删除dispatch/combine_setup/teardown aclnn接口，graph文件-ops-transformer-AtomGit

Mclassify_rule.yaml+0-14

Mdocs/zh/ascend950_op_list.md+8-8

	@@ -961,8 +961,8 @@ Ascend 950支持的算子分类和算子列表如下：
961	<td><a href="../../mc2/moe_distribute_combine_setup/README.md">moe_distribute_combine_setup</a></td>	961	<td><a href="../../mc2/moe_distribute_combine_setup/README.md">moe_distribute_combine_setup</a></td>
962	<td>✓</td>	962	<td>✓</td>
963	<td>✓</td>	963	<td>✓</td>
964	- <td>✓</td>	964	+ <td>✗</td>
965	- <td>✓</td>	965	+ <td>✗</td>
966	<td>AI Core</td>	966	<td>AI Core</td>
967	<td>MoeDistributeCombine的setup阶段，用于初始化通信资源。</td>	967	<td>MoeDistributeCombine的setup阶段，用于初始化通信资源。</td>
968	</tr>	968	</tr>
	@@ -971,8 +971,8 @@ Ascend 950支持的算子分类和算子列表如下：
971	<td><a href="../../mc2/moe_distribute_combine_teardown/README.md">moe_distribute_combine_teardown</a></td>	971	<td><a href="../../mc2/moe_distribute_combine_teardown/README.md">moe_distribute_combine_teardown</a></td>
972	<td>✓</td>	972	<td>✓</td>
973	<td>✓</td>	973	<td>✓</td>
974	- <td>✓</td>	974	+ <td>✗</td>
975	- <td>✓</td>	975	+ <td>✗</td>
976	<td>AI Core</td>	976	<td>AI Core</td>
977	<td>MoeDistributeCombine的teardown阶段，用于释放通信资源。</td>	977	<td>MoeDistributeCombine的teardown阶段，用于释放通信资源。</td>
978	</tr>	978	</tr>
	@@ -1011,8 +1011,8 @@ Ascend 950支持的算子分类和算子列表如下：
1011	<td><a href="../../mc2/moe_distribute_dispatch_setup/README.md">moe_distribute_dispatch_setup</a></td>	1011	<td><a href="../../mc2/moe_distribute_dispatch_setup/README.md">moe_distribute_dispatch_setup</a></td>
1012	<td>✓</td>	1012	<td>✓</td>
1013	<td>✓</td>	1013	<td>✓</td>
1014	- <td>✓</td>	1014	+ <td>✗</td>
1015	- <td>✓</td>	1015	+ <td>✗</td>
1016	<td>AI Core</td>	1016	<td>AI Core</td>
1017	<td>MoeDistributeDispatch的setup阶段，用于初始化通信资源。</td>	1017	<td>MoeDistributeDispatch的setup阶段，用于初始化通信资源。</td>
1018	</tr>	1018	</tr>
	@@ -1021,8 +1021,8 @@ Ascend 950支持的算子分类和算子列表如下：
1021	<td><a href="../../mc2/moe_distribute_dispatch_teardown/README.md">moe_distribute_dispatch_teardown</a></td>	1021	<td><a href="../../mc2/moe_distribute_dispatch_teardown/README.md">moe_distribute_dispatch_teardown</a></td>
1022	<td>✓</td>	1022	<td>✓</td>
1023	<td>✓</td>	1023	<td>✓</td>
1024	- <td>✓</td>	1024	+ <td>✗</td>
1025	- <td>✓</td>	1025	+ <td>✗</td>
1026	<td>AI Core</td>	1026	<td>AI Core</td>
1027	<td>MoeDistributeDispatch的teardown阶段，用于释放通信资源。</td>	1027	<td>MoeDistributeDispatch的teardown阶段，用于释放通信资源。</td>
1028	</tr>	1028	</tr>

Mdocs/zh/menu_aclnn_api.md+0-4

Mdocs/zh/op_api_list.md+0-4

Mmc2/common/op_graph/mc2_moe_gen_task_ops_utils.cpp+0-15

	@@ -36,10 +36,6 @@ const std::string DISTRIBUTE_BARRIER_OP_TYPE = "DistributeBarrier";
36	const std::string MOE_DISTRIBUTE_DISPATCH_V2_OP_TYPE = "MoeDistributeDispatchV2";	36	const std::string MOE_DISTRIBUTE_DISPATCH_V2_OP_TYPE = "MoeDistributeDispatchV2";
37	const std::string MOE_DISTRIBUTE_COMBINE_V2_OP_TYPE = "MoeDistributeCombineV2";	37	const std::string MOE_DISTRIBUTE_COMBINE_V2_OP_TYPE = "MoeDistributeCombineV2";
38	const std::string MOE_DISTRIBUTE_COMBINE_ADD_RMS_NORM_OP_TYPE = "MoeDistributeCombineAddRmsNorm";	38	const std::string MOE_DISTRIBUTE_COMBINE_ADD_RMS_NORM_OP_TYPE = "MoeDistributeCombineAddRmsNorm";
39	-const std::string MOE_DISTRIBUTE_COMBINE_SETUP_OP_TYPE = "MoeDistributeCombineSetup";
40	-const std::string MOE_DISTRIBUTE_COMBINE_TEARDOWN_OP_TYPE = "MoeDistributeCombineTeardown";
41	-const std::string MOE_DISTRIBUTE_DISPATCH_SETUP_OP_TYPE = "MoeDistributeDispatchSetup";
42	-const std::string MOE_DISTRIBUTE_DISPATCH_TEARDOWN_OP_TYPE = "MoeDistributeDispatchTeardown";
43	const std::string ALLTO_ALLV_GROUPED_MAT_MUL_OP_TYPE = "AlltoAllvGroupedMatMul";	39	const std::string ALLTO_ALLV_GROUPED_MAT_MUL_OP_TYPE = "AlltoAllvGroupedMatMul";
44	const std::string GROUPED_MAT_MUL_ALLTO_ALLV_OP_TYPE = "GroupedMatMulAlltoAllv";	40	const std::string GROUPED_MAT_MUL_ALLTO_ALLV_OP_TYPE = "GroupedMatMulAlltoAllv";
45	const std::string ALL_TO_ALLV_QUANT_GROUPED_MM_OP_TYPE = "AlltoAllvQuantGroupedMatMul";	41	const std::string ALL_TO_ALLV_QUANT_GROUPED_MM_OP_TYPE = "AlltoAllvQuantGroupedMatMul";
	@@ -48,7 +44,6 @@ const std::string ATTENTION_TO_FFN_OP_TYPE = "AttentionToFFN";
48	const std::string FFN_TO_ATTENTION_OP_TYPE = "FFNToAttention";	44	const std::string FFN_TO_ATTENTION_OP_TYPE = "FFNToAttention";
49	const int32_t GROUP_CNT_OF_DS_TRAINING = 1;	45	const int32_t GROUP_CNT_OF_DS_TRAINING = 1;
50	const int32_t GROUP_CNT_OF_MOE_DISTRIBUTE = 2;	46	const int32_t GROUP_CNT_OF_MOE_DISTRIBUTE = 2;
51	-const int32_t ONE_GROUP_CNT_OF_MOE_DISTRIBUTE = 1;
52	const int32_t GROUP_CNT_OF_DISTRIBUTE_BARRIER = 1;	47	const int32_t GROUP_CNT_OF_DISTRIBUTE_BARRIER = 1;
53	const int32_t GROUP_CNT_OF_ATTENTION_FFN = 1;	48	const int32_t GROUP_CNT_OF_ATTENTION_FFN = 1;
54	const int32_t GROUP_CNT = 2;	49	const int32_t GROUP_CNT = 2;
	@@ -70,10 +65,6 @@ static const std::map<const std::string, int32_t> GROUP_CNT_MAP{
70	{MOE_DISTRIBUTE_DISPATCH_V2_OP_TYPE, GROUP_CNT_OF_MOE_DISTRIBUTE},	65	{MOE_DISTRIBUTE_DISPATCH_V2_OP_TYPE, GROUP_CNT_OF_MOE_DISTRIBUTE},
71	{MOE_DISTRIBUTE_COMBINE_V2_OP_TYPE, GROUP_CNT_OF_MOE_DISTRIBUTE},	66	{MOE_DISTRIBUTE_COMBINE_V2_OP_TYPE, GROUP_CNT_OF_MOE_DISTRIBUTE},
72	{MOE_DISTRIBUTE_COMBINE_ADD_RMS_NORM_OP_TYPE, GROUP_CNT_OF_MOE_DISTRIBUTE},	67	{MOE_DISTRIBUTE_COMBINE_ADD_RMS_NORM_OP_TYPE, GROUP_CNT_OF_MOE_DISTRIBUTE},
73	- {MOE_DISTRIBUTE_COMBINE_SETUP_OP_TYPE, ONE_GROUP_CNT_OF_MOE_DISTRIBUTE},
74	- {MOE_DISTRIBUTE_COMBINE_TEARDOWN_OP_TYPE, ONE_GROUP_CNT_OF_MOE_DISTRIBUTE},
75	- {MOE_DISTRIBUTE_DISPATCH_SETUP_OP_TYPE, ONE_GROUP_CNT_OF_MOE_DISTRIBUTE},
76	- {MOE_DISTRIBUTE_DISPATCH_TEARDOWN_OP_TYPE, ONE_GROUP_CNT_OF_MOE_DISTRIBUTE},
77	{ALLTO_ALLV_GROUPED_MAT_MUL_OP_TYPE, GROUP_CNT_OF_DS_TRAINING},	68	{ALLTO_ALLV_GROUPED_MAT_MUL_OP_TYPE, GROUP_CNT_OF_DS_TRAINING},
78	{GROUPED_MAT_MUL_ALLTO_ALLV_OP_TYPE, GROUP_CNT_OF_DS_TRAINING},	69	{GROUPED_MAT_MUL_ALLTO_ALLV_OP_TYPE, GROUP_CNT_OF_DS_TRAINING},
79	{ALL_TO_ALLV_QUANT_GROUPED_MM_OP_TYPE, GROUP_CNT_OF_DS_TRAINING},	70	{ALL_TO_ALLV_QUANT_GROUPED_MM_OP_TYPE, GROUP_CNT_OF_DS_TRAINING},
	@@ -85,8 +76,6 @@ static const std::unordered_set<std::string> NO_AI_CPU_SET{MOE_DISTRIBUTE_DISPAT
85	MOE_DISTRIBUTE_COMBINE_OP_TYPE,	76	MOE_DISTRIBUTE_COMBINE_OP_TYPE,
86	MOE_DISTRIBUTE_DISPATCH_V2_OP_TYPE,	77	MOE_DISTRIBUTE_DISPATCH_V2_OP_TYPE,
87	MOE_DISTRIBUTE_COMBINE_V2_OP_TYPE,	78	MOE_DISTRIBUTE_COMBINE_V2_OP_TYPE,
88	- MOE_DISTRIBUTE_DISPATCH_TEARDOWN_OP_TYPE,
89	- MOE_DISTRIBUTE_COMBINE_TEARDOWN_OP_TYPE,
90	MOE_DISTRIBUTE_COMBINE_ADD_RMS_NORM_OP_TYPE,	79	MOE_DISTRIBUTE_COMBINE_ADD_RMS_NORM_OP_TYPE,
91	DISTRIBUTE_BARRIER_OP_TYPE,	80	DISTRIBUTE_BARRIER_OP_TYPE,
92	ATTENTION_TO_FFN_OP_TYPE,	81	ATTENTION_TO_FFN_OP_TYPE,
	@@ -300,10 +289,6 @@ ge::Status Mc2MoeGenTaskOpsUtils::Mc2MoeGenTaskCallback(const gert::ExeResGenera
300	return ge::GRAPH_SUCCESS;	289	return ge::GRAPH_SUCCESS;
301	}	290	}
302		291
303	- const std::string opTypeStr = opType; // 和V2的差异点，待确认，应该可以整合为一，用V2的if逻辑即可
304	- if (opTypeStr == MOE_DISTRIBUTE_COMBINE_TEARDOWN_OP_TYPE \|\| opTypeStr == MOE_DISTRIBUTE_DISPATCH_TEARDOWN_OP_TYPE) {
305	- return ge::GRAPH_SUCCESS;
306	- }
307	return Mc2MoeInsertTask(context, tasks, groupCnt);	292	return Mc2MoeInsertTask(context, tasks, groupCnt);
308	}	293	}
309		294

