "inputs": [
7
+ {
8
+ "name": "x",
9
+ "index": 0,
10
+ "dtype": "float16",
11
+ "format": "ND",
12
+ "paramType": "required",
13
+ "shape": [
14
+ -2
15
+ ]
16
+ },
17
+ {
18
+ "name": "cos",
19
+ "index": 1,
20
+ "dtype": "float16",
21
+ "format": "ND",
22
+ "paramType": "required",
23
+ "shape": [
24
+ -2
25
+ ]
26
+ },
27
+ {
28
+ "name": "sin",
29
+ "index": 2,
30
+ "dtype": "float16",
31
+ "format": "ND",
32
+ "paramType": "required",
33
+ "shape": [
34
+ -2
35
+ ]
36
+ }
37
+ ],
38
+ "outputs": [
39
+ {
40
+ "name": "y",
41
+ "index": 0,
42
+ "dtype": "float16",
43
+ "format": "ND",
44
+ "paramType": "required",
45
+ "shape": [
46
+ -2
47
+ ]
48
+ }
49
+ ]
50
+ },
51
+ {
52
+ "bin_filename": "InterleaveRope_5A99F786F3FB24A70D3CBC0A8D16CA01",
53
+ "inputs": [
54
+ {
55
+ "name": "x",
56
+ "index": 0,
57
+ "dtype": "bfloat16",
58
+ "format": "ND",
59
+ "paramType": "required",
60
+ "shape": [
61
+ -2
62
+ ]
63
+ },
64
+ {
65
+ "name": "cos",
66
+ "index": 1,
67
+ "dtype": "bfloat16",
68
+ "format": "ND",
69
+ "paramType": "required",
70
+ "shape": [
71
+ -2
72
+ ]
73
+ },
74
+ {
75
+ "name": "sin",
76
+ "index": 2,
77
+ "dtype": "bfloat16",
78
+ "format": "ND",
79
+ "paramType": "required",
80
+ "shape": [
81
+ -2
82
+ ]
83
+ }
84
+ ],
85
+ "outputs": [
86
+ {
87
+ "name": "y",
88
+ "index": 0,
89
+ "dtype": "bfloat16",
90
+ "format": "ND",
91
+ "paramType": "required",
92
+ "shape": [
93
+ -2
94
+ ]
95
+ }
96
+ ]
97
+ }
98
+ ]
99
+ }
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/config/ascend910b/interleave_rope_simplified_key.ini ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ; 该文件主要影响 opc 工具 编译二进制kernel时, --simplified_key_mode 选项中填写的值,格式如下所示:
2
+ ; [某算子]
3
+ ; default=xx
4
+ ; ascendxx=xx
5
+ ; 其中,default为默认mode,ascnedxx为可选mode,如果不同芯片有差异化要求时,需要配置;
6
+ ; 1)如果没有配置:非ascendC算子继续按空处理,即opc编译命令中不添加 --simplified_key_mode 选项,AscendC算子按照 simplified_key_mode=0 处理
7
+ ; 2)如果仅有default配置:各个版本按default配置
8
+ ; 3)如果仅有某些平台的配置,没有default配置:对应平台的按照配置的值传递,非对应平台的:非AscendC算子继续按空处理,AscendC算子按照 simplified_key_mode=0 处理
9
+ ; 4)如果default配置和平台配置都有:对应平台的使用平台的配置,非对应的平台的以default值配置。
10
+ ; 5)对于自定义simplified key的情况,需要在binary_simplified_key_mode.ini 文件中显式配置为None,不传入 --simplified_key_mode 选项,由opc工具和FE框架自行判断使用何种模式
11
+ ; 6)是否是AscendC算子,由 ops/build-in/tbe/op_info_cfg/parser/ascendc_config.json 中配置的算子名字和对于的平台决定
12
+ [InterleaveRope]
13
+ default=0
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/config/kirin9030/interleave_rope_binary.json ADDED
@@ -0,0 +1,52 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "op_type": "InterleaveRope",
3
+ "op_list": [
4
+ {
5
+ "bin_filename": "InterleaveRope_B8B1918CF4B1F919BC0F2A58138E37AB",
6
+ "inputs": [
7
+ {
8
+ "name": "x",
9
+ "index": 0,
10
+ "dtype": "float16",
11
+ "format": "ND",
12
+ "paramType": "required",
13
+ "shape": [
14
+ -2
15
+ ]
16
+ },
17
+ {
18
+ "name": "cos",
19
+ "index": 1,
20
+ "dtype": "float16",
21
+ "format": "ND",
22
+ "paramType": "required",
23
+ "shape": [
24
+ -2
25
+ ]
26
+ },
27
+ {
28
+ "name": "sin",
29
+ "index": 2,
30
+ "dtype": "float16",
31
+ "format": "ND",
32
+ "paramType": "required",
33
+ "shape": [
34
+ -2
35
+ ]
36
+ }
37
+ ],
38
+ "outputs": [
39
+ {
40
+ "name": "y",
41
+ "index": 0,
42
+ "dtype": "float16",
43
+ "format": "ND",
44
+ "paramType": "required",
45
+ "shape": [
46
+ -2
47
+ ]
48
+ }
49
+ ]
50
+ }
51
+ ]
52
+ }
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/config/kirin9030/interleave_rope_simplified_key.ini ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ; 该文件主要影响 opc 工具 编译二进制kernel时, --simplified_key_mode 选项中填写的值,格式如下所示:
2
+ ; [某算子]
3
+ ; default=xx
4
+ ; ascendxx=xx
5
+ ; 其中,default为默认mode,ascnedxx为可选mode,如果不同芯片有差异化要求时,需要配置;
6
+ ; 1)如果没有配置:非ascendC算子继续按空处理,即opc编译命令中不添加 --simplified_key_mode 选项,AscendC算子按照 simplified_key_mode=0 处理
7
+ ; 2)如果仅有default配置:各个版本按default配置
8
+ ; 3)如果仅有某些平台的配置,没有default配置:对应平台的按照配置的值传递,非对应平台的:非AscendC算子继续按空处理,AscendC算子按照 simplified_key_mode=0 处理
9
+ ; 4)如果default配置和平台配置都有:对应平台的使用平台的配置,非对应的平台的以default值配置。
10
+ ; 5)对于自定义simplified key的情况,需要在binary_simplified_key_mode.ini 文件中显式配置为None,不传入 --simplified_key_mode 选项,由opc工具和FE框架自行判断使用何种模式
11
+ ; 6)是否是AscendC算子,由 ops/build-in/tbe/op_info_cfg/parser/ascendc_config.json 中配置的算子名字和对于的平台决定
12
+ [InterleaveRope]
13
+ default=0
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/config/kirinx90/interleave_rope_binary.json ADDED
@@ -0,0 +1,52 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "op_type": "InterleaveRope",
3
+ "op_list": [
4
+ {
5
+ "bin_filename": "InterleaveRope_B8B1918CF4B1F919BC0F2A58138E37AB",
6
+ "inputs": [
7
+ {
8
+ "name": "x",
9
+ "index": 0,
10
+ "dtype": "float16",
11
+ "format": "ND",
12
+ "paramType": "required",
13
+ "shape": [
14
+ -2
15
+ ]
16
+ },
17
+ {
18
+ "name": "cos",
19
+ "index": 1,
20
+ "dtype": "float16",
21
+ "format": "ND",
22
+ "paramType": "required",
23
+ "shape": [
24
+ -2
25
+ ]
26
+ },
27
+ {
28
+ "name": "sin",
29
+ "index": 2,
30
+ "dtype": "float16",
31
+ "format": "ND",
32
+ "paramType": "required",
33
+ "shape": [
34
+ -2
35
+ ]
36
+ }
37
+ ],
38
+ "outputs": [
39
+ {
40
+ "name": "y",
41
+ "index": 0,
42
+ "dtype": "float16",
43
+ "format": "ND",
44
+ "paramType": "required",
45
+ "shape": [
46
+ -2
47
+ ]
48
+ }
49
+ ]
50
+ }
51
+ ]
52
+ }
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/config/kirinx90/interleave_rope_simplified_key.ini ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ; 该文件主要影响 opc 工具 编译二进制kernel时, --simplified_key_mode 选项中填写的值,格式如下所示:
2
+ ; [某算子]
3
+ ; default=xx
4
+ ; ascendxx=xx
5
+ ; 其中,default为默认mode,ascnedxx为可选mode,如果不同芯片有差异化要求时,需要配置;
6
+ ; 1)如果没有配置:非ascendC算子继续按空处理,即opc编译命令中不添加 --simplified_key_mode 选项,AscendC算子按照 simplified_key_mode=0 处理
7
+ ; 2)如果仅有default配置:各个版本按default配置
8
+ ; 3)如果仅有某些平台的配置,没有default配置:对应平台的按照配置的值传递,非对应平台的:非AscendC算子继续按空处理,AscendC算子按照 simplified_key_mode=0 处理
9
+ ; 4)如果default配置和平台配置都有:对应平台的使用平台的配置,非对应的平台的以default值配置。
10
+ ; 5)对于自定义simplified key的情况,需要在binary_simplified_key_mode.ini 文件中显式配置为None,不传入 --simplified_key_mode 选项,由opc工具和FE框架自行判断使用何种模式
11
+ ; 6)是否是AscendC算子,由 ops/build-in/tbe/op_info_cfg/parser/ascendc_config.json 中配置的算子名字和对于的平台决定
12
+ [InterleaveRope]
13
+ default=0
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/interleave_rope_def.cpp ADDED
@@ -0,0 +1,93 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ /*!
12
+ * \file interleave_rope_def.cpp
13
+ * \brief InterleaveRope op host config
14
+ */
15
+
16
+ #include "register/op_def_registry.h"
17
+
18
+ namespace ops {
19
+ class InterleaveRope : public OpDef {
20
+ public:
21
+ explicit InterleaveRope(const char* name) : OpDef(name)
22
+ {
23
+ this->Input("x")
24
+ .ParamType(REQUIRED)
25
+ .DataType({ge::DT_FLOAT16, ge::DT_BF16})
26
+ .Format({ge::FORMAT_ND, ge::FORMAT_ND})
27
+ .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND})
28
+ .AutoContiguous();
29
+ this->Input("cos")
30
+ .ParamType(REQUIRED)
31
+ .DataType({ge::DT_FLOAT16, ge::DT_BF16})
32
+ .Format({ge::FORMAT_ND, ge::FORMAT_ND})
33
+ .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND})
34
+ .AutoContiguous();
35
+ this->Input("sin")
36
+ .ParamType(REQUIRED)
37
+ .DataType({ge::DT_FLOAT16, ge::DT_BF16})
38
+ .Format({ge::FORMAT_ND, ge::FORMAT_ND})
39
+ .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND})
40
+ .AutoContiguous();
41
+ this->Output("y")
42
+ .ParamType(REQUIRED)
43
+ .DataType({ge::DT_FLOAT16, ge::DT_BF16})
44
+ .Format({ge::FORMAT_ND, ge::FORMAT_ND})
45
+ .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND});
46
+
47
+ this->AICore().AddConfig("ascend910b");
48
+ this->AICore().AddConfig("ascend910_93");
49
+
50
+ OpAICoreConfig config_kirin = GetKirinCoreConfig();
51
+ this->AICore().AddConfig("kirinx90", config_kirin);
52
+ this->AICore().AddConfig("kirin9030", config_kirin);
53
+ }
54
+
55
+ private:
56
+ OpAICoreConfig GetKirinCoreConfig() const
57
+ {
58
+ OpAICoreConfig config_kirin;
59
+ config_kirin.DynamicCompileStaticFlag(true)
60
+ .DynamicFormatFlag(true)
61
+ .DynamicRankSupportFlag(true)
62
+ .DynamicShapeSupportFlag(true)
63
+ .NeedCheckSupportFlag(false)
64
+ .PrecisionReduceFlag(true);
65
+ config_kirin.Input("x")
66
+ .ParamType(REQUIRED)
67
+ .DataType({ge::DT_FLOAT16})
68
+ .Format({ge::FORMAT_ND})
69
+ .UnknownShapeFormat({ge::FORMAT_ND})
70
+ .AutoContiguous();
71
+ config_kirin.Input("cos")
72
+ .ParamType(REQUIRED)
73
+ .DataType({ge::DT_FLOAT16})
74
+ .Format({ge::FORMAT_ND})
75
+ .UnknownShapeFormat({ge::FORMAT_ND})
76
+ .AutoContiguous();
77
+ config_kirin.Input("sin")
78
+ .ParamType(REQUIRED)
79
+ .DataType({ge::DT_FLOAT16})
80
+ .Format({ge::FORMAT_ND})
81
+ .UnknownShapeFormat({ge::FORMAT_ND})
82
+ .AutoContiguous();
83
+ config_kirin.Output("y")
84
+ .ParamType(REQUIRED)
85
+ .DataType({ge::DT_FLOAT16})
86
+ .Format({ge::FORMAT_ND})
87
+ .UnknownShapeFormat({ge::FORMAT_ND});
88
+ return config_kirin;
89
+ }
90
+ };
91
+
92
+ OP_ADD(InterleaveRope);
93
+ } // namespace ops
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/interleave_rope_infershape.cpp ADDED
@@ -0,0 +1,57 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ /*!
12
+ * \file interleave_rope_infer.cpp
13
+ * \brief
14
+ */
15
+ #include <sstream>
16
+ #include <string>
17
+ #include <vector>
18
+ #include "register/op_def_registry.h"
19
+ #include "log/log.h"
20
+ #include "util/math_util.h"
21
+ #include "util/shape_util.h"
22
+
23
+ using namespace ge;
24
+ namespace ops {
25
+ constexpr size_t INPUT_IDX_X = 0;
26
+ constexpr size_t INPUT_IDX_COS = 1;
27
+ constexpr size_t INPUT_IDX_SIN = 2;
28
+ constexpr size_t OUTPUT_IDX_Y = 0;
29
+
30
+ graphStatus InferShape4InterleaveRope(gert::InferShapeContext* context)
31
+ {
32
+ OP_LOGI(context, "Begin to do InferShape4InterleaveRope.");
33
+
34
+ const gert::Shape* xInputShape = context->GetInputShape(INPUT_IDX_X);
35
+ OP_CHECK_NULL_WITH_CONTEXT(context, xInputShape);
36
+ gert::Shape* yShape = context->GetOutputShape(OUTPUT_IDX_Y);
37
+ OP_CHECK_NULL_WITH_CONTEXT(context, yShape);
38
+ *yShape = *xInputShape;
39
+
40
+ OP_LOGD(context, "End to do InferShape4InterleaveRope.");
41
+ return GRAPH_SUCCESS;
42
+ }
43
+
44
+ graphStatus InferDtype4InterleaveRope(gert::InferDataTypeContext* context)
45
+ {
46
+ OP_LOGD(context, "InferDtype4InterleaveRope enter");
47
+
48
+ auto x_dtype = context->GetInputDataType(INPUT_IDX_X);
49
+ context->SetOutputDataType(OUTPUT_IDX_Y, x_dtype);
50
+
51
+ OP_LOGD(context, "InferDtype4InterleaveRope end");
52
+
53
+ return GRAPH_SUCCESS;
54
+ }
55
+
56
+ IMPL_OP_INFERSHAPE(InterleaveRope).InferShape(InferShape4InterleaveRope).InferDataType(InferDtype4InterleaveRope);
57
+ } // namespace ops
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/interleave_rope_tiling.cpp ADDED
@@ -0,0 +1,359 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ /*!
12
+ * \file interleave_rope_tiling.cpp
13
+ * \brief
14
+ */
15
+ #include "interleave_rope_tiling.h"
16
+ #include "register/op_def_registry.h"
17
+ #include "tiling/tiling_api.h"
18
+ #include "log/log.h"
19
+
20
+ namespace optiling {
21
+
22
+ constexpr int64_t X_INDEX = 0;
23
+ constexpr int64_t COS_INDEX = 1;
24
+ constexpr int64_t SIN_INDEX = 2;
25
+ constexpr int64_t INPUT_NUM = 3;
26
+ constexpr int64_t OUTPUT_NUM = 1;
27
+ constexpr int64_t SHAPE_IDX_B = 0;
28
+ constexpr int64_t SHAPE_IDX_N = 1;
29
+ constexpr int64_t SHAPE_IDX_S = 2;
30
+ constexpr int64_t SHAPE_IDX_D = 3;
31
+ constexpr int64_t DIM_SIZE = 4;
32
+ constexpr int64_t DIM_ONE = 1;
33
+ constexpr int64_t DEFAULT_WORKSPACE_SIZE = 32;
34
+ constexpr int64_t DEFAULT_HIDDEN_DIM = 64;
35
+ constexpr int64_t DEFAULT_BATCH_SIZE = 32;
36
+ constexpr int64_t DEFAULT_NUM_HEAD = 32;
37
+ constexpr int64_t DEFAULT_BATCH_PER_BLOCK = 4;
38
+ constexpr uint64_t SPLIT_BATCH = 0;
39
+ constexpr uint64_t SPLIT_NS = 1;
40
+ constexpr uint64_t SPLIT_N = 2;
41
+ constexpr uint64_t SPLIT_S = 3;
42
+ constexpr uint64_t SPLIT_BNS = 4;
43
+
44
+ constexpr int64_t NUM_TWO = 2;
45
+ constexpr int64_t NUM_FOUR = 4;
46
+ constexpr int64_t NUM_SIX = 6;
47
+ constexpr int64_t NUM_TEN = 10;
48
+ constexpr int64_t MIN_SEQUANCE_LEN = 512;
49
+
50
+ constexpr uint64_t FIXED_BNSD_B11D_TILINGKEY = 1000;
51
+ constexpr uint64_t B11D_TILINGKEY = 2000;
52
+ constexpr uint64_t B1SD_TILINGKEY = 3000;
53
+ constexpr uint64_t BN1D_TILINGKEY = 4000;
54
+ constexpr uint64_t BNSD_TILINGKEY = 5000;
55
+
56
+ static std::tuple<int64_t, int64_t, int64_t, int64_t> GetShapeTuple(
57
+ const gert::TilingContext* context, const int64_t index = 0)
58
+ {
59
+ const gert::StorageShape* shapePtr = context->GetInputShape(index);
60
+ OP_CHECK_IF(
61
+ shapePtr == nullptr, OP_LOGE(context, "Shape is nullptr."), return std::make_tuple(0, 0, 0, 0));
62
+
63
+ // check shape length is DIM_SIZE
64
+ OP_CHECK_IF(
65
+ shapePtr->GetStorageShape().GetDimNum() != DIM_SIZE,
66
+ OP_LOGE(context, "Shape must be (B,N,S,D)."), return std::make_tuple(0, 0, 0, 0));
67
+
68
+ return std::make_tuple(
69
+ shapePtr->GetStorageShape().GetDim(SHAPE_IDX_B), shapePtr->GetStorageShape().GetDim(SHAPE_IDX_N),
70
+ shapePtr->GetStorageShape().GetDim(SHAPE_IDX_S), shapePtr->GetStorageShape().GetDim(SHAPE_IDX_D));
71
+ }
72
+
73
+ ge::graphStatus InterleaveRopeTiling::GetPlatformInfo()
74
+ {
75
+ auto platformInfo = context_->GetPlatformInfo();
76
+ if (platformInfo == nullptr) {
77
+ auto compileInfoPtr = reinterpret_cast<const InterleaveRopeCompileInfo*>(context_->GetCompileInfo());
78
+ OP_CHECK_IF(
79
+ compileInfoPtr == nullptr, OP_LOGE(context_, "CompileInfo is nullptr."),
80
+ return ge::GRAPH_FAILED);
81
+ coreNum_ = compileInfoPtr->coreNum;
82
+ ubSize_ = compileInfoPtr->ubSize;
83
+ } else {
84
+ auto ascendcPlatform = platform_ascendc::PlatformAscendC(platformInfo);
85
+ coreNum_ = ascendcPlatform.GetCoreNumAiv();
86
+ uint64_t ubSize = 0;
87
+ ascendcPlatform.GetCoreMemSize(platform_ascendc::CoreMemType::UB, ubSize);
88
+ ubSize_ = ubSize;
89
+ }
90
+ return ge::GRAPH_SUCCESS;
91
+ }
92
+
93
+ ge::graphStatus InterleaveRopeTiling::GetShapeAttrsInfo()
94
+ {
95
+ auto xShapeTuple = GetShapeTuple(context_, X_INDEX);
96
+ auto cosShapeTuple = GetShapeTuple(context_, COS_INDEX);
97
+
98
+ batchSize_ = std::get<SHAPE_IDX_B>(xShapeTuple);
99
+ numHead_ = std::get<SHAPE_IDX_N>(xShapeTuple);
100
+ seqLength_ = std::get<SHAPE_IDX_S>(xShapeTuple);
101
+ hiddenDim_ = std::get<SHAPE_IDX_D>(xShapeTuple);
102
+ OP_CHECK_IF(
103
+ batchSize_ <= 0, OP_LOGE(context_, "batchSize should be greater than 0."),
104
+ return ge::GRAPH_FAILED);
105
+ OP_CHECK_IF(
106
+ numHead_ <= 0, OP_LOGE(context_, "numHead should be greater than 0."), return ge::GRAPH_FAILED);
107
+ OP_CHECK_IF(
108
+ seqLength_ <= 0, OP_LOGE(context_, "seqLength should be greater than 0."),
109
+ return ge::GRAPH_FAILED);
110
+ OP_CHECK_IF(
111
+ hiddenDim_ <= 0, OP_LOGE(context_, "hiddenDim should be greater than 0."),
112
+ return ge::GRAPH_FAILED);
113
+
114
+ tilingData_.set_batchSize(batchSize_);
115
+ tilingData_.set_numHead(numHead_);
116
+ tilingData_.set_seqLength(seqLength_);
117
+ tilingData_.set_hiddenDim(hiddenDim_);
118
+
119
+ // Check hiddenDim must be DEFAULT_HIDDEN_DIM
120
+ OP_CHECK_IF(
121
+ hiddenDim_ != DEFAULT_HIDDEN_DIM, OP_LOGE(context_, "hiddenDim must be 64."),
122
+ return ge::GRAPH_FAILED);
123
+ // Check cos sin shape must be same
124
+ OP_CHECK_IF(
125
+ cosShapeTuple != GetShapeTuple(context_, SIN_INDEX),
126
+ OP_LOGE(context_, "cos and sin shape must be the same."), return ge::GRAPH_FAILED);
127
+
128
+ // Check cosB must be batchSize
129
+ cosB_ = std::get<SHAPE_IDX_B>(cosShapeTuple);
130
+ OP_CHECK_IF(
131
+ cosB_ != batchSize_, OP_LOGE(context_, "cos or sin batchSize must same as x batchSize."),
132
+ return ge::GRAPH_FAILED);
133
+ // Check cosN must be 1
134
+ cosN_ = std::get<SHAPE_IDX_N>(cosShapeTuple);
135
+ OP_CHECK_IF(
136
+ cosN_ != 1, OP_LOGE(context_, "cos or sin numHead must be 1."), return ge::GRAPH_FAILED);
137
+ // Check cosS must 1
138
+ cosS_ = std::get<SHAPE_IDX_S>(cosShapeTuple);
139
+ OP_CHECK_IF(
140
+ cosS_ != seqLength_ && cosS_ != 1,
141
+ OP_LOGE(context_, "cos or sin seqLength must be 1 or the same as x seqLength."),
142
+ return ge::GRAPH_FAILED);
143
+ // Check cosD must be hiddenDim
144
+ int64_t cosD = std::get<SHAPE_IDX_D>(cosShapeTuple);
145
+ OP_CHECK_IF(
146
+ cosD != hiddenDim_, OP_LOGE(context_, "cos or sin hiddenDim must same as x hiddenDim."),
147
+ return ge::GRAPH_FAILED);
148
+
149
+ return ge::GRAPH_SUCCESS;
150
+ }
151
+
152
+ ge::graphStatus InterleaveRopeTiling::SplitBlockForFixBNS()
153
+ {
154
+ int64_t batchsPerBlock = DEFAULT_BATCH_PER_BLOCK;
155
+ tilingData_.set_batchsPerBlock(batchsPerBlock);
156
+ tilingData_.set_batchsLastBlock(batchsPerBlock);
157
+ tilingData_.set_numBlocks(batchSize_ / batchsPerBlock);
158
+ return ge::GRAPH_SUCCESS;
159
+ }
160
+
161
+ ge::graphStatus InterleaveRopeTiling::SplitBlockForBatch()
162
+ {
163
+ int64_t batchsPerBlock = Ops::Base::CeilDiv(batchSize_, static_cast<int64_t>(coreNum_));
164
+ int64_t needBlocks = Ops::Base::CeilDiv(batchSize_, batchsPerBlock);
165
+ int64_t batchsLastBlock = batchSize_ - (needBlocks - 1) * batchsPerBlock;
166
+ tilingData_.set_batchsPerBlock(batchsPerBlock);
167
+ tilingData_.set_batchsLastBlock(batchsLastBlock);
168
+ tilingData_.set_numBlocks(needBlocks);
169
+
170
+ if (tilingKey_ == B11D_TILINGKEY) {
171
+ int64_t NS = numHead_ * seqLength_;
172
+ tilingData_.set_hiddenDimCountPerBlock(NS);
173
+ tilingData_.set_hiddenDimCountLastBlock(NS);
174
+ SplitHiddenDim();
175
+ return ge::GRAPH_SUCCESS;
176
+ }
177
+
178
+ if (tilingKey_ == B1SD_TILINGKEY) {
179
+ tilingData_.set_hiddenDimCountPerBlock(seqLength_);
180
+ tilingData_.set_hiddenDimCountLastBlock(seqLength_);
181
+ SplitHiddenDim();
182
+ return ge::GRAPH_SUCCESS;
183
+ }
184
+
185
+ return ge::GRAPH_SUCCESS;
186
+ }
187
+
188
+ ge::graphStatus InterleaveRopeTiling::SplitBlockForNS()
189
+ {
190
+ tilingData_.set_batchsPerBlock(batchSize_);
191
+ tilingData_.set_batchsLastBlock(batchSize_);
192
+
193
+ int64_t NS = numHead_ * seqLength_;
194
+ int64_t NSPerBlock = Ops::Base::CeilDiv(NS, static_cast<int64_t>(coreNum_));
195
+ int64_t needBlocks = Ops::Base::CeilDiv(NS, NSPerBlock);
196
+ int64_t NSLastBlock = NS - (needBlocks - 1) * NSPerBlock;
197
+ tilingData_.set_numBlocks(needBlocks);
198
+
199
+ tilingData_.set_hiddenDimCountPerBlock(NSPerBlock);
200
+ tilingData_.set_hiddenDimCountLastBlock(NSLastBlock);
201
+ SplitHiddenDim();
202
+ return ge::GRAPH_SUCCESS;
203
+ }
204
+
205
+ ge::graphStatus InterleaveRopeTiling::SplitHiddenDim()
206
+ {
207
+ int64_t hiddenDimCountPerBlock = tilingData_.get_hiddenDimCountPerBlock();
208
+ int64_t hiddenDimCountLastBlock = tilingData_.get_hiddenDimCountLastBlock();
209
+
210
+ int64_t hiddenDimLoopsPerBlock;
211
+ int64_t hiddenDimCountPerLoopPerBlock;
212
+ int64_t hiddenDimCountLastLoopPerBlock;
213
+ SplitHiddenDimInblock(
214
+ hiddenDimCountPerBlock, hiddenDimLoopsPerBlock, hiddenDimCountPerLoopPerBlock, hiddenDimCountLastLoopPerBlock);
215
+ tilingData_.set_hiddenDimLoopsPerBlock(hiddenDimLoopsPerBlock);
216
+ tilingData_.set_hiddenDimCountPerLoopPerBlock(hiddenDimCountPerLoopPerBlock);
217
+ tilingData_.set_hiddenDimCountLastLoopPerBlock(hiddenDimCountLastLoopPerBlock);
218
+ OP_LOGD(
219
+ context_,
220
+ "hiddenDimLoopsPerBlock is: %ld, hiddenDimCountPerLoopPerBlock is: %ld, hiddenDimCountLastLoopPerBlock is: "
221
+ "%ld.",
222
+ hiddenDimLoopsPerBlock, hiddenDimCountPerLoopPerBlock, hiddenDimCountLastLoopPerBlock);
223
+
224
+ int64_t hiddenDimLoopsLastBlock;
225
+ int64_t hiddenDimCountPerLoopLastBlock;
226
+ int64_t hiddenDimCountLastLoopLastBlock;
227
+ SplitHiddenDimInblock(
228
+ hiddenDimCountLastBlock, hiddenDimLoopsLastBlock, hiddenDimCountPerLoopLastBlock,
229
+ hiddenDimCountLastLoopLastBlock);
230
+ tilingData_.set_hiddenDimLoopsLastBlock(hiddenDimLoopsLastBlock);
231
+ tilingData_.set_hiddenDimCountPerLoopLastBlock(hiddenDimCountPerLoopLastBlock);
232
+ tilingData_.set_hiddenDimCountLastLoopLastBlock(hiddenDimCountLastLoopLastBlock);
233
+ OP_LOGD(
234
+ context_,
235
+ "hiddenDimLoopsLastBlock is: %ld, hiddenDimCountPerLoopLastBlock is: %ld, hiddenDimCountLastLoopLastBlock is: "
236
+ "%ld.",
237
+ hiddenDimLoopsLastBlock, hiddenDimCountPerLoopLastBlock, hiddenDimCountLastLoopLastBlock);
238
+ return ge::GRAPH_SUCCESS;
239
+ }
240
+
241
+ ge::graphStatus InterleaveRopeTiling::SplitHiddenDimInblock(
242
+ int64_t hiddenDimCount, int64_t& hiddenDimLoops, int64_t& hiddenDimCountPerLoop, int64_t& hiddenDimCountLastLoop)
243
+ {
244
+ if (tilingKey_ == B11D_TILINGKEY) {
245
+ hiddenDimCountPerLoop =
246
+ (ubSize_ - hiddenDim_ * sizeof(float) * NUM_FOUR) / (hiddenDim_ * sizeof(float) * NUM_SIX) / NUM_TWO;
247
+ } else if (tilingKey_ == B1SD_TILINGKEY) {
248
+ hiddenDimCountPerLoop = (ubSize_) / (hiddenDim_ * sizeof(float) * NUM_TEN) / NUM_TWO;
249
+ }
250
+ hiddenDimCountPerLoop = hiddenDimCount > hiddenDimCountPerLoop ? hiddenDimCountPerLoop : hiddenDimCount;
251
+ hiddenDimLoops = Ops::Base::CeilDiv(hiddenDimCount, hiddenDimCountPerLoop);
252
+ hiddenDimCountLastLoop = hiddenDimCount - (hiddenDimLoops - 1) * hiddenDimCountPerLoop;
253
+ return ge::GRAPH_SUCCESS;
254
+ }
255
+
256
+ bool InterleaveRopeTiling::IsCapable()
257
+ {
258
+ return true;
259
+ }
260
+
261
+ ge::graphStatus InterleaveRopeTiling::DoOpTiling()
262
+ {
263
+ OP_LOGD(context_, "Start DoOpTiling.");
264
+
265
+ if (seqLength_ >= MIN_SEQUANCE_LEN && cosS_ == seqLength_) {
266
+ int64_t seqPerBlock = Ops::Base::CeilDiv(seqLength_, static_cast<int64_t>(coreNum_));
267
+ int64_t needBlocks = Ops::Base::CeilDiv(seqLength_, seqPerBlock);
268
+ tilingData_.set_numBlocks(needBlocks);
269
+ tilingKey_ = BN1D_TILINGKEY;
270
+ return ge::GRAPH_SUCCESS;
271
+ }
272
+
273
+ if (batchSize_ == DEFAULT_BATCH_SIZE && numHead_ == DEFAULT_NUM_HEAD && seqLength_ == 1 && cosN_ == 1 &&
274
+ cosS_ == 1) {
275
+ tilingKey_ = FIXED_BNSD_B11D_TILINGKEY;
276
+ SplitBlockForFixBNS();
277
+ } else if (cosN_ == 1 && cosS_ == 1) { // bnsd -> b11d
278
+ tilingKey_ = B11D_TILINGKEY;
279
+ if (batchSize_ >= static_cast<int64_t>(coreNum_) || batchSize_ >= numHead_ * seqLength_) {
280
+ tilingData_.set_splitAxis(SPLIT_BATCH);
281
+ SplitBlockForBatch();
282
+ } else {
283
+ tilingData_.set_splitAxis(SPLIT_NS);
284
+ SplitBlockForNS();
285
+ }
286
+ } else if (cosN_ == 1) { // bnsd -> b1sd
287
+ tilingKey_ = B1SD_TILINGKEY;
288
+ tilingData_.set_splitAxis(SPLIT_BATCH);
289
+ SplitBlockForBatch();
290
+ } else {
291
+ OP_LOGE(context_, "DoOpTiling failed, set tilingkey failed.");
292
+ return ge::GRAPH_FAILED;
293
+ }
294
+
295
+ OP_LOGD(context_, "DoOpTiling success.");
296
+ return ge::GRAPH_SUCCESS;
297
+ }
298
+
299
+ ge::graphStatus InterleaveRopeTiling::DoLibApiTiling()
300
+ {
301
+ return ge::GRAPH_SUCCESS;
302
+ }
303
+
304
+ ge::graphStatus InterleaveRopeTiling::GetWorkspaceSize()
305
+ {
306
+ return ge::GRAPH_SUCCESS;
307
+ }
308
+
309
+ uint64_t InterleaveRopeTiling::GetTilingKey() const
310
+ {
311
+ return tilingKey_;
312
+ }
313
+
314
+ ge::graphStatus InterleaveRopeTiling::PostTiling()
315
+ {
316
+ context_->SetTilingKey(GetTilingKey());
317
+ context_->SetBlockDim(tilingData_.get_numBlocks());
318
+ size_t* workspaces = context_->GetWorkspaceSizes(1);
319
+ workspaces[0] = DEFAULT_WORKSPACE_SIZE;
320
+ tilingData_.SaveToBuffer(context_->GetRawTilingData()->GetData(), context_->GetRawTilingData()->GetCapacity());
321
+ context_->GetRawTilingData()->SetDataSize(tilingData_.GetDataSize());
322
+ return ge::GRAPH_SUCCESS;
323
+ }
324
+ REGISTER_OPS_TILING_TEMPLATE(InterleaveRope, InterleaveRopeTiling, 1000);
325
+
326
+ ge::graphStatus Tiling4InterleaveRope(gert::TilingContext* context)
327
+ {
328
+ OP_LOGD(context, "TilingForInterleaveRope running.");
329
+ return Ops::Transformer::OpTiling::TilingRegistry::GetInstance().DoTilingImpl(context);
330
+ }
331
+
332
+ ge::graphStatus TilingPrepare4InterleaveRope(gert::TilingParseContext* context)
333
+ {
334
+ OP_LOGD(context, "TilingPrepare4InterleaveRope running.");
335
+ auto compileInfo = context->GetCompiledInfo<InterleaveRopeCompileInfo>();
336
+ OP_CHECK_NULL_WITH_CONTEXT(context, compileInfo);
337
+ auto platformInfo = context->GetPlatformInfo();
338
+ OP_CHECK_NULL_WITH_CONTEXT(context, platformInfo);
339
+ auto ascendcPlatform = platform_ascendc::PlatformAscendC(platformInfo);
340
+ compileInfo->coreNum = ascendcPlatform.GetCoreNumAiv();
341
+ OP_CHECK_IF(
342
+ compileInfo->coreNum <= 0, OP_LOGE(context, "coreNum must be greater than 0."),
343
+ return ge::GRAPH_FAILED);
344
+ uint64_t ubSize = 0;
345
+ ascendcPlatform.GetCoreMemSize(platform_ascendc::CoreMemType::UB, ubSize);
346
+ compileInfo->ubSize = ubSize;
347
+ OP_CHECK_IF(
348
+ compileInfo->ubSize <= 0, OP_LOGE(context, "ubSize must be greater than 0."),
349
+ return ge::GRAPH_FAILED);
350
+ OP_LOGD(context, "coreNum: %ld, ubSize: %ld", compileInfo->coreNum, compileInfo->ubSize);
351
+ OP_LOGD(context, "TilingPrepare4InterleaveRope success.");
352
+ return ge::GRAPH_SUCCESS;
353
+ }
354
+
355
+ IMPL_OP_OPTILING(InterleaveRope)
356
+ .Tiling(Tiling4InterleaveRope)
357
+ .TilingParse<InterleaveRopeCompileInfo>(TilingPrepare4InterleaveRope);
358
+
359
+ } // namespace optiling
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/interleave_rope_tiling.h ADDED
@@ -0,0 +1,104 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ /*!