Mmc2/grouped_mat_mul_all_reduce/tests/ut/op_host/test_grouped_mat_mul_all_reduce_infershape.cpp+5-17

	@@ -82,14 +82,13 @@ TEST_F(GroupedMatMulAllReduceInfershape, Runtime0)
82	outputShapeRef[3] = &yShape3;	82	outputShapeRef[3] = &yShape3;
83	auto contextHolder = gert::InferShapeContextFaker()	83	auto contextHolder = gert::InferShapeContextFaker()
84	.SetOpType("GroupedMatMulAllReduce")	84	.SetOpType("GroupedMatMulAllReduce")
85	- .NodeIoNum(8, 4)	85	+ .IrInstanceNum({4, 4, 0}, {4})
86	.InputTensors(inputShapeRef)	86	.InputTensors(inputShapeRef)
87	.OutputShapes(outputShapeRef)	87	.OutputShapes(outputShapeRef)
88	.Attr("splitItem", int64_t(0))	88	.Attr("splitItem", int64_t(0))
89	.Attr("group", AscendString("group"))	89	.Attr("group", AscendString("group"))
90	.Attr("reduceOp", AscendString("sum"))	90	.Attr("reduceOp", AscendString("sum"))
91	.Attr("commTurn", int64_t(0))	91	.Attr("commTurn", int64_t(0))
92	- .IrInstanceNum({4, 4, 0}, {4})
93	.Build();	92	.Build();
94	auto context = contextHolder.GetContext();	93	auto context = contextHolder.GetContext();
95	ASSERT_NE(context, nullptr);	94	ASSERT_NE(context, nullptr);
	@@ -141,7 +140,6 @@ TEST_F(GroupedMatMulAllReduceInfershape, Runtime0MultiDim)
141		140
142	auto contextHolder = gert::InferShapeContextFaker()	141	auto contextHolder = gert::InferShapeContextFaker()
143	.SetOpType("GroupedMatMulAllReduce")	142	.SetOpType("GroupedMatMulAllReduce")
144	- .NodeIoNum(8, 4)
145	.IrInstanceNum({4, 4, 0}, {4})	143	.IrInstanceNum({4, 4, 0}, {4})
146	.InputTensors(inputShapeRef)	144	.InputTensors(inputShapeRef)
147	.OutputShapes(outputShapeRef)	145	.OutputShapes(outputShapeRef)
	@@ -220,7 +218,6 @@ TEST_F(GroupedMatMulAllReduceInfershape, Runtime1)
220		218
221	auto contextHolder = gert::InferShapeContextFaker()	219	auto contextHolder = gert::InferShapeContextFaker()
222	.SetOpType("GroupedMatMulAllReduce")	220	.SetOpType("GroupedMatMulAllReduce")
223	- .NodeIoNum(6, 4)
224	.IrInstanceNum({1, 4, 0}, {4})	221	.IrInstanceNum({1, 4, 0}, {4})
225	.InputTensors(inputShapeRef)	222	.InputTensors(inputShapeRef)
226	.OutputShapes(outputShapeRef)	223	.OutputShapes(outputShapeRef)
	@@ -302,7 +299,6 @@ TEST_F(GroupedMatMulAllReduceInfershape, Runtime1NonIncreasingGroupList)
302		299
303	auto contextHolder = gert::InferShapeContextFaker()	300	auto contextHolder = gert::InferShapeContextFaker()
304	.SetOpType("GroupedMatMulAllReduce")	301	.SetOpType("GroupedMatMulAllReduce")
305	- .NodeIoNum(6, 4)
306	.IrInstanceNum({1, 4, 0}, {4})	302	.IrInstanceNum({1, 4, 0}, {4})
307	.InputTensors(inputShapeRef)	303	.InputTensors(inputShapeRef)
308	.OutputShapes(outputShapeRef)	304	.OutputShapes(outputShapeRef)
	@@ -351,7 +347,6 @@ TEST_F(GroupedMatMulAllReduceInfershape, Runtime2)
351		347
352	auto contextHolder = gert::InferShapeContextFaker()	348	auto contextHolder = gert::InferShapeContextFaker()
353	.SetOpType("GroupedMatMulAllReduce")	349	.SetOpType("GroupedMatMulAllReduce")
354	- .NodeIoNum(8, 1)
355	.IrInstanceNum({4, 4, 0}, {1})	350	.IrInstanceNum({4, 4, 0}, {1})
356	.InputTensors(inputShapeRef)	351	.InputTensors(inputShapeRef)
357	.OutputShapes(outputShapeRef)	352	.OutputShapes(outputShapeRef)
	@@ -414,7 +409,6 @@ TEST_F(GroupedMatMulAllReduceInfershape, Runtime3)
414		409
415	auto contextHolder = gert::InferShapeContextFaker()	410	auto contextHolder = gert::InferShapeContextFaker()
416	.SetOpType("GroupedMatMulAllReduce")	411	.SetOpType("GroupedMatMulAllReduce")
417	- .NodeIoNum(6, 1)
418	.IrInstanceNum({1, 4, 0}, {1})	412	.IrInstanceNum({1, 4, 0}, {1})
419	.InputTensors(inputShapeRef)	413	.InputTensors(inputShapeRef)
420	.OutputShapes(outputShapeRef)	414	.OutputShapes(outputShapeRef)
	@@ -453,8 +447,7 @@ TEST_F(GroupedMatMulAllReduceInfershape, InferShapeFor2dim)
453	inputShapeRef[3] = (gert::Tensor*)&groupListShape;	447	inputShapeRef[3] = (gert::Tensor*)&groupListShape;
454	auto contextHolder = gert::InferShapeContextFaker()	448	auto contextHolder = gert::InferShapeContextFaker()
455	.SetOpType("GroupedMatMulAllReduce")	449	.SetOpType("GroupedMatMulAllReduce")
456	- .NodeIoNum(4, 1)	450	+ .IrInstanceNum({1, 1, 1, 1}, {1})
457	- .IrInstanceNum({1, 1, 1}, {1})
458	.InputTensors(inputShapeRef)	451	.InputTensors(inputShapeRef)
459	.OutputShapes({&outputShape})	452	.OutputShapes({&outputShape})
460	.Attr("splitItem", int64_t(0))	453	.Attr("splitItem", int64_t(0))
	@@ -488,8 +481,7 @@ TEST_F(GroupedMatMulAllReduceInfershape, InferShapeFor3dim)
488		481
489	auto contextHolder = gert::InferShapeContextFaker()	482	auto contextHolder = gert::InferShapeContextFaker()
490	.SetOpType("GroupedMatMulAllReduce")	483	.SetOpType("GroupedMatMulAllReduce")
491	- .NodeIoNum(4, 1)	484	+ .IrInstanceNum({1, 1, 1, 1}, {1})
492	- .IrInstanceNum({1, 1, 1}, {1})
493	.InputTensors(inputShapeRef)	485	.InputTensors(inputShapeRef)
494	.OutputShapes({&outputShape})	486	.OutputShapes({&outputShape})
495	.Attr("splitItem", int64_t(0))	487	.Attr("splitItem", int64_t(0))
	@@ -522,8 +514,7 @@ TEST_F(GroupedMatMulAllReduceInfershape, InferShapeForinvalidK)
522	inputShapeRef[3] = (gert::Tensor*)&groupListShape;	514	inputShapeRef[3] = (gert::Tensor*)&groupListShape;
523	auto contextHolder = gert::InferShapeContextFaker()	515	auto contextHolder = gert::InferShapeContextFaker()
524	.SetOpType("GroupedMatMulAllReduce")	516	.SetOpType("GroupedMatMulAllReduce")
525	- .NodeIoNum(4, 1)	517	+ .IrInstanceNum({1, 1, 1, 1}, {1})
526	- .IrInstanceNum({1, 1, 1}, {1})
527	.InputTensors(inputShapeRef)	518	.InputTensors(inputShapeRef)
528	.OutputShapes({&outputShape})	519	.OutputShapes({&outputShape})
529	.Attr("splitItem", int64_t(0))	520	.Attr("splitItem", int64_t(0))
	@@ -555,8 +546,7 @@ TEST_F(GroupedMatMulAllReduceInfershape, InferShapeForzeroK)
555	inputShapeRef[3] = (gert::Tensor*)&groupListShape;	546	inputShapeRef[3] = (gert::Tensor*)&groupListShape;
556	auto contextHolder = gert::InferShapeContextFaker()	547	auto contextHolder = gert::InferShapeContextFaker()
557	.SetOpType("GroupedMatMulAllReduce")	548	.SetOpType("GroupedMatMulAllReduce")
558	- .NodeIoNum(4, 1)	549	+ .IrInstanceNum({1, 1, 1, 1}, {1})
559	- .IrInstanceNum({1, 1, 1}, {1})
560	.InputTensors(inputShapeRef)	550	.InputTensors(inputShapeRef)
561	.OutputShapes({&outputShape})	551	.OutputShapes({&outputShape})
562	.Attr("splitItem", int64_t(0))	552	.Attr("splitItem", int64_t(0))
	@@ -579,7 +569,6 @@ TEST_F(GroupedMatMulAllReduceInfershape, InferDtype)
579		569
580	auto contextHolder = gert::InferDataTypeContextFaker()	570	auto contextHolder = gert::InferDataTypeContextFaker()
581	.SetOpType("GroupedMatMulAllReduce")	571	.SetOpType("GroupedMatMulAllReduce")
582	- .NodeIoNum(2, 1)
583	.IrInstanceNum({1, 1}, {1})	572	.IrInstanceNum({1, 1}, {1})
584	.NodeInputTd(0, x1Dtype, ge::FORMAT_ND, ge::FORMAT_ND)	573	.NodeInputTd(0, x1Dtype, ge::FORMAT_ND, ge::FORMAT_ND)
585	.NodeInputTd(1, x2Dtype, ge::FORMAT_ND, ge::FORMAT_ND)	574	.NodeInputTd(1, x2Dtype, ge::FORMAT_ND, ge::FORMAT_ND)
	@@ -622,7 +611,6 @@ TEST_F(GroupedMatMulAllReduceInfershape, InferDtypeTestRuntime2)
622		611
623	auto contextHolder = gert::InferDataTypeContextFaker()	612	auto contextHolder = gert::InferDataTypeContextFaker()
624	.SetOpType("GroupedMatMulAllReduce")	613	.SetOpType("GroupedMatMulAllReduce")
625	- .NodeIoNum(2, 1)
626	.IrInstanceNum({1, 1}, {1})	614	.IrInstanceNum({1, 1}, {1})
627	.NodeInputTd(0, xDtype0, ge::FORMAT_ND, ge::FORMAT_ND)	615	.NodeInputTd(0, xDtype0, ge::FORMAT_ND, ge::FORMAT_ND)
628	.NodeInputTd(1, xDtype1, ge::FORMAT_ND, ge::FORMAT_ND)	616	.NodeInputTd(1, xDtype1, ge::FORMAT_ND, ge::FORMAT_ND)