12
+ * \file interleave_rope_tiling.h
13
+ * \brief
14
+ */
15
+ #ifndef OPS_BUILT_IN_OP_TILING_RUNTIME_INTERLEAVE_ROPE_H_
16
+ #define OPS_BUILT_IN_OP_TILING_RUNTIME_INTERLEAVE_ROPE_H_
17
+
18
+ #include "op_host/tiling_base.h"
19
+ #include "op_host/tiling_type.h"
20
+ #include "register/op_impl_registry.h"
21
+ #include "register/tilingdata_base.h"
22
+ #include "tiling/tiling_api.h"
23
+ #include "op_host/tiling_templates_registry.h"
24
+ #include "register/op_def_registry.h"
25
+ #include "log/log.h"
26
+ #include "platform/platform_infos_def.h"
27
+ #include "util/math_util.h"
28
+ #include "tiling/platform/platform_ascendc.h"
29
+
30
+ namespace optiling {
31
+ BEGIN_TILING_DATA_DEF(InterleaveRopeTilingData)
32
+ TILING_DATA_FIELD_DEF(int64_t, numBlocks);
33
+ TILING_DATA_FIELD_DEF(int64_t, splitAxis);
34
+
35
+ TILING_DATA_FIELD_DEF(int64_t, batchSize);
36
+ TILING_DATA_FIELD_DEF(int64_t, numHead);
37
+ TILING_DATA_FIELD_DEF(int64_t, seqLength);
38
+ TILING_DATA_FIELD_DEF(int64_t, hiddenDim);
39
+
40
+ TILING_DATA_FIELD_DEF(int64_t, batchsPerBlock);
41
+ TILING_DATA_FIELD_DEF(int64_t, batchsLastBlock);
42
+ TILING_DATA_FIELD_DEF(int64_t, batchLoops);
43
+ TILING_DATA_FIELD_DEF(int64_t, batchPerLoop);
44
+ TILING_DATA_FIELD_DEF(int64_t, batchLastLoop);
45
+
46
+ TILING_DATA_FIELD_DEF(int64_t, hiddenDimCountPerBlock);
47
+ TILING_DATA_FIELD_DEF(int64_t, hiddenDimCountLastBlock);
48
+ TILING_DATA_FIELD_DEF(int64_t, hiddenDimLoopsPerBlock);
49
+ TILING_DATA_FIELD_DEF(int64_t, hiddenDimCountPerLoopPerBlock);
50
+ TILING_DATA_FIELD_DEF(int64_t, hiddenDimCountLastLoopPerBlock);
51
+ TILING_DATA_FIELD_DEF(int64_t, hiddenDimLoopsLastBlock);
52
+ TILING_DATA_FIELD_DEF(int64_t, hiddenDimCountPerLoopLastBlock);
53
+ TILING_DATA_FIELD_DEF(int64_t, hiddenDimCountLastLoopLastBlock);
54
+
55
+ END_TILING_DATA_DEF;
56
+
57
+ REGISTER_TILING_DATA_CLASS(InterleaveRope, InterleaveRopeTilingData)
58
+
59
+ struct InterleaveRopeCompileInfo {
60
+ int64_t coreNum = 0;
61
+ int64_t ubSize = 0;
62
+ };
63
+
64
+ class InterleaveRopeTiling : public Ops::Transformer::OpTiling::TilingBaseClass {
65
+ public:
66
+ explicit InterleaveRopeTiling(gert::TilingContext* p_context_) : TilingBaseClass(p_context_)
67
+ {}
68
+ ~InterleaveRopeTiling() override = default;
69
+
70
+ uint64_t coreNum_ = 0;
71
+ uint64_t ubSize_ = 0;
72
+ int64_t batchSize_ = 0;
73
+ int64_t numHead_ = 0;
74
+ int64_t seqLength_ = 0;
75
+ int64_t hiddenDim_ = 0;
76
+ int64_t cosB_ = 0;
77
+ int64_t cosN_ = 0;
78
+ int64_t cosS_ = 0;
79
+ int64_t cosD_ = 0;
80
+
81
+ protected:
82
+ bool IsCapable() override;
83
+ ge::graphStatus DoOpTiling() override;
84
+ ge::graphStatus DoLibApiTiling() override;
85
+ uint64_t GetTilingKey() const override;
86
+ ge::graphStatus GetPlatformInfo() override;
87
+ ge::graphStatus GetShapeAttrsInfo() override;
88
+ ge::graphStatus GetWorkspaceSize() override;
89
+ ge::graphStatus PostTiling() override;
90
+ ge::graphStatus SplitBlockForFixBNS();
91
+ ge::graphStatus SplitBlockForBatch();
92
+ ge::graphStatus SplitBlockForNS();
93
+ ge::graphStatus SplitHiddenDim();
94
+ ge::graphStatus SplitHiddenDimInblock(
95
+ int64_t hiddenDimCount, int64_t& hiddenDimLoops, int64_t& hiddenDimCountPerLoop,
96
+ int64_t& hiddenDimCountLastLoop);
97
+
98
+ private:
99
+ InterleaveRopeTilingData tilingData_;
100
+ };
101
+
102
+ } // namespace optiling
103
+
104
+ #endif // OPS_BUILT_IN_OP_TILING_RUNTIME_INTERLEAVE_ROPE_H_
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/op_api/aclnn_interleave_rope.cpp ADDED
@@ -0,0 +1,52 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ #include "aclnn_interleave_rope.h"
12
+ #include "opdev/make_op_executor.h"
13
+ #include "opdev/op_dfx.h"
14
+ #include "opdev/op_executor.h"
15
+ #include "op_api/op_api_def.h"
16
+ #include "aclnn_kernels/common/op_error_check.h"
17
+ #include "external/aclnn_kernels/aclnn_platform.h"
18
+ #include "posembedding/rotary_position_embedding/op_host/op_api/aclnn_rotary_position_embedding.h"
19
+ #include "aclnnInner_interleave_rope.h"
20
+
21
+ using namespace op;
22
+
23
+ #ifdef __cplusplus
24
+ extern "C" {
25
+ #endif
26
+
27
+ constexpr int64_t HALF_INTERLEAVE_MODE = 3;
28
+
29
+ aclnnStatus aclnnInterleaveRopeGetWorkspaceSize(
30
+ const aclTensor* x, const aclTensor* cos, const aclTensor* sin, aclTensor* out, uint64_t* workspaceSize,
31
+ aclOpExecutor** executor)
32
+ {
33
+ if (Ops::Transformer::AclnnUtil::IsRegbase()) {
34
+ return aclnnRotaryPositionEmbeddingGetWorkspaceSize(
35
+ x, cos, sin, HALF_INTERLEAVE_MODE, out, workspaceSize, executor);
36
+ } else {
37
+ return aclnnInnerInterleaveRopeGetWorkspaceSize(x, cos, sin, out, workspaceSize, executor);
38
+ }
39
+ }
40
+
41
+ aclnnStatus aclnnInterleaveRope(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, aclrtStream stream)
42
+ {
43
+ if (Ops::Transformer::AclnnUtil::IsRegbase()) {
44
+ return aclnnRotaryPositionEmbedding(workspace, workspaceSize, executor, stream);
45
+ } else {
46
+ return aclnnInnerInterleaveRope(workspace, workspaceSize, executor, stream);
47
+ }
48
+ }
49
+
50
+ #ifdef __cplusplus
51
+ }
52
+ #endif
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/op_api/aclnn_interleave_rope.h ADDED
@@ -0,0 +1,58 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ #ifndef OP_API_INC_INTERLEAVE_ROPE_H_
12
+ #define OP_API_INC_INTERLEAVE_ROPE_H_
13
+
14
+ #include "aclnn/aclnn_base.h"
15
+ #include "aclnn_util.h"
16
+
17
+ #ifdef __cplusplus
18
+ extern "C" {
19
+ #endif
20
+
21
+ /**
22
+ * @brief aclnnInterleaveRope的第一段接口,根据具体的计算流程,计算workspace大小。
23
+ * @domain aclnn_ops_infer
24
+ *
25
+ * @param [in] x: npu
26
+ * device侧的aclTensor, 数据类型支持FLOAT32,FLOAT16,BFLOAT16,支持非连续的Tensor,数据格式支持ND。
27
+ * @param [in] sin: npu
28
+ * device侧的aclTensor, 数据类型支持FLOAT32,FLOAT16,BFLOAT16,数据类型与x一致,shape与x满足broadcast关系
29
+ * 支持非连续的Tensor,数据格式支持ND。
30
+ * @param [in] cos: npu
31
+ * device侧的aclTensor, 数据类型支持FLOAT32,FLOAT16,BFLOAT16,数据类型与x一致,shape与sin一致
32
+ * 支持非连续的Tensor,数据格式支持ND。
33
+ * @param [in] out: npu
34
+ * device侧的aclTensor, 数据类型支持FLOAT32,FLOAT16,BFLOAT16,且数据类型与x一致,
35
+ * shape与x相同, 数据格式支持ND, 且数据格式需要与x一致。
36
+ * @param [out] workspaceSize: 返回用户需要在npu device侧申请的workspace大小.
37
+ * @param [out] executor: 返回op执行器,包含算子计算流程。
38
+ * @return aclnnStatus: 返回状态码。
39
+ */
40
+ ACLNN_API aclnnStatus aclnnInterleaveRopeGetWorkspaceSize(
41
+ const aclTensor* x, const aclTensor* cos, const aclTensor* sin, aclTensor* out, uint64_t* workspaceSize,
42
+ aclOpExecutor** executor);
43
+
44
+ /* @brief aclnnInterleaveRope的第二段接口,用于执行计算.
45
+ * @param [in] workspace: 在npu device侧申请的workspace内存起址
46
+ * @param [in] workspace_size: 在npu device侧申请的workspace大小,由第一段接口aclnnBitwiseNotGetWorkspaceSize获取.
47
+ * @param [in] executor: op执行器,包含了算子计算流程.
48
+ * @param [in] stream: acl stream流.
49
+ * @return aclnnStatus: 返回状态码。
50
+ */
51
+ ACLNN_API aclnnStatus
52
+ aclnnInterleaveRope(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, aclrtStream stream);
53
+
54
+ #ifdef __cplusplus
55
+ }
56
+ #endif
57
+
58
+ #endif
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_host/op_api/aclnn_rotary_position_embedding.h ADDED
@@ -0,0 +1,58 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ #ifndef OP_API_INC_ROTARY_POSITION_EMBEDDING_H_
12
+ #define OP_API_INC_ROTARY_POSITION_EMBEDDING_H_
13
+
14
+ #include "aclnn/aclnn_base.h"
15
+ #include "aclnn_util.h"
16
+
17
+ #ifdef __cplusplus
18
+ extern "C" {
19
+ #endif
20
+
21
+ /**
22
+ * @brief aclnnRotaryPositionEmbedding的第一段接口,根据具体的计算流程,计算workspace大小。
23
+ * @domain aclnn_ops_infer
24
+ *
25
+ * @param [in] x: npu
26
+ * device侧的aclTensor, 数据类型支持FLOAT32,FLOAT16,BFLOAT16,支持非连续的Tensor,数据格式支持ND。
27
+ * @param [in] sin: npu
28
+ * device侧的aclTensor, 数据类型支持FLOAT32,FLOAT16,BFLOAT16,数据类型与x一致,shape与x满足broadcast关系
29
+ * 支持非连续的Tensor,数据格式支持ND。
30
+ * @param [in] cos: npu
31
+ * device侧的aclTensor, 数据类型支持FLOAT32,FLOAT16,BFLOAT16,数据类型与x一致,shape与sin一致
32
+ * 支持非连续的Tensor,数据格式支持ND。
33
+ * @param [in] mode: 支持int64_t类型,取值范围[0, 3],分别表示half,quarter,interleave,half-interleave四种模式.
34
+ * @param [in] out: npu
35
+ * device侧的aclTensor, 数据类型支持FLOAT32,FLOAT16,BFLOAT16,且数据类型与x一致,
36
+ * shape与x相同, 数据格式支持ND, 且数据格式需要与x一致。
37
+ * @param [out] workspaceSize: 返回用户需要在npu device侧申请的workspace大小。
38
+ * @param [out] executor: 返回op执行器,包含算子计算流程。
39
+ * @return aclnnStatus: 返回状态码。
40
+ */
41
+ ACLNN_API aclnnStatus aclnnRotaryPositionEmbeddingGetWorkspaceSize(const aclTensor* x, const aclTensor* cos,
42
+ const aclTensor* sin, int64_t mode, aclTensor* out,
43
+ uint64_t* workspaceSize, aclOpExecutor** executor);
44
+ /* @brief aclnnRotaryPositionEmbedding的第二段接口,用于执行计算.
45
+ * @param [in] workspace: 在npu device侧申请的workspace内存起址
46
+ * @param [in] workspace_size: 在npu device侧申请的workspace大小,由第一段接口aclnnBitwiseNotGetWorkspaceSize获取。
47
+ * @param [in] executor: op执行器,包含了算子计算流程。
48
+ * @param [in] stream: acl stream流。
49
+ * @return aclnnStatus: 返回状态码。
50
+ */
51
+ ACLNN_API aclnnStatus aclnnRotaryPositionEmbedding(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor,
52
+ aclrtStream stream);
53
+
54
+ #ifdef __cplusplus
55
+ }
56
+ #endif
57
+
58
+ #endif
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_kernel/interleave_rope.cpp ADDED
@@ -0,0 +1,50 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ /*!
12
+ * \file interleave_rope.cpp
13
+ * \brief
14
+ */
15
+
16
+ #include "interleave_rope_fixed_bnsd_b11d.h"
17
+ #include "interleave_rope_b11d.h"
18
+ #include "interleave_rope_b1sd.h"
19
+ #include "interleave_rope_split_s.h"
20
+
21
+ using namespace InterleaveRope;
22
+
23
+ #define INTERLEAVE_ROPE_FIXED_BNSD_B11D 1000
24
+ #define INTERLEAVE_ROPE_B11D 2000
25
+ #define INTERLEAVE_ROPE_B1SD 3000
26
+ #define INTERLEAVE_ROPE_SPLIT_S 4000
27
+
28
+ extern "C" __global__ __aicore__ void interleave_rope(
29
+ GM_ADDR x, GM_ADDR cos, GM_ADDR sin, GM_ADDR y, GM_ADDR workspace, GM_ADDR tiling)
30
+ {
31
+ TPipe pipe;
32
+ GET_TILING_DATA(tilingData, tiling);
33
+ if (TILING_KEY_IS(INTERLEAVE_ROPE_FIXED_BNSD_B11D)) {
34
+ KernelInterleaveRopeFixBNSD<DTYPE_X> op(&pipe, &tilingData);
35
+ op.Init(x, cos, sin, y);
36
+ op.Process();
37
+ } else if (TILING_KEY_IS(INTERLEAVE_ROPE_B11D)) {
38
+ KernelInterleaveRopeB11D<DTYPE_X> op(&pipe, &tilingData);
39
+ op.Init(x, cos, sin, y);
40
+ op.Process();
41
+ } else if (TILING_KEY_IS(INTERLEAVE_ROPE_B1SD)) {
42
+ KernelInterleaveRopeB1SD<DTYPE_X> op(&pipe, &tilingData);
43
+ op.Init(x, cos, sin, y);
44
+ op.Process();
45
+ } else if (TILING_KEY_IS(INTERLEAVE_ROPE_SPLIT_S)) {
46
+ KernelInterleaveRopeSplitS<DTYPE_X> op(&pipe, &tilingData);
47
+ op.Init(x, cos, sin, y);
48
+ op.Process();
49
+ }
50
+ }
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_kernel/interleave_rope_b11d.h ADDED
@@ -0,0 +1,204 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ /*!
12
+ * \file interleave_rope_b11d.h
13
+ * \brief
14
+ */
15
+ #ifndef _INTERLEAVE_ROPE_B11D_H_
16
+ #define _INTERLEAVE_ROPE_B11D_H_
17
+ #include "platform/platform_infos_def.h"
18
+ #include "kernel_operator.h"
19
+
20
+ namespace InterleaveRope {
21
+ using namespace AscendC;
22
+
23
+ constexpr uint64_t SPLIT_BATCH = 0;
24
+ constexpr uint64_t SPLIT_NS = 1;
25
+
26
+ template <typename T>
27
+ class KernelInterleaveRopeB11D {
28
+ public:
29
+ __aicore__ inline KernelInterleaveRopeB11D(TPipe* pipe, const InterleaveRopeTilingData* tiling)
30
+ : pipe_(pipe), tilingData_(tiling)
31
+ {}
32
+
33
+ __aicore__ inline void Init(GM_ADDR x, GM_ADDR cos, GM_ADDR sin, GM_ADDR y)
34
+ {
35
+ /*
36
+ * For each block, process
37
+ * x: [B, N, S, 64]
38
+ * cos: [B, 1, 1, 64]
39
+ * sin: [B, 1, 1, 64]
40
+ * y: [B, N, S, 64]
41
+ */
42
+
43
+ numHead_ = tilingData_->numHead;
44
+ seqLength_ = tilingData_->seqLength;
45
+ NS_ = numHead_ * seqLength_;
46
+
47
+ batchsPerBlock_ = tilingData_->batchsPerBlock;
48
+ curBlockBatchs_ = tilingData_->batchsPerBlock;
49
+ batchLoops_ = tilingData_->batchLoops;
50
+ batchPerLoop_ = tilingData_->batchPerLoop;
51
+ batchLastLoop_ = tilingData_->batchLastLoop;
52
+ hiddenDimLoops_ = tilingData_->hiddenDimLoopsPerBlock;
53
+ hiddenDimCountPerLoop_ = tilingData_->hiddenDimCountPerLoopPerBlock;
54
+ hiddenDimCountLastLoop_ = tilingData_->hiddenDimCountLastLoopPerBlock;
55
+ if (GetBlockIdx() == GetBlockNum() - 1) {
56
+ curBlockBatchs_ = tilingData_->batchsLastBlock;
57
+ hiddenDimLoops_ = tilingData_->hiddenDimLoopsLastBlock;
58
+ hiddenDimCountPerLoop_ = tilingData_->hiddenDimCountPerLoopLastBlock;
59
+ hiddenDimCountLastLoop_ = tilingData_->hiddenDimCountLastLoopLastBlock;
60
+ }
61
+
62
+ if (tilingData_->splitAxis == SPLIT_BATCH) {
63
+ sinCosIdxOffsetBase_ = GetBlockIdx() * batchsPerBlock_;
64
+ }
65
+
66
+ xGm.SetGlobalBuffer((__gm__ T*)x);
67
+ yGm.SetGlobalBuffer((__gm__ T*)y);
68
+ cosGm.SetGlobalBuffer((__gm__ T*)cos);
69
+ sinGm.SetGlobalBuffer((__gm__ T*)sin);
70
+
71
+ // init pipe
72
+ pipe_->InitBuffer(inQueueX, 1, hiddenDimCountPerLoop_ * hiddenDim * sizeof(T));
73
+ pipe_->InitBuffer(outQueueY, 1, hiddenDimCountPerLoop_ * hiddenDim * sizeof(float) * numTwo);
74
+ pipe_->InitBuffer(inQueueCos, 1, hiddenDim * sizeof(float) * numTwo);
75
+ pipe_->InitBuffer(inQueueSin, 1, hiddenDim * sizeof(float) * numTwo);
76
+
77
+ pipe_->InitBuffer(bufferReal, hiddenDimCountPerLoop_ * hiddenDimHalf * sizeof(float) * numTwo);
78
+ pipe_->InitBuffer(bufferImag, hiddenDimCountPerLoop_ * hiddenDimHalf * sizeof(float) * numTwo);
79
+ pipe_->InitBuffer(buffer_, hiddenDimCountPerLoop_ * hiddenDim * sizeof(float));
80
+ }
81
+
82
+ __aicore__ inline void Process()
83
+ {
84
+ for (int64_t i = 0; i < curBlockBatchs_; i++) {
85
+ Rope(i);
86
+ }
87
+ }
88
+
89
+ __aicore__ inline void Rope(int64_t idx)
90
+ {
91
+ int64_t sinCosIdxOffset = sinCosIdxOffsetBase_ + idx;
92
+ // rope
93
+ LocalTensor<float> cosLocal = inQueueCos.AllocTensor<float>();
94
+ DataCopy(cosLocal[hiddenDim].template ReinterpretCast<T>(), cosGm[sinCosIdxOffset * hiddenDim], hiddenDim);
95
+
96
+ inQueueCos.EnQue(cosLocal);
97
+ cosLocal = inQueueCos.DeQue<float>();
98
+ Cast(cosLocal, cosLocal[hiddenDim].template ReinterpretCast<T>(), RoundMode::CAST_NONE, hiddenDim);
99
+
100
+ LocalTensor<float> sinLocal = inQueueSin.AllocTensor<float>();
101
+ DataCopy(sinLocal[hiddenDim].template ReinterpretCast<T>(), sinGm[sinCosIdxOffset * hiddenDim], hiddenDim);
102
+ inQueueSin.EnQue(sinLocal);
103
+ sinLocal = inQueueSin.DeQue<float>();
104
+ Cast(sinLocal, sinLocal[hiddenDim].template ReinterpretCast<T>(), RoundMode::CAST_NONE, hiddenDim);
105
+ int64_t factor = hiddenDimCountPerLoop_;
106
+ int64_t batchOffset = tilingData_->splitAxis == SPLIT_BATCH ?
107
+ (GetBlockIdx() * batchsPerBlock_ + idx) * NS_ :
108
+ idx * NS_ + GetBlockIdx() * tilingData_->hiddenDimCountPerBlock;
109
+ for (int i = 0; i < hiddenDimLoops_; i++) {
110
+ if (i == hiddenDimLoops_ - 1) {
111
+ factor = hiddenDimCountLastLoop_;
112
+ }
113
+ int64_t xOffset = (batchOffset + i * hiddenDimCountPerLoop_) * hiddenDim;
114
+
115
+ // load xLocal
116
+ LocalTensor<T> xLocal = inQueueX.AllocTensor<T>();
117
+ DataCopy(xLocal, xGm[xOffset], factor * hiddenDim);
118
+ inQueueX.EnQue(xLocal);
119
+ xLocal = inQueueX.DeQue<T>();
120
+
121
+ // split Real and Imag
122
+ LocalTensor<float> realLocal = bufferReal.Get<float>();
123
+ LocalTensor<float> imagLocal = bufferImag.Get<float>();
124
+ LocalTensor<float> buf_ = buffer_.Get<float>();
125
+ uint64_t rsvdCnt = 0;
126
+ GatherMask(
127
+ realLocal[hiddenDimCountPerLoop_ * hiddenDimHalf].template ReinterpretCast<T>(), xLocal, 1, true,
128
+ factor * hiddenDim, {1, 1, 8, 0}, rsvdCnt);
129
+ GatherMask(
130
+ imagLocal[hiddenDimCountPerLoop_ * hiddenDimHalf].template ReinterpretCast<T>(), xLocal, numTwo, true,
131
+ factor * hiddenDim, {1, 1, 8, 0}, rsvdCnt);
132
+ Cast(
133
+ realLocal, realLocal[hiddenDimCountPerLoop_ * hiddenDimHalf].template ReinterpretCast<T>(),
134
+ RoundMode::CAST_NONE, factor * hiddenDimHalf);
135
+ Cast(
136
+ imagLocal, imagLocal[hiddenDimCountPerLoop_ * hiddenDimHalf].template ReinterpretCast<T>(),
137
+ RoundMode::CAST_NONE, factor * hiddenDimHalf);
138
+ inQueueX.FreeTensor(xLocal);
139
+
140
+ uint64_t mask[numTwo] = {0xffffffff, 0}; // mask hiddenDimHalf Elements
141
+ LocalTensor<float> outLocal = outQueueY.AllocTensor<float>();
142
+ Mul(outLocal, realLocal, cosLocal, mask, factor * numTwo, {1, 1, 1, 8, 4, 0});
143
+ Mul(outLocal[hiddenDimHalf], imagLocal, cosLocal[hiddenDimHalf], mask, factor * numTwo, {1, 1, 1, 8, 4, 0});
144
+ PipeBarrier<PIPE_V>();
145
+
146
+ Muls<float>(imagLocal, imagLocal, -1.0f, factor * hiddenDimHalf);
147
+ PipeBarrier<PIPE_V>();
148
+ Mul(buf_, imagLocal, sinLocal, mask, factor * numTwo, {1, 1, 1, 8, 4, 0});
149
+ Mul(buf_[hiddenDimHalf], realLocal, sinLocal[hiddenDimHalf], mask, factor * numTwo, {1, 1, 1, 8, 4, 0});
150
+ PipeBarrier<PIPE_V>();
151
+
152
+ Add(outLocal, outLocal, buf_, factor * hiddenDim);
153
+ PipeBarrier<PIPE_V>();
154
+ Cast(
155
+ outLocal[hiddenDimCountPerLoop_ * hiddenDim].template ReinterpretCast<T>(), outLocal,
156
+ RoundMode::CAST_RINT, factor * hiddenDim);
157
+ PipeBarrier<PIPE_V>();
158
+
159
+ outQueueY.EnQue(outLocal);
160
+ outLocal = outQueueY.DeQue<float>();
161
+ DataCopy(
162
+ yGm[xOffset], outLocal[hiddenDimCountPerLoop_ * hiddenDim].template ReinterpretCast<T>(),
163
+ factor * hiddenDim);
164
+ outQueueY.FreeTensor(outLocal);
165
+ }
166
+ inQueueCos.FreeTensor(cosLocal);
167
+ inQueueSin.FreeTensor(sinLocal);
168
+ }
169
+
170
+ private:
171
+ TPipe* pipe_ = nullptr;
172
+ const InterleaveRopeTilingData* tilingData_;
173
+ GlobalTensor<T> xGm;
174
+ GlobalTensor<T> yGm;
175
+ GlobalTensor<T> cosGm;
176
+ GlobalTensor<T> sinGm;
177
+
178
+ TQue<QuePosition::VECIN, 1> inQueueX;
179
+ TQue<QuePosition::VECIN, 1> inQueueCos;
180
+ TQue<QuePosition::VECIN, 1> inQueueSin;
181
+ TQue<QuePosition::VECOUT, 1> outQueueY;
182
+ TBuf<TPosition::VECCALC> bufferReal;
183
+ TBuf<TPosition::VECCALC> bufferImag;
184
+ TBuf<TPosition::VECCALC> buffer_;
185
+ int64_t numHead_ = 0;
186
+ int64_t seqLength_ = 0;
187
+ int64_t NS_ = 0;
188
+
189
+ constexpr static int64_t hiddenDim = 64;
190
+ constexpr static int64_t hiddenDimHalf = 32;
191
+ constexpr static int64_t numTwo = 2;
192
+ int64_t batchLoops_ = 0;
193
+ int64_t batchPerLoop_ = 0;
194
+ int64_t batchLastLoop_ = 0;
195
+ int64_t batchsPerBlock_ = 0;
196
+ int64_t curBlockBatchs_ = 0;
197
+ int64_t hiddenDimLoops_ = 0;
198
+ int64_t hiddenDimCountPerLoop_ = 0;
199
+ int64_t hiddenDimCountLastLoop_ = 0;
200
+ int64_t sinCosIdxOffsetBase_ = 0;
201
+ };
202
+ } // namespace InterleaveRope
203
+
204
+ #endif // _INTERLEAVE_ROPE_B11D_H_
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_kernel/interleave_rope_b1sd.h ADDED
@@ -0,0 +1,205 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ /*!