Mmc2/moe_distribute_combine_setup/README.md+1-1

	@@ -1,3 +1,3 @@
1	# MoeDistributeCombineSetup	1	# MoeDistributeCombineSetup
2		2
3	-本目录仅包含MoeDistributeCombineSetup算子对应的aclnn接口；如您想要贡献该算子的Ascend C实现，请参考[贡献流程](../../CONTRIBUTING.md)。	3	+本目录包含MoeDistributeCombineSetup算子的实现，不提供公开的aclnn接口；如您想要为该算子贡献代码，请参考[贡献流程](../../CONTRIBUTING.md)。

Dmc2/moe_distribute_combine_setup/docs/aclnnMoeDistributeCombineSetup.md+0-1031

Dmc2/moe_distribute_combine_setup/examples/test_aclnn_moe_distribute_combine_setup.cpp+0-356

	@@ -1,356 +0,0 @@
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 test_aclnn_moe_distribute_combine_setup.cpp
13	- * \brief aclnn直调测试脚本
14	- */
15	-
16	-#include <thread>
17	-#include <iostream>
18	-#include <string>
19	-#include <vector>
20	-#include <cstring>
21	-#include "acl/acl.h"
22	-#include "hccl/hccl.h"
23	-#include "aclnnop/aclnn_moe_distribute_combine_setup.h"
24	-#include "aclnnop/aclnn_moe_distribute_combine_teardown.h"
25	-
26	-#define CHECK_RET(cond, return_expr) \
27	- do { \
28	- if (!(cond)) { \
29	- return_expr; \
30	- } \
31	- } while (0)
32	-
33	-#define LOG_PRINT(message, ...) \
34	- do { \
35	- printf(message, ##__VA_ARGS__); \
36	- } while (0)
37	-
38	-constexpr int DEV_NUM = 2;
39	-
40	-template <typename Func>
41	-class Guard {
42	-public:
43	- explicit Guard(Func& func)
44	- : func_(func)
45	- {}
46	- ~Guard()
47	- {
48	- func_();
49	- }
50	-
51	-private:
52	- Func& func_;
53	-};
54	-
55	-int64_t AlignN(int64_t x, int64_t n)
56	-{
57	- return (x + n - 1) / n * n;
58	-}
59	-
60	-int64_t GetShapeSize(const std::vector<int64_t>& shape)
61	-{
62	- int64_t shape_size = 1;
63	- for (auto i : shape) {
64	- shape_size *= i;
65	- }
66	- return shape_size;
67	-}
68	-
69	-struct Args {
70	- uint32_t rankId;
71	- uint32_t epRankId;
72	- HcclComm hcclEpComm;
73	- aclrtStream combinesetupstream;
74	- aclrtStream combineteardownstream;
75	- aclrtContext context;
76	-};
77	-
78	-template <typename T>
79	-int CreateAclTensor(const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr,
80	- aclDataType dataType, aclTensor** tensor)
81	-{
82	- auto size = GetShapeSize(shape) * sizeof(T);
83	- auto ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
84	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclrtMalloc failed. ret: %d\n", ret); return ret);
85	-
86	- ret = aclrtMemcpy(*deviceAddr, size, hostData.data(), size, ACL_MEMCPY_HOST_TO_DEVICE);
87	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclrtMemcpy failed. ret: %d\n", ret); return ret);
88	- std::vector<int64_t> strides(shape.size(), 1);
89	- for (int64_t i = shape.size() - 2; i >= 0; i--) {
90	- strides[i] = shape[i + 1] * strides[i + 1];
91	- }
92	- *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND,
93	- shape.data(), shape.size(), *deviceAddr);
94	- return 0;
95	-}
96	-
97	-void DestroyTensor(aclTensor* tensor)
98	-{
99	- if (tensor != nullptr) {
100	- aclDestroyTensor(tensor);
101	- }
102	-}
103	-
104	-void FreeDeviceAddr(void* deviceAddr)
105	-{
106	- if (deviceAddr != nullptr) {
107	- aclrtFree(deviceAddr);
108	- }
109	-}
110	-
111	-int LaunchOneProcess(Args& args)
112	-{
113	- int ret = aclrtSetCurrentContext(args.context);
114	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclrtSetCurrentContext failed. ret: %d\n", ret); return ret);
115	-
116	- char hcomEpName[128] = {0};
117	- ret = HcclGetCommName(args.hcclEpComm, hcomEpName);
118	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] HcclGetCommName failed. ret: %d\n", ret); return ret);
119	-
120	- auto destroyFunc = [&args]() {
121	- std::cout << "== begin to destroy " << std::endl;
122	- HcclCommDestroy(args.hcclEpComm);
123	- aclrtDestroyStream(args.combinesetupstream);
124	- aclrtDestroyStream(args.combineteardownstream);
125	- aclrtDestroyContext(args.context);
126	- aclrtResetDevice(args.rankId);
127	- };
128	- auto guard = Guard<decltype(destroyFunc)>(destroyFunc);
129	-
130	- // 场景参数
131	- int64_t bs = 16;
132	- int64_t h = 4096;
133	- int64_t k = 6;
134	- int64_t expertShardType = 0;
135	- int64_t sharedExpertNum = 0;
136	- int64_t sharedExpertRankNum = 0;
137	- int64_t moeExpertNum = 32;
138	- int64_t commQuantMode = 0;
139	- int64_t tpWorldSize = 1;
140	- int64_t epWorldSize = DEV_NUM;
141	- int64_t commType = 2;
142	- int64_t timeOut = 100000000;
143	-
144	- int64_t globalBS = bs * epWorldSize;
145	- int64_t localExpertNum;
146	- int64_t localToken;
147	-
148	- if (args.epRankId < sharedExpertRankNum) {
149	- localExpertNum = 1;
150	- localToken = globalBS / sharedExpertRankNum;
151	- } else {
152	- localExpertNum = moeExpertNum / (epWorldSize - sharedExpertRankNum);
153	- localToken = globalBS * (localExpertNum < k ? localExpertNum : k);
154	- }
155	-
156	- uint64_t assistInfoForCombineOutSize = localToken * 128;
157	-
158	- // 定义 shape
159	- std::vector<int64_t> expandXShape{tpWorldSize * localToken, h};
160	- std::vector<int64_t> expertIdsShape{bs, k};
161	- std::vector<int64_t> expertScalesShape{bs, k};
162	- std::vector<int64_t> expandIdxShape{bs * k};
163	- std::vector<int64_t> assistInfoForCombineOutShape{assistInfoForCombineOutSize};
164	- std::vector<int64_t> quantExpandXOutShape{tpWorldSize * localToken,
165	- AlignN(AlignN(h, 32) + AlignN(h, 8) / 8 * sizeof(float), 512)};
166	- std::vector<int64_t> commCmdInfoOutShapeforcombine{(localToken + epWorldSize) * 16};
167	- std::vector<int64_t> xOutShape{bs, h};
168	-
169	- int64_t expandXShapeSize = GetShapeSize(expandXShape);
170	- int64_t expertIdsShapeSize = GetShapeSize(expertIdsShape);
171	- int64_t expertScalesShapeSize = GetShapeSize(expertScalesShape);
172	- int64_t expandIdxShapeSize = GetShapeSize(expandIdxShape);
173	- int64_t assistInfoForCombineOutShapeSize = GetShapeSize(assistInfoForCombineOutShape);
174	- int64_t quantExpandXOutShapeSize = GetShapeSize(quantExpandXOutShape);
175	- int64_t commCmdInfoOutShapeSizeforcombine = GetShapeSize(commCmdInfoOutShapeforcombine);
176	- int64_t xOutShapeSize = GetShapeSize(xOutShape);
177	-
178	- // 构造 host 数据
179	- std::vector<int16_t> expandXHostData(expandXShapeSize, 0);
180	- std::vector<int32_t> expertIdsHostData;
181	- for (int32_t token_id = 0; token_id < expertIdsShape[0]; token_id++) {
182	- for (int32_t k_id = 0; k_id < expertIdsShape[1]; k_id++) {
183	- expertIdsHostData.push_back(k_id);
184	- }
185	- }
186	- std::vector<float> expertScalesHostData(expertScalesShapeSize, 0);
187	- std::vector<int32_t> expandIdxHostData(expandIdxShapeSize, 0);
188	- std::vector<int32_t> assistInfoForCombineOutHostData(assistInfoForCombineOutShapeSize, 0);
189	- std::vector<int8_t> quantExpandXOutHostData(quantExpandXOutShapeSize, 0);
190	- std::vector<int32_t> commCmdInfoOutforCombineHostData(commCmdInfoOutShapeSizeforcombine, 0);
191	- std::vector<float> xOutHostData(xOutShapeSize, 0);
192	-
193	- // 声明 device 地址和 tensor
194	- void* expandXDeviceAddr = nullptr;
195	- void* expertIdsDeviceAddr = nullptr;
196	- void* expertScalesDeviceAddr = nullptr;
197	- void* expandIdxDeviceAddr = nullptr;
198	- void* assistInfoForCombineOutDeviceAddr = nullptr;
199	- void* quantExpandXOutDeviceAddr = nullptr;
200	- void* commCmdInfoOutforCombineDeviceAddr = nullptr;
201	- void* xOutDeviceAddr = nullptr;
202	-
203	- aclTensor* expandX = nullptr;
204	- aclTensor* expertIds = nullptr;
205	- aclTensor* expertScales = nullptr;
206	- aclTensor* expandIdx = nullptr;
207	- aclTensor* assistInfoForCombineOut = nullptr;
208	- aclTensor* quantExpandXOut = nullptr;
209	- aclTensor* commCmdInfoOutforCombine = nullptr;
210	- aclTensor* xOut = nullptr;
211	-
212	- // 创建 tensor
213	- ret = CreateAclTensor(expandXHostData, expandXShape, &expandXDeviceAddr, aclDataType::ACL_FLOAT16, &expandX);
214	- CHECK_RET(ret == ACL_SUCCESS, return ret);
215	- ret = CreateAclTensor(expertIdsHostData, expertIdsShape, &expertIdsDeviceAddr, aclDataType::ACL_INT32, &expertIds);
216	- CHECK_RET(ret == ACL_SUCCESS, return ret);
217	- ret = CreateAclTensor(expertScalesHostData, expertScalesShape, &expertScalesDeviceAddr, aclDataType::ACL_FLOAT,
218	- &expertScales);
219	- CHECK_RET(ret == ACL_SUCCESS, return ret);
220	- ret = CreateAclTensor(expandIdxHostData, expandIdxShape, &expandIdxDeviceAddr, aclDataType::ACL_INT32, &expandIdx);
221	- CHECK_RET(ret == ACL_SUCCESS, return ret);
222	- ret = CreateAclTensor(assistInfoForCombineOutHostData, assistInfoForCombineOutShape,
223	- &assistInfoForCombineOutDeviceAddr, aclDataType::ACL_INT32, &assistInfoForCombineOut);
224	- CHECK_RET(ret == ACL_SUCCESS, return ret);
225	- ret = CreateAclTensor(quantExpandXOutHostData, quantExpandXOutShape, &quantExpandXOutDeviceAddr,
226	- aclDataType::ACL_INT8, &quantExpandXOut);
227	- CHECK_RET(ret == ACL_SUCCESS, return ret);
228	- ret = CreateAclTensor(commCmdInfoOutforCombineHostData, commCmdInfoOutShapeforcombine,
229	- &commCmdInfoOutforCombineDeviceAddr, aclDataType::ACL_INT32, &commCmdInfoOutforCombine);
230	- CHECK_RET(ret == ACL_SUCCESS, return ret);
231	- ret = CreateAclTensor(xOutHostData, xOutShape, &xOutDeviceAddr, aclDataType::ACL_FLOAT16, &xOut);
232	- CHECK_RET(ret == ACL_SUCCESS, return ret);
233	-
234	- // 声明算子执行必需变量
235	- uint64_t workspaceSize = 0;
236	- aclOpExecutor* executor = nullptr;
237	- void* workspaceAddr = nullptr;
238	-
239	- /****************************调用combine_setup******************************************/
240	- ret = aclnnMoeDistributeCombineSetupGetWorkspaceSize(
241	- expandX, expertIds, assistInfoForCombineOut, hcomEpName, epWorldSize, args.epRankId, moeExpertNum,
242	- expertShardType, sharedExpertNum, sharedExpertRankNum, globalBS, commQuantMode, commType, nullptr,
243	- quantExpandXOut, commCmdInfoOutforCombine, &workspaceSize, &executor);
244	- CHECK_RET(ret == ACL_SUCCESS,
245	- LOG_PRINT("[ERROR] aclnnMoeDistributeCombineSetupGetWorkspaceSize failed. ret = %d \n", ret);
246	- return ret);
247	-
248	- if (workspaceSize > 0) {
249	- ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
250	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] CombineSetup aclrtMalloc failed. ret = %d\n", ret);
251	- return ret);
252	- }
253	-
254	- ret = aclnnMoeDistributeCombineSetup(workspaceAddr, workspaceSize, executor, args.combinesetupstream);
255	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclnnMoeDistributeCombineSetup failed. ret = %d \n", ret);
256	- return ret);
257	-
258	- ret = aclrtSynchronizeStreamWithTimeout(args.combinesetupstream, timeOut);
259	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclrtSynchronizeStreamWithTimeout failed. ret = %d \n", ret);
260	- return ret);
261	- LOG_PRINT("[INFO] device_%d aclnnMoeDistributeCombineSetup execute successfully.\n", args.rankId);
262	-
263	- /****************************调用combine_teardown******************************************/
264	- ret = aclnnMoeDistributeCombineTeardownGetWorkspaceSize(
265	- expandX, quantExpandXOut, expertIds, expandIdx, expertScales, commCmdInfoOutforCombine, nullptr, nullptr,
266	- hcomEpName, epWorldSize, args.epRankId, moeExpertNum, expertShardType, sharedExpertNum, sharedExpertRankNum,
267	- globalBS, commQuantMode, commType, nullptr, xOut, &workspaceSize, &executor);
268	- CHECK_RET(ret == ACL_SUCCESS,
269	- LOG_PRINT("[ERROR] aclnnMoeDistributeCombineTeardownGetWorkspaceSize failed. ret = %d \n", ret);
270	- return ret);
271	-
272	- if (workspaceSize > 0) {
273	- ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
274	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] CombineTeardown aclrtMalloc failed. ret = %d\n", ret);
275	- return ret);
276	- }
277	-
278	- ret = aclnnMoeDistributeCombineTeardown(workspaceAddr, workspaceSize, executor, args.combineteardownstream);
279	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclnnMoeDistributeCombineTeardown failed. ret = %d \n", ret);
280	- return ret);
281	-
282	- ret = aclrtSynchronizeStreamWithTimeout(args.combineteardownstream, timeOut);
283	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclrtSynchronizeStreamWithTimeout failed. ret = %d \n", ret);
284	- return ret);
285	- LOG_PRINT("[INFO] device_%d aclnnMoeDistributeCombineTeardown execute successfully.\n", args.rankId);
286	-
287	- // 释放资源
288	- if (workspaceSize > 0) {
289	- aclrtFree(workspaceAddr);
290	- }
291	- DestroyTensor(expandX);
292	- DestroyTensor(expertIds);
293	- DestroyTensor(expertScales);
294	- DestroyTensor(expandIdx);
295	- DestroyTensor(assistInfoForCombineOut);
296	- DestroyTensor(quantExpandXOut);
297	- DestroyTensor(commCmdInfoOutforCombine);
298	- DestroyTensor(xOut);
299	-
300	- FreeDeviceAddr(expandXDeviceAddr);
301	- FreeDeviceAddr(expertIdsDeviceAddr);
302	- FreeDeviceAddr(expertScalesDeviceAddr);
303	- FreeDeviceAddr(expandIdxDeviceAddr);
304	- FreeDeviceAddr(assistInfoForCombineOutDeviceAddr);
305	- FreeDeviceAddr(quantExpandXOutDeviceAddr);
306	- FreeDeviceAddr(commCmdInfoOutforCombineDeviceAddr);
307	- FreeDeviceAddr(xOutDeviceAddr);
308	-
309	- return 0;
310	-}
311	-
312	-int main(int argc, char* argv[])
313	-{
314	- int ret = aclInit(nullptr);
315	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclInit failed. ret = %d \n", ret); return ret);
316	-
317	- aclrtStream combineSetupStream[DEV_NUM];
318	- aclrtStream combineTeardownStream[DEV_NUM];
319	- aclrtContext context[DEV_NUM];
320	- for (uint32_t rankId = 0; rankId < DEV_NUM; rankId++) {
321	- ret = aclrtSetDevice(rankId);
322	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclrtSetDevice failed. ret = %d \n", ret); return ret);
323	- ret = aclrtCreateContext(&context[rankId], rankId);
324	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclrtCreateContext failed. ret = %d \n", ret); return ret);
325	- ret = aclrtCreateStream(&combineSetupStream[rankId]);
326	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclrtCreateStream failed. ret = %d \n", ret); return ret);
327	- ret = aclrtCreateStream(&combineTeardownStream[rankId]);
328	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclrtCreateStream failed. ret = %d \n", ret); return ret);
329	- }
330	-
331	- int32_t devices[DEV_NUM];
332	- for (int i = 0; i < DEV_NUM; i++) {
333	- devices[i] = i;
334	- }
335	-
336	- HcclComm comms[DEV_NUM];
337	- ret = HcclCommInitAll(DEV_NUM, devices, comms);
338	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] HcclCommInitAll failed. ret = %d \n", ret); return ret);
339	-
340	- Args args[DEV_NUM];
341	- std::vector<std::unique_ptr<std::thread>> threads(DEV_NUM);
342	- for (uint32_t rankId = 0; rankId < DEV_NUM; rankId++) {
343	- args[rankId].rankId = rankId;
344	- args[rankId].epRankId = rankId;
345	- args[rankId].hcclEpComm = comms[rankId];
346	- args[rankId].combinesetupstream = combineSetupStream[rankId];
347	- args[rankId].combineteardownstream = combineTeardownStream[rankId];
348	- args[rankId].context = context[rankId];
349	- threads[rankId].reset(new (std::nothrow) std::thread(&LaunchOneProcess, std::ref(args[rankId])));
350	- }
351	- for (uint32_t rankId = 0; rankId < DEV_NUM; rankId++) {
352	- threads[rankId]->join();
353	- }
354	- aclFinalize();
355	- return 0;
356	-}