12
+ * \file interleave_rope_b1sd.h
13
+ * \brief
14
+ */
15
+ #ifndef _INTERLEAVE_ROPE_B1SD_H_
16
+ #define _INTERLEAVE_ROPE_B1SD_H_
17
+ #include "platform/platform_infos_def.h"
18
+ #include "kernel_operator.h"
19
+
20
+ namespace InterleaveRope {
21
+ using namespace AscendC;
22
+
23
+ template <typename T>
24
+ class KernelInterleaveRopeB1SD {
25
+ public:
26
+ __aicore__ inline KernelInterleaveRopeB1SD(TPipe* pipe, const InterleaveRopeTilingData* tiling)
27
+ : pipe_(pipe), tilingData_(tiling)
28
+ {}
29
+
30
+ __aicore__ inline void Init(GM_ADDR x, GM_ADDR cos, GM_ADDR sin, GM_ADDR y)
31
+ {
32
+ /*
33
+ * For each block, process
34
+ * x: [B, N, S, 64]
35
+ * cos: [B, 1, S, 64]
36
+ * sin: [B, 1, S, 64]
37
+ * y: [B, N, S, 64]
38
+ */
39
+
40
+ numHead_ = tilingData_->numHead;
41
+ seqLength_ = tilingData_->seqLength;
42
+ NS_ = numHead_ * seqLength_;
43
+
44
+ batchsPerBlock_ = tilingData_->batchsPerBlock;
45
+ curBlockBatchs_ = tilingData_->batchsPerBlock;
46
+ batchLoops_ = tilingData_->batchLoops;
47
+ batchPerLoop_ = tilingData_->batchPerLoop;
48
+ batchLastLoop_ = tilingData_->batchLastLoop;
49
+ hiddenDimLoops_ = tilingData_->hiddenDimLoopsPerBlock;
50
+ hiddenDimCountPerLoop_ = tilingData_->hiddenDimCountPerLoopPerBlock;
51
+ hiddenDimCountLastLoop_ = tilingData_->hiddenDimCountLastLoopPerBlock;
52
+ if (GetBlockIdx() == GetBlockNum() - 1) {
53
+ curBlockBatchs_ = tilingData_->batchsLastBlock;
54
+ hiddenDimLoops_ = tilingData_->hiddenDimLoopsLastBlock;
55
+ hiddenDimCountPerLoop_ = tilingData_->hiddenDimCountPerLoopLastBlock;
56
+ hiddenDimCountLastLoop_ = tilingData_->hiddenDimCountLastLoopLastBlock;
57
+ }
58
+
59
+ xGm.SetGlobalBuffer((__gm__ T*)x);
60
+ yGm.SetGlobalBuffer((__gm__ T*)y);
61
+ cosGm.SetGlobalBuffer((__gm__ T*)cos);
62
+ sinGm.SetGlobalBuffer((__gm__ T*)sin);
63
+
64
+ // init pipe
65
+ pipe_->InitBuffer(inQueueX, 1, hiddenDimCountPerLoop_ * hiddenDim * sizeof(T));
66
+ pipe_->InitBuffer(outQueueY, 1, hiddenDimCountPerLoop_ * hiddenDim * sizeof(float) * numTwo);
67
+ pipe_->InitBuffer(inQueueCos, 1, hiddenDimCountPerLoop_ * hiddenDim * sizeof(float) * numTwo);
68
+ pipe_->InitBuffer(inQueueSin, 1, hiddenDimCountPerLoop_ * hiddenDim * sizeof(float) * numTwo);
69
+
70
+ pipe_->InitBuffer(bufferReal, hiddenDimCountPerLoop_ * hiddenDimHalf * sizeof(float) * numTwo);
71
+ pipe_->InitBuffer(bufferImag, hiddenDimCountPerLoop_ * hiddenDimHalf * sizeof(float) * numTwo);
72
+ pipe_->InitBuffer(buffer_, hiddenDimCountPerLoop_ * hiddenDim * sizeof(float));
73
+ }
74
+
75
+ __aicore__ inline void Process()
76
+ {
77
+ for (int64_t i = 0; i < curBlockBatchs_; i++) {
78
+ Rope(i);
79
+ }
80
+ }
81
+
82
+ __aicore__ inline void Rope(int64_t idx)
83
+ {
84
+ int64_t hiddenDimCount = hiddenDimCountPerLoop_;
85
+ for (int64_t hiddenDimLoop = 0; hiddenDimLoop < hiddenDimLoops_; hiddenDimLoop++) {
86
+ if (hiddenDimLoop == hiddenDimLoops_ - 1) {
87
+ hiddenDimCount = hiddenDimCountLastLoop_;
88
+ }
89
+ int64_t sinCosIdxOffset = ((GetBlockIdx() * tilingData_->batchsPerBlock + idx) * seqLength_ +
90
+ hiddenDimLoop * hiddenDimCountPerLoop_) *
91
+ hiddenDim;
92
+ LocalTensor<float> cosLocal = inQueueCos.AllocTensor<float>();
93
+ DataCopy(
94
+ cosLocal[hiddenDimCount * hiddenDim].template ReinterpretCast<T>(), cosGm[sinCosIdxOffset],
95
+ hiddenDimCount * hiddenDim);
96
+
97
+ inQueueCos.EnQue(cosLocal);
98
+ cosLocal = inQueueCos.DeQue<float>();
99
+ Cast(
100
+ cosLocal, cosLocal[hiddenDimCount * hiddenDim].template ReinterpretCast<T>(), RoundMode::CAST_NONE,
101
+ hiddenDimCount * hiddenDim);
102
+
103
+ LocalTensor<float> sinLocal = inQueueSin.AllocTensor<float>();
104
+ DataCopy(
105
+ sinLocal[hiddenDimCount * hiddenDim].template ReinterpretCast<T>(), sinGm[sinCosIdxOffset],
106
+ hiddenDimCount * hiddenDim);
107
+ inQueueSin.EnQue(sinLocal);
108
+ sinLocal = inQueueSin.DeQue<float>();
109
+ Cast(
110
+ sinLocal, sinLocal[hiddenDimCount * hiddenDim].template ReinterpretCast<T>(), RoundMode::CAST_NONE,
111
+ hiddenDimCount * hiddenDim);
112
+
113
+ int64_t batchOffset = (GetBlockIdx() * tilingData_->batchsPerBlock + idx) * numHead_ * seqLength_;
114
+ for (int64_t nIdx = 0; nIdx < numHead_; nIdx++) {
115
+ int64_t xOffset =
116
+ (batchOffset + nIdx * seqLength_ + hiddenDimLoop * hiddenDimCountPerLoop_) * hiddenDim;
117
+ // load xLocal
118
+ LocalTensor<T> xLocal = inQueueX.AllocTensor<T>();
119
+ DataCopy(xLocal, xGm[xOffset], hiddenDimCount * hiddenDim);
120
+ inQueueX.EnQue(xLocal);
121
+ xLocal = inQueueX.DeQue<T>();
122
+
123
+ // split Real and Imag
124
+ LocalTensor<float> realLocal = bufferReal.Get<float>();
125
+ LocalTensor<float> imagLocal = bufferImag.Get<float>();
126
+ LocalTensor<float> buf_ = buffer_.Get<float>();
127
+ uint64_t rsvdCnt = 0;
128
+ GatherMask(
129
+ realLocal[hiddenDimCount * hiddenDimHalf].template ReinterpretCast<T>(), xLocal, 1, true,
130
+ hiddenDimCount * hiddenDim, {1, 1, 8, 0}, rsvdCnt);
131
+ GatherMask(
132
+ imagLocal[hiddenDimCount * hiddenDimHalf].template ReinterpretCast<T>(), xLocal, numTwo, true,
133
+ hiddenDimCount * hiddenDim, {1, 1, 8, 0}, rsvdCnt);
134
+ Cast(
135
+ realLocal, realLocal[hiddenDimCount * hiddenDimHalf].template ReinterpretCast<T>(),
136
+ RoundMode::CAST_NONE, hiddenDimCount * hiddenDimHalf);
137
+ Cast(
138
+ imagLocal, imagLocal[hiddenDimCount * hiddenDimHalf].template ReinterpretCast<T>(),
139
+ RoundMode::CAST_NONE, hiddenDimCount * hiddenDimHalf);
140
+ inQueueX.FreeTensor(xLocal);
141
+
142
+ uint64_t mask[numTwo] = {0xffffffff, 0}; // mask hiddenDimHalf Elements
143
+ LocalTensor<float> outLocal = outQueueY.AllocTensor<float>();
144
+ Mul(outLocal, realLocal, cosLocal, mask, hiddenDimCount, {1, 1, 1, 8, 4, 8});
145
+ Mul(outLocal[hiddenDimHalf], imagLocal, cosLocal[hiddenDimHalf], mask, hiddenDimCount,
146
+ {1, 1, 1, 8, 4, 8});
147
+ PipeBarrier<PIPE_V>();
148
+ Muls<float>(imagLocal, imagLocal, -1.0f, hiddenDimCount * hiddenDimHalf);
149
+ PipeBarrier<PIPE_V>();
150
+ Mul(buf_, imagLocal, sinLocal, mask, hiddenDimCount, {1, 1, 1, 8, 4, 8});
151
+ Mul(buf_[hiddenDimHalf], realLocal, sinLocal[hiddenDimHalf], mask, hiddenDimCount, {1, 1, 1, 8, 4, 8});
152
+ PipeBarrier<PIPE_V>();
153
+ Add(outLocal, outLocal, buf_, hiddenDimCount * hiddenDim);
154
+ PipeBarrier<PIPE_V>();
155
+ Cast(
156
+ outLocal[hiddenDimCount * hiddenDim].template ReinterpretCast<T>(), outLocal, RoundMode::CAST_RINT,
157
+ hiddenDimCount * hiddenDim);
158
+ PipeBarrier<PIPE_V>();
159
+ outQueueY.EnQue(outLocal);
160
+ outLocal = outQueueY.DeQue<float>();
161
+ DataCopy(
162
+ yGm[xOffset], outLocal[hiddenDimCount * hiddenDim].template ReinterpretCast<T>(),
163
+ hiddenDimCount * hiddenDim);
164
+ outQueueY.FreeTensor(outLocal);
165
+ }
166
+ inQueueCos.FreeTensor(cosLocal);
167
+ inQueueSin.FreeTensor(sinLocal);
168
+ }
169
+ }
170
+
171
+ private:
172
+ TPipe* pipe_ = nullptr;
173
+ const InterleaveRopeTilingData* tilingData_;
174
+ GlobalTensor<T> xGm;
175
+ GlobalTensor<T> yGm;
176
+ GlobalTensor<T> cosGm;
177
+ GlobalTensor<T> sinGm;
178
+
179
+ TQue<QuePosition::VECIN, 1> inQueueX;
180
+ TQue<QuePosition::VECIN, 1> inQueueCos;
181
+ TQue<QuePosition::VECIN, 1> inQueueSin;
182
+ TQue<QuePosition::VECOUT, 1> outQueueY;
183
+ TBuf<TPosition::VECCALC> bufferReal;
184
+ TBuf<TPosition::VECCALC> bufferImag;
185
+ TBuf<TPosition::VECCALC> buffer_;
186
+ int64_t numHead_ = 0;
187
+ int64_t seqLength_ = 0;
188
+ int64_t NS_ = 0;
189
+
190
+ constexpr static int64_t hiddenDim = 64;
191
+ constexpr static int64_t hiddenDimHalf = 32;
192
+ constexpr static int64_t numTwo = 2;
193
+ int64_t batchLoops_ = 0;
194
+ int64_t batchPerLoop_ = 0;
195
+ int64_t batchLastLoop_ = 0;
196
+
197
+ int64_t batchsPerBlock_ = 0;
198
+ int64_t curBlockBatchs_ = 0;
199
+ int64_t hiddenDimLoops_ = 0;
200
+ int64_t hiddenDimCountPerLoop_ = 0;
201
+ int64_t hiddenDimCountLastLoop_ = 0;
202
+ };
203
+ } // namespace InterleaveRope
204
+
205
+ #endif // _INTERLEAVE_ROPE_B1SD_H_
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_kernel/interleave_rope_fixed_bnsd_b11d.h ADDED
@@ -0,0 +1,188 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ /*!
12
+ * \file interleave_rope_fixed_bnsd_b11d.h
13
+ * \brief
14
+ */
15
+ #ifndef _INTERLEAVE_ROPE_FIX_BNSD_B11D_H_
16
+ #define _INTERLEAVE_ROPE_FIX_BNSD_B11D_H_
17
+ #include "platform/platform_infos_def.h"
18
+ #include "kernel_operator.h"
19
+
20
+ namespace InterleaveRope {
21
+ using namespace AscendC;
22
+
23
+ template <typename T>
24
+ class KernelInterleaveRopeFixBNSD {
25
+ public:
26
+ __aicore__ inline KernelInterleaveRopeFixBNSD(TPipe* pipe, const InterleaveRopeTilingData* tiling)
27
+ : pipe_(pipe), tilingData_(tiling)
28
+ {}
29
+
30
+ __aicore__ inline void Init(GM_ADDR x, GM_ADDR cos, GM_ADDR sin, GM_ADDR y)
31
+ {
32
+ /*
33
+ * original shape:
34
+ * x: [32, 32, 1, 64]
35
+ * cos: [32, 1, 1, 64]
36
+ * sin: [32, 1, 1, 64]
37
+ * y: [32, 32, 1, 64]
38
+ * For each block, process
39
+ * x: [4, 32, 1, 64]
40
+ * cos: [4, 1, 1, 64]
41
+ * sin: [4, 1, 1, 64]
42
+ * y: [4, 32, 1, 64]
43
+ */
44
+
45
+ // init global memory
46
+ xGm.SetGlobalBuffer(
47
+ (__gm__ T*)x + GetBlockIdx() * tilingData_->batchsPerBlock * numHead * hiddenDim,
48
+ tilingData_->batchsPerBlock * numHead * hiddenDim);
49
+ cosGm.SetGlobalBuffer(
50
+ (__gm__ T*)cos + GetBlockIdx() * tilingData_->batchsPerBlock * hiddenDim,
51
+ tilingData_->batchsPerBlock * hiddenDim);
52
+ sinGm.SetGlobalBuffer(
53
+ (__gm__ T*)sin + GetBlockIdx() * tilingData_->batchsPerBlock * hiddenDim,
54
+ tilingData_->batchsPerBlock * hiddenDim);
55
+ yGm.SetGlobalBuffer(
56
+ (__gm__ T*)y + GetBlockIdx() * tilingData_->batchsPerBlock * numHead * hiddenDim,
57
+ tilingData_->batchsPerBlock * numHead * hiddenDim);
58
+
59
+ // init pipe
60
+ pipe_->InitBuffer(inQueueX, 1, tilingData_->batchsPerBlock * numHead * hiddenDim * sizeof(T));
61
+ pipe_->InitBuffer(outQueueY, 1, tilingData_->batchsPerBlock * numHead * hiddenDim * sizeof(float) * numTwo);
62
+
63
+ pipe_->InitBuffer(inQueueCos, 1, tilingData_->batchsPerBlock * hiddenDim * sizeof(float) * numTwo);
64
+ pipe_->InitBuffer(inQueueSin, 1, tilingData_->batchsPerBlock * hiddenDim * sizeof(float) * numTwo);
65
+
66
+ pipe_->InitBuffer(bufferReal, tilingData_->batchsPerBlock * numHead * hiddenDimHalf * sizeof(float) * numTwo);
67
+ pipe_->InitBuffer(bufferImag, tilingData_->batchsPerBlock * numHead * hiddenDimHalf * sizeof(float) * numTwo);
68
+ pipe_->InitBuffer(buffer_, tilingData_->batchsPerBlock * numHead * hiddenDim * sizeof(float) * numTwo);
69
+ }
70
+
71
+ __aicore__ inline void Process()
72
+ {
73
+ Rope();
74
+ }
75
+
76
+ __aicore__ inline void Rope()
77
+ {
78
+ // load
79
+ LocalTensor<T> xLocal = inQueueX.AllocTensor<T>();
80
+ DataCopy(xLocal, xGm, tilingData_->batchsPerBlock * numHead * hiddenDim);
81
+ inQueueX.EnQue(xLocal);
82
+ xLocal = inQueueX.DeQue<T>();
83
+
84
+ // split Real and Imag
85
+ LocalTensor<float> realLocal = bufferReal.Get<float>();
86
+ LocalTensor<float> imagLocal = bufferImag.Get<float>();
87
+ LocalTensor<float> buf_ = buffer_.Get<float>();
88
+ uint64_t rsvdCnt = 0;
89
+ GatherMask(
90
+ realLocal[tilingData_->batchsPerBlock * numHead * hiddenDimHalf].template ReinterpretCast<T>(), xLocal, 1,
91
+ true, tilingData_->batchsPerBlock * numHead * hiddenDim, {1, 1, 8, 0}, rsvdCnt);
92
+ GatherMask(
93
+ imagLocal[tilingData_->batchsPerBlock * numHead * hiddenDimHalf].template ReinterpretCast<T>(), xLocal,
94
+ numTwo, true, tilingData_->batchsPerBlock * numHead * hiddenDim, {1, 1, 8, 0}, rsvdCnt);
95
+ PipeBarrier<PIPE_V>();
96
+ Cast(
97
+ realLocal, realLocal[tilingData_->batchsPerBlock * numHead * hiddenDimHalf].template ReinterpretCast<T>(),
98
+ RoundMode::CAST_NONE, tilingData_->batchsPerBlock * numHead * hiddenDimHalf);
99
+ Cast(
100
+ imagLocal, imagLocal[tilingData_->batchsPerBlock * numHead * hiddenDimHalf].template ReinterpretCast<T>(),
101
+ RoundMode::CAST_NONE, tilingData_->batchsPerBlock * numHead * hiddenDimHalf);
102
+
103
+ inQueueX.FreeTensor(xLocal);
104
+
105
+ // rope
106
+ LocalTensor<float> cosLocal = inQueueCos.AllocTensor<float>();
107
+ DataCopy(
108
+ cosLocal[tilingData_->batchsPerBlock * hiddenDim].template ReinterpretCast<T>(), cosGm,
109
+ tilingData_->batchsPerBlock * hiddenDim);
110
+
111
+ inQueueCos.EnQue(cosLocal);
112
+ cosLocal = inQueueCos.DeQue<float>();
113
+
114
+ Cast(
115
+ cosLocal, cosLocal[tilingData_->batchsPerBlock * hiddenDim].template ReinterpretCast<T>(),
116
+ RoundMode::CAST_NONE, tilingData_->batchsPerBlock * hiddenDim);
117
+ PipeBarrier<PIPE_V>();
118
+ uint64_t mask[numTwo] = {0xffffffff, 0};
119
+ LocalTensor<float> outLocal = outQueueY.AllocTensor<float>();
120
+ for (int64_t i = 0; i < tilingData_->batchsPerBlock; i++) {
121
+ Mul(outLocal[i * numHead * hiddenDim], realLocal[i * numHead * hiddenDimHalf], cosLocal[i * hiddenDim],
122
+ mask, hiddenDim, {1, 1, 1, 8, 4, 0});
123
+ Mul(outLocal[i * numHead * hiddenDim + numHead], imagLocal[i * numHead * hiddenDimHalf],
124
+ cosLocal[i * hiddenDim + hiddenDimHalf], mask, hiddenDim, {1, 1, 1, 8, 4, 0});
125
+ }
126
+ PipeBarrier<PIPE_V>();
127
+ inQueueCos.FreeTensor(cosLocal);
128
+
129
+ LocalTensor<float> sinLocal = inQueueSin.AllocTensor<float>();
130
+
131
+ DataCopy(
132
+ sinLocal[tilingData_->batchsPerBlock * hiddenDim].template ReinterpretCast<T>(), sinGm,
133
+ tilingData_->batchsPerBlock * hiddenDim);
134
+
135
+ inQueueSin.EnQue(sinLocal);
136
+ sinLocal = inQueueSin.DeQue<float>();
137
+ Cast(
138
+ sinLocal, sinLocal[tilingData_->batchsPerBlock * hiddenDim].template ReinterpretCast<T>(),
139
+ RoundMode::CAST_NONE, tilingData_->batchsPerBlock * hiddenDim);
140
+ PipeBarrier<PIPE_V>();
141
+ Muls<float>(imagLocal, imagLocal, -1.0f, tilingData_->batchsPerBlock * numHead * hiddenDimHalf);
142
+ PipeBarrier<PIPE_V>();
143
+ for (int64_t i = 0; i < tilingData_->batchsPerBlock; i++) {
144
+ Mul(buf_[i * numHead * hiddenDim], imagLocal[i * numHead * hiddenDimHalf], sinLocal[i * hiddenDim], mask,
145
+ hiddenDim, {1, 1, 1, 8, 4, 0});
146
+ Mul(buf_[i * numHead * hiddenDim + hiddenDimHalf], realLocal[i * numHead * hiddenDimHalf],
147
+ sinLocal[i * hiddenDim + hiddenDimHalf], mask, hiddenDim, {1, 1, 1, 8, 4, 0});
148
+ }
149
+ PipeBarrier<PIPE_V>();
150
+ inQueueSin.FreeTensor(sinLocal);
151
+ Add(outLocal, outLocal, buf_, tilingData_->batchsPerBlock * numHead * hiddenDim);
152
+ PipeBarrier<PIPE_V>();
153
+ Cast(
154
+ outLocal[tilingData_->batchsPerBlock * numHead * hiddenDim].template ReinterpretCast<T>(), outLocal,
155
+ RoundMode::CAST_RINT, tilingData_->batchsPerBlock * numHead * hiddenDim);
156
+ outQueueY.EnQue(outLocal);
157
+ outLocal = outQueueY.DeQue<float>();
158
+ DataCopy(
159
+ yGm, outLocal[tilingData_->batchsPerBlock * numHead * hiddenDim].template ReinterpretCast<T>(),
160
+ tilingData_->batchsPerBlock * numHead * hiddenDim);
161
+ outQueueY.FreeTensor(outLocal);
162
+ }
163
+
164
+ private:
165
+ TPipe* pipe_ = nullptr;
166
+ const InterleaveRopeTilingData* tilingData_;
167
+ GlobalTensor<T> xGm;
168
+ GlobalTensor<T> yGm;
169
+ GlobalTensor<T> cosGm;
170
+ GlobalTensor<T> sinGm;
171
+
172
+ TQue<QuePosition::VECIN, 1> inQueueX;
173
+ TQue<QuePosition::VECIN, 1> inQueueCos;
174
+ TQue<QuePosition::VECIN, 1> inQueueSin;
175
+ TQue<QuePosition::VECOUT, 1> outQueueY;
176
+ TBuf<TPosition::VECCALC> bufferReal;
177
+ TBuf<TPosition::VECCALC> bufferImag;
178
+ TBuf<TPosition::VECCALC> buffer_;
179
+
180
+ constexpr static int64_t hiddenDim = 64;
181
+ constexpr static int64_t hiddenDimHalf = 32;
182
+ constexpr static int64_t numHead = 32;
183
+ constexpr static int64_t seqLength = 1;
184
+ constexpr static int64_t numTwo = 2;
185
+ };
186
+ } // namespace InterleaveRope
187
+
188
+ #endif // _INTERLEAVE_ROPE_FIX_BNSD_B11D_H_
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/op_kernel/interleave_rope_split_s.h ADDED
@@ -0,0 +1,249 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ /*!
12
+ * \file interleave_rope_split_s.h
13
+ * \brief
14
+ */
15
+ #ifndef _INTERLEAVE_ROPE_SPLIT_S_H_
16
+ #define _INTERLEAVE_ROPE_SPLIT_S_H_
17
+ #include "platform/platform_infos_def.h"
18
+ #include "kernel_operator.h"
19
+
20
+ namespace InterleaveRope {
21
+ using namespace AscendC;
22
+
23
+ template <typename T>
24
+ class KernelInterleaveRopeSplitS {
25
+ public:
26
+ __aicore__ inline KernelInterleaveRopeSplitS(TPipe* pipe, const InterleaveRopeTilingData* tiling)
27
+ : pipe_(pipe), tilingData_(tiling)
28
+ {}
29
+
30
+ __aicore__ inline void Init(GM_ADDR x, GM_ADDR cos, GM_ADDR sin, GM_ADDR y)
31
+ {
32
+ /*
33
+ * For each block, process
34
+ * x: [B, N, S, 64]
35
+ * cos: [B, 1, S, 64]
36
+ * sin: [B, 1, S, 64]
37
+ * y: [B, N, S, 64]
38
+ */
39
+ // tilingData
40
+ int64_t blockFactor = (tilingData_->seqLength + tilingData_->numBlocks - 1) / tilingData_->numBlocks;
41
+ int64_t currentBlockFactor = blockFactor;
42
+ if (GetBlockIdx() == tilingData_->numBlocks - 1) {
43
+ currentBlockFactor = tilingData_->seqLength - GetBlockIdx() * blockFactor;
44
+ }
45
+ ubFactor = 32; // ubfactor: 32
46
+ ubLoop = currentBlockFactor / ubFactor;
47
+ ubTail = currentBlockFactor - ubFactor * ubLoop;
48
+
49
+ xGm.SetGlobalBuffer((__gm__ T*)x + GetBlockIdx() * blockFactor * ROPE_LENGTH);
50
+ yGm.SetGlobalBuffer((__gm__ T*)y + GetBlockIdx() * blockFactor * ROPE_LENGTH);
51
+ cosGm.SetGlobalBuffer((__gm__ T*)cos + GetBlockIdx() * blockFactor * ROPE_LENGTH);
52
+ sinGm.SetGlobalBuffer((__gm__ T*)sin + GetBlockIdx() * blockFactor * ROPE_LENGTH);
53
+
54
+ // init pipe
55
+ pipe_->InitBuffer(inQueueX, 2, ubFactor * ROPE_LENGTH * sizeof(T)); // 2*ubFactor*64*2=256*ubFactor
56
+ pipe_->InitBuffer(inQueuePosEmb, 2, 2 * ubFactor * ROPE_LENGTH * sizeof(T)); // 2*2*ubFactor*64*2=512*ubFactor
57
+ pipe_->InitBuffer(outQueue, 2, ubFactor * ROPE_LENGTH * sizeof(T)); // 2*ubFactor*64*2=256*ubFactor
58
+ pipe_->InitBuffer(wsBuffer, 8 * ubFactor * ROPE_LENGTH * sizeof(float)); // 8*ubFactor*64*4=2048*ubFactor
59
+ }
60
+
61
+ __aicore__ inline void Process()
62
+ {
63
+ LocalTensor<float> workspaceBuffer = wsBuffer.Get<float>();
64
+
65
+ DataCopyPadExtParams<T> padParams{false, 0, 0, 0};
66
+ DataCopyExtParams copyParamsContinguous;
67
+ copyParamsContinguous.blockCount = 1;
68
+ int64_t xOffset = 0;
69
+ int64_t freqOffset = 0;
70
+ for (int64_t batchId = 0; batchId < tilingData_->batchSize; batchId++) {
71
+ copyParamsContinguous.blockLen = ubFactor * ROPE_LENGTH * sizeof(T);
72
+ for (int64_t loopIdx = 0; loopIdx < ubLoop; ++loopIdx) {
73
+ freqOffset = batchId * tilingData_->seqLength * ROPE_LENGTH + loopIdx * ubFactor * ROPE_LENGTH;
74
+ /*
75
+ * cosLocal = [batchId, 0, :ubFactor:, 64]
76
+ * sinLocal = [batchId, 0, :ubFactor:, 64]
77
+ */
78
+ LocalTensor<T> cosLocal = inQueuePosEmb.AllocTensor<T>();
79
+ LocalTensor<T> sinLocal = cosLocal[ubFactor * ROPE_LENGTH];
80
+ DataCopyPad(cosLocal, cosGm[freqOffset], copyParamsContinguous, padParams);
81
+ DataCopyPad(sinLocal, sinGm[freqOffset], copyParamsContinguous, padParams);
82
+ inQueuePosEmb.EnQue(cosLocal);
83
+ cosLocal = inQueuePosEmb.DeQue<T>();
84
+ sinLocal = cosLocal[ubFactor * ROPE_LENGTH];
85
+
86
+ for (int64_t headId = 0; headId < tilingData_->numHead; ++headId) {
87
+ xOffset = batchId * tilingData_->numHead * tilingData_->seqLength * ROPE_LENGTH +
88
+ loopIdx * ubFactor * ROPE_LENGTH + headId * tilingData_->seqLength * ROPE_LENGTH;
89
+ /*
90
+ * xLocal = [batchId, headId, seqId: seqId+ubFactor, 64]
91
+ * yLocal = [batchId, headId, seqId: seqId+ubFactor, 64]
92
+ */
93
+ LocalTensor<T> xLocal = inQueueX.AllocTensor<T>();
94
+ DataCopyPad(xLocal, xGm[xOffset], copyParamsContinguous, padParams);
95
+ inQueueX.EnQue(xLocal);
96
+ xLocal = inQueueX.DeQue<T>();
97
+ LocalTensor<T> yLocal = outQueue.AllocTensor<T>();
98
+ RoPE<T, true, ROPE_LENGTH>(yLocal, xLocal, cosLocal, sinLocal, workspaceBuffer, ubFactor);
99
+ inQueueX.FreeTensor(xLocal);
100
+ outQueue.EnQue(yLocal);
101
+ yLocal = outQueue.DeQue<T>();
102
+ DataCopyPad(yGm[xOffset], yLocal, copyParamsContinguous);
103
+ outQueue.FreeTensor(yLocal);
104
+ }
105
+ inQueuePosEmb.FreeTensor(cosLocal);
106
+ }
107
+ copyParamsContinguous.blockLen = ubTail * ROPE_LENGTH * sizeof(T);
108
+ if (ubTail > 0) {
109
+ freqOffset = batchId * tilingData_->seqLength * ROPE_LENGTH + ubLoop * ubFactor * ROPE_LENGTH;
110
+ /*
111
+ * cosLocal = [batchId, 0, :ubTail:, 64]
112
+ * sinLocal = [batchId, 0, :ubTail:, 64]
113
+ */
114
+ LocalTensor<T> cosLocal = inQueuePosEmb.AllocTensor<T>();
115
+ LocalTensor<T> sinLocal = cosLocal[ubFactor * ROPE_LENGTH];
116
+ DataCopyPad(cosLocal, cosGm[freqOffset], copyParamsContinguous, padParams);
117
+ DataCopyPad(sinLocal, sinGm[freqOffset], copyParamsContinguous, padParams);
118
+ inQueuePosEmb.EnQue(cosLocal);
119
+ cosLocal = inQueuePosEmb.DeQue<T>();
120
+ sinLocal = cosLocal[ubFactor * ROPE_LENGTH];
121
+
122
+ for (int64_t headId = 0; headId < tilingData_->numHead; ++headId) {
123
+ xOffset = batchId * tilingData_->numHead * tilingData_->seqLength * ROPE_LENGTH +
124
+ ubLoop * ubFactor * ROPE_LENGTH + headId * tilingData_->seqLength * ROPE_LENGTH;
125
+ /*
126
+ * xLocal = [batchId, headId, seqId: seqId+ubFactor, 64]
127
+ * yLocal = [batchId, headId, seqId: seqId+ubFactor, 64]
128
+ */
129
+ LocalTensor<T> xLocal = inQueueX.AllocTensor<T>();
130
+ DataCopyPad(xLocal, xGm[xOffset], copyParamsContinguous, padParams);
131
+ inQueueX.EnQue(xLocal);
132
+ xLocal = inQueueX.DeQue<T>();
133
+ LocalTensor<T> yLocal = outQueue.AllocTensor<T>();
134
+ RoPE<T, true, ROPE_LENGTH>(yLocal, xLocal, cosLocal, sinLocal, workspaceBuffer, ubTail);
135
+ inQueueX.FreeTensor(xLocal);
136
+ outQueue.EnQue(yLocal);
137
+ yLocal = outQueue.DeQue<T>();
138
+ DataCopyPad(yGm[xOffset], yLocal, copyParamsContinguous);
139
+ outQueue.FreeTensor(yLocal);
140
+ }
141
+ inQueuePosEmb.FreeTensor(cosLocal);
142
+ }
143
+ }
144
+ }
145
+
146
+ template <typename T1, bool isElementWise = true, int64_t headSize>
147
+ __aicore__ inline void RoPE(
148
+ const LocalTensor<T1>& outLocal, const LocalTensor<T1>& xLocal, const LocalTensor<T1>& cosLocal,
149
+ const LocalTensor<T1>& sinLocal, const LocalTensor<float>& wsLocal, int64_t rows)
150
+ {
151
+ constexpr static int64_t NUM_ONE = 1;
152
+ constexpr static int64_t NUM_TWO = 2;
153
+ constexpr static int64_t NUM_THREE = 3;
154
+ constexpr static int64_t NUM_FOUR = 4;
155
+ constexpr static int64_t NUM_EIGHT = 8;
156
+ if constexpr (isElementWise) {
157
+ /**
158
+ * cosLocalFp32 : [rows * headSize * 0, rows * headSize * 1]
159
+ * sinLocalFp32 : [rows * headSize * 1, rows * headSize * 2]
160
+ * y0 : [rows * headSize * 2, rows * headSize * 3]
161
+ * y1 : [rows * headSize * 3, rows * headSize * 4]
162
+ * realLocalFp32: [rows * headSize * 4, rows * headSize * 5]
163
+ * imagLocalFp32: [rows * headSize * 5, rows * headSize * 6]
164
+ * realLocal : [rows * headSize * 6, rows * headSize * 7]
165
+ * imagLocal : [rows * headSize * 7, rows * headSize * 8]
166
+ */
167
+ int64_t cosLocalFp32Offset = rows * headSize * 0;
168
+ int64_t sinLocalFp32Offset = rows * headSize * 1;
169
+ int64_t y0Offset = rows * headSize * 2;
170
+ int64_t y1Offset = rows * headSize * 3;
171
+ int64_t realLocalFp32Offset = rows * headSize * 4;
172
+ int64_t imagLocalFp32Offset = rows * headSize * 5;
173
+ int64_t realLocalOffset = rows * headSize * 6;
174
+ int64_t imagLocalOffset = rows * headSize * 7;
175
+
176
+ LocalTensor<float> cosLocalFp32 = wsLocal[cosLocalFp32Offset];
177
+ LocalTensor<float> sinLocalFp32 = wsLocal[sinLocalFp32Offset];
178
+ LocalTensor<float> y0 = wsLocal[y0Offset];
179
+ LocalTensor<float> y1 = wsLocal[y1Offset];
180
+ LocalTensor<float> realLocalFp32 = wsLocal[realLocalFp32Offset];
181
+ LocalTensor<float> imagLocalFp32 = wsLocal[imagLocalFp32Offset];
182
+ LocalTensor<T1> realLocal = wsLocal[realLocalOffset].template ReinterpretCast<T1>();
183
+ LocalTensor<T1> imagLocal = wsLocal[imagLocalOffset].template ReinterpretCast<T1>();
184
+
185
+ // step #1: cast cosLocal and sinLocal to fp32
186
+ Cast(cosLocalFp32, cosLocal, RoundMode::CAST_NONE, rows * headSize);
187
+ Cast(sinLocalFp32, sinLocal, RoundMode::CAST_NONE, rows * headSize);
188
+ PipeBarrier<PIPE_V>();
189
+
190
+ // step #2: Gather out real and imag
191
+ uint64_t rsvdCnt = 0;
192
+ GatherMask(realLocal, xLocal, NUM_ONE, true, rows * headSize, {1, 1, NUM_EIGHT, 0}, rsvdCnt);
193
+ GatherMask(imagLocal, xLocal, NUM_TWO, true, rows * headSize, {1, 1, NUM_EIGHT, 0}, rsvdCnt);
194
+ PipeBarrier<PIPE_V>();
195
+
196
+ // step #3: Cast realLocal and imagLocal to Fp32
197
+ Cast(realLocalFp32, realLocal, RoundMode::CAST_NONE, rows * (headSize >> 1));
198
+ Cast(imagLocalFp32, imagLocal, RoundMode::CAST_NONE, rows * (headSize >> 1));
199
+ PipeBarrier<PIPE_V>();
200
+
201
+ // step #4: y0 = (realLocalFp32, imagLocalFp32) * cosLocalFp32
202
+ Mul(y0, realLocalFp32, cosLocalFp32, /* mask */ (headSize >> 1), /* repeat */ rows,
203
+ {1, 1, 1, NUM_EIGHT, NUM_FOUR, NUM_EIGHT});
204
+ Mul(y0[(headSize >> 1)], imagLocalFp32, cosLocalFp32[(headSize >> 1)], /* mask */ (headSize >> 1),
205
+ /* repeat */ rows, {1, 1, 1, NUM_EIGHT, NUM_FOUR, NUM_EIGHT});
206
+ PipeBarrier<PIPE_V>();
207
+
208
+ // step #5: y1 = (-imagLocalFp32, realLocalFp32) * sinLocalFp32
209
+ Muls<float>(imagLocalFp32, imagLocalFp32, -1.0f, rows * (headSize >> 1));
210
+ PipeBarrier<PIPE_V>();
211
+ Mul(y1, imagLocalFp32, sinLocalFp32, /* mask */ (headSize >> 1), /* repeat */ rows,
212
+ {1, 1, 1, NUM_EIGHT, NUM_FOUR, NUM_EIGHT});
213
+ Mul(y1[(headSize >> 1)], realLocalFp32, sinLocalFp32[(headSize >> 1)], /* mask */ (headSize >> 1),
214
+ /* repeat */ rows, {1, 1, 1, NUM_EIGHT, NUM_FOUR, NUM_EIGHT});
215
+ PipeBarrier<PIPE_V>();
216
+
217
+ // step #6: y0 = y0 + y1
218
+ Add(y0, y0, y1, rows * headSize);
219
+ PipeBarrier<PIPE_V>();
220
+
221
+ // step #7: outLocal = Cast(y0, T1)
222
+ if constexpr (std::is_same<T1, bfloat16_t>::value) {
223
+ Cast(outLocal, y0, RoundMode::CAST_RINT, rows * headSize);
224
+ } else if constexpr (std::is_same<T1, half>::value) {
225
+ Cast(outLocal, y0, RoundMode::CAST_NONE, rows * headSize);
226
+ }
227
+ }
228
+ }
229
+
230
+ private:
231
+ TPipe* pipe_ = nullptr;
232
+ const InterleaveRopeTilingData* tilingData_;
233
+ constexpr static int64_t ROPE_LENGTH = 64;
234
+ GlobalTensor<T> xGm;
235
+ GlobalTensor<T> yGm;
236
+ GlobalTensor<T> cosGm;
237
+ GlobalTensor<T> sinGm;
238
+
239
+ TQue<QuePosition::VECIN, 1> inQueueX;
240
+ TQue<QuePosition::VECIN, 1> inQueuePosEmb;
241
+ TQue<QuePosition::VECOUT, 1> outQueue;
242
+ TBuf<TPosition::VECCALC> wsBuffer;
243
+ int64_t ubFactor = 1;
244
+ int64_t ubTail = 0;
245
+ int64_t ubLoop = 0;
246
+ };
247
+ } // namespace InterleaveRope
248
+
249
+ #endif // _INTERLEAVE_ROPE_SPLIT_S_H_
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/test_data/cases_manifest.json ADDED
@@ -0,0 +1,139 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "op_name": "interleave_rope",
3
+ "op_camel": "InterleaveRope",
4
+ "output_keys": [
5
+ "y"
6
+ ],
7
+ "cases": [
8
+ {
9
+ "id": 0,
10
+ "dtype": "float16",
11
+ "shape": [
12
+ 32,
13
+ 1,
14
+ 32,
15
+ 64
16
+ ],
17
+ "attrs": {
18
+ "cos_s": 1
19
+ }
20
+ },
21
+ {
22
+ "id": 1,
23
+ "dtype": "float16",
24
+ "shape": [
25
+ 32,
26
+ 1,
27
+ 64,
28
+ 64
29
+ ],
30
+ "attrs": {
31
+ "cos_s": 1
32
+ }
33
+ },
34
+ {
35
+ "id": 2,
36
+ "dtype": "float16",
37
+ "shape": [
38
+ 32,
39
+ 1,
40
+ 128,
41
+ 64
42
+ ],
43
+ "attrs": {
44
+ "cos_s": 1
45
+ }
46
+ },
47
+ {
48
+ "id": 3,
49
+ "dtype": "bfloat16",
50
+ "shape": [
51
+ 64,
52
+ 1,
53
+ 32,
54
+ 64
55
+ ],
56
+ "attrs": {
57
+ "cos_s": 1
58
+ }
59
+ },
60
+ {
61
+ "id": 4,
62
+ "dtype": "bfloat16",
63
+ "shape": [
64
+ 64,
65
+ 1,
66
+ 64,
67
+ 64
68
+ ],
69
+ "attrs": {
70
+ "cos_s": 1
71
+ }
72
+ },
73
+ {
74
+ "id": 5,
75
+ "dtype": "float16",
76
+ "shape": [
77
+ 128,
78
+ 1,
79
+ 32,
80
+ 64
81
+ ],
82
+ "attrs": {
83
+ "cos_s": 1
84
+ }
85
+ },
86
+ {
87
+ "id": 6,
88
+ "dtype": "float16",
89
+ "shape": [
90
+ 32,
91
+ 1,
92
+ 512,
93
+ 64
94
+ ],
95
+ "attrs": {
96
+ "cos_s": 512
97
+ }
98
+ },
99
+ {
100
+ "id": 7,
101
+ "dtype": "bfloat16",
102
+ "shape": [
103
+ 128,
104
+ 1,
105
+ 64,
106
+ 64
107
+ ],
108
+ "attrs": {
109
+ "cos_s": 1
110
+ }
111
+ },
112
+ {
113
+ "id": 8,
114
+ "dtype": "float16",
115
+ "shape": [
116
+ 64,
117
+ 1,
118
+ 1024,
119
+ 64
120
+ ],
121
+ "attrs": {
122
+ "cos_s": 1024
123
+ }
124
+ },
125
+ {
126
+ "id": 9,
127
+ "dtype": "bfloat16",
128
+ "shape": [
129
+ 32,
130
+ 1,
131
+ 1,
132
+ 64
133
+ ],
134
+ "attrs": {
135
+ "cos_s": 1
136
+ }
137
+ }
138
+ ]
139
+ }
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/CMakeLists.txt ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # -----------------------------------------------------------------------------------------------------------
2
+ # Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+ # This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+ # CANN Open Software License Agreement Version 2.0 (the "License").
5
+ # Please refer to the License for details. You may not use this file except in compliance with the License.
6
+ # THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+ # INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+ # See LICENSE in the root of the software repository for the full text of the License.
9
+ # -----------------------------------------------------------------------------------------------------------
10
+
11
+ message(STATUS "=== Debug: start ops.posembedding.interleave_rope.tests.CMakeLists.txt ")
12
+ file(GLOB CURRENT_DIRS RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_SOURCE_DIR}/*)
13
+ foreach(SUB_DIR ${CURRENT_DIRS})
14
+ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/${SUB_DIR}/CMakeLists.txt")
15
+ add_subdirectory(${SUB_DIR})
16
+ endif()
17
+ endforeach()
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/ut/CMakeLists.txt ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # -----------------------------------------------------------------------------------------------------------
2
+ # Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+ # This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+ # CANN Open Software License Agreement Version 2.0 (the "License").
5
+ # Please refer to the License for details. You may not use this file except in compliance with the License.
6
+ # THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+ # INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+ # See LICENSE in the root of the software repository for the full text of the License.
9
+ # -----------------------------------------------------------------------------------------------------------
10
+
11
+ # 每个目录下需要生成的可执行文件,具体参考:ops/built-in/test/CMakeLists.txt: 50~124
12
+ file(GLOB CURRENT_DIRS RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_SOURCE_DIR}/*)
13
+ foreach(SUB_DIR ${CURRENT_DIRS})
14
+ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/${SUB_DIR}/CMakeLists.txt")
15
+ add_subdirectory(${SUB_DIR})
16
+ endif()
17
+ endforeach()
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/ut/op_host/CMakeLists.txt ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # -----------------------------------------------------------------------------------------------------------
2
+ # Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+ # This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+ # CANN Open Software License Agreement Version 2.0 (the "License").
5
+ # Please refer to the License for details. You may not use this file except in compliance with the License.
6
+ # THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+ # INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+ # See LICENSE in the root of the software repository for the full text of the License.
9
+ # -----------------------------------------------------------------------------------------------------------
10
+
11
+ if(UT_TEST_ALL OR OP_HOST_UT)
12
+ add_modules_ut_sources(UT_NAME ${OP_TILING_MODULE_NAME} MODE PRIVATE DIR ${CMAKE_CURRENT_SOURCE_DIR})
13
+ add_modules_ut_sources(UT_NAME ${OP_INFERSHAPE_MODULE_NAME} MODE PRIVATE DIR ${CMAKE_CURRENT_SOURCE_DIR})
14
+ endif()
15
+
16
+ file(GLOB CURRENT_DIRS RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_SOURCE_DIR}/*)
17
+ foreach(SUB_DIR ${CURRENT_DIRS})
18
+ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/${SUB_DIR}/CMakeLists.txt")
19
+ add_subdirectory(${SUB_DIR})
20
+ endif()
21
+ endforeach()
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/ut/op_host/op_api/CMakeLists.txt ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # -----------------------------------------------------------------------------------------------------------
2
+ # Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+ # This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+ # CANN Open Software License Agreement Version 2.0 (the "License").
5
+ # Please refer to the License for details. You may not use this file except in compliance with the License.
6
+ # THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+ # INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+ # See LICENSE in the root of the software repository for the full text of the License.
9
+ # -----------------------------------------------------------------------------------------------------------
10
+
11
+ message(STATUS "=== Debug: target_sources add test_interleave_rope")
12
+
13
+ if(UT_TEST_ALL OR OP_API_UT)
14
+ add_modules_ut_sources(UT_NAME ${OP_API_MODULE_NAME} MODE PRIVATE DIR ${CMAKE_CURRENT_SOURCE_DIR})
15
+ endif()
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/ut/op_host/op_api/test_aclnn_interleave_rope.cpp ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ #include <vector>
12
+ #include <array>
13
+ #include <float.h>
14
+ #include "gtest/gtest.h"
15
+ #include "../../../../op_host/op_api/aclnn_interleave_rope.h"
16
+ #include "opdev/platform.h"
17
+ #include "op_api_ut_common/tensor_desc.h"
18
+ #include "op_api_ut_common/scalar_desc.h"
19
+ #include "op_api_ut_common/op_api_ut.h"
20
+
21
+ using namespace std;
22
+
23
+ class l2_interleave_rope_test : public testing::Test
24
+ {
25
+ protected:
26
+ static void SetUpTestCase()
27
+ {
28
+ cout << "l2_interleave_rope_test SetUp" << endl;
29
+ }
30
+
31
+ static void TearDownTestCase()
32
+ {
33
+ cout << "l2_interleave_rope_test TearDown" << endl;
34
+ }
35
+ };
36
+
37
+ // dtype fp32
38
+ TEST_F(l2_interleave_rope_test, Ascend910B2_interleave_rope_fp32)
39
+ {
40
+ auto x = TensorDesc({8, 1, 1, 64}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
41
+ auto cos = TensorDesc({8, 1, 1, 64}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(0, 8);
42
+ auto sin = TensorDesc({8, 1, 1, 64}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(0, 8);
43
+ auto out = TensorDesc({8, 1, 1, 64}, ACL_FLOAT16, ACL_FORMAT_ND);
44
+ auto ut = OP_API_UT(aclnnInterleaveRope, INPUT(x, cos, sin), OUTPUT(out));
45
+ uint64_t workspaceSize = 0;
46
+ aclOpExecutor* executor = nullptr;
47
+ aclnnStatus getWorkspaceResult = ut.TestGetWorkspaceSizeWithNNopbaseInner(&workspaceSize, executor);
48
+ }
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/ut/op_host/test_interleave_rope_infershape.cpp ADDED
@@ -0,0 +1,92 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ #include <gtest/gtest.h>
12
+ #include <iostream>
13
+ #include "infer_shape_context_faker.h"
14
+ #include "infer_shape_case_executor.h"
15
+ #include "base/registry/op_impl_space_registry_v2.h"
16
+ #include "infer_datatype_context_faker.h"
17
+
18
+ class InterleaveRope : public testing::Test {
19
+ protected:
20
+ static void SetUpTestCase()
21
+ {
22
+ std::cout << "InterleaveRope SetUp" << std::endl;
23
+ }
24
+
25
+ static void TearDownTestCase()
26
+ {
27
+ std::cout << "InterleaveRope TearDown" << std::endl;
28
+ }
29
+ };
30
+
31
+ TEST_F(InterleaveRope, InterleaveRope_infer_shape_00)
32
+ {
33
+ gert::InfershapeContextPara infershapeContextPara(
34
+ "InterleaveRope",
35
+ {
36
+ {{{32, 32, 1, 64}, {32, 32, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
37
+ {{{32, 1, 1, 64}, {32, 1, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
38
+ {{{32, 1, 1, 64}, {32, 1, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
39
+ },
40
+ {
41
+ {{{}, {}}, ge::DT_FLOAT16, ge::FORMAT_ND},
42
+ });
43
+ std::vector<std::vector<int64_t>> expectOutputShape = {{32, 32, 1, 64}};
44
+ ExecuteTestCase(infershapeContextPara, ge::GRAPH_SUCCESS, expectOutputShape);
45
+ }
46
+
47
+ TEST_F(InterleaveRope, InterleaveRope_infer_shape_01)
48
+ {
49
+ gert::InfershapeContextPara infershapeContextPara(
50
+ "InterleaveRope",
51
+ {
52
+ {{{32, 32, 4, 64}, {32, 32, 4, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
53
+ {{{32, 1, 4, 64}, {32, 1, 4, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
54
+ {{{32, 1, 4, 64}, {32, 1, 4, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
55
+ },
56
+ {
57
+ {{{}, {}}, ge::DT_FLOAT16, ge::FORMAT_ND},
58
+ });
59
+ std::vector<std::vector<int64_t>> expectOutputShape = {{32, 32, 4, 64}};
60
+ ExecuteTestCase(infershapeContextPara, ge::GRAPH_SUCCESS, expectOutputShape);
61
+ }
62
+
63
+ TEST_F(InterleaveRope, InterleaveRope_inferdtype)
64
+ {
65
+ ASSERT_NE(gert::DefaultOpImplSpaceRegistryV2::GetInstance().GetSpaceRegistry()->GetOpImpl("InterleaveRope"),
66
+ nullptr);
67
+ auto data_type_func = gert::DefaultOpImplSpaceRegistryV2::GetInstance()
68
+ .GetSpaceRegistry()
69
+ ->GetOpImpl("InterleaveRope")
70
+ ->infer_datatype;
71
+ ASSERT_NE(data_type_func, nullptr);
72
+
73
+ ge::DataType input_0 = ge::DT_FLOAT16;
74
+ ge::DataType output_0 = ge::DT_FLOAT16;
75
+ auto context_holder = gert::InferDataTypeContextFaker()
76
+ .IrInputNum(3)
77
+ .NodeIoNum(3, 1)
78
+ .IrInstanceNum({1, 1, 1})
79
+ .NodeInputTd(0, ge::DT_FLOAT16, ge::FORMAT_ND, ge::FORMAT_ND)
80
+ .NodeInputTd(1, ge::DT_FLOAT16, ge::FORMAT_ND, ge::FORMAT_ND)
81
+ .NodeInputTd(2, ge::DT_FLOAT16, ge::FORMAT_ND, ge::FORMAT_ND)
82
+ .NodeOutputTd(0, ge::FORMAT_ND, ge::FORMAT_ND)
83
+ .InputDataTypes({&input_0, &input_0, &input_0})
84
+ .OutputDataTypes({&output_0})
85
+ .Build();
86
+
87
+ auto context = context_holder.GetContext<gert::InferDataTypeContext>();
88
+ EXPECT_EQ(data_type_func(context), ge::GRAPH_SUCCESS);
89
+ ASSERT_NE(context, nullptr);
90
+
91
+ EXPECT_EQ(context->GetOutputDataType(0), ge::DT_FLOAT16);
92
+ }
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/ut/op_host/test_interleave_rope_tiling.cpp ADDED
@@ -0,0 +1,220 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ /* !
12
+ * \file test_interleave_rope_tiling.cpp
13
+ * \brief
14
+ */
15
+
16
+ #include <iostream>
17
+ #include <gtest/gtest.h>
18
+ #include "tiling_context_faker.h"
19
+ #include "tiling_case_executor.h"
20
+ #include "../../../op_host/interleave_rope_tiling.h"
21
+
22
+ using namespace std;
23
+
24
+ class InterleaveRopeTiling : public testing::Test
25
+ {
26
+ protected:
27
+ static void SetUpTestCase()
28
+ {
29
+ std::cout << "InterleaveRopeTiling SetUp" << std::endl;
30
+ }
31
+
32
+ static void TearDownTestCase()
33
+ {
34
+ std::cout << "InterleaveRopeTiling TearDown" << std::endl;
35
+ }
36
+ };
37
+
38
+ static string TilingData2Str(const gert::TilingData* tiling_data)
39
+ {
40
+ auto data = tiling_data->GetData();
41
+ string result;
42
+ for (size_t i = 0; i < tiling_data->GetDataSize(); i += sizeof(int64_t)) {
43
+ result += std::to_string((reinterpret_cast<const int64_t*>(tiling_data->GetData())[i / sizeof(int64_t)]));
44
+ result += " ";
45
+ }
46
+
47
+ return result;
48
+ }
49
+
50
+ TEST_F(InterleaveRopeTiling, interleave_rope_tiling_000)
51
+ {
52
+ optiling::InterleaveRopeCompileInfo compileInfo = {};
53
+ gert::TilingContextPara tilingContextPara("InterleaveRope",
54
+ {
55
+ {{{32, 32, 1, 64}, {32, 32, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
56
+ {{{32, 1, 1, 64}, {32, 1, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
57
+ {{{32, 1, 1, 64}, {32, 1, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
58
+ },
59
+ {
60
+ {{{32, 32, 1, 64}, {32, 32, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
61
+ },
62
+ &compileInfo);
63
+ uint64_t expectTilingKey = 1000;
64
+ string expectTilingData = "8 0 32 32 1 64 4 4 0 0 0 0 0 0 0 0 0 0 0 ";
65
+ std::vector<size_t> expectWorkspaces = {32};
66
+ ExecuteTestCase(tilingContextPara, ge::GRAPH_SUCCESS, expectTilingKey, expectTilingData, expectWorkspaces);
67
+ }
68
+
69
+ TEST_F(InterleaveRopeTiling, interleave_rope_tiling_001)
70
+ {
71
+ optiling::InterleaveRopeCompileInfo compileInfo = {};
72
+ gert::TilingContextPara tilingContextPara("InterleaveRope",
73
+ {
74
+ {{{32, 32, 4, 64}, {32, 32, 4, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
75
+ {{{32, 1, 1, 64}, {32, 1, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
76
+ {{{32, 1, 1, 64}, {32, 1, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
77
+ },
78
+ {
79
+ {{{32, 32, 4, 64}, {32, 32, 4, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
80
+ },
81
+ &compileInfo);
82
+ uint64_t expectTilingKey = 2000;
83
+ string expectTilingData = "64 1 32 32 4 64 32 32 0 0 0 2 2 1 2 2 1 2 2 ";
84
+ std::vector<size_t> expectWorkspaces = {32};
85
+ ExecuteTestCase(tilingContextPara, ge::GRAPH_SUCCESS, expectTilingKey, expectTilingData, expectWorkspaces);
86
+ }
87
+
88
+ TEST_F(InterleaveRopeTiling, interleave_rope_tiling_002)
89
+ {
90
+ optiling::InterleaveRopeCompileInfo compileInfo = {};
91
+ gert::TilingContextPara tilingContextPara("InterleaveRope",
92
+ {
93
+ {{{32, 32, 4, 64}, {32, 32, 4, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
94
+ {{{32, 1, 4, 64}, {32, 1, 4, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
95
+ {{{32, 1, 4, 64}, {32, 1, 4, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
96
+ },
97
+ {
98
+ {{{32, 32, 4, 64}, {32, 32, 4, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
99
+ },
100
+ &compileInfo);
101
+ uint64_t expectTilingKey = 3000;
102
+ string expectTilingData = "32 0 32 32 4 64 1 1 0 0 0 4 4 1 4 4 1 4 4 ";
103
+ std::vector<size_t> expectWorkspaces = {32};
104
+ ExecuteTestCase(tilingContextPara, ge::GRAPH_SUCCESS, expectTilingKey, expectTilingData, expectWorkspaces);
105
+ }
106
+
107
+ TEST_F(InterleaveRopeTiling, interleave_rope_tiling_hidden_dim_not_64)
108
+ {
109
+ optiling::InterleaveRopeCompileInfo compileInfo = {};
110
+ gert::TilingContextPara tilingContextPara("InterleaveRope",
111
+ {
112
+ {{{32, 32, 1, 63}, {32, 32, 1, 63}}, ge::DT_FLOAT16, ge::FORMAT_ND},
113
+ {{{32, 1, 1, 64}, {32, 1, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
114
+ {{{32, 1, 1, 64}, {32, 1, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
115
+ },
116
+ {
117
+ {{{32, 32, 1, 64}, {32, 32, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
118
+ },
119
+ &compileInfo);
120
+ uint64_t expectTilingKey = 0;
121
+ string expectTilingData = "";
122
+ std::vector<size_t> expectWorkspaces = {};
123
+ ExecuteTestCase(tilingContextPara, ge::GRAPH_FAILED, expectTilingKey, expectTilingData, expectWorkspaces);
124
+ }
125
+
126
+ TEST_F(InterleaveRopeTiling, interleave_rope_tiling_shape_len_not_4)
127
+ {
128
+ optiling::InterleaveRopeCompileInfo compileInfo = {};
129
+ gert::TilingContextPara tilingContextPara("InterleaveRope",
130
+ {
131
+ {{{32, 32, 1}, {32, 32, 1}}, ge::DT_FLOAT16, ge::FORMAT_ND},
132
+ {{{32, 1, 1, 64}, {32, 1, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
133
+ {{{32, 1, 1, 64}, {32, 1, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
134
+ },
135
+ {
136
+ {{{32, 32, 1, 64}, {32, 32, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
137
+ },
138
+ &compileInfo);
139
+ uint64_t expectTilingKey = 0;
140
+ string expectTilingData = "";
141
+ std::vector<size_t> expectWorkspaces = {};
142
+ ExecuteTestCase(tilingContextPara, ge::GRAPH_FAILED, expectTilingKey, expectTilingData, expectWorkspaces);
143
+ }
144
+
145
+ TEST_F(InterleaveRopeTiling, interleave_rope_tiling_sin_not_eq_cos)
146
+ {
147
+ optiling::InterleaveRopeCompileInfo compileInfo = {};
148
+ gert::TilingContextPara tilingContextPara("InterleaveRope",
149
+ {
150
+ {{{32, 32, 1, 63}, {32, 32, 1, 63}}, ge::DT_FLOAT16, ge::FORMAT_ND},
151
+ {{{32, 1, 1, 64}, {32, 1, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
152
+ {{{33, 1, 1, 64}, {33, 1, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
153
+ },
154
+ {
155
+ {{{32, 32, 1, 64}, {32, 32, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
156
+ },
157
+ &compileInfo);
158
+ uint64_t expectTilingKey = 0;
159
+ string expectTilingData = "";
160
+ std::vector<size_t> expectWorkspaces = {};
161
+ ExecuteTestCase(tilingContextPara, ge::GRAPH_FAILED, expectTilingKey, expectTilingData, expectWorkspaces);
162
+ }
163
+
164
+ TEST_F(InterleaveRopeTiling, interleave_rope_tiling_cosB_not_eq_batch_size)
165
+ {
166
+ optiling::InterleaveRopeCompileInfo compileInfo = {};
167
+ gert::TilingContextPara tilingContextPara("InterleaveRope",
168
+ {
169
+ {{{32, 32, 1, 64}, {32, 32, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
170
+ {{{3, 1, 1, 64}, {3, 1, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
171
+ {{{3, 1, 1, 64}, {3, 1, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
172
+ },
173
+ {
174
+ {{{32, 32, 1, 64}, {32, 32, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
175
+ },
176
+ &compileInfo);
177
+ uint64_t expectTilingKey = 0;
178
+ string expectTilingData = "";
179
+ std::vector<size_t> expectWorkspaces = {};
180
+ ExecuteTestCase(tilingContextPara, ge::GRAPH_FAILED, expectTilingKey, expectTilingData, expectWorkspaces);
181
+ }
182
+
183
+ TEST_F(InterleaveRopeTiling, interleave_rope_tiling_cosN_not_eq_1)
184
+ {
185
+ optiling::InterleaveRopeCompileInfo compileInfo = {};
186
+ gert::TilingContextPara tilingContextPara("InterleaveRope",
187
+ {
188
+ {{{32, 32, 1, 64}, {32, 32, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
189
+ {{{32, 3, 1, 64}, {32, 3, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
190
+ {{{33, 3, 1, 64}, {33, 3, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
191
+ },
192
+ {
193
+ {{{32, 32, 1, 64}, {32, 32, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
194
+ },
195
+ &compileInfo);
196
+ uint64_t expectTilingKey = 0;
197
+ string expectTilingData = "";
198
+ std::vector<size_t> expectWorkspaces = {};
199
+ ExecuteTestCase(tilingContextPara, ge::GRAPH_FAILED, expectTilingKey, expectTilingData, expectWorkspaces);
200
+ }
201
+
202
+ TEST_F(InterleaveRopeTiling, interleave_rope_tiling_cosD_not_eq_HiddenDim)
203
+ {
204
+ optiling::InterleaveRopeCompileInfo compileInfo = {};
205
+ gert::TilingContextPara tilingContextPara("InterleaveRope",
206
+ {
207
+ {{{32, 32, 1, 64}, {32, 32, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
208
+ {{{32, 1, 1, 62}, {32, 1, 1, 62}}, ge::DT_FLOAT16, ge::FORMAT_ND},
209
+ {{{33, 1, 1, 62}, {33, 1, 1, 62}}, ge::DT_FLOAT16, ge::FORMAT_ND},
210
+ },
211
+ {
212
+ {{{32, 32, 1, 64}, {32, 32, 1, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},
213
+ },
214
+ &compileInfo);
215
+ uint64_t expectTilingKey = 0;
216
+ string expectTilingData = "";
217
+ std::vector<size_t> expectWorkspaces = {};
218
+ ExecuteTestCase(tilingContextPara, ge::GRAPH_FAILED, expectTilingKey, expectTilingData, expectWorkspaces);
219
+ }
220
+
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/ut/op_kernel/CMakeLists.txt ADDED
@@ -0,0 +1,31 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # -----------------------------------------------------------------------------------------------------------
2
+ # Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+ # This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+ # CANN Open Software License Agreement Version 2.0 (the "License").
5
+ # Please refer to the License for details. You may not use this file except in compliance with the License.
6
+ # THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+ # INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+ # See LICENSE in the root of the software repository for the full text of the License.
9
+ # -----------------------------------------------------------------------------------------------------------
10
+
11
+ file(GLOB CURRENT_SOURCE_DIRS LIST_DIRECTORIES true ${CMAKE_CURRENT_SOURCE_DIR}/*)
12
+ message(STATUS "=== Debug: CURRENT_SOURCE_DIRS =${CURRENT_SOURCE_DIRS} ")
13
+ foreach(SUB_DIR ${CURRENT_SOURCE_DIRS})
14
+ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/${SUB_DIR}/CMakeLists.txt")
15
+ add_subdirectory(${SUB_DIR})
16
+ endif()
17
+ endforeach()
18
+
19
+ if (UT_TEST_ALL OR OP_KERNEL_UT)
20
+ # 算子自己的tiling文件路径
21
+ set(interleave_rope_tiling_files
22
+ ${CMAKE_CURRENT_SOURCE_DIR}/../../../op_host/interleave_rope_tiling.cpp
23
+ )
24
+ # 使用AddOpTestCase
25
+ # param1:算子名称,以kernel方式命名
26
+ # param2:soc版本,多个以分号分隔,例如:"Ascend950PR_9599;AscendB1"
27
+ # param3:自定义编译选项,一般填写测试的一种典型数据类型组合,不需要则传入空字符串,例如:"-DDTYPE_X=float",多个使用空格分隔,例如:"-DDTYPE_X=float -DDTYPE_Y=float"
28
+ # param4:该算子依赖的所有tiling源码文件
29
+
30
+ AddOpTestCase(interleave_rope "ascend910B1" "-DDTYPE_X=float" "${interleave_rope_tiling_files}")
31
+ endif()
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/ut/op_kernel/test_interleave_rope.cpp ADDED
@@ -0,0 +1,148 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ /*!
12
+ * \file test_interleave_rope.cpp
13
+ * \brief
14
+ */
15
+
16
+ #include <array>
17
+ #include <vector>
18
+ #include <iostream>
19
+ #include <string>
20
+ #include <cstdint>
21
+ #include "gtest/gtest.h"
22
+ #include "tikicpulib.h"
23
+ #include "test_interleave_rope.h"
24
+ #include "data_utils.h"
25
+ #include "tiling_case_executor.h"
26
+
27
+ using namespace std;
28
+
29
+ extern "C" __global__ __aicore__ void interleave_rope(
30
+ GM_ADDR x, GM_ADDR cos, GM_ADDR sin, GM_ADDR y, GM_ADDR workspace, GM_ADDR tiling);
31
+
32
+ class interleave_rope_test : public testing::Test
33
+ {
34
+ protected:
35
+ static void SetUpTestCase()
36
+ {
37
+ cout << "interleave_rope_test SetUp\n" << endl;
38
+ }
39
+ static void TearDownTestCase()
40
+ {
41
+ cout << "interleave_rope_test TearDown\n" << endl;
42
+ }
43
+ };
44
+
45
+ TEST_F(interleave_rope_test, test_case_0001)
46
+ {
47
+ size_t xSize = 32 * 32 * 64 * sizeof(half);
48
+ size_t ySize = 32 * 32 * 64 * sizeof(half);
49
+ size_t cosSize = 32 * 1 * 64 * sizeof(half);
50
+ size_t sinSize = 32 * 1 * 64 * sizeof(half);
51
+ size_t tiling_data_size = sizeof(InterleaveRopeTilingData);
52
+ uint32_t blockDim = 8;
53
+
54
+ uint8_t* x = (uint8_t*)AscendC::GmAlloc(xSize);
55
+ uint8_t* y = (uint8_t*)AscendC::GmAlloc(ySize);
56
+ uint8_t* cos = (uint8_t*)AscendC::GmAlloc(cosSize);
57
+ uint8_t* sin = (uint8_t*)AscendC::GmAlloc(sinSize);
58
+ uint8_t* workspace = (uint8_t*)AscendC::GmAlloc(16 * 2);
59
+ uint8_t* tiling = (uint8_t*)AscendC::GmAlloc(tiling_data_size);
60
+
61
+ char* path_ = get_current_dir_name();
62
+ string path(path_);
63
+
64
+ InterleaveRopeTilingData* tilingDatafromBin = reinterpret_cast<InterleaveRopeTilingData*>(tiling);
65
+
66
+ tilingDatafromBin->blockDim = 8;
67
+
68
+ ICPU_SET_TILING_KEY(1000);
69
+ // ICPU_RUN_KF(interleave_rope, blockDim, x, cos, sin, y, workspace, (uint8_t*)(tilingDatafromBin));
70
+
71
+ AscendC::GmFree(x);
72
+ AscendC::GmFree(y);
73
+ AscendC::GmFree(cos);
74
+ AscendC::GmFree(sin);
75
+ AscendC::GmFree(workspace);
76
+ AscendC::GmFree(tiling);
77
+ free(path_);
78
+ }
79
+
80
+ TEST_F(interleave_rope_test, test_case_0002)
81
+ {
82
+ size_t xSize = 32 * 32 * 64 * sizeof(half);
83
+ size_t ySize = 32 * 32 * 64 * sizeof(half);
84
+ size_t cosSize = 32 * 1 * 64 * sizeof(half);
85
+ size_t sinSize = 32 * 1 * 64 * sizeof(half);
86
+ size_t tiling_data_size = sizeof(InterleaveRopeTilingData);
87
+ uint32_t blockDim = 8;
88
+
89
+ uint8_t* x = (uint8_t*)AscendC::GmAlloc(xSize);
90
+ uint8_t* y = (uint8_t*)AscendC::GmAlloc(ySize);
91
+ uint8_t* cos = (uint8_t*)AscendC::GmAlloc(cosSize);
92
+ uint8_t* sin = (uint8_t*)AscendC::GmAlloc(sinSize);
93
+ uint8_t* workspace = (uint8_t*)AscendC::GmAlloc(16 * 2);
94
+ uint8_t* tiling = (uint8_t*)AscendC::GmAlloc(tiling_data_size);
95
+
96
+ char* path_ = get_current_dir_name();
97
+ string path(path_);
98
+
99
+ InterleaveRopeTilingData* tilingDatafromBin = reinterpret_cast<InterleaveRopeTilingData*>(tiling);
100
+
101
+ tilingDatafromBin->blockDim = 8;
102
+
103
+ ICPU_SET_TILING_KEY(2000);
104
+ // ICPU_RUN_KF(interleave_rope, blockDim, x, cos, sin, y, workspace, (uint8_t*)(tilingDatafromBin));
105
+
106
+ AscendC::GmFree(x);
107
+ AscendC::GmFree(y);
108
+ AscendC::GmFree(cos);
109
+ AscendC::GmFree(sin);
110
+ AscendC::GmFree(workspace);
111
+ AscendC::GmFree(tiling);
112
+ free(path_);
113
+ }
114
+
115
+ TEST_F(interleave_rope_test, test_case_0003)
116
+ {
117
+ size_t xSize = 32 * 32 * 64 * sizeof(half);
118
+ size_t ySize = 32 * 32 * 64 * sizeof(half);
119
+ size_t cosSize = 32 * 1 * 64 * sizeof(half);
120
+ size_t sinSize = 32 * 1 * 64 * sizeof(half);
121
+ size_t tiling_data_size = sizeof(InterleaveRopeTilingData);
122
+ uint32_t blockDim = 8;
123
+
124
+ uint8_t* x = (uint8_t*)AscendC::GmAlloc(xSize);
125
+ uint8_t* y = (uint8_t*)AscendC::GmAlloc(ySize);
126
+ uint8_t* cos = (uint8_t*)AscendC::GmAlloc(cosSize);
127
+ uint8_t* sin = (uint8_t*)AscendC::GmAlloc(sinSize);
128
+ uint8_t* workspace = (uint8_t*)AscendC::GmAlloc(16 * 2);
129
+ uint8_t* tiling = (uint8_t*)AscendC::GmAlloc(tiling_data_size);
130
+
131
+ char* path_ = get_current_dir_name();
132
+ string path(path_);
133
+
134
+ InterleaveRopeTilingData* tilingDatafromBin = reinterpret_cast<InterleaveRopeTilingData*>(tiling);
135
+
136
+ tilingDatafromBin->blockDim = 8;
137
+
138
+ ICPU_SET_TILING_KEY(3000);
139
+ // ICPU_RUN_KF(interleave_rope, blockDim, x, cos, sin, y, workspace, (uint8_t*)(tilingDatafromBin));
140
+
141
+ AscendC::GmFree(x);
142
+ AscendC::GmFree(y);
143
+ AscendC::GmFree(cos);
144
+ AscendC::GmFree(sin);
145
+ AscendC::GmFree(workspace);
146
+ AscendC::GmFree(tiling);
147
+ free(path_);
148
+ }
datasets/ops_src/ops_transformer/test_kernels/kernels/interleave_rope/tests/ut/op_kernel/test_interleave_rope.h ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ /*!