Dmc2/moe_distribute_combine_setup/op_api/aclnn_moe_distribute_combine_setup.cpp+0-224

	@@ -1,224 +0,0 @@
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 aclnn_moe_distribute_combine_setup.cpp
13	- * \brief
14	- */
15	-#include "aclnn_moe_distribute_combine_setup.h"
16	-#include <algorithm>
17	-#include "common/utils/op_mc2.h"
18	-#include "common/op_host/op_api/mc2_3rd_matmul_util.h"
19	-#include "common/utils/op_mc2_def.h"
20	-#include "aclnn_kernels/common/op_error_check.h"
21	-
22	-#include "aclnnInner_moe_distribute_combine_setup.h"
23	-#include "common/op_api/mc2_context.h"
24	-
25	-#include "opdev/op_log.h"
26	-#include "opdev/common_types.h"
27	-#include "opdev/platform.h"
28	-#include "op_host/util/op_const_def.h"
29	-#include "mc2_log_compat.h"
30	-
31	-namespace {
32	-
33	-using namespace Mc2Aclnn;
34	-using namespace op;
35	-
36	-static inline bool CheckEmptyTensor(const aclTensor* tensor, const char* name)
37	-{
38	- const auto& shape = tensor->GetViewShape();
39	- if (shape.GetDimNum() == 0) {
40	- OP_LOGE_FOR_INVALID_SHAPE_WITH_REASON("aclnnMoeDistributeCombineSetup", name, "[]",
41	- (std::string("The shape of ") + name + " cannot be empty.").c_str());
42	- return false;
43	- }
44	- for (size_t i = 0; i < shape.GetDimNum(); ++i) {
45	- if (shape.GetDim(i) <= 0) {
46	- OP_LOGE_FOR_INVALID_SHAPEDIM_WITH_REASON("aclnnMoeDistributeCombineSetup", name,
47	- std::to_string(shape.GetDim(i)).c_str(),
48	- "dimension must be positive");
49	- return false;
50	- }
51	- }
52	- return true;
53	-}
54	-
55	-#define OP_CHECK_EMPTY_TENSOR(tensor, retExpr) \
56	- if (!CheckEmptyTensor(tensor, #tensor)) { \
57	- retExpr; \
58	- }
59	-
60	-static constexpr size_t TWO_DIM = 2;
61	-static constexpr int64_t ALIGN_8 = 8;
62	-static constexpr int64_t ALIGN_32 = 32;
63	-static constexpr int64_t ALIGN_512 = 512;
64	-static constexpr int64_t COMM_CMD_INFO_BASE = 16;
65	-
66	-enum class NnopbaseHcclServerType : uint32_t {
67	- NNOPBASE_HCCL_SERVER_TYPE_AICPU = 0,
68	- NNOPBASE_HCCL_SERVER_TYPE_MTE,
69	- NNOPBASE_HCCL_SERVER_TYPE_CCU,
70	- NNOPBASE_HCCL_SERVER_TYPE_END
71	-};
72	-
73	-static inline int64_t AlignUp(int64_t x, int64_t base)
74	-{
75	- if (base == 0) {
76	- OP_LOGD("AlignUp: base cannot be zero");
77	- return 0;
78	- }
79	- return ((x + base - 1) / base) * base;
80	-}
81	-
82	-// check nullptr
83	-static bool CheckNotNull(const aclTensor* expandX, const aclTensor* expertIds, const aclTensor* assistInfoForCombine,
84	- const char* groupEp, aclTensor* quantExpandXOut, aclTensor* commCmdInfoOut)
85	-{
86	- OP_LOGD("aclnn_moe_distribute_combine_setup CheckNotNull start");
87	- OP_CHECK_NULL(expandX, return false);
88	- OP_CHECK_NULL(expertIds, return false);
89	- OP_CHECK_NULL(assistInfoForCombine, return false);
90	- OP_CHECK_NULL(quantExpandXOut, return false);
91	- OP_CHECK_NULL(commCmdInfoOut, return false);
92	- if ((groupEp == nullptr) \|\| (strnlen(groupEp, HCCL_GROUP_NAME_MAX) == 0)) {
93	- OP_LOGE_WITH_INVALID_INPUT("aclnnMoeDistributeCombineSetup", "groupEp");
94	- return false;
95	- }
96	- OP_LOGD("aclnn_moe_distribute_combine_setup CheckNotNull success");
97	- return true;
98	-}
99	-
100	-// 入参校验
101	-static aclnnStatus CheckParams(const aclTensor* expandX, const aclTensor* expertIds,
102	- const aclTensor* assistInfoForCombine, const char* groupEp, int64_t epWorldSize,
103	- int64_t epRankId, int64_t moeExpertNum, int64_t expertShardType, int64_t sharedExpertNum,
104	- int64_t sharedExpertRankNum, int64_t globalBs, int64_t commQuantMode, int64_t commType,
105	- const char* commAlg, aclTensor* quantExpandXOut, aclTensor* commCmdInfoOut)
106	-{
107	- (void)epWorldSize;
108	- (void)epRankId;
109	- (void)moeExpertNum;
110	- (void)expertShardType;
111	- (void)sharedExpertNum;
112	- (void)sharedExpertRankNum;
113	- (void)globalBs;
114	- (void)commQuantMode;
115	- (void)commType;
116	- (void)commAlg;
117	- OP_LOGD("aclnn_moe_distribute_combine_setup CheckParams start");
118	- CHECK_RET(CheckNotNull(expandX, expertIds, assistInfoForCombine, groupEp, quantExpandXOut, commCmdInfoOut),
119	- ACLNN_ERR_PARAM_NULLPTR);
120	- OP_CHECK_EMPTY_TENSOR(expandX, return ACLNN_ERR_PARAM_INVALID);
121	- OP_CHECK_EMPTY_TENSOR(expertIds, return ACLNN_ERR_PARAM_INVALID);
122	- OP_CHECK_EMPTY_TENSOR(assistInfoForCombine, return ACLNN_ERR_PARAM_INVALID);
123	- OP_CHECK_EMPTY_TENSOR(quantExpandXOut, return ACLNN_ERR_PARAM_INVALID);
124	- OP_CHECK_EMPTY_TENSOR(commCmdInfoOut, return ACLNN_ERR_PARAM_INVALID);
125	- if (strnlen(groupEp, HCCL_GROUP_NAME_MAX) >= HCCL_GROUP_NAME_MAX) {
126	- OP_LOGE_FOR_INVALID_VALUE_WITH_REASON("aclnnMoeDistributeCombineSetup", "groupEp",
127	- "length exceeds " + std::to_string(HCCL_GROUP_NAME_MAX),
128	- "groupEp name too long");
129	- return ACLNN_ERR_PARAM_INVALID;
130	- }
131	- OP_LOGD("aclnn_moe_distribute_combine_setup CheckParams success");