12
+ * \file test_interleave_rope.h
13
+ * \brief
14
+ */
15
+
16
+ #ifndef _INTERLEAVE_ROPE_TILING_H_
17
+ #define _INTERLEAVE_ROPE_TILING_H_
18
+
19
+ #include "kernel_tiling/kernel_tiling.h"
20
+
21
+ #define DT_BF16 bfloat16_t
22
+ #define ORIG_DTYPE_START DT_BF16
23
+ #define __CCE_UT_TEST__
24
+
25
+ #pragma pack(1)
26
+
27
+ #pragma pack()
28
+
29
+ #define CONVERT_TILING_DATA(tilingStruct, tilingDataPointer, tilingPointer) \
30
+ __ubuf__ tilingStruct* tilingDataPointer = \
31
+ reinterpret_cast<__ubuf__ tilingStruct*>((__ubuf__ uint8_t*)(tilingPointer));
32
+
33
+ #define INIT_TILING_DATA(tilingStruct, tilingDataPointer, tilingPointer) \
34
+ CONVERT_TILING_DATA(tilingStruct, tilingDataPointer, tilingPointer);
35
+
36
+ #define GET_TILING_DATA(tilingData, tilingPointer) \
37
+ InterleaveRopeTilingData tilingData; \
38
+ INIT_TILING_DATA(InterleaveRopeTilingData, tilingDataPointer, tilingPointer); \
39
+ (tilingData).blockDim = tilingDataPointer->blockDim; \
40
+ (tilingData).splitAxis = tilingDataPointer->splitAxis; \
41
+ (tilingData).batchSize = tilingDataPointer->batchSize; \
42
+ (tilingData).numHead = tilingDataPointer->numHead; \
43
+ (tilingData).seqLength = tilingDataPointer->seqLength; \
44
+ (tilingData).hiddenDim = tilingDataPointer->hiddenDim; \
45
+ (tilingData).batchsPerBlock = tilingDataPointer->batchsPerBlock; \
46
+ (tilingData).batchsLastBlock = tilingDataPointer->batchsLastBlock; \
47
+ (tilingData).batchLoops = tilingDataPointer->batchLoops; \
48
+ (tilingData).batchPerLoop = tilingDataPointer->batchPerLoop; \
49
+ (tilingData).batchLastLoop = tilingDataPointer->batchLastLoop; \
50
+ (tilingData).hiddenDimCountPerBlock = tilingDataPointer->hiddenDimCountPerBlock; \
51
+ (tilingData).hiddenDimCountLastBlock = tilingDataPointer->hiddenDimCountLastBlock; \
52
+ (tilingData).hiddenDimLoopsPerBlock = tilingDataPointer->hiddenDimLoopsPerBlock; \
53
+ (tilingData).hiddenDimCountPerLoopPerBlock = tilingDataPointer->hiddenDimCountPerLoopPerBlock; \
54
+ (tilingData).hiddenDimCountLastLoopPerBlock = tilingDataPointer->hiddenDimCountLastLoopPerBlock; \
55
+ (tilingData).hiddenDimLoopsLastBlock = tilingDataPointer->hiddenDimLoopsLastBlock; \
56
+ (tilingData).hiddenDimCountPerLoopLastBlock = tilingDataPointer->hiddenDimCountPerLoopLastBlock; \
57
+ (tilingData).hiddenDimCountLastLoopLastBlock = tilingDataPointer->hiddenDimCountLastLoopLastBlock;
58
+
59
+ #define DTYPE_X half
60
+
61
+ #endif
datasets/ops_src/ops_transformer/test_kernels/kernels/qkv_rms_norm_rope_cache/CMakeLists.txt ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # This program is free software, you can redistribute it and/or modify.
2
+ # Copyright (c) 2026 Huawei Technologies Co., Ltd.
3
+ # This file is a part of the CANN Open Software.
4
+ # Licensed under CANN Open Software License Agreement Version 2.0 (the "License").
5
+ # Please refer to the License for details. You may not use this file except in compliance with the License.
6
+ # THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
7
+ # See LICENSE in the root of the software repository for the full text of the License.
8
+ # ============================================================================
9
+
10
+ message(STATUS "=== Debug: start ops.posembedding.qkv_rms_norm_rope_cache.CMakeLists.txt ")
11
+ file(GLOB CURRENT_DIRS RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_SOURCE_DIR}/*)
12
+ if(NOT ENABLE_TEST AND NOT BENCHMARK)
13
+ list(REMOVE_ITEM CURRENT_DIRS tests)
14
+ endif()
15
+ foreach(SUB_DIR ${CURRENT_DIRS})
16
+ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/${SUB_DIR}/CMakeLists.txt")
17
+ add_subdirectory(${SUB_DIR})
18
+ endif()
19
+ endforeach()
datasets/ops_src/ops_transformer/test_kernels/kernels/qkv_rms_norm_rope_cache/README.md ADDED
@@ -0,0 +1,302 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # QkvRmsNormRopeCache
2
+
3
+ ## 产品支持情况
4
+
5
+ |产品 | 是否支持 |
6
+ |:-------------------------|:----------:|
7
+ | <term>Ascend 950PR/Ascend 950DT</term> | × |
8
+ | <term>Atlas A3 训练系列产品/Atlas A3 推理系列产品</term> | √ |
9
+ | <term>Atlas A2 训练系列产品/Atlas A2 推理系列产品</term> | √ |
10
+ | <term>Atlas 200I/500 A2 推理产品</term> | × |
11
+ | <term>Atlas 推理系列产品</term> | × |
12
+ | <term>Atlas 训练系列产品</term> | × |
13
+
14
+ ## 功能说明
15
+
16
+ - 算子功能:输入qkv融合张量,通过SplitVD拆分q、k、v张量,执行RmsNorm、ApplyRotaryPosEmb、Quant、Scatter融合操作,输出q_out、k_cache、v_cache、q_out_before_quant(可选)、k_out_before_quant(可选)、v_out_before_quant(可选)。
17
+ - 本算子目前支持的场景如下表:
18
+
19
+ | 场景类型 | 情况概要 |
20
+ |:----------------------|:-----------------------------|
21
+ |<ul><li>cache_mode为PA_NZ</li><li>q无量化</li><li>k和v支持无量化、对称量化和非对称量化</li><li>q_out_before_quant/k_out_before_quant/v_out_before_quant不输出</li></ul>|qkv Shape为[$B_{qkv}$ * $S_{qkv}$, $N_{qkv}$ * $D_{qkv}$],q、k、v具有完全相同的D维度。主要计算过程与输出对应关系:<br><ul><li>qkv经过SplitVD->q、k、v</li><li>q经过RmsNorm、RoPE->q_out</li><li>k经过RmsNorm、RoPE、Quant(可选)、Scatter->k_cache</li><li>v经过Quant(可选)、Scatter->v_cache</li></ul>|
22
+
23
+ - 计算公式:
24
+
25
+ (1) SplitVD:
26
+
27
+ 下式中,$N_q$、$N_k$、$N_v$分别表示q、k、v分量的注意力头数量,必须满足:
28
+
29
+ $$
30
+ \begin{cases}
31
+ N_k = N_v \\
32
+ N_{qkv} = N_k + N_v + N_q \\
33
+ D_{qkv} = D_q = D_k = D_v
34
+ \end{cases}
35
+ $$
36
+
37
+ $$
38
+ \begin{aligned}
39
+ q &= qkv[..., [:N_q] * D_{qkv}] \\
40
+ k &= qkv[..., [N_q:-N_v] * D_{qkv}] \\
41
+ v &= qkv[..., [-N_v:] * D_{qkv}]
42
+ \end{aligned}
43
+ $$
44
+
45
+ (2) RmsNorm:
46
+
47
+ 此处x和y分别表示RmsNorm的输入张量和输出张量,归一化沿最后一维(feature dimension)进行,该计算规则通用于q、k分量。
48
+
49
+ $$
50
+ squareX = x * x
51
+ $$
52
+
53
+ $$
54
+ meanSquareX = squareX.mean(dim = -1, keepdim = True)
55
+ $$
56
+
57
+ $$
58
+ rms = \sqrt{meanSquareX + epsilon}
59
+ $$
60
+
61
+ $$
62
+ y = (x / rms) * gamma
63
+ $$
64
+
65
+ (3) RoPE (Half-and-Half):
66
+
67
+ 此处的y指代完成RmsNorm计算的输出结果。
68
+
69
+ $$
70
+ y1 = y[\ldots, :d/2]
71
+ $$
72
+
73
+ $$
74
+ y2 = y[\ldots, d/2:]
75
+ $$
76
+
77
+ $$
78
+ y\_RoPE = torch.cat((-y2, y1), dim = -1)
79
+ $$
80
+
81
+ $$
82
+ y\_embed = (y * cos) + y\_RoPE * sin
83
+ $$
84
+
85
+ (4) Quant:
86
+
87
+ 无量化:
88
+
89
+ $$
90
+ kQuant = kRoPE \\
91
+ vQuant = v
92
+ $$
93
+
94
+ 对称量化部分:
95
+
96
+ $$
97
+ kQuant = kRoPE / kScale \\
98
+ vQuant = v / vScale
99
+ $$
100
+
101
+ 非对称量化部分:
102
+
103
+ $$
104
+ kQuant = kRoPE / kScale + kOffset \\
105
+ vQuant = v / vScale + vOffset
106
+ $$
107
+
108
+ ## 参数说明
109
+
110
+ <table style="undefined;table-layout: fixed; width: 1576px"><colgroup>
111
+ <col style="width: 170px">
112
+ <col style="width: 170px">
113
+ <col style="width: 312px">
114
+ <col style="width: 213px">
115
+ <col style="width: 100px">
116
+ </colgroup>
117
+ <thead>
118
+ <tr>
119
+ <th>参数名</th>
120
+ <th>输入/输出/属性</th>
121
+ <th>描述</th>
122
+ <th>数据类型</th>
123
+ <th>数据格式</th>
124
+ </tr></thead>
125
+ <tbody>
126
+ <tr>
127
+ <td>qkv</td>
128
+ <td>输入</td>
129
+ <td>用于切分出q、k、v的输入数据,对应公式中的qkv。shape为[B<sub>qkv</sub> * S<sub>qkv</sub>, N<sub>qkv</sub> * D<sub>qkv</sub>]。</td>
130
+ <td>FLOAT16、BFLOAT16</td>
131
+ <td>ND</td>
132
+ </tr>
133
+ <tr>
134
+ <td>q_gamma</td>
135
+ <td>输入</td>
136
+ <td>用于q的rms_norm计算的输入数据,对应公式中的gamma。与输入qkv的数据类型相同,shape为[D<sub>qkv</sub>]。</td>
137
+ <td>FLOAT16、BFLOAT16</td>
138
+ <td>ND</td>
139
+ </tr>
140
+ <tr>
141
+ <td>k_gamma</td>
142
+ <td>输入</td>
143
+ <td>用于k的rms_norm计算的输入数据,对应公式中的gamma。与输入qkv的数据类型相同,shape为[D<sub>qkv</sub>]。</td>
144
+ <td>FLOAT16、BFLOAT16</td>
145
+ <td>ND</td>
146
+ </tr>
147
+ <tr>
148
+ <td>cos</td>
149
+ <td>输入</td>
150
+ <td>用于rope计算的输入数据,对输入张量进行余弦变换,对应公式中的cos。与输入qkv的数据类型相同,shape为[B<sub>qkv</sub> * S<sub>qkv</sub>, 1 * D_rope],D_rope = D<sub>qkv</sub>且要求(D<sub>qkv</sub> * qkv数据类型所占字节数)可以被32整除。</td>
151
+ <td>FLOAT16、BFLOAT16</td>
152
+ <td>ND</td>
153
+ </tr>
154
+ <tr>
155
+ <td>sin</td>
156
+ <td>输入</td>
157
+ <td>用于rope计算的输入数据,对输入张量进行正弦变换,对应公式中的sin。与输入cos的数据类型、格式保持一致。</td>
158
+ <td>FLOAT16、BFLOAT16</td>
159
+ <td>ND</td>
160
+ </tr>
161
+ <tr>
162
+ <td>index</td>
163
+ <td>输入</td>
164
+ <td>用于指定写入cache的具体索引位置。shape为[B<sub>qkv</sub> * S<sub>qkv</sub>]。</td>
165
+ <td>INT64</td>
166
+ <td>ND</td>
167
+ </tr>
168
+ <tr>
169
+ <td>q_out</td>
170
+ <td>输入/输出</td>
171
+ <td>提前申请的cache,输入输出同地址复用。与输入qkv的数据类型相同,shape为[B<sub>qkv</sub> * S<sub>qkv</sub>, N<sub>q</sub> * D<sub>qkv</sub>]。</td>
172
+ <td>FLOAT16、BFLOAT16</td>
173
+ <td>ND</td>
174
+ </tr>
175
+ <tr>
176
+ <td>k_cache</td>
177
+ <td>输入/输出</td>
178
+ <td>提前申请的cache,输入输出同地址复用。与输入qkv的数据类型相同(k不量化),或者INT8(k量化)。shape为[BlockNum, N<sub>k</sub> * D<sub>qkv</sub> // 16, BlockSize, 16](k不量化),或者[BlockNum, N<sub>k</sub> * D<sub>qkv</sub> // 32, BlockSize, 32](k量化)。</td>
179
+ <td>FLOAT16、BFLOAT16、INT8</td>
180
+ <td>ND</td>
181
+ </tr>
182
+ <tr>
183
+ <td>v_cache</td>
184
+ <td>输入/输出</td>
185
+ <td>提前申请的cache,输入输出同地址复用。与输入qkv的数据类型相同(v不量化),或者INT8(v量化)。shape为[BlockNum, N<sub>k</sub> * D<sub>qkv</sub> // 16, BlockSize, 16](v不量化),或者[BlockNum, N<sub>k</sub> * D<sub>qkv</sub> // 32, BlockSize, 32](v量化)。</td>
186
+ <td>FLOAT16、BFLOAT16、INT8</td>
187
+ <td>ND</td>
188
+ </tr>
189
+ <tr>
190
+ <td>k_scale</td>
191
+ <td>可选输入</td>
192
+ <td>当k_cache数据类型为INT8时需要此输入参数,对应公式中的kScale。shape为[N<sub>k</sub>, D<sub>qkv</sub>]。</td>
193
+ <td>FLOAT32</td>
194
+ <td>ND</td>
195
+ </tr>
196
+ <tr>
197
+ <td>v_scale</td>
198
+ <td>可选输入</td>
199
+ <td>当v_cache数据类型为INT8时需要此输入参数,对应公式中的vScale。shape为[N<sub>v</sub>, D<sub>qkv</sub>]。</td>
200
+ <td>FLOAT32</td>
201
+ <td>ND</td>
202
+ </tr>
203
+ <tr>
204
+ <td>k_offset</td>
205
+ <td>可选输入</td>
206
+ <td>当k_cache数据类型为INT8且对应的k_scale输入存在并量化场景为非对称量化时,需要此参数输入,对应公式中的kOffset。shape为[N<sub>k</sub>, D<sub>qkv</sub>]。</td>
207
+ <td>FLOAT32</td>
208
+ <td>ND</td>
209
+ </tr>
210
+ <tr>
211
+ <td>v_offset</td>
212
+ <td>可选输入</td>
213
+ <td>当v_cache数据类型为INT8且对应的v_scale输入存在并量化场景为非对称量化时,需要此参数输入,对应公式中的vOffset。shape为[N<sub>v</sub>, D<sub>qkv</sub>]。</td>
214
+ <td>FLOAT32</td>
215
+ <td>ND</td>
216
+ </tr>
217
+ <tr>
218
+ <td>q_out_before_quant</td>
219
+ <td>可选输出</td>
220
+ <td>即将写入到q_out中的数据。</td>
221
+ <td>FLOAT16、BFLOAT16</td>
222
+ <td>ND</td>
223
+ </tr>
224
+ <tr>
225
+ <td>k_out_before_quant</td>
226
+ <td>可选输出</td>
227
+ <td>即将写入到k_cache中的数据,在未经量化和Scatter前的中间计算结果。</td>
228
+ <td>FLOAT16、BFLOAT16</td>
229
+ <td>ND</td>
230
+ </tr>
231
+ <tr>
232
+ <td>v_out_before_quant</td>
233
+ <td>可选输出</td>
234
+ <td>即将写入到v_cache中的数据,在未经量化和Scatter前的中间计算结果。</td>
235
+ <td>FLOAT16、BFLOAT16</td>
236
+ <td>ND</td>
237
+ </tr>
238
+ <tr>
239
+ <td>qkv_size</td>
240
+ <td>属性</td>
241
+ <td>按[B<sub>qkv</sub>, S<sub>qkv</sub>, N<sub>qkv</sub>, D<sub>qkv</sub>]顺序传入,提供输入参数qkv矩阵的B,S,N,D维度具体尺寸。</td>
242
+ <td>INT64</td>
243
+ <td>-</td>
244
+ </tr>
245
+ <tr>
246
+ <td>head_nums</td>
247
+ <td>属性</td>
248
+ <td>按[N<sub>q</sub>, N<sub>k</sub>, N<sub>v</sub>]顺序传入,提供输入参数qkv矩阵中,qkv分量单元中分的N维度具体尺寸。</td>
249
+ <td>INT64</td>
250
+ <td>-</td>
251
+ </tr>
252
+ <tr>
253
+ <td>epsilon</td>
254
+ <td>可选属性</td>
255
+ <td><ul>
256
+ <li>用于防止RmsNorm计算除0错误,对应公式中的epsilon。</li>
257
+ <li>默认值为1e-6。</li>
258
+ </ul></td>
259
+ <td>FLOAT32</td>
260
+ <td>-</td>
261
+ </tr>
262
+ <tr>
263
+ <td>cache_mode</td>
264
+ <td>可选属性</td>
265
+ <td><ul>
266
+ <li>cache格式的选择标记,目前只支持PA_NZ。</li>
267
+ <li>默认值为PA_NZ。</li>
268
+ </ul></td>
269
+ <td>CHAR*</td>
270
+ <td>-</td>
271
+ </tr>
272
+ <tr>
273
+ <td>is_output_qkv</td>
274
+ <td>可选属性</td>
275
+ <td><ul>
276
+ <li>表示是否需要输出各cache输出中对应内容在未经量化和Scatter前的原始值。</li>
277
+ <li>默认值为false。</li>
278
+ </ul></td>
279
+ <td>BOOL</td>
280
+ <td>-</td>
281
+ </tr>
282
+
283
+ </tbody></table>
284
+
285
+ ## 约束说明
286
+
287
+ * 输入shape限制:
288
+ * B<sub>qkv</sub>为输入qkv的batch_size,S<sub>qkv</sub>为输入qkv的sequence length,大小由qkvSize决定。
289
+ * N<sub>qkv</sub>为输入qkv的head number。D<sub>qkv</sub>为输入qkv的head dim,目前仅支持128。D<sub>q</sub>、D<sub>k</sub>和D<sub>k</sub>分别为q、k、v的head dim,要求D<sub>qkv</sub> = D<sub>q</sub> = D<sub>k</sub> = D<sub>v</sub>,D<sub>qkv</sub>需要满足(D<sub>qkv</sub>*qkv数据类型占字节数)可以被32整除。
290
+ * 根据rope规则,D<sub>k</sub>和D<sub>q</sub>为偶数。若cache_mode为PA_NZ场景下,D<sub>k</sub>、D<sub>q</sub>需32B对齐;BlockSize需32B对齐。
291
+ * 关于上述32B对齐的情形,对齐值由cache的数据类型决定。以BlockSize为例,若cache的数据类型为int8,则需要满足BlockSize % 32 = 0;若cache的数据类型为float16,则需要满足BlockSize % 16 = 0;若k_cache与v_cache参数的dtype不一致,BlockSize需同时满足BlockSize % 32 = 0和BlockSize % 16 = 0。
292
+ * BlockNum为写入cache的内存块数,大小由用户输入场景决定,要求BlockNum >= Ceil(S<sub>qkv</sub> / BlockSize) * B<sub>qkv</sub>。
293
+ * 使用requireMemory表示存放数据所需的空间大小,需满足:requireMemory >= (B<sub>qkv</sub> * S<sub>qkv</sub> * N<sub>qkv</sub> * D<sub>qkv</sub> + 2 * D<sub>qkv</sub> + 2 * B<sub>qkv</sub> * S<sub>qkv</sub> * D<sub>qkv</sub> + B<sub>qkv</sub> * S<sub>qkv</sub> * N<sub>q</sub> * D<sub>qkv</sub> + BlockNum * BlockSize * N<sub>v</sub> * D<sub>qkv</sub> + BlockNum * BlockSize * N<sub>k</sub> * D<sub>qkv</sub>) * sizeof(FLOAT16) + B<sub>qkv</sub> * S<sub>qkv</sub> * sizeof(INT64) + (2 * N<sub>k</sub> * D<sub>qkv</sub> + 2 * N<sub>v</sub>) * sizeof(FLOAT),当计算出requireMemory的大小超过当前AI处理器的GM空间总大小,不支持使用该算子。
294
+ * 其他限制:
295
+ * 对于index,要求index的value值范围为[-1, BlockNum * BlockSize)。value数值不可以重复,index为-1时,代表跳过更新。
296
+ * k_scale, v_scale表示对称量化的缩放因子,因此若传参,则值不能为0。
297
+
298
+ ## 调用说明
299
+
300
+ | 调用方式 | 样例代码 | 说明 |
301
+ | ---------------- | --------------------------- | --------------------------------------------------- |
302
+ | aclnn接口 | [test_aclnn_qkv_rms_norm_rope_cache](examples/test_aclnn_qkv_rms_norm_rope_cache.cpp) | 通过[aclnnQkvRmsNormRopeCache](docs/aclnnQkvRmsNormRopeCache.md)接口方式调用QkvRmsNormRopeCache算子。 |
datasets/ops_src/ops_transformer/test_kernels/kernels/qkv_rms_norm_rope_cache/docs/aclnnQkvRmsNormRopeCache.md ADDED
@@ -0,0 +1,724 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # aclnnQkvRmsNormRopeCache
2
+
3
+ ## 产品支持情况
4
+
5
+ | 产品 | 是否支持 |
6
+ | :----------------------------------------------------------- | :------: |
7
+ | <term>Ascend 950PR/Ascend 950DT</term> | × |
8
+ | <term>Atlas A3 训练系列产品/Atlas A3 推理系列产品</term> | √ |
9
+ | <term>Atlas A2 训练系列产品/Atlas A2 推理系列产品</term> | √ |
10
+ | <term>Atlas 200I/500 A2 推理产品</term> | × |
11
+ | <term>Atlas 推理系列产品</term> | × |
12
+ | <term>Atlas 训练系列产品</term> | × |
13
+
14
+ ## 功能说明
15
+
16
+ - 接口功能:输入qkv融合张量,通过SplitVD拆分q、k、v张量,执行RmsNorm、ApplyRotaryPosEmb、Quant、Scatter融合操作,输出qOut、kCache、vCache、qBeforeQuant(可选)、kBeforeQuant(可选)、vBeforeQuant(可选)。
17
+ - 本接口目前支持的场景如下表:
18
+
19
+ |场景类型|情况概要|
20
+ |:---|:---|
21
+ |<ul><li>cacheMode为PA_NZ</li><li>q无量化</li><li>k和v支持无量化、对称量化和非对称量化</li><li>qBeforeQuant/kBeforeQuant/vBeforeQuant不输出</li></ul>|qkv Shape为[$B_{qkv}$ * $S_{qkv}$, $N_{qkv}$ * $D_{qkv}$],q、k、v具有完全相同的D维度。主要计算过程与输出对应关系:<br><ul><li>qkv经过SplitVD->q、k、v</li><li>q经过RmsNorm、RoPE->qOut</li><li>k经过RmsNorm、RoPE、Quant(可选)、Scatter->kCache</li><li>v经过Quant(可选)、Scatter->vCache</li></ul>|
22
+
23
+ - 计算公式:
24
+
25
+ (1) SplitVD:
26
+
27
+ 下式中,$N_q$、$N_k$、$N_v$分别表示q、k、v分量的注意力头数量,必须满足:
28
+
29
+ $$
30
+ \begin{cases}
31
+ N_k = N_v \\
32
+ N_{qkv} = N_k + N_v + N_q \\
33
+ D_{qkv} = D_q = D_k = D_v
34
+ \end{cases}
35
+ $$
36
+
37
+ $$
38
+ \begin{aligned}
39
+ q &= qkv[..., [:N_q] * D_{qkv}] \\
40
+ k &= qkv[..., [N_q:-N_v] * D_{qkv}] \\
41
+ v &= qkv[..., [-N_v:] * D_{qkv}]
42
+ \end{aligned}
43
+ $$
44
+
45
+ (2) RmsNorm:
46
+
47
+ 此处x和y分别表示RmsNorm的输入张量和输出张量,归一化沿最后一维(feature dimension)进行,该计算规则通用于q、k分量。
48
+
49
+ $$
50
+ squareX = x * x
51
+ $$
52
+
53
+ $$
54
+ meanSquareX = squareX.mean(dim = -1, keepdim = True)
55
+ $$
56
+
57
+ $$
58
+ rms = \sqrt{meanSquareX + epsilon}
59
+ $$
60
+
61
+ $$
62
+ y = (x / rms) * gamma
63
+ $$
64
+
65
+ (3) RoPE (Half-and-Half):
66
+
67
+ 此处的y指代完成RmsNorm计算的输出结果。
68
+
69
+ $$
70
+ y1 = y[\ldots, :d/2]
71
+ $$
72
+
73
+ $$
74
+ y2 = y[\ldots, d/2:]
75
+ $$
76
+
77
+ $$
78
+ y\_RoPE = torch.cat((-y2, y1), dim = -1)
79
+ $$
80
+
81
+ $$
82
+ y\_embed = (y * cos) + y\_RoPE * sin
83
+ $$
84
+
85
+ (4) Quant:
86
+
87
+ 无量化:
88
+
89
+ $$
90
+ kQuant = kRoPE \\
91
+ vQuant = v
92
+ $$
93
+
94
+ 对称量化部分:
95
+
96
+ $$
97
+ kQuant = kRoPE / kScale \\
98
+ vQuant = v / vScale
99
+ $$
100
+
101
+ 非对称量化部分:
102
+
103
+ $$
104
+ kQuant = kRoPE / kScale + kOffset \\
105
+ vQuant = v / vScale + vOffset
106
+ $$
107
+
108
+ ## 函数原型
109
+
110
+ 该算子采用两段式接口,必须先调用“aclnnQkvRmsNormRopeCacheGetWorkspaceSize”接口获得入参并根据流程计算所需workspace大小,再调用“aclnnQkvRmsNormRopeCache”接口执行计算。
111
+
112
+ ```Cpp
113
+ aclnnStatus aclnnQkvRmsNormRopeCacheGetWorkspaceSize(
114
+ const aclTensor *qkv,
115
+ const aclTensor *qGamma,
116
+ const aclTensor *kGamma,
117
+ const aclTensor *cos,
118
+ const aclTensor *sin,
119
+ const aclTensor *index,
120
+ aclTensor *qOut,
121
+ aclTensor *kCache,
122
+ aclTensor *vCache,
123
+ const aclTensor *kScaleOptional,
124
+ const aclTensor *vScaleOptional,
125
+ const aclTensor *kOffsetOptional,
126
+ const aclTensor *vOffsetOptional,
127
+ const aclIntArray *qkvSize,
128
+ const aclIntArray *headNums,
129
+ double epsilon,
130
+ char *cacheModeOptional,
131
+ const aclTensor *qOutBeforeQuant,
132
+ const aclTensor *kOutBeforeQuant,
133
+ const aclTensor *vOutBeforeQuant,
134
+ uint64_t *workspaceSize,
135
+ aclOpExecutor **executor)
136
+ ```
137
+
138
+ ```Cpp
139
+ aclnnStatus aclnnQkvRmsNormRopeCache(
140
+ void *workspace,
141
+ uint64_t workspaceSize,
142
+ aclOpExecutor *executor,
143
+ aclrtStream stream)
144
+ ```
145
+
146
+ ## aclnnQkvRmsNormRopeCacheGetWorkspaceSize
147
+
148
+ - **参数说明:**
149
+
150
+ <table style="undefined;table-layout: fixed; width: 1503px"><colgroup>
151
+ <col style="width: 146px">
152
+ <col style="width: 120px">
153
+ <col style="width: 271px">
154
+ <col style="width: 392px">
155
+ <col style="width: 228px">
156
+ <col style="width: 101px">
157
+ <col style="width: 100px">
158
+ <col style="width: 145px">
159
+ </colgroup>
160
+ <thead>
161
+ <tr>
162
+ <th>参数名</th>
163
+ <th>输入/输出</th>
164
+ <th>描述</th>
165
+ <th>使用说明</th>
166
+ <th>数据类型</th>
167
+ <th>数据格式</th>
168
+ <th>维度(shape)</th>
169
+ <th>非连续Tensor</th>
170
+ </tr></thead>
171
+ <tbody>
172
+ <tr>
173
+ <td>qkv</td>
174
+ <td>输入</td>
175
+ <td>用于切分出q、k、v的输入数据,对应计算公式中的qkv。</td>
176
+ <td><ul><li>不支持空Tensor。</li><li>shape为[B<sub>qkv</sub> * S<sub>qkv</sub>, N<sub>qkv</sub> * D<sub>qkv</sub>]。</li></ul></td>
177
+ <td>FLOAT16、BFLOAT16</td>
178
+ <td>ND</td>
179
+ <td>2</td>
180
+ <td>√</td>
181
+ </tr>
182
+ <tr>
183
+ <td>qGamma</td>
184
+ <td>输入</td>
185
+ <td>用于q的rms_norm计算的输入数据,对应计算公式中的gamma。</td>
186
+ <td><ul><li>不支持空Tensor。</li><li>与输入qkv的数据类型相同。</li><li>shape为[D<sub>qkv</sub>]。</li></ul></td>
187
+ <td>FLOAT16、BFLOAT16</td>
188
+ <td>ND</td>
189
+ <td>1</td>
190
+ <td>√</td>
191
+ </tr>
192
+ <tr>
193
+ <td>kGamma</td>
194
+ <td>输入</td>
195
+ <td>用于k的rms_norm计算的输入数据,对应计算公式中的gamma。</td>
196
+ <td><ul><li>不支持空Tensor。</li><li>与输入qkv的数据类型相同。</li><li>shape为[D<sub>qkv</sub>]。</li></ul></td>
197
+ <td>FLOAT16、BFLOAT16</td>
198
+ <td>ND</td>
199
+ <td>1</td>
200
+ <td>√</td>
201
+ </tr>
202
+ <tr>
203
+ <td>cos</td>
204
+ <td>输入</td>
205
+ <td>用于rope计算的输入数据,对输入张量进行余弦变换,对应计算公式中的cos。</td>
206
+ <td><ul><li>不支持空Tensor。</li><li>与输入qkv的数据类型相同。</li><li>shape为[B<sub>qkv</sub> * S<sub>qkv</sub>, 1 * D_rope],D_rope = D<sub>qkv</sub>且要求(D<sub>qkv</sub> * qkv数据类型所占字节数)可以被32整除。</li></ul></td>
207
+ <td>FLOAT16、BFLOAT16</td>
208
+ <td>ND</td>
209
+ <td>2</td>
210
+ <td>√</td>
211
+ </tr>
212
+ <tr>
213
+ <td>sin</td>
214
+ <td>输入</td>
215
+ <td>用于rope计算的输入数据,对输入张量进行正弦变换,对应计算公式中的sin。</td>
216
+ <td><ul><li>不支持空Tensor。</li><li>与输入cos的数据类型、格式保持一致。</li></ul></td>
217
+ <td>FLOAT16、BFLOAT16</td>
218
+ <td>ND</td>
219
+ <td>2</td>
220
+ <td>√</td>
221
+ </tr>
222
+ <tr>
223
+ <td>index</td>
224
+ <td>输入</td>
225
+ <td>用于指定写入cache的具体索引位置,对应计算公式中的index。</td>
226
+ <td><ul><li>不支持空Tensor。</li><li>与输入qkv的数据类型相同。</li><li>shape为[B<sub>qkv</sub> * S<sub>qkv</sub>],要求index的value不能重复,value的数值范围为[-1, BlockNum * BlockSize]。</li></ul></td>
227
+ <td>INT64</td>
228
+ <td>ND</td>
229
+ <td>1</td>
230
+ <td>√</td>
231
+ </tr>
232
+ <tr>
233
+ <td>qOut</td>
234
+ <td>输入\输出</td>
235
+ <td>提请申请的cache,输入输出同地址复用。</td>
236
+ <td><ul><li>不支持空Tensor。</li><li>与输入qkv的数据类型相同。</li><li>shape为[B<sub>qkv</sub> * S<sub>qkv</sub>, N<sub>q</sub> * D<sub>qkv</sub>]。</li></ul></td>
237
+ <td>FLOAT16、BFLOAT16</td>
238
+ <td>ND</td>
239
+ <td>2</td>
240
+ <td>√</td>
241
+ </tr>
242
+ <tr>
243
+ <td>kCache</td>
244
+ <td>输入\输出</td>
245
+ <td>提请申请的cache,输入输出同地址复用。</td>
246
+ <td><ul><li>不支持空Tensor。</li><li>与输入qkv的数据类型相同(k不量化),或者INT8(k量化)。</li><li>shape为[BlockNum, N<sub>k</sub> * D<sub>qkv</sub> // 16, BlockSize, 16](k不量化),或者[BlockNum, N<sub>k</sub> * D<sub>qkv</sub> // 32, BlockSize, 32](k量化)。</li></ul></td>
247
+ <td>FLOAT16、BFLOAT16、INT8</td>
248
+ <td>ND</td>
249
+ <td>4</td>
250
+ <td>√</td>
251
+ </tr>
252
+ <tr>
253
+ <td>vCache</td>
254
+ <td>输入\输出</td>
255
+ <td>提请申请的cache,输入输出同地址复用。</td>
256
+ <td><ul><li>不支持空Tensor。</li><li>与输入qkv的数据类型相同(v不量化),或者INT8(v量化)。</li><li>shape为[BlockNum, N<sub>k</sub> * D<sub>qkv</sub> // 16, BlockSize, 16](v不量化),或者[BlockNum, N<sub>k</sub> * D<sub>qkv</sub> // 32, BlockSize, 32](v量化)。</li></ul></td>
257
+ <td>FLOAT16、BFLOAT16、INT8</td>
258
+ <td>ND</td>
259
+ <td>4</td>
260
+ <td>√</td>
261
+ </tr>
262
+ <tr>
263
+ <td>kScaleOptional</td>
264
+ <td>可选输入</td>
265
+ <td>当kCache为INT8数据类型时需要存在。对应计算公式中的kScale。</td>
266
+ <td>shape为[N<sub>k</sub>, D<sub>qkv</sub>]。</td>
267
+ <td>FLOAT32</td>
268
+ <td>ND</td>
269
+ <td>2</td>
270
+ <td>√</td>
271
+ </tr>
272
+ <tr>
273
+ <td>vScaleOptional</td>
274
+ <td>可选输入</td>
275
+ <td>当vCache为INT8数据类型时需要存在。对应计算公式中的vScale。</td>
276
+ <td>shape为[N<sub>v</sub>, D<sub>qkv</sub>]。</td>
277
+ <td>FLOAT32</td>
278
+ <td>ND</td>
279
+ <td>2</td>
280
+ <td>√</td>
281
+ </tr>
282
+ <tr>
283
+ <td>kOffsetOptional</td>
284
+ <td>可选输入</td>
285
+ <td>对应计算公式中的kOffset。</td>
286
+ <td><ul><li>当kCache数据类型为INT8且对应的kScaleOptional输入存在并量化场景为非对称量化时,需要此参数输入。</li><li>shape为[N<sub>k</sub>, D<sub>qkv</sub>]。</li></ul></td>
287
+ <td>FLOAT32</td>
288
+ <td>ND</td>
289
+ <td>2</td>
290
+ <td>√</td>
291
+ </tr>
292
+ <tr>
293
+ <td>vOffsetOptional</td>
294
+ <td>可选输入</td>
295
+ <td>对应计算公式中的vOffset。</td>
296
+ <td><ul><li>当vCache数据类型为INT8且对应的vScaleOptional输入存在并量化场景为非对称量化时,需要此参数输入。</li><li>shape为[N<sub>v</sub>, D<sub>qkv</sub>]。</li></ul></td>
297
+ <td>FLOAT32</td>
298
+ <td>ND</td>
299
+ <td>2</td>
300
+ <td>√</td>
301
+ </tr>
302
+ <tr>
303
+ <td>qkvSize</td>
304
+ <td>输入</td>
305
+ <td>按[B<sub>qkv</sub>, S<sub>qkv</sub>, N<sub>qkv</sub>, D<sub>qkv</sub>]顺序传入,提供输入参数qkv矩阵的B,S,N,D维度具体尺寸。</td>
306
+ <td><ul><li>要求N<sub>qkv</sub> = N<sub>q</sub> + N<sub>k</sub> + N<sub>v</sub>。</li><li>元素数量严格等于4,各元素必须为非零有效正整数。</li></ul></td>
307
+ <td>INT64</td>
308
+ <td>-</td>
309
+ <td>-</td>
310
+ <td>-</td>
311
+ </tr>
312
+ <tr>
313
+ <td>headNums</td>
314
+ <td>输入</td>
315
+ <td>按[N<sub>q</sub>, N<sub>k</sub>, N<sub>v</sub>]顺序传入,提供输入参数qkv矩阵中,qkv分量单元中分的N维度具体尺寸。</td>
316
+ <td>[N<sub>q</sub>, N<sub>k</sub>, N<sub>v</sub>]必须为实际原始值,不可进行约化。</td>
317
+ <td>INT64</td>
318
+ <td>-</td>
319
+ <td>-</td>
320
+ <td>-</td>
321
+ </tr>
322
+ <tr>
323
+ <td>epsilon</td>
324
+ <td>输入</td>
325
+ <td>RmsNorm计算中防止除0。对应计算公式中的epsilon。</td>
326
+ <td>建议值为1e-6</td>
327
+ <td>FLOAT32</td>
328
+ <td>-</td>
329
+ <td>-</td>
330
+ <td>-</td>
331
+ </tr>
332
+ <tr>
333
+ <td>cacheModeOptional</td>
334
+ <td>输入</td>
335
+ <td>表示cache的格式。</td>
336
+ <td>建议值为"PA_NZ"。</td>
337
+ <td>CHAR*</td>
338
+ <td>-</td>
339
+ <td>-</td>
340
+ <td>-</td>
341
+ </tr>
342
+ <tr>
343
+ <td>qOutBeforeQuant</td>
344
+ <td>输出</td>
345
+ <td>表示即将写入到qOut中的数据。</td>
346
+ <td><ul><li>分型、数据类型,需要与输入参数qkv保持一致。</li><li>shape为[B<sub>qkv</sub> * S<sub>qkv</sub>, N<sub>q</sub> * D<sub>qkv</sub>]。</li></ul></td>
347
+ <td>FLOAT16、BFLOAT16</td>
348
+ <td>ND</td>
349
+ <td>2</td>
350
+ <td>√</td>
351
+ </tr>
352
+ <tr>
353
+ <td>kOutBeforeQuant</td>
354
+ <td>输出</td>
355
+ <td>表示即将写入到vCache中的数据,在未经量化和Scatter前的中间计算结果。</td>
356
+ <td><ul><li>分型、数据类型,需要与输入参数qkv保持一致。</li><li>shape为[B<sub>qkv</sub> * S<sub>qkv</sub>, N<sub>q</sub> * D<sub>qkv</sub>]。</li></ul></td>
357
+ <td>FLOAT16、BFLOAT16</td>
358
+ <td>ND</td>
359
+ <td>2</td>
360