132	- return ACLNN_SUCCESS;
133	-}
134	-
135	-} // namespace
136	-
137	-extern "C" void __attribute__((weak)) NnopbaseSetHcclServerType(void* executor, NnopbaseHcclServerType sType);
138	-
139	-extern "C" aclnnStatus aclnnMoeDistributeCombineSetupGetWorkspaceSize(
140	- const aclTensor* expandX, const aclTensor* expertIds, const aclTensor* assistInfoForCombine, const char* groupEp,
141	- int64_t epWorldSize, int64_t epRankId, int64_t moeExpertNum, int64_t expertShardType, int64_t sharedExpertNum,
142	- int64_t sharedExpertRankNum, int64_t globalBs, int64_t commQuantMode, int64_t commType, const char* commAlg,
143	- aclTensor* quantExpandXOut, aclTensor* commCmdInfoOut, uint64_t* workspaceSize, aclOpExecutor** executor)
144	-{
145	- OP_LOGD("aclnnMoeDistributeCombineSetupGetWorkspaceSize start.");
146	- if (GetCurrentPlatformInfo().GetCurNpuArch() != Ops::Base::DAV_3510) {
147	- OP_LOGE_FOR_INVALID_VALUE_WITH_REASON(
148	- "aclnnMoeDistributeCombineSetup", "npuArch",
149	- std::to_string(static_cast<int64_t>(GetCurrentPlatformInfo().GetCurNpuArch())).c_str(),
150	- "only support Ops::Base::DAV_3510");
151	- return ACLNN_ERR_PARAM_INVALID;
152	- }
153	-
154	- auto retParam = CheckParams(expandX, expertIds, assistInfoForCombine, groupEp, epWorldSize, epRankId, moeExpertNum,
155	- expertShardType, sharedExpertNum, sharedExpertRankNum, globalBs, commQuantMode,
156	- commType, commAlg, quantExpandXOut, commCmdInfoOut);
157	- CHECK_RET(retParam == ACLNN_SUCCESS, retParam);
158	-
159	- aclnnStatus retStatus;
160	- uint64_t hcclBuffSize = 0;
161	- aclTensor* mc2Context = nullptr;
162	- const char* opName = "moe_distribute_a2av_setup_teardown";
163	- retStatus = Mc2Aclnn::Mc2Context::GetMc2ContextTensor(groupEp, opName, hcclBuffSize, mc2Context,
164	- CommProtocol::COMM_PROTOCOL_UBC_CTP);
165	- CHECK_RET(retStatus == ACLNN_SUCCESS, retStatus);
166	- retStatus = aclnnInnerMoeDistributeCombineSetupGetWorkspaceSize(
167	- mc2Context, expandX, expertIds, assistInfoForCombine, epWorldSize, epRankId, moeExpertNum, hcclBuffSize,
168	- expertShardType, sharedExpertNum, sharedExpertRankNum, globalBs, commQuantMode, commType,
169	- const_cast<char*>(commAlg), quantExpandXOut, commCmdInfoOut, workspaceSize, executor);
170	- return retStatus;
171	-}
172	-
173	-extern "C" aclnnStatus aclnnMoeDistributeCombineSetupTeardownCalcOutputSize(
174	- const aclTensor* expandX, const aclTensor* expertIds, const aclTensor* assistInfoForCombine, const char* groupEp,
175	- int64_t epWorldSize, int64_t epRankId, int64_t moeExpertNum, int64_t expertShardType, int64_t sharedExpertNum,
176	- int64_t sharedExpertRankNum, int64_t globalBs, int64_t commQuantMode, int64_t commType, const char* commAlg,
177	- uint64_t& tokenMsgSize, uint64_t& commCmdInfoOutSize)
178	-{
179	- (void)expertIds;
180	- (void)assistInfoForCombine;
181	- (void)groupEp;
182	- (void)epRankId;
183	- (void)moeExpertNum;
184	- (void)expertShardType;
185	- (void)sharedExpertNum;
186	- (void)sharedExpertRankNum;
187	- (void)globalBs;
188	- (void)commQuantMode;
189	- (void)commType;
190	- (void)commAlg;
191	- OP_CHECK_NULL(expandX, return ACLNN_ERR_PARAM_NULLPTR);
192	- OP_CHECK_EMPTY_TENSOR(expandX, return ACLNN_ERR_PARAM_INVALID);
193	- if (expandX->GetViewShape().GetDimNum() != TWO_DIM) {
194	- OP_LOGE_FOR_INVALID_SHAPEDIM_WITH_REASON("aclnnMoeDistributeCombineSetup", "expandX",
195	- "dim=" + std::to_string(expandX->GetViewShape().GetDimNum()),
196	- "The shape dim of expandX must be 2D.");
197	- return ACLNN_ERR_PARAM_INVALID;
198	- }
199	- int64_t a = expandX->GetViewShape().GetDim(0);
200	- int64_t h = expandX->GetViewShape().GetDim(1);
201	- if (a <= 0 \|\| h <= 0) {
202	- OP_LOGE_FOR_INVALID_SHAPE_WITH_REASON("aclnnMoeDistributeCombineSetup", "expandX",
203	- "[" + std::to_string(a) + "," + std::to_string(h) + "]",
204	- "dimensions must be greater than zero");
205	- return ACLNN_ERR_PARAM_INVALID;
206	- }
207	-
208	- tokenMsgSize =
209	- static_cast<uint64_t>(AlignUp(AlignUp(h, ALIGN_32) + AlignUp(h, ALIGN_8) / ALIGN_8 * sizeof(float), ALIGN_512));
210	- commCmdInfoOutSize = static_cast<uint64_t>((a + epWorldSize) * COMM_CMD_INFO_BASE);
211	- return ACLNN_SUCCESS;
212	-}
213	-
214	-extern "C" aclnnStatus aclnnMoeDistributeCombineSetup(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor,
215	- aclrtStream stream)
216	-{
217	- OP_LOGD("aclnnMoeDistributeCombineSetup start");
218	- if (NnopbaseSetHcclServerType) {
219	- NnopbaseSetHcclServerType(executor, NnopbaseHcclServerType::NNOPBASE_HCCL_SERVER_TYPE_MTE);
220	- }
221	- aclnnStatus ret = aclnnInnerMoeDistributeCombineSetup(workspace, workspaceSize, executor, stream);
222	- OP_LOGD("aclnnMoeDistributeCombineSetup success");
223	- return ret;
224	-}

Dmc2/moe_distribute_combine_setup/op_api/aclnn_moe_distribute_combine_setup.h+0-103