+ <td>√</td>
361
+ </tr>
362
+ <tr>
363
+ <td>vOutBeforeQuant</td>
364
+ <td>输出</td>
365
+ <td>表示即将写入到vCache中的数据,在未经量化和Scatter前的中间计算结果。</td>
366
+ <td><ul><li>分型、数据类型,需要与输入参数qkv保持一致。</li><li>shape为[B<sub>qkv</sub> * S<sub>qkv</sub>, N<sub>q</sub> * D<sub>qkv</sub>]。</li></ul></td>
367
+ <td>FLOAT16、BFLOAT16</td>
368
+ <td>ND</td>
369
+ <td>2</td>
370
+ <td>√</td>
371
+ </tr>
372
+ <tr>
373
+ <td>workspaceSize</td>
374
+ <td>输出</td>
375
+ <td>返回需要在Device侧申请的workspace大小。</td>
376
+ <td>-</td>
377
+ <td>-</td>
378
+ <td>-</td>
379
+ <td>-</td>
380
+ <td>-</td>
381
+ </tr>
382
+ <tr>
383
+ <td>executor</td>
384
+ <td>输出</td>
385
+ <td>返回op执行器,包含了算子计算流程。</td>
386
+ <td>-</td>
387
+ <td>-</td>
388
+ <td>-</td>
389
+ <td>-</td>
390
+ <td>-</td>
391
+ </tr>
392
+ </tbody>
393
+ </table>
394
+
395
+ - **返回值:**
396
+
397
+ aclnnStatus:返回算子的执行状态码。
398
+
399
+ 第一段接口完成入参校验,出现以下场景时报错:
400
+ <table style="undefined;table-layout: fixed;width: 1155px"><colgroup>
401
+ <col style="width: 253px">
402
+ <col style="width: 140px">
403
+ <col style="width: 762px">
404
+ </colgroup>
405
+ <thead>
406
+ <tr>
407
+ <th>返回值</th>
408
+ <th>错误码</th>
409
+ <th>描述</th>
410
+ </tr>
411
+ </thead>
412
+ <tbody>
413
+ <tr>
414
+ <td rowspan="1">ACLNN_ERR_PARAM_NULLPTR</td>
415
+ <td rowspan="1">161001</td>
416
+ <td>传入的qkv、qGamma、kGamma、cos、sin等为空指针。</td>
417
+ </tr>
418
+ <tr>
419
+ <td rowspan="3">ACLNN_ERR_PARAM_INVALID</td>
420
+ <td rowspan="3">161002</td>
421
+ <td>输入或输出的数据类型不在支持的范围内。</td>
422
+ </tr>
423
+ <tr>
424
+ <td>输入或输出的参数维度不在支持的范围内。</td>
425
+ </tr>
426
+ <tr>
427
+ <td>dim不在指定的取值范围内。</td>
428
+ </tr>
429
+ </tbody></table>
430
+
431
+ ## aclnnQkvRmsNormRopeCache
432
+
433
+ - **参数说明:**
434
+
435
+ <table style="undefined;table-layout: fixed; width: 953px"><colgroup>
436
+ <col style="width: 173px">
437
+ <col style="width: 112px">
438
+ <col style="width: 668px">
439
+ </colgroup>
440
+ <thead>
441
+ <tr>
442
+ <th>参数名</th>
443
+ <th>输入/输出</th>
444
+ <th>描述</th>
445
+ </tr></thead>
446
+ <tbody>
447
+ <tr>
448
+ <td>workspace</td>
449
+ <td>输入</td>
450
+ <td>在Device侧申请的workspace内存地址。</td>
451
+ </tr>
452
+ <tr>
453
+ <td>workspaceSize</td>
454
+ <td>输入</td>
455
+ <td>在Device侧申请的workspace大小,由第一段接口aclnnQkvRmsNormRopeCacheGetWorkspaceSize获取。</td>
456
+ </tr>
457
+ <tr>
458
+ <td>executor</td>
459
+ <td>输入</td>
460
+ <td>op执行器,包含了算子计算流程。</td>
461
+ </tr>
462
+ <tr>
463
+ <td>stream</td>
464
+ <td>输入</td>
465
+ <td>指定执行任务的Stream。</td>
466
+ </tr>
467
+ </tbody>
468
+ </table>
469
+
470
+ - **返回值:**
471
+
472
+ aclnnStatus:返回算子的执行状态码。
473
+
474
+ ## 约束说明
475
+
476
+ - 确定性计算:
477
+ - aclnnQkvRmsNormRopeCache默认确定性实现。
478
+
479
+ - 输入shape限制:
480
+ * B<sub>qkv</sub>为输入qkv的batch_size,S<sub>qkv</sub>为输入qkv的sequence length,大小由qkvSize决定。
481
+ * N<sub>qkv</sub>为输入qkv的head number。D<sub>qkv</sub>为输入qkv的head dim,目前仅支持128。D<sub>q</sub>、D<sub>k</sub>和D<sub>k</sub>分别为q、k、v的head dim,要求D<sub>qkv</sub> = D<sub>q</sub> = D<sub>k</sub> = D<sub>v</sub>,D<sub>qkv</sub>需要满足(D<sub>qkv</sub>*qkv数据类型占字节数)可以被32整除。
482
+ * 根据rope规则,D<sub>k</sub>和D<sub>q</sub>为偶数。若cacheMode为PA_NZ场景下,D<sub>k</sub>、D<sub>q</sub>需32B对齐;BlockSize需32B对齐。
483
+ * 关于上述32B对齐的情形,对齐值由cache的数据类型决定。以BlockSize为例,若cache的数据类型为int8,则需要满足BlockSize % 32 = 0;若cache的数据类型为float16,则需要满足BlockSize % 16 = 0;若kCache与vCache参数的dtype不一致,BlockSize需同时满足BlockSize % 32 = 0和BlockSize % 16 = 0。
484
+ * BlockNum为写入cache的内存块数,大小由用户输入场景决定,要求BlockNum >= Ceil(S<sub>qkv</sub> / BlockSize) * B<sub>qkv</sub>。
485
+ * 使用requireMemory表示存放数据所需的空间大小,需满足:requireMemory >= (B<sub>qkv</sub> * S<sub>qkv</sub> * N<sub>qkv</sub> * D<sub>qkv</sub> + 2 * D<sub>qkv</sub> + 2 * B<sub>qkv</sub> * S<sub>qkv</sub> * D<sub>qkv</sub> + B<sub>qkv</sub> * S<sub>qkv</sub> * N<sub>q</sub> * D<sub>qkv</sub> + BlockNum * BlockSize * N<sub>v</sub> * D<sub>qkv</sub> + BlockNum * BlockSize * N<sub>k</sub> * D<sub>qkv</sub>) * sizeof(FLOAT16) + B<sub>qkv</sub> * S<sub>qkv</sub> * sizeof(INT64) + (2 * N<sub>k</sub> * D<sub>qkv</sub> + 2 * N<sub>v</sub>) * sizeof(FLOAT),当计算出requireMemory的大小超过当前AI处理器的GM空间总大小,不支持使用该接口。
486
+ - 其他限制:
487
+ * 对于index,要求index的value值范围为[-1, BlockNum * BlockSize)。value数值不可以重复,index为-1时,代表跳过更新。
488
+ * kScaleOptional, vScaleOptional表示对称量化的缩放因子,因此若传参,则值不能为0。
489
+
490
+ ## 调用示例
491
+
492
+ 以下示例代码展示接口调用流程,可供编译和运行时参考。
493
+
494
+ ```Cpp
495
+ #include <iostream>
496
+ #include <vector>
497
+ #include "acl/acl.h"
498
+ #include "aclnnop/aclnn_qkv_rms_norm_rope_cache.h"
499
+
500
+ #define CHECK_RET(cond, return_expr) \
501
+ do { \
502
+ if (!(cond)) { \
503
+ return_expr; \
504
+ } \
505
+ } while (0)
506
+
507
+ #define LOG_PRINT(message, ...) \
508
+ do { \
509
+ printf(message, ##__VA_ARGS__); \
510
+ } while (0)
511
+
512
+ int64_t GetShapeSize(const std::vector<int64_t>& shape)
513
+ {
514
+ int64_t shapeSize = 1;
515
+ for (auto i : shape) {
516
+ shapeSize *= i;
517
+ }
518
+ return shapeSize;
519
+ }
520
+
521
+ void PrintOutResult(std::vector<int64_t>& shape, void** deviceAddr)
522
+ {
523
+ auto size = GetShapeSize(shape);
524
+ std::vector<int8_t> resultData(size, 0);
525
+ auto ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]), *deviceAddr, size * sizeof(resultData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
526
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return);
527
+ for (int64_t i = 0; i < size; i++) {
528
+ LOG_PRINT("result[%ld] is: %d\n", i, resultData[i]);
529
+ }
530
+ }
531
+
532
+ int Init(int32_t deviceId, aclrtStream* stream)
533
+ {
534
+ // 固定写法,资源初始化
535
+ auto ret = aclInit(nullptr);
536
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclInit failed. ERROR: %d\n", ret); return ret);
537
+ ret = aclrtSetDevice(deviceId);
538
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret); return ret);
539
+ ret = aclrtCreateStream(stream);
540
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateStream failed. ERROR: %d\n", ret); return ret);
541
+ return 0;
542
+ }
543
+
544
+ template <typename T>
545
+ int CreateAclTensor(
546
+ const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr, aclDataType dataType,
547
+ aclTensor** tensor)
548
+ {
549
+ auto size = GetShapeSize(shape) * sizeof(T);
550
+ // 调用aclrtMalloc申请device侧内存
551
+ auto ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
552
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMalloc failed. ERROR: %d\n", ret); return ret);
553
+ // 调用aclrtMemcpy将host侧数据拷贝到device侧内存上
554
+ ret = aclrtMemcpy(*deviceAddr, size, hostData.data(), size, ACL_MEMCPY_HOST_TO_DEVICE);
555
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMemcpy failed. ERROR: %d\n", ret); return ret);
556
+
557
+ // 计算连续tensor的strides
558
+ std::vector<int64_t> strides(shape.size(), 1);
559
+ for (int64_t i = shape.size() - 2; i >= 0; i--) {
560
+ strides[i] = shape[i + 1] * strides[i + 1];
561
+ }
562
+
563
+ // 调用aclCreateTensor接口创建aclTensor
564
+ *tensor = aclCreateTensor(
565
+ shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND, shape.data(), shape.size(),
566
+ *deviceAddr);
567
+ return 0;
568
+ }
569
+
570
+ int main()
571
+ {
572
+ // 1. 固定写法,device/stream初始化,参考acl API
573
+ // 根据自己的实际device填写deviceId
574
+ int32_t deviceId = 0;
575
+ aclrtStream stream;
576
+ auto ret = Init(deviceId, &stream);
577
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);
578
+
579
+ // 2. 构造输入与输出,需要根据API的接口定义构造
580
+ std::vector<int64_t> qkvShape = {48, 2304};
581
+ std::vector<int64_t> qGammaShape = {128};
582
+ std::vector<int64_t> kGammaShape = {128};
583
+ std::vector<int64_t> cosShape = {48, 128};
584
+ std::vector<int64_t> sinShape = {48, 128};
585
+ std::vector<int64_t> indexShape = {48};
586
+ std::vector<int64_t> qOutShape = {48, 2048};
587
+ std::vector<int64_t> kCacheShape = {16, 4, 128, 32};
588
+ std::vector<int64_t> vCacheShape = {16, 4, 128, 32};
589
+ std::vector<int64_t> kScaleShape = {1, 128};
590
+ std::vector<int64_t> vScaleShape = {1, 128};
591
+ std::vector<int64_t> qkv_size_list = {16, 3, 18, 128};
592
+ std::vector<int64_t> head_nums_list = {16, 1, 1};
593
+
594
+ std::vector<int16_t> qkvHostData(48 * 2304, 0);
595
+ std::vector<int16_t> qGammaHostData(128, 0);
596
+ std::vector<int16_t> kGammaHostData(128, 0);
597
+ std::vector<int16_t> cosHostData(48 * 128, 0);
598
+ std::vector<int16_t> sinHostData(48 * 128, 0);
599
+ std::vector<int64_t> indexHostData(48, 0);
600
+ std::vector<int16_t> qOutHostData(48 * 2048, 0);
601
+ std::vector<int16_t> kCacheHostData(16 * 4 * 128 * 32, 0);
602
+ std::vector<int16_t> vCacheHostData(16 * 4 * 128 * 32, 0);
603
+ std::vector<int16_t> kScaleHostData(1 * 128, 0);
604
+ std::vector<int16_t> vScaleHostData(1 * 128, 0);
605
+
606
+ void* qkvDeviceAddr = nullptr;
607
+ void* qGammaDeviceAddr = nullptr;
608
+ void* kGammaDeviceAddr = nullptr;
609
+ void* cosDeviceAddr = nullptr;
610
+ void* sinDeviceAddr = nullptr;
611
+ void* indexDeviceAddr = nullptr;
612
+ void* qOutDeviceAddr = nullptr;
613
+ void* kCacheDeviceAddr = nullptr;
614
+ void* vCacheDeviceAddr = nullptr;
615
+ void* kScaleDeviceAddr = nullptr;
616
+ void* vScaleDeviceAddr = nullptr;
617
+
618
+ aclTensor* qkv = nullptr;
619
+ aclTensor* qGamma = nullptr;
620
+ aclTensor* kGamma = nullptr;
621
+ aclTensor* cos = nullptr;
622
+ aclTensor* sin = nullptr;
623
+ aclTensor* index = nullptr;
624
+ aclTensor* qOut = nullptr;
625
+ aclTensor* kCache = nullptr;
626
+ aclTensor* vCache = nullptr;
627
+ aclTensor* kScale = nullptr;
628
+ aclTensor* vScale = nullptr;
629
+
630
+ aclIntArray *qkv_size = aclCreateIntArray(qkv_size_list.data(), qkv_size_list.size());
631
+ aclIntArray *head_nums = aclCreateIntArray(head_nums_list.data(), head_nums_list.size());
632
+
633
+ double epsilon = 1e-6;
634
+ char* cacheMode = "PA_NZ";
635
+ bool isOutputQkv = false;
636
+
637
+ ret = CreateAclTensor(qkvHostData, qkvShape, &qkvDeviceAddr, aclDataType::ACL_FLOAT16, &qkv);
638
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
639
+ ret = CreateAclTensor(qGammaHostData, qGammaShape, &qGammaDeviceAddr, aclDataType::ACL_FLOAT16, &qGamma);
640
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
641
+ ret = CreateAclTensor(kGammaHostData, kGammaShape, &kGammaDeviceAddr, aclDataType::ACL_FLOAT16, &kGamma);
642
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
643
+ ret = CreateAclTensor(cosHostData, cosShape, &cosDeviceAddr, aclDataType::ACL_FLOAT16, &cos);
644
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
645
+ ret = CreateAclTensor(sinHostData, sinShape, &sinDeviceAddr, aclDataType::ACL_FLOAT16, &sin);
646
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
647
+ ret = CreateAclTensor(indexHostData, indexShape, &indexDeviceAddr, aclDataType::ACL_INT64, &index);
648
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
649
+ ret = CreateAclTensor(qOutHostData, qOutShape, &qOutDeviceAddr, aclDataType::ACL_FLOAT16, &qOut);
650
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
651
+ ret = CreateAclTensor(kCacheHostData, kCacheShape, &kCacheDeviceAddr, aclDataType::ACL_INT8, &kCache);
652
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
653
+ ret = CreateAclTensor(vCacheHostData, vCacheShape, &vCacheDeviceAddr, aclDataType::ACL_INT8, &vCache);
654
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
655
+ ret = CreateAclTensor(kScaleHostData, kScaleShape, &kScaleDeviceAddr, aclDataType::ACL_FLOAT, &kScale);
656
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
657
+ ret = CreateAclTensor(vScaleHostData, vScaleShape, &vScaleDeviceAddr, aclDataType::ACL_FLOAT, &vScale);
658
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
659
+
660
+ // 3. 调用CANN算子库API,需要修改为具体的API
661
+ uint64_t workspaceSize = 0;
662
+ aclOpExecutor* executor;
663
+
664
+ // 调用aclnnQkvRmsNormRopeCache第一段接口
665
+ ret = aclnnQkvRmsNormRopeCacheGetWorkspaceSize(
666
+ qkv, qGamma, kGamma, cos, sin, index, qOut, kCache, vCache, kScale, vScale, nullptr, nullptr, qkv_size,
667
+ head_nums, epsilon, cacheMode, nullptr, nullptr, nullptr, &workspaceSize, &executor);
668
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnQkvRmsNormRopeCacheGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
669
+
670
+ // 根据第一段接口计算出的workspaceSize申请device内存
671
+ void* workspaceAddr = nullptr;
672
+ if (workspaceSize > 0) {
673
+ ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
674
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("allocate workspace failed. ERROR: %d\n", ret); return ret);
675
+ }
676
+
677
+ // 调用aclnnQkvRmsNormRopeCache第二段接口
678
+ ret = aclnnQkvRmsNormRopeCache(workspaceAddr, workspaceSize, executor, stream);
679
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnQkvRmsNormRopeCache failed. ERROR: %d\n", ret); return ret);
680
+
681
+ // 4. 固定写法,同步等待任务执行结束
682
+ ret = aclrtSynchronizeStream(stream);
683
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);
684
+
685
+ // 5. 获取输出的值,将device侧内存上的结果拷贝至host侧,需要根据具体API的接口定义修改
686
+ PrintOutResult(kCacheShape, &kCacheDeviceAddr);
687
+ PrintOutResult(vCacheShape, &vCacheDeviceAddr);
688
+
689
+ // 6. 释放aclTensor和aclScalar,需要根据具体API的接口定义修改
690
+ aclDestroyTensor(qkv);
691
+ aclDestroyTensor(qGamma);
692
+ aclDestroyTensor(kGamma);
693
+ aclDestroyTensor(cos);
694
+ aclDestroyTensor(sin);
695
+ aclDestroyTensor(index);
696
+ aclDestroyTensor(qOut);
697
+ aclDestroyTensor(kCache);
698
+ aclDestroyTensor(vCache);
699
+ aclDestroyTensor(kScale);
700
+ aclDestroyTensor(vScale);
701
+
702
+ // 7. 释放device资源
703
+ aclrtFree(qkvDeviceAddr);
704
+ aclrtFree(qGammaDeviceAddr);
705
+ aclrtFree(kGammaDeviceAddr);
706
+ aclrtFree(cosDeviceAddr);
707
+ aclrtFree(sinDeviceAddr);
708
+ aclrtFree(indexDeviceAddr);
709
+ aclrtFree(qOutDeviceAddr);
710
+ aclrtFree(kCacheDeviceAddr);
711
+ aclrtFree(vCacheDeviceAddr);
712
+ aclrtFree(kScaleDeviceAddr);
713
+ aclrtFree(vScaleDeviceAddr);
714
+
715
+ if (workspaceSize > 0) {
716
+ aclrtFree(workspaceAddr);
717
+ }
718
+ aclrtDestroyStream(stream);
719
+ aclrtResetDevice(deviceId);
720
+ aclFinalize();
721
+
722
+ return 0;
723
+ }
724
+ ```
datasets/ops_src/ops_transformer/test_kernels/kernels/qkv_rms_norm_rope_cache/examples/test_aclnn_qkv_rms_norm_rope_cache.cpp ADDED
@@ -0,0 +1,239 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ #include <iostream>
12
+ #include <vector>
13
+ #include "acl/acl.h"
14
+ #include "aclnnop/aclnn_qkv_rms_norm_rope_cache.h"
15
+
16
+ #define CHECK_RET(cond, return_expr) \
17
+ do { \
18
+ if (!(cond)) { \
19
+ return_expr; \
20
+ } \
21
+ } while (0)
22
+
23
+ #define LOG_PRINT(message, ...) \
24
+ do { \
25
+ printf(message, ##__VA_ARGS__); \
26
+ } while (0)
27
+
28
+ int64_t GetShapeSize(const std::vector<int64_t>& shape)
29
+ {
30
+ int64_t shapeSize = 1;
31
+ for (auto i : shape) {
32
+ shapeSize *= i;
33
+ }
34
+ return shapeSize;
35
+ }
36
+
37
+ void PrintOutResult(std::vector<int64_t>& shape, void** deviceAddr)
38
+ {
39
+ auto size = GetShapeSize(shape);
40
+ std::vector<int8_t> resultData(size, 0);
41
+ auto ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]), *deviceAddr, size * sizeof(resultData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
42
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return);
43
+ for (int64_t i = 0; i < size; i++) {
44
+ LOG_PRINT("result[%ld] is: %d\n", i, resultData[i]);
45
+ }
46
+ }
47
+
48
+ int Init(int32_t deviceId, aclrtStream* stream)
49
+ {
50
+ // 固定写法,资源初始化
51
+ auto ret = aclInit(nullptr);
52
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclInit failed. ERROR: %d\n", ret); return ret);
53
+ ret = aclrtSetDevice(deviceId);
54
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret); return ret);
55
+ ret = aclrtCreateStream(stream);
56
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateStream failed. ERROR: %d\n", ret); return ret);
57
+ return 0;
58
+ }
59
+
60
+ template <typename T>
61
+ int CreateAclTensor(
62
+ const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr, aclDataType dataType,
63
+ aclTensor** tensor)
64
+ {
65
+ auto size = GetShapeSize(shape) * sizeof(T);
66
+ // 调用aclrtMalloc申请device侧内存
67
+ auto ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
68
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMalloc failed. ERROR: %d\n", ret); return ret);
69
+ // 调用aclrtMemcpy将host侧数据拷贝到device侧内存上
70
+ ret = aclrtMemcpy(*deviceAddr, size, hostData.data(), size, ACL_MEMCPY_HOST_TO_DEVICE);
71
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMemcpy failed. ERROR: %d\n", ret); return ret);
72
+
73
+ // 计算连续tensor的strides
74
+ std::vector<int64_t> strides(shape.size(), 1);
75
+ for (int64_t i = shape.size() - 2; i >= 0; i--) {
76
+ strides[i] = shape[i + 1] * strides[i + 1];
77
+ }
78
+
79
+ // 调用aclCreateTensor接口创建aclTensor
80
+ *tensor = aclCreateTensor(
81
+ shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND, shape.data(), shape.size(),
82
+ *deviceAddr);
83
+ return 0;
84
+ }
85
+
86
+ int main()
87
+ {
88
+ // 1. 固定写法,device/stream初始化, 参考acl API
89
+ // 根据自己的实际device填写deviceId
90
+ int32_t deviceId = 0;
91
+ aclrtStream stream;
92
+ auto ret = Init(deviceId, &stream);
93
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);
94
+
95
+ // 2. 构造输入与输出,需要根据API的接口定义构造
96
+ std::vector<int64_t> qkvShape = {48, 2304};
97
+ std::vector<int64_t> qGammaShape = {128};
98
+ std::vector<int64_t> kGammaShape = {128};
99
+ std::vector<int64_t> cosShape = {48, 128};
100
+ std::vector<int64_t> sinShape = {48, 128};
101
+ std::vector<int64_t> indexShape = {48};
102
+ std::vector<int64_t> qOutShape = {48, 2048};
103
+ std::vector<int64_t> kCacheShape = {16, 4, 128, 32};
104
+ std::vector<int64_t> vCacheShape = {16, 4, 128, 32};
105
+ std::vector<int64_t> kScaleShape = {1, 128};
106
+ std::vector<int64_t> vScaleShape = {1, 128};
107
+ std::vector<int64_t> qkv_size_list = {16, 3, 18, 128};
108
+ std::vector<int64_t> head_nums_list = {16, 1, 1};
109
+
110
+ std::vector<int16_t> qkvHostData(48 * 2304, 0);
111
+ std::vector<int16_t> qGammaHostData(128, 0);
112
+ std::vector<int16_t> kGammaHostData(128, 0);
113
+ std::vector<int16_t> cosHostData(48 * 128, 0);
114
+ std::vector<int16_t> sinHostData(48 * 128, 0);
115
+ std::vector<int64_t> indexHostData(48, 0);
116
+ std::vector<int16_t> qOutHostData(48 * 2048, 0);
117
+ std::vector<int16_t> kCacheHostData(16 * 4 * 128 * 32, 0);
118
+ std::vector<int16_t> vCacheHostData(16 * 4 * 128 * 32, 0);
119
+ std::vector<int16_t> kScaleHostData(1 * 128, 0);
120
+ std::vector<int16_t> vScaleHostData(1 * 128, 0);
121
+
122
+ void* qkvDeviceAddr = nullptr;
123
+ void* qGammaDeviceAddr = nullptr;
124
+ void* kGammaDeviceAddr = nullptr;
125
+ void* cosDeviceAddr = nullptr;
126
+ void* sinDeviceAddr = nullptr;
127
+ void* indexDeviceAddr = nullptr;
128
+ void* qOutDeviceAddr = nullptr;
129
+ void* kCacheDeviceAddr = nullptr;
130
+ void* vCacheDeviceAddr = nullptr;
131
+ void* kScaleDeviceAddr = nullptr;
132
+ void* vScaleDeviceAddr = nullptr;
133
+
134
+ aclTensor* qkv = nullptr;
135
+ aclTensor* qGamma = nullptr;
136
+ aclTensor* kGamma = nullptr;
137
+ aclTensor* cos = nullptr;
138
+ aclTensor* sin = nullptr;
139
+ aclTensor* index = nullptr;
140
+ aclTensor* qOut = nullptr;
141
+ aclTensor* kCache = nullptr;
142
+ aclTensor* vCache = nullptr;
143
+ aclTensor* kScale = nullptr;
144
+ aclTensor* vScale = nullptr;
145
+
146
+ aclIntArray *qkv_size = aclCreateIntArray(qkv_size_list.data(), qkv_size_list.size());
147
+ aclIntArray *head_nums = aclCreateIntArray(head_nums_list.data(), head_nums_list.size());
148
+
149
+ double epsilon = 1e-6;
150
+ char* cacheMode = "PA_NZ";
151
+ bool isOutputQkv = false;
152
+
153
+ ret = CreateAclTensor(qkvHostData, qkvShape, &qkvDeviceAddr, aclDataType::ACL_FLOAT16, &qkv);
154
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
155
+ ret = CreateAclTensor(qGammaHostData, qGammaShape, &qGammaDeviceAddr, aclDataType::ACL_FLOAT16, &qGamma);
156
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
157
+ ret = CreateAclTensor(kGammaHostData, kGammaShape, &kGammaDeviceAddr, aclDataType::ACL_FLOAT16, &kGamma);
158
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
159
+ ret = CreateAclTensor(cosHostData, cosShape, &cosDeviceAddr, aclDataType::ACL_FLOAT16, &cos);
160
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
161
+ ret = CreateAclTensor(sinHostData, sinShape, &sinDeviceAddr, aclDataType::ACL_FLOAT16, &sin);
162
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
163
+ ret = CreateAclTensor(indexHostData, indexShape, &indexDeviceAddr, aclDataType::ACL_INT64, &index);
164
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
165
+ ret = CreateAclTensor(qOutHostData, qOutShape, &qOutDeviceAddr, aclDataType::ACL_FLOAT16, &qOut);
166
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
167
+ ret = CreateAclTensor(kCacheHostData, kCacheShape, &kCacheDeviceAddr, aclDataType::ACL_INT8, &kCache);
168
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
169
+ ret = CreateAclTensor(vCacheHostData, vCacheShape, &vCacheDeviceAddr, aclDataType::ACL_INT8, &vCache);
170
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
171
+ ret = CreateAclTensor(kScaleHostData, kScaleShape, &kScaleDeviceAddr, aclDataType::ACL_FLOAT, &kScale);
172
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
173
+ ret = CreateAclTensor(vScaleHostData, vScaleShape, &vScaleDeviceAddr, aclDataType::ACL_FLOAT, &vScale);
174
+ CHECK_RET(ret == ACL_SUCCESS, return ret);
175
+
176
+ // 3. 调用CANN算子库API,需要修改为具体的API
177
+ uint64_t workspaceSize = 0;
178
+ aclOpExecutor* executor;
179
+
180
+ // 调用aclnnQkvRmsNormRopeCache第一段接口
181
+ ret = aclnnQkvRmsNormRopeCacheGetWorkspaceSize(
182
+ qkv, qGamma, kGamma, cos, sin, index, qOut, kCache, vCache, kScale, vScale, nullptr, nullptr, qkv_size,
183
+ head_nums, epsilon, cacheMode, nullptr, nullptr, nullptr, &workspaceSize, &executor);
184
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnQkvRmsNormRopeCacheGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
185
+
186
+ // 根据第一段接口计算出的workspaceSize申请device内存
187
+ void* workspaceAddr = nullptr;
188
+ if (workspaceSize > 0) {
189
+ ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
190
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("allocate workspace failed. ERROR: %d\n", ret); return ret);
191
+ }
192
+
193
+ // 调用aclnnQkvRmsNormRopeCache第二段接口
194
+ ret = aclnnQkvRmsNormRopeCache(workspaceAddr, workspaceSize, executor, stream);
195
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnQkvRmsNormRopeCache failed. ERROR: %d\n", ret); return ret);
196
+
197
+ // 4. 固定写法,同步等待任务执行结束
198
+ ret = aclrtSynchronizeStream(stream);
199
+ CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);
200
+
201
+ // 5. 获取输出的值,将device侧内存上的结果拷贝至host侧,需要根据具体API的接口定义修改
202
+ PrintOutResult(kCacheShape, &kCacheDeviceAddr);
203
+ PrintOutResult(vCacheShape, &vCacheDeviceAddr);
204
+
205
+ // 6. 释放aclTensor和aclScalar,需要根据具体API的接口定义修改
206
+ aclDestroyTensor(qkv);
207
+ aclDestroyTensor(qGamma);
208
+ aclDestroyTensor(kGamma);
209
+ aclDestroyTensor(cos);
210
+ aclDestroyTensor(sin);
211
+ aclDestroyTensor(index);
212
+ aclDestroyTensor(qOut);
213
+ aclDestroyTensor(kCache);
214
+ aclDestroyTensor(vCache);
215
+ aclDestroyTensor(kScale);
216
+ aclDestroyTensor(vScale);
217
+
218
+ // 7. 释放device 资源
219
+ aclrtFree(qkvDeviceAddr);
220
+ aclrtFree(qGammaDeviceAddr);
221
+ aclrtFree(kGammaDeviceAddr);
222
+ aclrtFree(cosDeviceAddr);
223
+ aclrtFree(sinDeviceAddr);
224
+ aclrtFree(indexDeviceAddr);
225
+ aclrtFree(qOutDeviceAddr);
226
+ aclrtFree(kCacheDeviceAddr);
227
+ aclrtFree(vCacheDeviceAddr);
228
+ aclrtFree(kScaleDeviceAddr);
229
+ aclrtFree(vScaleDeviceAddr);
230
+
231
+ if (workspaceSize > 0) {
232
+ aclrtFree(workspaceAddr);
233
+ }
234
+ aclrtDestroyStream(stream);
235
+ aclrtResetDevice(deviceId);
236
+ aclFinalize();
237
+
238
+ return 0;
239
+ }
datasets/ops_src/ops_transformer/test_kernels/kernels/qkv_rms_norm_rope_cache/op_graph/CMakeLists.txt ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # This program is free software, you can redistribute it and/or modify.
2
+ # Copyright (c) 2026 Huawei Technologies Co., Ltd.
3
+ # This file is a part of the CANN Open Software.
4
+ # Licensed under CANN Open Software License Agreement Version 2.0 (the "License").
5
+ # Please refer to the License for details. You may not use this file except in compliance with the License.
6
+ # THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
7
+ # See LICENSE in the root of the software repository for the full text of the License.
8
+ # ============================================================================
9
+
10
+ message(STATUS "=== Debug: start ops.posembedding.qkv_rms_norm_rope_cache.graph_plugin.CMakeLists.txt ")
11
+ add_graph_plugin_sources()
datasets/ops_src/ops_transformer/test_kernels/kernels/qkv_rms_norm_rope_cache/op_graph/qkv_rms_norm_rope_cache_proto.h ADDED
@@ -0,0 +1,121 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /**
2
+  * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3
+  * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4
+  * CANN Open Software License Agreement Version 2.0 (the "License").
5
+  * Please refer to the License for details. You may not use this file except in compliance with the License.
6
+  * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7
+  * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8
+  * See LICENSE in the root of the software repository for the full text of the License.
9
+  */
10
+
11
+ /*!
12
+ * \file qkv_rms_norm_rope_cache_proto.h
13
+ * \brief
14
+ */
15
+ #ifndef OPS_OP_PROTO_INC_QKV_RMS_NORM_ROPE_CACHE_H_
16
+ #define OPS_OP_PROTO_INC_QKV_RMS_NORM_ROPE_CACHE_H_
17
+
18
+ #include "graph/operator_reg.h"
19
+ #include "graph/types.h"
20
+
21
+ namespace ge {
22
+ /**
23
+ * @brief The fusion operator of SpiltVD, RMSNorm, RotaryPositionEmbedding, Quant and Update KVCache.
24
+ *
25
+ * @par Inputs:
26
+ * @li qkv: A tensor. The type support float16 and bfloat16. Its shape is [Bqkv*Sqkv,Nqkv*Dqkv]. Format: ND.
27
+ * @li q_gamma: A tensor, used in RMS Norm. Its type is consistent with qkv. Its shape is [Dqkv,]. Format: ND.
28
+ * @li k_gamma: A tensor, used in RMS Norm. Its type is consistent with qkv. Its shape is [Dqkv,]. Format: ND.
29
+ * @li cos: A tensor, from position embedding. Its type is consistent with kv. Its shape is [Bqkv*Sqkv,Dqkv]. Format: ND.
30
+ * @li sin: A tensor, from position embedding. Its type is consistent with qkv. [Bqkv*Sqkv,Dqkv].
31
+ * The shape of sin should be consistent with that of cos. Format: ND.
32
+ * @li index: A tensor. The type support int64. When the cache_mode is PA_NZ, its shape is [Bqkv*Sqkv] Format: ND.
33
+ * @li q_out: A tensor. Its type is consistent with qkv. Its shape is [Bqkv*Sqkv, Nq*Dqkv]. Format: ND.
34
+ * @li k_cache: A tensor. Its type is consistent with qkv or int8. When the type is int8,
35
+ * its shape is [BlockNum, Nk * Dqkv // 32, BlockSize, 32]. when the type consistent with qkv,
36
+ * its shape is [BlockNum, Nk * Dqkv // 16, BlockSize, 16]. Format: ND.
37
+ * @li v_cache: A tensor. Its type is consistent with qkv or int8. When the type is int8,
38
+ * its shape is [BlockNum, Nv * Dqkv // 32, BlockSize, 32]. when the type consistent with qkv,
39
+ * its shape is [BlockNum, Nv * Dqkv // 16, BlockSize, 16]. Format: ND.
40
+ * @li k_scale: A tensor. The type support float32. This input parameter is required when the data type of k_cache
41
+ * is int8. Its shape can be [Nk,Dqkv]. Format: ND.
42
+ * @li v_scale: A tensor. The type support float32. This input parameter is required when the data type of v_cache
43
+ * is int8. Its shape can be [Nv,Dqkv]. Format: ND.
44
+ * @li k_offset: A tensor. The type support float32. When the data type of k_cache is int8 and the corresponding
45
+ * input exists with the quantization scenario being asymmetric quantization, this parameter input is required. Its
46
+ * shape can be [Nk,Dqkv]. Format: ND.
47
+ * @li v_offset: A tensor. The type support float32. When the data type of v_cache is int8 and the corresponding
48
+ * input exists with the quantization scenario being asymmetric quantization, this parameter input is required. Its
49
+ * shape can be [Nv,Dqkv]. Format: ND.
50
+ *
51
+ * @par Outputs:
52
+ * @li q_out: A tensor. Its type is consistent with qkv. Its shape is [Bqkv*Sqkv, Nq*Dqkv]. Format: ND.
53
+ * @li k_cache: A tensor. Its type is consistent with qkv or int8. When the type is int8,
54
+ * its shape is [BlockNum, Nk * Dqkv // 32, BlockSize, 32]. when the type consistent with qkv,
55
+ * its shape is [BlockNum, Nk * Dqkv // 16, BlockSize, 16]. Format: ND.
56
+ * @li v_cache: A tensor. Its type is consistent with qkv or int8. When the type is int8,
57
+ * its shape is [BlockNum, Nk * Dqkv // 32, BlockSize, 32]. when the type consistent with qkv,
58
+ * its shape is [BlockNum, Nk * Dqkv // 16, BlockSize, 16]. Format: ND.
59
+ * @li q_out_before_quant: A tensor. Its type is consistent with qkv. If is_output_qkv is true, this parameter output is required.
60
+ * Its shape is [Bqkv*Sqkv, Nq*Dqkv]. Format: ND.
61
+ * @li k_out_before_quant: A tensor. Its type is consistent with qkv. If is_output_qkv is true, this parameter output is required.
62
+ * Its shape is [Bqkv*Sqkv, Nk*Dqkv]. Format: ND.
63
+ * @li v_out_before_quant: A tensor. Its type is consistent with qkv. If is_output_qkv is true, this parameter output is required.
64
+ * Its shape is [Bqkv*Sqkv, Nv*Dqkv]. Format: ND.
65
+ *
66
+ * @par Attributes:
67
+ * @li qkv_size: A Listint. Provide the specific dimension sizes of the input of qkv matrices in the order of [Bqkv, Sqkv, Nqkv, Dqkv].
68
+ * @li head_nums: A Listint. Specifying the N-dim,ensional sizes of the q, k, v components in the qkv input matrix, in the order [Nq, Nk, Nv].
69
+ * @li epsilon: A float32. The epsilon value for RMSNorm. Default: 1e-6.
70
+ * @li cache_mode: A string. The cache dtype. There is one types:PA_NZ. Default: PA_NZ.
71
+ * @li is_output_qkv: A bool. Control whether k_cache_proto and v_cache_proto are output. If it is true, then k_out_before_quant and v_out_before_quant need to be
72
+ * output. If it is false, then k_cache_proto and v_cache_proto do not need to be output. Default: false.
73
+ *
74
+ * @attention Constraints:
75
+ * @li The shape format description in the parameter description:
76
+ * - Bqkv represents the batch size of the input qkv, and Skv represents the sequence length of the input qkv. The
77
+ * size is determined by the user's input scenario and has no explicit restrictions.
78
+ * - Nqkv is the head number of the input qkv. This operator is strongly related to the Qwen3-Moe 235b network structure.
79
+ * - Dqkv is the head dim of the input qkv. The data Dqkv required for rms norm calculation and the data Dqkv required
80
+ * for rope calculation are split from the Dqkv of the input qkv. In rope calculation, Drope=Dqkv.
81
+ * Therefore, Dqkv must comply with the rope rules. According to the rope rule, Dqkv is an even number.
82
+ * If cache_mode is an NZ scenario(cahce_mode is PA_NZ), Dqkv needs to be 32B aligned.
83
+ * - If cache_mode is a PA scenario, the BlockSize needs to be 32B aligned.
84
+ * - Regarding the above-mentioned 32B alignment situation, the alignment value is determined by the data type
85
+ * of the cache. Take BlockSize as an example. If the data type of the cache is int8, then BlockSize%32=0 is
86
+ * required. If the data type of the cache is flaot16, then BlockSize%16=0 is required.
87
+ * - BlockNum is the number of memory blocks written to the cache, and its size is determined by the user's input
88
+ * scenario without any definite limit.
89
+ * @li The index-related constraints:
90
+ * When cache_mode is PA_NZ, the shape is [Bqkv*Sqkv], and the value range of the index must be
91
+ * [-1,BlockNum*BlockSize). The values cannot be repeated.
92
+ */
93
+ REG_OP(QkvRmsNormRopeCache)
94
+ .INPUT(qkv, TensorType({DT_FLOAT16, DT_BF16}))
95
+ .INPUT(q_gamma, TensorType({DT_FLOAT16, DT_BF16}))
96
+ .INPUT(k_gamma, TensorType({DT_FLOAT16, DT_BF16}))
97
+ .INPUT(cos, TensorType({DT_FLOAT16, DT_BF16}))
98
+ .INPUT(sin, TensorType({DT_FLOAT16, DT_BF16}))
99
+ .INPUT(index, TensorType({DT_INT64}))
100
+ .INPUT(q_out, TensorType({DT_FLOAT16, DT_BF16}))
101
+ .INPUT(k_cache, TensorType({DT_FLOAT16, DT_BF16, DT_INT8}))
102
+ .INPUT(v_cache, TensorType({DT_FLOAT16, DT_BF16, DT_INT8}))
103
+ .OPTIONAL_INPUT(k_scale, TensorType({DT_FLOAT}))
104
+ .OPTIONAL_INPUT(v_scale, TensorType({DT_FLOAT}))
105
+ .OPTIONAL_INPUT(k_offset, TensorType({DT_FLOAT}))
106
+ .OPTIONAL_INPUT(v_offset, TensorType({DT_FLOAT}))
107
+ .OUTPUT(q_out, TensorType({DT_FLOAT16, DT_BF16}))
108
+ .OUTPUT(k_cache, TensorType({DT_FLOAT16, DT_BF16, DT_INT8}))
109
+ .OUTPUT(v_cache, TensorType({DT_FLOAT16, DT_BF16, DT_INT8}))
110
+ .OUTPUT(q_out_before_quant, TensorType({DT_FLOAT16, DT_BF16}))
111
+ .OUTPUT(k_out_before_quant, TensorType({DT_FLOAT16, DT_BF16}))
112
+ .OUTPUT(v_out_before_quant, TensorType({DT_FLOAT16, DT_BF16}))
113
+ .REQUIRED_ATTR(qkv_size, ListInt)
114
+ .REQUIRED_ATTR(head_nums, ListInt)
115
+ .ATTR(epsilon, Float, 1e-6)
116
+ .ATTR(cache_mode, String, "PA_NZ")
117
+ .ATTR(is_output_qkv, Bool, false)
118
+ .OP_END_FACTORY_REG(QkvRmsNormRopeCache)
119
+ } // namespace ge
120
+
121
+ #endif // OPS_OP_PROTO_INC_QKV_RMS_NORM_ROPE_CACHE_H_
datasets/ops_src/ops_transformer/test_kernels/kernels/qkv_rms_norm_rope_cache/op_host/CMakeLists.txt ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # This program is free software, you can redistribute it and/or modify.
2
+ # Copyright (c) 2026 Huawei Technologies Co., Ltd.
3
+ # This file is a part of the CANN Open Software.
4
+ # Licensed under CANN Open Software License Agreement Version 2.0 (the "License").
5
+ # Please refer to the License for details. You may not use this file except in compliance with the License.
6
+ # THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
7
+ # See LICENSE in the root of the software repository for the full text of the License.
8
+ # ============================================================================
9
+
10
+ message(STATUS "=== Debug: start ops.posembedding.qkv_rms_norm_rope_cache.op_host.CMakeLists.txt ")
11
+
12
+ if (BUILD_OPEN_PROJECT)
13
+ message(STATUS "=== Debug: BUILD is BUILD_OPEN_PROJECT")
14
+ target_sources(op_host_aclnnExc PRIVATE
15
+ qkv_rms_norm_rope_cache_def.cpp
16
+ )
17
+ add_ops_compile_options(
18
+ OP_NAME QkvRmsNormRopeCache
19
+ OPTIONS --cce-auto-sync=on
20
+ -Wno-deprecated-declarations
21
+ -Werror
22
+ )
23
+ endif()
24
+
25
+ if(NOT BUILD_OPS_RTY_KERNEL)
26
+ message(STATUS "=== Debug: BUILD is NOT BUILD_OPS_RTY_KERNEL")
27
+ add_modules_sources(OPTYPE qkv_rms_norm_rope_cache ACLNNTYPE aclnn_exclude)
28
+ endif()
datasets/ops_src/ops_transformer/test_kernels/kernels/qkv_rms_norm_rope_cache/op_host/config/ascend910_93/qkv_rms_norm_rope_cache_binary.json ADDED
@@ -0,0 +1,1949 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "op_type": "QkvRmsNormRopeCache",
3
+ "op_list": [
4
+ {
5
+ "bin_filename": "QkvRmsNormRopeCache_fp16",
6
+ "inputs": [
7
+ {
8
+ "name": "qkv",
9
+ "index": 0,
10
+ "dtype": "float16",
11
+ "format": "ND",
12
+ "paramType": "required",
13
+ "shape": [
14
+ -2
15
+ ],
16
+ "format_match_mode": "FormatAgnostic"
17
+ },
18
+ {
19
+ "name": "q_gamma",
20
+ "index": 1,
21
+ "dtype": "float16",
22
+ "format": "ND",
23
+ "paramType": "required",
24
+ "shape": [
25
+ -2
26
+ ],
27
+ "format_match_mode": "FormatAgnostic"
28
+ },
29
+ {
30
+ "name": "k_gamma",
31
+ "index": 2,
32
+ "dtype": "float16",
33
+ "format": "ND",
34
+ "paramType": "required",
35
+ "shape": [
36
+ -2
37
+ ],
38
+ "format_match_mode": "FormatAgnostic"
39
+ },
40
+ {
41
+ "name": "cos",
42
+ "index": 3,
43
+ "dtype": "float16",
44
+ "format": "ND",
45
+ "paramType": "required",
46
+ "shape": [
47
+ -2
48
+ ],
49
+ "format_match_mode": "FormatAgnostic"
50
+ },
51
+ {
52
+ "name": "sin",
53
+ "index": 4,
54
+ "dtype": "float16",
55
+ "format": "ND",
56
+ "paramType": "required",
57
+ "shape": [
58
+ -2
59
+ ],
60
+ "format_match_mode": "FormatAgnostic"
61
+ },
62
+ {
63
+ "name": "index",
64
+ "index": 5,
65
+ "dtype": "int64",
66
+ "format": "ND",
67
+ "paramType": "required",
68
+ "shape": [
69
+ -2
70
+ ],
71
+ "format_match_mode": "FormatAgnostic"
72
+ },
73
+ {
74
+ "name": "q_out",
75
+ "index": 6,
76
+ "dtype": "float16",
77
+ "format": "ND",
78
+ "paramType": "required",
79
+ "shape": [
80
+ -2
81
+ ],
82
+ "format_match_mode": "FormatAgnostic"
83
+ },
84
+ {
85
+ "name": "k_cache",
86
+ "index": 7,
87
+ "dtype": "float16",
88
+ "format": "ND",
89
+ "paramType": "required",
90
+ "shape": [
91
+ -2
92
+ ],
93
+ "format_match_mode": "FormatAgnostic"
94
+ },
95
+ {
96
+ "name": "v_cache",
97
+ "index": 8,
98
+ "dtype": "float16",
99
+ "format": "ND",
100
+ "paramType": "required",
101
+ "shape": [
102
+ -2
103
+ ],
104
+ "format_match_mode": "FormatAgnostic"
105
+ },
106
+ {
107
+ "name": "k_scale",
108
+ "index": 9,
109
+ "dtype": "float32",
110
+ "format": "ND",
111
+ "paramType": "optional",
112
+ "shape": [
113
+ -2
114
+ ],
115
+ "format_match_mode": "FormatAgnostic"
116
+ },
117
+ {
118
+ "name": "v_scale",
119
+ "index": 10,
120
+ "dtype": "float32",
121
+ "format": "ND",
122
+ "paramType": "optional",
123
+ "shape": [
124
+ -2
125
+ ],
126
+ "format_match_mode": "FormatAgnostic"
127
+ },
128
+ {
129
+ "name": "k_offset",
130
+ "index": 11,
131
+ "dtype": "float32",
132
+ "format": "ND",
133
+ "paramType": "optional",
134
+ "shape": [
135
+ -2
136
+ ],
137
+ "format_match_mode": "FormatAgnostic"
138
+ },
139
+ {
140
+ "name": "v_offset",
141
+ "index": 12,
142
+ "dtype": "float32",
143
+ "format": "ND",
144
+ "paramType": "optional",
145
+ "shape": [
146
+ -2
147
+ ],
148
+ "format_match_mode": "FormatAgnostic"
149
+ }
150
+ ],
151
+ "outputs": [
152
+ {
153
+ "name": "q_out",
154
+ "index": 0,
155
+ "dtype": "float16",
156
+ "format": "ND",
157
+ "paramType": "required",
158
+ "shape": [
159
+ -2
160
+ ],
161
+ "format_match_mode": "FormatAgnostic"
162
+ },
163
+ {
164
+ "name": "k_cache",
165
+ "index": 1,
166
+ "dtype": "float16",
167
+ "format": "ND",
168
+ "paramType": "required",
169
+ "shape": [
170
+ -2
171
+ ],
172
+ "format_match_mode": "FormatAgnostic"
173
+ },
174
+ {
175
+ "name": "v_cache",
176
+ "index": 2,
177
+ "dtype": "float16",
178
+ "format": "ND",
179
+ "paramType": "required",
180
+ "shape": [
181
+ -2
182
+ ],
183
+ "format_match_mode": "FormatAgnostic"
184
+ },
185
+ {
186
+ "name": "q_out_before_quant",
187
+ "index": 3,
188
+ "dtype": "float16",
189
+ "format": "ND",
190
+ "paramType": "required",
191
+ "shape": [
192
+ -2
193
+ ],
194
+ "format_match_mode": "FormatAgnostic"
195
+ },
196
+ {
197
+ "name": "k_out_before_quant",
198
+ "index": 4,
199
+ "dtype": "float16",
200
+ "format": "ND",
201
+ "paramType": "required",
202
+ "shape": [
203
+ -2
204
+ ],
205
+ "format_match_mode": "FormatAgnostic"
206
+ },
207
+ {
208
+ "name": "v_out_before_quant",
209
+ "index": 5,
210
+ "dtype": "float16",
211
+ "format": "ND",
212
+ "paramType": "required",
213
+ "shape": [
214
+ -2
215
+ ],
216
+ "format_match_mode": "FormatAgnostic"
217
+ }
218
+ ],
219
+ "attrs": [
220
+ {
221
+ "name": "qkv_size",
222
+ "dtype": "list_int",
223
+ "value": []
224
+ },
225
+ {
226
+ "name": "head_nums",
227
+ "dtype": "list_int",
228
+ "value": []
229
+ },
230
+ {
231
+ "name": "epsilon",
232
+ "dtype": "float",
233
+ "value": 0.000001
234
+ },
235
+ {
236
+ "name": "cache_mode",
237
+ "dtype": "string",
238
+ "value": "PA_NZ"
239
+ },
240
+ {
241
+ "name": "is_output_qkv",
242
+ "dtype": "bool",
243
+ "value": false
244
+ }
245
+ ]
246
+ },
247
+ {
248
+ "bin_filename": "QkvRmsNormRopeCache_bf16",
249
+ "inputs": [
250
+ {
251
+ "name": "qkv",
252
+ "index": 0,
253
+ "dtype": "bfloat16",
254
+ "format": "ND",
255
+ "paramType": "required",
256
+ "shape": [
257
+ -2
258
+ ],
259
+ "format_match_mode": "FormatAgnostic"
260
+ },
261
+ {
262
+ "name": "q_gamma",
263
+ "index": 1,
264
+ "dtype": "bfloat16",
265
+ "format": "ND",
266
+ "paramType": "required",
267
+ "shape": [
268
+ -2
269
+ ],
270
+ "format_match_mode": "FormatAgnostic"
271
+ },
272
+ {
273
+ "name": "k_gamma",
274
+ "index": 2,
275
+ "dtype": "bfloat16",
276
+ "format": "ND",
277
+ "paramType": "required",
278
+ "shape": [
279
+ -2
280
+ ],
281
+ "format_match_mode": "FormatAgnostic"
282
+ },
283
+ {
284
+ "name": "cos",
285
+ "index": 3,
286
+ "dtype": "bfloat16",
287
+ "format": "ND",
288
+ "paramType": "required",
289
+ "shape": [
290
+ -2
291
+ ],
292
+ "format_match_mode": "FormatAgnostic"
293
+ },
294
+ {
295
+ "name": "sin",
296
+ "index": 4,
297
+ "dtype": "bfloat16",
298
+ "format": "ND",
299
+ "paramType": "required",
300
+ "shape": [
301
+ -2
302
+ ],
303
+ "format_match_mode": "FormatAgnostic"
304
+ },
305
+ {
306
+ "name": "index",
307
+ "index": 5,
308
+ "dtype": "int64",
309
+ "format": "ND",
310
+ "paramType": "required",
311
+ "shape": [
312
+ -2
313
+ ],
314
+ "format_match_mode": "FormatAgnostic"
315
+ },
316
+ {
317
+ "name": "q_out",
318
+ "index": 6,
319
+ "dtype": "bfloat16",
320
+ "format": "ND",
321
+ "paramType": "required",
322
+ "shape": [
323
+ -2
324
+ ],
325
+ "format_match_mode": "FormatAgnostic"
326
+ },
327
+ {
328
+ "name": "k_cache",
329
+ "index": 7,
330
+ "dtype": "bfloat16",
331
+ "format": "ND",
332
+ "paramType": "required",
333
+ "shape": [
334
+ -2
335
+ ],
336
+ "format_match_mode": "FormatAgnostic"
337
+ },
338
+ {
339
+ "name": "v_cache",
340
+ "index": 8,
341
+ "dtype": "bfloat16",
342
+ "format": "ND",
343
+ "paramType": "required",
344
+ "shape": [
345
+ -2
346
+ ],
347
+ "format_match_mode": "FormatAgnostic"
348
+ },
349
+ {
350
+ "name": "k_scale",
351
+ "index": 9,
352
+ "dtype": "float32",
353
+ "format": "ND",
354
+ "paramType": "optional",
355
+ "shape": [
356
+ -2
357
+ ],
358
+ "format_match_mode": "FormatAgnostic"
359
+ },
360
+ {
361
+ "name": "v_scale",
362
+ "index": 10,
363
+ "dtype": "float32",
364
+ "format": "ND",
365
+ "paramType": "optional",
366
+ "shape": [
367
+ -2
368
+ ],
369
+ "format_match_mode": "FormatAgnostic"
370
+ },
371
+ {
372
+ "name": "k_offset",
373
+ "index": 11,
374
+ "dtype": "float32",
375
+ "format": "ND",
376
+ "paramType": "optional",
377
+ "shape": [
378
+ -2
379
+ ],
380
+ "format_match_mode": "FormatAgnostic"
381
+ },
382
+ {
383
+ "name": "v_offset",
384
+ "index": 12,
385
+ "dtype": "float32",
386
+ "format": "ND",
387
+ "paramType": "optional",
388
+ "shape": [
389
+ -2
390
+ ],
391
+ "format_match_mode": "FormatAgnostic"
392
+ }
393
+ ],
394
+ "outputs": [
395
+ {
396
+ "name": "q_out",
397
+ "index": 0,
398
+ "dtype": "bfloat16",
399
+ "format": "ND",
400
+ "paramType": "required",
401
+ "shape": [
402
+ -2
403
+ ],
404
+ "format_match_mode": "FormatAgnostic"
405
+ },
406
+ {
407
+ "name": "k_cache",
408
+ "index": 1,
409
+ "dtype": "bfloat16",
410
+ "format": "ND",
411
+ "paramType": "required",
412
+ "shape": [
413
+ -2
414
+ ],
415
+ "format_match_mode": "FormatAgnostic"
416
+ },
417
+ {
418
+ "name": "v_cache",
419
+ "index": 2,
420
+ "dtype": "bfloat16",
421
+ "format": "ND",
422
+ "paramType": "required",
423
+ "shape": [
424
+ -2
425
+ ],
426
+ "format_match_mode": "FormatAgnostic"
427
+ },
428
+ {
429
+ "name": "q_out_before_quant",
430
+ "index": 3,
431
+ "dtype": "bfloat16",
432
+ "format": "ND",
433
+ "paramType": "required",
434
+ "shape": [
435
+ -2
436
+ ],
437
+ "format_match_mode": "FormatAgnostic"
438
+ },
439
+ {
440
+ "name": "k_out_before_quant",
441
+ "index": 4,
442
+ "dtype": "bfloat16",
443
+ "format": "ND",
444
+ "paramType": "required",
445
+ "shape": [
446
+ -2
447
+ ],
448
+ "format_match_mode": "FormatAgnostic"
449
+ },
450
+ {
451
+ "name": "v_out_before_quant",
452
+ "index": 5,
453
+ "dtype": "bfloat16",
454
+ "format": "ND",
455
+ "paramType": "required",
456
+ "shape": [
457
+ -2
458
+ ],
459
+ "format_match_mode": "FormatAgnostic"
460
+ }
461
+ ],
462
+ "attrs": [
463
+ {
464
+ "name": "qkv_size",
465
+ "dtype": "list_int",
466
+ "value": []
467
+ },
468
+ {
469
+ "name": "head_nums",
470
+ "dtype": "list_int",
471
+ "value": []
472
+ },
473
+ {
474
+ "name": "epsilon",
475
+ "dtype": "float",
476
+ "value": 0.000001
477
+ },
478
+ {
479
+ "name": "cache_mode",
480
+ "dtype": "string",
481
+ "value": "PA_NZ"
482
+ },
483
+ {
484
+ "name": "is_output_qkv",
485
+ "dtype": "bool",
486
+ "value": false
487
+ }
488
+ ]
489
+ },
490
+ {
491
+ "bin_filename": "QkvRmsNormRopeCache_fp16_Kquant_Vquant",
492
+ "inputs": [
493
+ {
494
+ "name": "qkv",
495
+ "index": 0,
496
+ "dtype": "float16",
497
+ "format": "ND",
498
+ "paramType": "required",
499
+ "shape": [
500
+ -2
501
+ ],
502
+ "format_match_mode": "FormatAgnostic"
503
+ },
504
+ {
505
+ "name": "q_gamma",
506
+ "index": 1,
507
+ "dtype": "float16",
508
+ "format": "ND",
509
+ "paramType": "required",
510
+ "shape": [
511
+ -2
512
+ ],
513
+ "format_match_mode": "FormatAgnostic"
514
+ },
515
+ {
516
+ "name": "k_gamma",
517
+ "index": 2,
518
+ "dtype": "float16",
519
+ "format": "ND",
520
+ "paramType": "required",
521
+ "shape": [
522
+ -2
523
+ ],
524
+ "format_match_mode": "FormatAgnostic"
525
+ },
526
+ {
527
+ "name": "cos",
528
+ "index": 3,
529
+ "dtype": "float16",
530
+ "format": "ND",
531
+ "paramType": "required",
532
+ "shape": [
533
+ -2
534
+ ],
535
+ "format_match_mode": "FormatAgnostic"
536
+ },
537
+ {
538
+ "name": "sin",
539
+ "index": 4,
540
+ "dtype": "float16",
541
+ "format": "ND",
542
+ "paramType": "required",
543
+ "shape": [
544
+ -2
545
+ ],
546
+ "format_match_mode": "FormatAgnostic"
547
+ },
548
+ {
549
+ "name": "index",
550
+ "index": 5,
551
+ "dtype": "int64",
552
+ "format": "ND",
553
+ "paramType": "required",
554
+ "shape": [
555
+ -2
556
+ ],
557
+ "format_match_mode": "FormatAgnostic"
558
+ },
559
+ {
560
+ "name": "q_out",
561
+ "index": 6,
562
+ "dtype": "float16",
563
+ "format": "ND",
564
+ "paramType": "required",
565
+ "shape": [
566
+ -2
567
+ ],
568
+ "format_match_mode": "FormatAgnostic"
569
+ },
570
+ {
571
+ "name": "k_cache",
572
+ "index": 7,
573
+ "dtype": "int8",
574
+ "format": "ND",
575
+ "paramType": "required",
576
+ "shape": [
577
+ -2
578
+ ],
579
+ "format_match_mode": "FormatAgnostic"
580
+ },
581
+ {
582
+ "name": "v_cache",
583
+ "index": 8,
584
+ "dtype": "int8",
585
+ "format": "ND",
586
+ "paramType": "required",
587
+ "shape": [
588
+ -2
589
+ ],
590
+ "format_match_mode": "FormatAgnostic"
591
+ },
592
+ {
593
+ "name": "k_scale",
594
+ "index": 9,
595
+ "dtype": "float32",
596
+ "format": "ND",
597
+ "paramType": "optional",
598
+ "shape": [
599
+ -2
600
+ ],
601
+ "format_match_mode": "FormatAgnostic"
602
+ },
603
+ {
604
+ "name": "v_scale",
605
+ "index": 10,
606
+ "dtype": "float32",
607
+ "format": "ND",
608
+ "paramType": "optional",
609
+ "shape": [
610
+ -2
611
+ ],
612
+ "format_match_mode": "FormatAgnostic"
613
+ },
614
+ {
615
+ "name": "k_offset",
616
+ "index": 11,
617
+ "dtype": "float32",
618
+ "format": "ND",
619
+ "paramType": "optional",
620
+ "shape": [
621
+ -2
622
+ ],
623
+ "format_match_mode": "FormatAgnostic"
624
+ },
625
+ {
626
+ "name": "v_offset",
627
+ "index": 12,
628
+ "dtype": "float32",
629
+ "format": "ND",
630
+ "paramType": "optional",
631
+ "shape": [
632
+ -2
633
+ ],
634
+ "format_match_mode": "FormatAgnostic"
635
+ }
636
+ ],
637
+ "outputs": [
638
+ {
639
+ "name": "q_out",
640
+ "index": 0,
641
+ "dtype": "float16",
642
+ "format": "ND",
643
+ "paramType": "required",
644
+ "shape": [
645
+ -2
646
+ ],
647
+ "format_match_mode": "FormatAgnostic"
648
+ },
649
+ {
650
+ "name": "k_cache",
651
+ "index": 1,
652
+ "dtype": "int8",
653
+ "format": "ND",
654
+ "paramType": "required",
655
+ "shape": [
656
+ -2
657
+ ],
658
+ "format_match_mode": "FormatAgnostic"
659
+ },
660
+ {
661
+ "name": "v_cache",
662
+ "index": 2,
663
+ "dtype": "int8",
664
+ "format": "ND",
665
+ "paramType": "required",
666
+ "shape": [
667
+ -2
668
+ ],
669
+ "format_match_mode": "FormatAgnostic"
670
+ },
671
+ {
672
+ "name": "q_out_before_quant",
673
+ "index": 3,
674
+ "dtype": "float16",
675
+ "format": "ND",
676
+ "paramType": "required",
677
+ "shape": [
678
+ -2
679
+ ],
680
+ "format_match_mode": "FormatAgnostic"
681
+ },
682
+ {
683
+ "name": "k_out_before_quant",
684
+ "index": 4,
685
+ "dtype": "float16",
686
+ "format": "ND",
687
+ "paramType": "required",
688
+ "shape": [
689
+ -2
690
+ ],
691
+ "format_match_mode": "FormatAgnostic"
692
+ },
693
+ {
694
+ "name": "v_out_before_quant",
695
+ "index": 5,
696
+ "dtype": "float16",
697
+ "format": "ND",
698
+ "paramType": "required",
699
+ "shape": [
700
+ -2
701
+ ],
702
+ "format_match_mode": "FormatAgnostic"
703
+ }
704
+ ],
705
+ "attrs": [
706
+ {
707
+ "name": "qkv_size",
708
+ "dtype": "list_int",
709
+ "value": []
710
+ },
711
+ {
712
+ "name": "head_nums",
713
+ "dtype": "list_int",
714
+ "value": []
715
+ },
716
+ {
717
+ "name": "epsilon",
718
+ "dtype": "float",
719
+ "value": 0.000001
720
+ },
721
+ {