	@@ -1,103 +0,0 @@
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 aclnn_moe_distribute_combine_setup.h
13	- * \brief
14	- */
15	-#ifndef OP_API_INC_MOE_DISTRIBUTE_COMBINE_SETUP
16	-#define OP_API_INC_MOE_DISTRIBUTE_COMBINE_SETUP
17	-
18	-#include <string>
19	-
20	-#include "aclnn/aclnn_base.h"
21	-#include "aclnn_util.h"
22	-
23	-#ifdef __cplusplus
24	-extern "C" {
25	-#endif
26	-
27	-/**
28	- * 算子功能：进行AlltoAllv通信，将数据写入对端GM。
29	- * @brief aclnnMoeDistributeCombineSetup的第一段接口，根据具体的计算流程，计算workspace大小。
30	- * @domain aclnn_ops_infer
31	- * @param [in] expandX: 计算输入，Tensor，数据类型float16，bfloat16，必须为2维，数据格式支持ND。
32	- * @param [in] expertIds: 计算输入，Tensor，数据类型int32，必须为2维，数据格式支持ND。
33	- * @param [in] assistInfoForCombine: 计算输入，Tensor，数据类型int32，必须为1维，数据格式支持ND。
34	- * @param [in] groupEp: 计算输入，str。ep通信域名称，专家并行的通信域。不能和groupTp相同。
35	- * @param [in] epWorldSize: 计算输入，int。ep通信域size。
36	- * @param [in] epRankId: 计算输入，int。ep本卡Id。同一个EP通信域中各卡的epRankId不重复。
37	- * @param [in] moeExpertNum: 计算输入，int。MOE专家数量。
38	- * @param [in] expertShardType: 计算可选输入，int。共享专家卡分布类型。当前仅支持传0。
39	- * @param [in] sharedExpertNum: 计算可选输入，int。共享专家数量。
40	- * @param [in] sharedExpertRankNum: 计算可选输入，int。共享专家卡数量。
41	- * @param [in] globalBs: 计算可选输入，int。
42	- * @param [in] commQuantMode: 计算可选输入，int。通信量化类型。预留参数，暂未使用，传0即可。
43	- * @param [in] commType: 计算可选输入，int。通信方案选择。预留参数，暂未使用，传0即可。
44	- * @param [in] commAlg: 计算可选输入，str。通信算法选择。预留参数，暂未使用。
45	- * @param [out] quantExpandXOut: 计算输出，Tensor，数据类型支持int8，仅支持2维，数据格式支持ND。表示量化处理后的token。
46	- * @param [out] commCmdInfoOut: 计算输出，Tensor，数据类型支持int32，仅支持1维，数据格式支持ND。表示通信的cmd信息。
47	- * @param [out] workspaceSize: 出参，返回需要在npu device侧申请的workspace大小。
48	- * @param [out] executor: 出参，返回op执行器，包含了算子计算流程。
49	- * @return aclnnStatus: 返回值，返回状态码.
50	- *
51	- */
52	-ACLNN_API aclnnStatus aclnnMoeDistributeCombineSetupGetWorkspaceSize(
53	- const aclTensor* expandX, const aclTensor* expertIds, const aclTensor* assistInfoForCombine, const char* groupEp,
54	- int64_t epWorldSize, int64_t epRankId, int64_t moeExpertNum, int64_t expertShardType, int64_t sharedExpertNum,
55	- int64_t sharedExpertRankNum, int64_t globalBs, int64_t commQuantMode, int64_t commType, const char* commAlg,
56	- aclTensor* quantExpandXOut, aclTensor* commCmdInfoOut, uint64_t* workspaceSize, aclOpExecutor** executor);
57	-
58	-/**
59	- * 算子功能：进行AlltoAllv通信，将数据写入对端GM。
60	- * @brief 计算aclnnMoeDistributeCombineSetup部分输出的size大小
61	- * @domain aclnn_ops_infer
62	- * @param [in] expandX: 计算输入，Tensor，数据类型float16，bfloat16，必须为2维，数据格式支持ND。
63	- * @param [in] expertIds: 计算输入，Tensor，数据类型int32，必须为2维，数据格式支持ND。
64	- * @param [in] assistInfoForCombine: 计算输入，Tensor，数据类型int32，必须为1维，数据格式支持ND。
65	- * @param [in] groupEp: 计算输入，str。ep通信域名称，专家并行的通信域。不能和groupTp相同。
66	- * @param [in] epWorldSize: 计算输入，int。ep通信域size。
67	- * @param [in] epRankId: 计算输入，int。ep本卡Id。同一个EP通信域中各卡的epRankId不重复。
68	- * @param [in] moeExpertNum: 计算输入，int。MOE专家数量。
69	- * @param [in] expertShardType: 计算可选输入，int。共享专家卡分布类型。当前仅支持传0。
70	- * @param [in] sharedExpertNum: 计算可选输入，int。共享专家数量。
71	- * @param [in] sharedExpertRankNum: 计算可选输入，int。共享专家卡数量。
72	- * @param [in] globalBs: 计算可选输入，int。
73	- * @param [in] commQuantMode: 计算可选输入，int。通信量化类型。预留参数，暂未使用，传0即可。
74	- * @param [in] commType: 计算可选输入，int。通信方案选择。预留参数，暂未使用，传0即可。
75	- * @param [in] commAlg: 计算可选输入，str。通信算法选择。预留参数，暂未使用。
76	- * @param [out] tokenMsgSize: 计算输出，int。aclnnMoeDistributeCombineSetup接口quantExpandXOut第二维的大小。
77	- * @param [out] commCmdInfoOutSize: 计算输出，int。aclnnMoeDistributeCombineSetup接口commCmdInfoOut的大小。
78	- * @return aclnnStatus: 返回值，返回状态码
79	- *
80	- */
81	-ACLNN_API aclnnStatus aclnnMoeDistributeCombineSetupTeardownCalcOutputSize(
82	- const aclTensor* expandX, const aclTensor* expertIds, const aclTensor* assistInfoForCombine, const char* groupEp,
83	- int64_t epWorldSize, int64_t epRankId, int64_t moeExpertNum, int64_t expertShardType, int64_t sharedExpertNum,
84	- int64_t sharedExpertRankNum, int64_t globalBs, int64_t commQuantMode, int64_t commType, const char* commAlg,
85	- uint64_t& tokenMsgSize, uint64_t& commCmdInfoOutSize);
86	-
87	-/**
88	- * @brief aclnnMoeDistributeCombineSetup的第二段接口，用于执行计算。
89	- * @param [in] workspace: 在npu device侧申请的workspace内存起址。
90	- * @param [in] workspaceSize: 在npu device侧申请的workspace大小，
91	- * 由第一段接口aclnnMoeDistributeCombineSetupGetWorkspaceSize获取。
92	- * @param [in] executor: op执行器，包含了算子计算流程。
93	- * @param [in] stream: acl stream流。
94	- * @return aclnnStatus: 返回状态码。
95	- */
96	-ACLNN_API aclnnStatus aclnnMoeDistributeCombineSetup(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor,
97	- aclrtStream stream);
98	-
99	-#ifdef __cplusplus
100	-}
101	-#endif
102	-
103	-#endif // OP_API_INC_MOE_DISTRIBUTE_COMBINE_SETUP

Dmc2/moe_distribute_combine_setup/op_graph/CMakeLists.txt+0-13

	@@ -1,13 +0,0 @@
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	-if (BUILD_OPEN_PROJECT)
12	- add_graph_plugin_sources()
13	-endif()

Dmc2/moe_distribute_combine_setup/op_graph/fusion_pass/.gitkeep+0-0

The file is empty

Dmc2/moe_distribute_combine_setup/op_graph/moe_distribute_combine_setup_gen_task.cpp+0-45

	@@ -1,45 +0,0 @@
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 moe_distribute_combine_setup_gen_task.cpp
13	- * \brief
14	- */
15	-#include <vector>
16	-
17	-#include "common/utils/op_mc2.h"
18	-#include "platform/platform_info.h"
19	-
20	-#include "op_graph/mc2_gen_task_ops_utils.h"
21	-#include "op_graph/mc2_moe_gen_task_ops_utils.h"
22	-#include "register/op_impl_registry.h"
23	-#include "mc2_log.h"
24	-namespace ops {
25	-
26	-ge::Status MoeDistributeCombineSetupCalcParamFunc(gert::ExeResGenerationContext* context)
27	-{
28	- const ge::AscendString name = "aicpu kfc server";
29	- const ge::AscendString reuseKey = "kfc_stream";
30	- OPS_LOG_D(context->GetNodeName(), "Do A3 aicpu CalcParam.");
31	- return Mc2GenTaskOpsUtils::CommonKFCMc2CalcParamFunc(context, name, reuseKey);
32	-}
33	-
34	-ge::Status MoeDistributeCombineSetupGenTaskFunc(const gert::ExeResGenerationContext* context,
35	- std::vector<std::vector<uint8_t>>& tasks)
36	-{
37	- OPS_LOG_I(context->GetNodeName(), "Do A3 aicpu GenTask.");
38	- return Mc2MoeGenTaskOpsUtils::Mc2MoeGenTaskCallbackV2(context, tasks);
39	-}
40	-
41	-// new ver
42	-IMPL_OP(MoeDistributeCombineSetup)
43	- .CalcOpParam(MoeDistributeCombineSetupCalcParamFunc)
44	- .GenerateTask(MoeDistributeCombineSetupGenTaskFunc);
45	-} // namespace ops

Dmc2/moe_distribute_combine_setup/op_graph/moe_distribute_combine_setup_proto.h+0-79