722
+ "name": "cache_mode",
723
+ "dtype": "string",
724
+ "value": "PA_NZ"
725
+ },
726
+ {
727
+ "name": "is_output_qkv",
728
+ "dtype": "bool",
729
+ "value": false
730
+ }
731
+ ]
732
+ },
733
+ {
734
+ "bin_filename": "QkvRmsNormRopeCache_bf16_Kquant_Vquant",
735
+ "inputs": [
736
+ {
737
+ "name": "qkv",
738
+ "index": 0,
739
+ "dtype": "bfloat16",
740
+ "format": "ND",
741
+ "paramType": "required",
742
+ "shape": [
743
+ -2
744
+ ],
745
+ "format_match_mode": "FormatAgnostic"
746
+ },
747
+ {
748
+ "name": "q_gamma",
749
+ "index": 1,
750
+ "dtype": "bfloat16",
751
+ "format": "ND",
752
+ "paramType": "required",
753
+ "shape": [
754
+ -2
755
+ ],
756
+ "format_match_mode": "FormatAgnostic"
757
+ },
758
+ {
759
+ "name": "k_gamma",
760
+ "index": 2,
761
+ "dtype": "bfloat16",
762
+ "format": "ND",
763
+ "paramType": "required",
764
+ "shape": [
765
+ -2
766
+ ],
767
+ "format_match_mode": "FormatAgnostic"
768
+ },
769
+ {
770
+ "name": "cos",
771
+ "index": 3,
772
+ "dtype": "bfloat16",
773
+ "format": "ND",
774
+ "paramType": "required",
775
+ "shape": [
776
+ -2
777
+ ],
778
+ "format_match_mode": "FormatAgnostic"
779
+ },
780
+ {
781
+ "name": "sin",
782
+ "index": 4,
783
+ "dtype": "bfloat16",
784
+ "format": "ND",
785
+ "paramType": "required",
786
+ "shape": [
787
+ -2
788
+ ],
789
+ "format_match_mode": "FormatAgnostic"
790
+ },
791
+ {
792
+ "name": "index",
793
+ "index": 5,
794
+ "dtype": "int64",
795
+ "format": "ND",
796
+ "paramType": "required",
797
+ "shape": [
798
+ -2
799
+ ],
800
+ "format_match_mode": "FormatAgnostic"
801
+ },
802
+ {
803
+ "name": "q_out",
804
+ "index": 6,
805
+ "dtype": "bfloat16",
806
+ "format": "ND",
807
+ "paramType": "required",
808
+ "shape": [
809
+ -2
810
+ ],
811
+ "format_match_mode": "FormatAgnostic"
812
+ },
813
+ {
814
+ "name": "k_cache",
815
+ "index": 7,
816
+ "dtype": "int8",
817
+ "format": "ND",
818
+ "paramType": "required",
819
+ "shape": [
820
+ -2
821
+ ],
822
+ "format_match_mode": "FormatAgnostic"
823
+ },
824
+ {
825
+ "name": "v_cache",
826
+ "index": 8,
827
+ "dtype": "int8",
828
+ "format": "ND",
829
+ "paramType": "required",
830
+ "shape": [
831
+ -2
832
+ ],
833
+ "format_match_mode": "FormatAgnostic"
834
+ },
835
+ {
836
+ "name": "k_scale",
837
+ "index": 9,
838
+ "dtype": "float32",
839
+ "format": "ND",
840
+ "paramType": "optional",
841
+ "shape": [
842
+ -2
843
+ ],
844
+ "format_match_mode": "FormatAgnostic"
845
+ },
846
+ {
847
+ "name": "v_scale",
848
+ "index": 10,
849
+ "dtype": "float32",
850
+ "format": "ND",
851
+ "paramType": "optional",
852
+ "shape": [
853
+ -2
854
+ ],
855
+ "format_match_mode": "FormatAgnostic"
856
+ },
857
+ {
858
+ "name": "k_offset",
859
+ "index": 11,
860
+ "dtype": "float32",
861
+ "format": "ND",
862
+ "paramType": "optional",
863
+ "shape": [
864
+ -2
865
+ ],
866
+ "format_match_mode": "FormatAgnostic"
867
+ },
868
+ {
869
+ "name": "v_offset",
870
+ "index": 12,
871
+ "dtype": "float32",
872
+ "format": "ND",
873
+ "paramType": "optional",
874
+ "shape": [
875
+ -2
876
+ ],
877
+ "format_match_mode": "FormatAgnostic"
878
+ }
879
+ ],
880
+ "outputs": [
881
+ {
882
+ "name": "q_out",
883
+ "index": 0,
884
+ "dtype": "bfloat16",
885
+ "format": "ND",
886
+ "paramType": "required",
887
+ "shape": [
888
+ -2
889
+ ],
890
+ "format_match_mode": "FormatAgnostic"
891
+ },
892
+ {
893
+ "name": "k_cache",
894
+ "index": 1,
895
+ "dtype": "int8",
896
+ "format": "ND",
897
+ "paramType": "required",
898
+ "shape": [
899
+ -2
900
+ ],
901
+ "format_match_mode": "FormatAgnostic"
902
+ },
903
+ {
904
+ "name": "v_cache",
905
+ "index": 2,
906
+ "dtype": "int8",
907
+ "format": "ND",
908
+ "paramType": "required",
909
+ "shape": [
910
+ -2
911
+ ],
912
+ "format_match_mode": "FormatAgnostic"
913
+ },
914
+ {
915
+ "name": "q_out_before_quant",
916
+ "index": 3,
917
+ "dtype": "bfloat16",
918
+ "format": "ND",
919
+ "paramType": "required",
920
+ "shape": [
921
+ -2
922
+ ],
923
+ "format_match_mode": "FormatAgnostic"
924
+ },
925
+ {
926
+ "name": "k_out_before_quant",
927
+ "index": 4,
928
+ "dtype": "bfloat16",
929
+ "format": "ND",
930
+ "paramType": "required",
931
+ "shape": [
932
+ -2
933
+ ],
934
+ "format_match_mode": "FormatAgnostic"
935
+ },
936
+ {
937
+ "name": "v_out_before_quant",
938
+ "index": 5,
939
+ "dtype": "bfloat16",
940
+ "format": "ND",
941
+ "paramType": "required",
942
+ "shape": [
943
+ -2
944
+ ],
945
+ "format_match_mode": "FormatAgnostic"
946
+ }
947
+ ],
948
+ "attrs": [
949
+ {
950
+ "name": "qkv_size",
951
+ "dtype": "list_int",
952
+ "value": []
953
+ },
954
+ {
955
+ "name": "head_nums",
956
+ "dtype": "list_int",
957
+ "value": []
958
+ },
959
+ {
960
+ "name": "epsilon",
961
+ "dtype": "float",
962
+ "value": 0.000001
963
+ },
964
+ {
965
+ "name": "cache_mode",
966
+ "dtype": "string",
967
+ "value": "PA_NZ"
968
+ },
969
+ {
970
+ "name": "is_output_qkv",
971
+ "dtype": "bool",
972
+ "value": false
973
+ }
974
+ ]
975
+ },
976
+ {
977
+ "bin_filename": "QkvRmsNormRopeCache_fp16_Kquant",
978
+ "inputs": [
979
+ {
980
+ "name": "qkv",
981
+ "index": 0,
982
+ "dtype": "float16",
983
+ "format": "ND",
984
+ "paramType": "required",
985
+ "shape": [
986
+ -2
987
+ ],
988
+ "format_match_mode": "FormatAgnostic"
989
+ },
990
+ {
991
+ "name": "q_gamma",
992
+ "index": 1,
993
+ "dtype": "float16",
994
+ "format": "ND",
995
+ "paramType": "required",
996
+ "shape": [
997
+ -2
998
+ ],
999
+ "format_match_mode": "FormatAgnostic"
1000
+ },
1001
+ {
1002
+ "name": "k_gamma",
1003
+ "index": 2,
1004
+ "dtype": "float16",
1005
+ "format": "ND",
1006
+ "paramType": "required",
1007
+ "shape": [
1008
+ -2
1009
+ ],
1010
+ "format_match_mode": "FormatAgnostic"
1011
+ },
1012
+ {
1013
+ "name": "cos",
1014
+ "index": 3,
1015
+ "dtype": "float16",
1016
+ "format": "ND",
1017
+ "paramType": "required",
1018
+ "shape": [
1019
+ -2
1020
+ ],
1021
+ "format_match_mode": "FormatAgnostic"
1022
+ },
1023
+ {
1024
+ "name": "sin",
1025
+ "index": 4,
1026
+ "dtype": "float16",
1027
+ "format": "ND",
1028
+ "paramType": "required",
1029
+ "shape": [
1030
+ -2
1031
+ ],
1032
+ "format_match_mode": "FormatAgnostic"
1033
+ },
1034
+ {
1035
+ "name": "index",
1036
+ "index": 5,
1037
+ "dtype": "int64",
1038
+ "format": "ND",
1039
+ "paramType": "required",
1040
+ "shape": [
1041
+ -2
1042
+ ],
1043
+ "format_match_mode": "FormatAgnostic"
1044
+ },
1045
+ {
1046
+ "name": "q_out",
1047
+ "index": 6,
1048
+ "dtype": "float16",
1049
+ "format": "ND",
1050
+ "paramType": "required",
1051
+ "shape": [
1052
+ -2
1053
+ ],
1054
+ "format_match_mode": "FormatAgnostic"
1055
+ },
1056
+ {
1057
+ "name": "k_cache",
1058
+ "index": 7,
1059
+ "dtype": "int8",
1060
+ "format": "ND",
1061
+ "paramType": "required",
1062
+ "shape": [
1063
+ -2
1064
+ ],
1065
+ "format_match_mode": "FormatAgnostic"
1066
+ },
1067
+ {
1068
+ "name": "v_cache",
1069
+ "index": 8,
1070
+ "dtype": "float16",
1071
+ "format": "ND",
1072
+ "paramType": "required",
1073
+ "shape": [
1074
+ -2
1075
+ ],
1076
+ "format_match_mode": "FormatAgnostic"
1077
+ },
1078
+ {
1079
+ "name": "k_scale",
1080
+ "index": 9,
1081
+ "dtype": "float32",
1082
+ "format": "ND",
1083
+ "paramType": "optional",
1084
+ "shape": [
1085
+ -2
1086
+ ],
1087
+ "format_match_mode": "FormatAgnostic"
1088
+ },
1089
+ {
1090
+ "name": "v_scale",
1091
+ "index": 10,
1092
+ "dtype": "float32",
1093
+ "format": "ND",
1094
+ "paramType": "optional",
1095
+ "shape": [
1096
+ -2
1097
+ ],
1098
+ "format_match_mode": "FormatAgnostic"
1099
+ },
1100
+ {
1101
+ "name": "k_offset",
1102
+ "index": 11,
1103
+ "dtype": "float32",
1104
+ "format": "ND",
1105
+ "paramType": "optional",
1106
+ "shape": [
1107
+ -2
1108
+ ],
1109
+ "format_match_mode": "FormatAgnostic"
1110
+ },
1111
+ {
1112
+ "name": "v_offset",
1113
+ "index": 12,
1114
+ "dtype": "float32",
1115
+ "format": "ND",
1116
+ "paramType": "optional",
1117
+ "shape": [
1118
+ -2
1119
+ ],
1120
+ "format_match_mode": "FormatAgnostic"
1121
+ }
1122
+ ],
1123
+ "outputs": [
1124
+ {
1125
+ "name": "q_out",
1126
+ "index": 0,
1127
+ "dtype": "float16",
1128
+ "format": "ND",
1129
+ "paramType": "required",
1130
+ "shape": [
1131
+ -2
1132
+ ],
1133
+ "format_match_mode": "FormatAgnostic"
1134
+ },
1135
+ {
1136
+ "name": "k_cache",
1137
+ "index": 1,
1138
+ "dtype": "int8",
1139
+ "format": "ND",
1140
+ "paramType": "required",
1141
+ "shape": [
1142
+ -2
1143
+ ],
1144
+ "format_match_mode": "FormatAgnostic"
1145
+ },
1146
+ {
1147
+ "name": "v_cache",
1148
+ "index": 2,
1149
+ "dtype": "float16",
1150
+ "format": "ND",
1151
+ "paramType": "required",
1152
+ "shape": [
1153
+ -2
1154
+ ],
1155
+ "format_match_mode": "FormatAgnostic"
1156
+ },
1157
+ {
1158
+ "name": "q_out_before_quant",
1159
+ "index": 3,
1160
+ "dtype": "float16",
1161
+ "format": "ND",
1162
+ "paramType": "required",
1163
+ "shape": [
1164
+ -2
1165
+ ],
1166
+ "format_match_mode": "FormatAgnostic"
1167
+ },
1168
+ {
1169
+ "name": "k_out_before_quant",
1170
+ "index": 4,
1171
+ "dtype": "float16",
1172
+ "format": "ND",
1173
+ "paramType": "required",
1174
+ "shape": [
1175
+ -2
1176
+ ],
1177
+ "format_match_mode": "FormatAgnostic"
1178
+ },
1179
+ {
1180
+ "name": "v_out_before_quant",
1181
+ "index": 5,
1182
+ "dtype": "float16",
1183
+ "format": "ND",
1184
+ "paramType": "required",
1185
+ "shape": [
1186
+ -2
1187
+ ],
1188
+ "format_match_mode": "FormatAgnostic"
1189
+ }
1190
+ ],
1191
+ "attrs": [
1192
+ {
1193
+ "name": "qkv_size",
1194
+ "dtype": "list_int",
1195
+ "value": []
1196
+ },
1197
+ {
1198
+ "name": "head_nums",
1199
+ "dtype": "list_int",
1200
+ "value": []
1201
+ },
1202
+ {
1203
+ "name": "epsilon",
1204
+ "dtype": "float",
1205
+ "value": 0.000001
1206
+ },
1207
+ {
1208
+ "name": "cache_mode",
1209
+ "dtype": "string",
1210
+ "value": "PA_NZ"
1211
+ },
1212
+ {
1213
+ "name": "is_output_qkv",
1214
+ "dtype": "bool",
1215
+ "value": false
1216
+ }
1217
+ ]
1218
+ },
1219
+ {
1220
+ "bin_filename": "QkvRmsNormRopeCache_bf16_Kquant",
1221
+ "inputs": [
1222
+ {
1223
+ "name": "qkv",
1224
+ "index": 0,
1225
+ "dtype": "bfloat16",
1226
+ "format": "ND",
1227
+ "paramType": "required",
1228
+ "shape": [
1229
+ -2
1230
+ ],
1231
+ "format_match_mode": "FormatAgnostic"
1232
+ },
1233
+ {
1234
+ "name": "q_gamma",
1235
+ "index": 1,
1236
+ "dtype": "bfloat16",
1237
+ "format": "ND",
1238
+ "paramType": "required",
1239
+ "shape": [
1240
+ -2
1241
+ ],
1242
+ "format_match_mode": "FormatAgnostic"
1243
+ },
1244
+ {
1245
+ "name": "k_gamma",
1246
+ "index": 2,
1247
+ "dtype": "bfloat16",
1248
+ "format": "ND",
1249
+ "paramType": "required",
1250
+ "shape": [
1251
+ -2
1252
+ ],
1253
+ "format_match_mode": "FormatAgnostic"
1254
+ },
1255
+ {
1256
+ "name": "cos",
1257
+ "index": 3,
1258
+ "dtype": "bfloat16",
1259
+ "format": "ND",
1260
+ "paramType": "required",
1261
+ "shape": [
1262
+ -2
1263
+ ],
1264
+ "format_match_mode": "FormatAgnostic"
1265
+ },
1266
+ {
1267
+ "name": "sin",
1268
+ "index": 4,
1269
+ "dtype": "bfloat16",
1270
+ "format": "ND",
1271
+ "paramType": "required",
1272
+ "shape": [
1273
+ -2
1274
+ ],
1275
+ "format_match_mode": "FormatAgnostic"
1276
+ },
1277
+ {
1278
+ "name": "index",
1279
+ "index": 5,
1280
+ "dtype": "int64",
1281
+ "format": "ND",
1282
+ "paramType": "required",
1283
+ "shape": [
1284
+ -2
1285
+ ],
1286
+ "format_match_mode": "FormatAgnostic"
1287
+ },
1288
+ {
1289
+ "name": "q_out",
1290
+ "index": 6,
1291
+ "dtype": "bfloat16",
1292
+ "format": "ND",
1293
+ "paramType": "required",
1294
+ "shape": [
1295
+ -2
1296
+ ],
1297
+ "format_match_mode": "FormatAgnostic"
1298
+ },
1299
+ {
1300
+ "name": "k_cache",
1301
+ "index": 7,
1302
+ "dtype": "int8",
1303
+ "format": "ND",
1304
+ "paramType": "required",
1305
+ "shape": [
1306
+ -2
1307
+ ],
1308
+ "format_match_mode": "FormatAgnostic"
1309
+ },
1310
+ {
1311
+ "name": "v_cache",
1312
+ "index": 8,
1313
+ "dtype": "bfloat16",
1314
+ "format": "ND",
1315
+ "paramType": "required",
1316
+ "shape": [
1317
+ -2
1318
+ ],
1319
+ "format_match_mode": "FormatAgnostic"
1320
+ },
1321
+ {
1322
+ "name": "k_scale",
1323
+ "index": 9,
1324
+ "dtype": "float32",
1325
+ "format": "ND",
1326
+ "paramType": "optional",
1327
+ "shape": [
1328
+ -2
1329
+ ],
1330
+ "format_match_mode": "FormatAgnostic"
1331
+ },
1332
+ {
1333
+ "name": "v_scale",
1334
+ "index": 10,
1335
+ "dtype": "float32",
1336
+ "format": "ND",
1337
+ "paramType": "optional",
1338
+ "shape": [
1339
+ -2
1340
+ ],
1341
+ "format_match_mode": "FormatAgnostic"
1342
+ },
1343
+ {
1344
+ "name": "k_offset",
1345
+ "index": 11,
1346
+ "dtype": "float32",
1347
+ "format": "ND",
1348
+ "paramType": "optional",
1349
+ "shape": [
1350
+ -2
1351
+ ],
1352
+ "format_match_mode": "FormatAgnostic"
1353
+ },
1354
+ {
1355
+ "name": "v_offset",
1356
+ "index": 12,
1357
+ "dtype": "float32",
1358
+ "format": "ND",
1359
+ "paramType": "optional",
1360
+ "shape": [
1361
+ -2
1362
+ ],
1363
+ "format_match_mode": "FormatAgnostic"
1364
+ }
1365
+ ],
1366
+ "outputs": [
1367
+ {
1368
+ "name": "q_out",
1369
+ "index": 0,
1370
+ "dtype": "bfloat16",
1371
+ "format": "ND",
1372
+ "paramType": "required",
1373
+ "shape": [
1374
+ -2
1375
+ ],
1376
+ "format_match_mode": "FormatAgnostic"
1377
+ },
1378
+ {
1379
+ "name": "k_cache",
1380
+ "index": 1,
1381
+ "dtype": "int8",
1382
+ "format": "ND",
1383
+ "paramType": "required",
1384
+ "shape": [
1385
+ -2
1386
+ ],
1387
+ "format_match_mode": "FormatAgnostic"
1388
+ },
1389
+ {
1390
+ "name": "v_cache",
1391
+ "index": 2,
1392
+ "dtype": "bfloat16",
1393
+ "format": "ND",
1394
+ "paramType": "required",
1395
+ "shape": [
1396
+ -2
1397
+ ],
1398
+ "format_match_mode": "FormatAgnostic"
1399
+ },
1400
+ {
1401
+ "name": "q_out_before_quant",
1402
+ "index": 3,
1403
+ "dtype": "bfloat16",
1404
+ "format": "ND",
1405
+ "paramType": "required",
1406
+ "shape": [
1407
+ -2
1408
+ ],
1409
+ "format_match_mode": "FormatAgnostic"
1410
+ },
1411
+ {
1412
+ "name": "k_out_before_quant",
1413
+ "index": 4,
1414
+ "dtype": "bfloat16",
1415
+ "format": "ND",
1416
+ "paramType": "required",
1417
+ "shape": [
1418
+ -2
1419
+ ],
1420
+ "format_match_mode": "FormatAgnostic"
1421
+ },
1422
+ {
1423
+ "name": "v_out_before_quant",
1424
+ "index": 5,
1425
+ "dtype": "bfloat16",
1426
+ "format": "ND",
1427
+ "paramType": "required",
1428
+ "shape": [
1429
+ -2
1430
+ ],
1431
+ "format_match_mode": "FormatAgnostic"
1432
+ }
1433
+ ],
1434
+ "attrs": [
1435
+ {
1436
+ "name": "qkv_size",
1437
+ "dtype": "list_int",
1438
+ "value": []
1439
+ },
1440
+ {
1441
+ "name": "head_nums",
1442
+ "dtype": "list_int",
1443
+ "value": []
1444
+ },
1445
+ {
1446
+ "name": "epsilon",
1447
+ "dtype": "float",
1448
+ "value": 0.000001
1449
+ },
1450
+ {
1451
+ "name": "cache_mode",
1452
+ "dtype": "string",
1453
+ "value": "PA_NZ"
1454
+ },
1455
+ {
1456
+ "name": "is_output_qkv",
1457
+ "dtype": "bool",
1458
+ "value": false
1459
+ }
1460
+ ]
1461
+ },
1462
+ {
1463
+ "bin_filename": "QkvRmsNormRopeCache_fp16_Vquant",
1464
+ "inputs": [
1465
+ {
1466
+ "name": "qkv",
1467
+ "index": 0,
1468
+ "dtype": "float16",
1469
+ "format": "ND",
1470
+ "paramType": "required",
1471
+ "shape": [
1472
+ -2
1473
+ ],
1474
+ "format_match_mode": "FormatAgnostic"
1475
+ },
1476
+ {
1477
+ "name": "q_gamma",
1478
+ "index": 1,
1479
+ "dtype": "float16",
1480
+ "format": "ND",
1481
+ "paramType": "required",
1482
+ "shape": [
1483
+ -2
1484
+ ],
1485
+ "format_match_mode": "FormatAgnostic"
1486
+ },
1487
+ {
1488
+ "name": "k_gamma",
1489
+ "index": 2,
1490
+ "dtype": "float16",
1491
+ "format": "ND",
1492
+ "paramType": "required",
1493
+ "shape": [
1494
+ -2
1495
+ ],
1496
+ "format_match_mode": "FormatAgnostic"
1497
+ },
1498
+ {
1499
+ "name": "cos",
1500
+ "index": 3,
1501
+ "dtype": "float16",
1502
+ "format": "ND",
1503
+ "paramType": "required",
1504
+ "shape": [
1505
+ -2
1506
+ ],
1507
+ "format_match_mode": "FormatAgnostic"
1508
+ },
1509
+ {
1510
+ "name": "sin",
1511
+ "index": 4,
1512
+ "dtype": "float16",
1513
+ "format": "ND",
1514
+ "paramType": "required",
1515
+ "shape": [
1516
+ -2
1517
+ ],
1518
+ "format_match_mode": "FormatAgnostic"
1519
+ },
1520
+ {
1521
+ "name": "index",
1522
+ "index": 5,
1523
+ "dtype": "int64",
1524
+ "format": "ND",
1525
+ "paramType": "required",
1526
+ "shape": [
1527
+ -2
1528
+ ],
1529
+ "format_match_mode": "FormatAgnostic"
1530
+ },
1531
+ {
1532
+ "name": "q_out",
1533
+ "index": 6,
1534
+ "dtype": "float16",
1535
+ "format": "ND",
1536
+ "paramType": "required",
1537
+ "shape": [
1538
+ -2
1539
+ ],
1540
+ "format_match_mode": "FormatAgnostic"
1541
+ },
1542
+ {
1543
+ "name": "k_cache",
1544
+ "index": 7,
1545
+ "dtype": "float16",
1546
+ "format": "ND",
1547
+ "paramType": "required",
1548
+ "shape": [
1549
+ -2
1550
+ ],
1551
+ "format_match_mode": "FormatAgnostic"
1552
+ },
1553
+ {
1554
+ "name": "v_cache",
1555
+ "index": 8,
1556
+ "dtype": "int8",
1557
+ "format": "ND",
1558
+ "paramType": "required",
1559
+ "shape": [
1560
+ -2
1561
+ ],
1562
+ "format_match_mode": "FormatAgnostic"
1563
+ },
1564
+ {
1565
+ "name": "k_scale",
1566
+ "index": 9,
1567
+ "dtype": "float32",
1568
+ "format": "ND",
1569
+ "paramType": "optional",
1570
+ "shape": [
1571
+ -2
1572
+ ],
1573
+ "format_match_mode": "FormatAgnostic"
1574
+ },
1575
+ {
1576
+ "name": "v_scale",
1577
+ "index": 10,
1578
+ "dtype": "float32",
1579
+ "format": "ND",
1580
+ "paramType": "optional",
1581
+ "shape": [
1582
+ -2
1583
+ ],
1584
+ "format_match_mode": "FormatAgnostic"
1585
+ },
1586
+ {
1587
+ "name": "k_offset",
1588
+ "index": 11,
1589
+ "dtype": "float32",
1590
+ "format": "ND",
1591
+ "paramType": "optional",
1592
+ "shape": [
1593
+ -2
1594
+ ],
1595
+ "format_match_mode": "FormatAgnostic"
1596
+ },
1597
+ {
1598
+ "name": "v_offset",
1599
+ "index": 12,
1600
+ "dtype": "float32",
1601
+ "format": "ND",
1602
+ "paramType": "optional",
1603
+ "shape": [
1604
+ -2
1605
+ ],
1606
+ "format_match_mode": "FormatAgnostic"
1607
+ }
1608
+ ],
1609
+ "outputs": [
1610
+ {
1611
+ "name": "q_out",
1612
+ "index": 0,
1613
+ "dtype": "float16",
1614
+ "format": "ND",
1615
+ "paramType": "required",
1616
+ "shape": [
1617
+ -2
1618
+ ],
1619
+ "format_match_mode": "FormatAgnostic"
1620
+ },
1621
+ {
1622
+ "name": "k_cache",
1623
+ "index": 1,
1624
+ "dtype": "float16",
1625
+ "format": "ND",
1626
+ "paramType": "required",
1627
+ "shape": [
1628
+ -2
1629
+ ],
1630
+ "format_match_mode": "FormatAgnostic"
1631
+ },
1632
+ {
1633
+ "name": "v_cache",
1634
+ "index": 2,
1635
+ "dtype": "int8",
1636
+ "format": "ND",
1637
+ "paramType": "required",
1638
+ "shape": [
1639
+ -2
1640
+ ],
1641
+ "format_match_mode": "FormatAgnostic"
1642
+ },
1643
+ {
1644
+ "name": "q_out_before_quant",
1645
+ "index": 3,
1646
+ "dtype": "float16",
1647
+ "format": "ND",
1648
+ "paramType": "required",
1649
+ "shape": [
1650
+ -2
1651
+ ],
1652
+ "format_match_mode": "FormatAgnostic"
1653
+ },
1654
+ {
1655
+ "name": "k_out_before_quant",
1656
+ "index": 4,
1657
+ "dtype": "float16",
1658
+ "format": "ND",
1659
+ "paramType": "required",
1660
+ "shape": [
1661
+ -2
1662
+ ],
1663
+ "format_match_mode": "FormatAgnostic"
1664
+ },
1665
+ {
1666
+ "name": "v_out_before_quant",
1667
+ "index": 5,
1668
+ "dtype": "float16",
1669
+ "format": "ND",
1670
+ "paramType": "required",
1671
+ "shape": [
1672
+ -2
1673
+ ],
1674
+ "format_match_mode": "FormatAgnostic"
1675
+ }
1676
+ ],
1677
+ "attrs": [
1678
+ {
1679
+ "name": "qkv_size",
1680
+ "dtype": "list_int",
1681
+ "value": []
1682
+ },
1683
+ {
1684
+ "name": "head_nums",
1685
+ "dtype": "list_int",
1686
+ "value": []
1687
+ },
1688
+ {
1689
+ "name": "epsilon",
1690
+ "dtype": "float",
1691
+ "value": 0.000001
1692
+ },
1693
+ {
1694
+ "name": "cache_mode",
1695
+ "dtype": "string",
1696
+ "value": "PA_NZ"
1697
+ },
1698
+ {
1699
+ "name": "is_output_qkv",
1700
+ "dtype": "bool",
1701
+ "value": false
1702
+ }
1703
+ ]
1704
+ },
1705
+ {
1706
+ "bin_filename": "QkvRmsNormRopeCache_bf16_Vquant",
1707
+ "inputs": [
1708
+ {
1709
+ "name": "qkv",
1710
+ "index": 0,
1711
+ "dtype": "bfloat16",
1712
+ "format": "ND",
1713
+ "paramType": "required",
1714
+ "shape": [
1715
+ -2
1716
+ ],
1717
+ "format_match_mode": "FormatAgnostic"
1718
+ },
1719
+ {
1720
+ "name": "q_gamma",
1721
+ "index": 1,
1722
+ "dtype": "bfloat16",
1723
+ "format": "ND",
1724
+ "paramType": "required",
1725
+ "shape": [
1726
+ -2
1727
+ ],
1728
+ "format_match_mode": "FormatAgnostic"
1729
+ },
1730
+ {
1731
+ "name": "k_gamma",
1732
+ "index": 2,
1733
+ "dtype": "bfloat16",
1734
+ "format": "ND",
1735
+ "paramType": "required",
1736
+ "shape": [
1737
+ -2
1738
+ ],
1739
+ "format_match_mode": "FormatAgnostic"
1740
+ },
1741
+ {
1742
+ "name": "cos",
1743
+ "index": 3,
1744
+ "dtype": "bfloat16",
1745
+ "format": "ND",
1746
+ "paramType": "required",
1747
+ "shape": [
1748
+ -2
1749
+ ],
1750
+ "format_match_mode": "FormatAgnostic"
1751
+ },
1752
+ {
1753
+ "name": "sin",
1754
+ "index": 4,
1755
+ "dtype": "bfloat16",
1756
+ "format": "ND",
1757
+ "paramType": "required",
1758
+ "shape": [
1759
+ -2
1760
+ ],
1761
+ "format_match_mode": "FormatAgnostic"
1762
+ },
1763
+ {
1764
+ "name": "index",
1765
+ "index": 5,
1766
+ "dtype": "int64",
1767
+ "format": "ND",
1768
+ "paramType": "required",
1769
+ "shape": [
1770
+ -2
1771
+ ],
1772
+ "format_match_mode": "FormatAgnostic"
1773
+ },
1774
+ {
1775
+ "name": "q_out",
1776
+ "index": 6,
1777
+ "dtype": "bfloat16",
1778
+ "format": "ND",
1779
+ "paramType": "required",
1780
+ "shape": [
1781
+ -2
1782
+ ],
1783
+ "format_match_mode": "FormatAgnostic"
1784
+ },
1785
+ {
1786
+ "name": "k_cache",
1787
+ "index": 7,
1788
+ "dtype": "bfloat16",
1789
+ "format": "ND",
1790
+ "paramType": "required",
1791
+ "shape": [
1792
+ -2
1793
+ ],
1794
+ "format_match_mode": "FormatAgnostic"
1795
+ },
1796
+ {
1797
+ "name": "v_cache",
1798
+ "index": 8,
1799
+ "dtype": "int8",
1800
+ "format": "ND",
1801
+ "paramType": "required",
1802
+ "shape": [
1803
+ -2
1804
+ ],
1805
+ "format_match_mode": "FormatAgnostic"
1806
+ },
1807
+ {
1808
+ "name": "k_scale",
1809
+ "index": 9,
1810
+ "dtype": "float32",
1811
+ "format": "ND",
1812
+ "paramType": "optional",
1813
+ "shape": [
1814
+ -2
1815
+ ],
1816
+ "format_match_mode": "FormatAgnostic"
1817
+ },
1818
+ {
1819
+ "name": "v_scale",
1820
+ "index": 10,
1821
+ "dtype": "float32",
1822
+ "format": "ND",
1823
+ "paramType": "optional",
1824
+ "shape": [
1825
+ -2
1826
+ ],
1827
+ "format_match_mode": "FormatAgnostic"
1828
+ },
1829
+ {
1830
+ "name": "k_offset",
1831
+ "index": 11,
1832
+ "dtype": "float32",
1833
+ "format": "ND",
1834
+ "paramType": "optional",
1835
+ "shape": [
1836
+ -2
1837
+ ],
1838
+ "format_match_mode": "FormatAgnostic"
1839
+ },
1840
+ {
1841
+ "name": "v_offset",
1842
+ "index": 12,
1843
+ "dtype": "float32",
1844
+ "format": "ND",
1845
+ "paramType": "optional",
1846
+ "shape": [
1847
+ -2
1848
+ ],
1849
+ "format_match_mode": "FormatAgnostic"
1850
+ }
1851
+ ],
1852
+ "outputs": [
1853
+ {
1854
+ "name": "q_out",
1855
+ "index": 0,
1856
+ "dtype": "bfloat16",
1857
+ "format": "ND",
1858
+ "paramType": "required",
1859
+ "shape": [
1860
+ -2
1861
+ ],
1862
+ "format_match_mode": "FormatAgnostic"
1863
+ },
1864
+ {
1865
+ "name": "k_cache",
1866
+ "index": 1,
1867
+ "dtype": "bfloat16",
1868
+ "format": "ND",
1869
+ "paramType": "required",
1870
+ "shape": [
1871
+ -2
1872
+ ],
1873
+ "format_match_mode": "FormatAgnostic"
1874
+ },
1875
+ {
1876
+ "name": "v_cache",
1877
+ "index": 2,
1878
+ "dtype": "int8",
1879
+ "format": "ND",
1880
+ "paramType": "required",
1881
+ "shape": [
1882
+ -2
1883
+ ],
1884
+ "format_match_mode": "FormatAgnostic"
1885
+ },
1886
+ {
1887
+ "name": "q_out_before_quant",
1888
+ "index": 3,
1889
+ "dtype": "bfloat16",
1890
+ "format": "ND",
1891
+ "paramType": "required",
1892
+ "shape": [
1893
+ -2
1894
+ ],
1895
+ "format_match_mode": "FormatAgnostic"
1896
+ },
1897
+ {
1898
+ "name": "k_out_before_quant",
1899
+ "index": 4,
1900
+ "dtype": "bfloat16",
1901
+ "format": "ND",
1902
+ "paramType": "required",
1903
+ "shape": [
1904
+ -2
1905
+ ],
1906
+ "format_match_mode": "FormatAgnostic"
1907
+ },
1908
+ {
1909
+ "name": "v_out_before_quant",
1910
+ "index": 5,
1911
+ "dtype": "bfloat16",
1912
+ "format": "ND",
1913
+ "paramType": "required",
1914
+ "shape": [
1915
+ -2
1916
+ ],
1917
+ "format_match_mode": "FormatAgnostic"
1918
+ }
1919
+ ],
1920
+ "attrs": [
1921
+ {
1922
+ "name": "qkv_size",
1923
+ "dtype": "list_int",
1924
+ "value": []
1925
+ },
1926
+ {
1927
+ "name": "head_nums",
1928
+ "dtype": "list_int",
1929
+ "value": []
1930
+ },
1931
+ {
1932
+ "name": "epsilon",
1933
+ "dtype": "float",
1934
+ "value": 0.000001
1935
+ },
1936
+ {
1937
+ "name": "cache_mode",
1938
+ "dtype": "string",
1939
+ "value": "PA_NZ"
1940
+ },
1941
+ {
1942
+ "name": "is_output_qkv",
1943
+ "dtype": "bool",
1944
+ "value": false
1945
+ }
1946
+ ]
1947
+ }
1948
+ ]
1949
+ }