	@@ -1,79 +0,0 @@
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 moe_distribute_combine_setup_proto.h
13	- * \brief 图模式原型定义
14	- */
15	-
16	-#ifndef MOE_DISTRIBUTE_COMBINE_SETUP_PROTO_H_
17	-#define MOE_DISTRIBUTE_COMBINE_SETUP_PROTO_H_
18	-
19	-#include "graph/operator_reg.h"
20	-
21	-namespace ge {
22	-
23	-/**
24	- * @brief MoeDistributeCombineSetup operator interface implementation.
25	- * @par Inputs
26	- * Three inputs, including:
27	- * @li context: A tensor. Support dtype: int32, dimension must be 1. Shape supports (2052, ), support format: ND.
28	- * @li expand_x: A tensor. Support dtype: float16, bfloat16, dimension must be 2. Shape supports (A, H), support format:
29	- * ND.
30	- * @li expert_ids: A tensor. Support dtype: int32, indicates top k experts of each token, dimension must be 2. Shape
31	- * supports (BS, K), support format: ND.
32	- * @li assist_info_for_combine: A tensor. Support dtype: int32, dimension must be 1. Shape supports (N, ), support
33	- * format: ND.
34	- *
35	- * @par Attributes
36	- * @li ep_world_size: Required. Input ep comm world size, dtype: int64.
37	- * @li ep_rank_id: Required. Input ep comm rank Id, dtype: int64.
38	- * @li moe_expert_num: Required. Input moe expert num, dtype: int64.
39	- * @li ccl_buffer_size: Required, Input ccl buffer size, Support Range: [0, MAX_INT32)，MAX_INT32 = 2^31 - 1, dtype:
40	- * Int64.
41	- * @li expert_shard_type: Input moe shard type, dtype: int64. Default: 0.
42	- * @li shared_expert_num: Input shared expert num, dtype: int64. Default: 1.
43	- * @li shared_expert_rank_num: Input shared expert rank num, dtype: int64. Default: 0.
44	- * @li comm_quant_mode: Input quant mode. The options are 0 (non-quantization), 1 (static quantization), and 2 (dynamic
45	- * quantization). dtype: int64. Default: 0.
46	- * @li global_bs: Input global batch size, dtype: int64. Default: 0.
47	- * @li comm_type: Input comm type, dtype: int64. Default: 0.
48	- * @li comm_alg: Input comm algorithm, dtype: String. Default: "".
49	- *
50	- * @par Outputs
51	- * Two outputs, including:
52	- * @li quant_expand_x: A tensor. Result of quantized expand_x. Support dtype: int8, dimension must be 2. Shape supports
53	- * (A, H), support format: ND.
54	- * @li comm_cmd_info: A tensor. Communication command info. Support dtype: int32, dimension must be 1. Shape supports
55	- * (N, ), support format: ND.
56	- */
57	-REG_OP(MoeDistributeCombineSetup)
58	- .INPUT(context, TensorType({DT_INT32}))
59	- .INPUT(expand_x, TensorType({DT_BF16, DT_FLOAT16}))
60	- .INPUT(expert_ids, TensorType({DT_INT32}))
61	- .INPUT(assist_info_for_combine, TensorType({DT_INT32}))
62	- .OUTPUT(quant_expand_x, TensorType({DT_INT8}))
63	- .OUTPUT(comm_cmd_info, TensorType({DT_INT32}))
64	- .REQUIRED_ATTR(ep_world_size, Int)
65	- .REQUIRED_ATTR(ep_rank_id, Int)
66	- .REQUIRED_ATTR(moe_expert_num, Int)
67	- .REQUIRED_ATTR(ccl_buffer_size, Int)
68	- .ATTR(expert_shard_type, Int, 0)
69	- .ATTR(shared_expert_num, Int, 1)
70	- .ATTR(shared_expert_rank_num, Int, 0)
71	- .ATTR(global_bs, Int, 0)
72	- .ATTR(comm_quant_mode, Int, 0)
73	- .ATTR(comm_type, Int, 0)
74	- .ATTR(comm_alg, String, "")
75	- .OP_END_FACTORY_REG(MoeDistributeCombineSetup)
76	-
77	-} // namespace ge
78	-
79	-#endif // MOE_DISTRIBUTE_COMBINE_SETUP_PROTO_H_

Mmc2/moe_distribute_combine_setup/op_host/CMakeLists.txt+2-4

	@@ -13,14 +13,12 @@ if (BUILD_OPEN_PROJECT) # custom
13	moe_distribute_combine_setup_def.cpp	13	moe_distribute_combine_setup_def.cpp
14	)	14	)
15	add_modules_sources_with_soc(	15	add_modules_sources_with_soc(
16	- OP_API_INDEPENDENT ON
17	- OP_API_DIR ${CMAKE_CURRENT_SOURCE_DIR}/../op_api
18	OP_MC2_ENABLE ON	16	OP_MC2_ENABLE ON
19	- OPTYPE moe_distribute_combine_setup moe_distribute_combine_setup ACLNNTYPE aclnn aclnn_inner)	17	+ OPTYPE moe_distribute_combine_setup ACLNNTYPE aclnn_inner)
20	set(SUB_MC2_COMPILE TRUE PARENT_SCOPE)	18	set(SUB_MC2_COMPILE TRUE PARENT_SCOPE)
21	set(MC2_OPT ON PARENT_SCOPE)	19	set(MC2_OPT ON PARENT_SCOPE)
22	set(moe_distribute_combine_setup_depends mc2/common mc2/3rd PARENT_SCOPE)	20	set(moe_distribute_combine_setup_depends mc2/common mc2/3rd PARENT_SCOPE)
23	else() # 回黄host	21	else() # 回黄host
24	- add_mc2_modules_sources(OPTYPE moe_distribute_combine_setup moe_distribute_combine_setup ACLNNTYPE aclnn aclnn_inner)	22	+ add_mc2_modules_sources(OPTYPE moe_distribute_combine_setup ACLNNTYPE aclnn_inner)
25	set(SUB_MC2_COMPILE TRUE PARENT_SCOPE)	23	set(SUB_MC2_COMPILE TRUE PARENT_SCOPE)
26	endif()	24	endif()

Dmc2/moe_distribute_combine_setup/tests/ut/op_api/CMakeLists.txt+0-13

	@@ -1,13 +0,0 @@
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	-if(UT_TEST_ALL OR OP_API_UT)
12	- add_modules_ut_sources(UT_NAME ${OP_API_MODULE_NAME} MODE PRIVATE DIR ${CMAKE_CURRENT_SOURCE_DIR})
13	-endif()

Dmc2/moe_distribute_combine_setup/tests/ut/op_api/test_aclnn_moe_distribute_combine_setup.cpp+0-697

	@@ -1,697 +0,0 @@
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 test_aclnn_moe_distribute_combine_setup.cpp
13	- * \brief aclnn ut
14	- */
15	-
16	-#include <float.h>
17	-#include <array>
18	-#include <vector>
19	-#include <string>
20	-#include <gtest/gtest.h>
21	-#include <gmock/gmock.h>
22	-#include "../../../op_api/aclnn_moe_distribute_combine_setup.h"
23	-#include "op_api_ut_common/tensor_desc.h"
24	-#include "op_api_ut_common/op_api_ut.h"
25	-#include "opdev/platform.h"
26	-
27	-using namespace op;
28	-
29	-namespace MoeDistributeCombineSetupUT {
30	-static aclTensor* CreateAclTensor(const std::vector<int64_t> shape, aclDataType dataType, aclFormat format)
31	-{
32	- void* storage_data = nullptr;
33	- std::vector<int64_t> strides(shape.size(), 1);
34	- for (int64_t i = shape.size() - 2; i >= 0; i--) {
35	- strides[i] = shape[i + 1] * strides[i + 1];
36	- }
37	- aclTensor* tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, format, shape.data(),
38	- shape.size(), storage_data);
39	- assert(tensor != nullptr);
40	- return tensor;
41	-}
42	-
43	-static aclTensor* CreateAclTensorOrNull(const std::vector<int64_t> shape, aclDataType dataType, aclFormat format)
44	-{
45	- if (shape.empty()) {
46	- return nullptr;
47	- }
48	- return CreateAclTensor(shape, dataType, format);
49	-}
50	-
51	-class TestAclnnMoeDistributeCombineSetup : public testing::Test {
52	-protected:
53	- static void SetUpTestCase()
54	- {
55	- op::SetPlatformNpuArch(Ops::Base::DAV_3510);
56	- std::cout << "TestAclnnMoeDistributeCombineSetup SetUp" << std::endl;
57	- }
58	- static void TearDownTestCase()
59	- {
60	- op::SetPlatformSocVersion(op::SocVersion::ASCEND910B);
61	- std::cout << "TestAclnnMoeDistributeCombineSetup TearDown" << std::endl;
62	- }
63	-};
64	-
65	-struct MoeDistributeCombineSetupAclnnTestParam {
66	- string caseName;
67	-
68	- // shape
69	- std::vector<int64_t> expandXShape;
70	- std::vector<int64_t> expertIdsShape;
71	- std::vector<int64_t> assistInfoForCombineShape;
72	- std::vector<int64_t> quantExpandXOutShape;
73	- std::vector<int64_t> commCmdInfoOutShape;
74	-
75	- // 通信域标识
76	- char* groupEp;
77	- int64_t epWorldSize;
78	- int64_t epRankId;
79	- int64_t moeExpertNum;
80	- int64_t expertShardType;
81	- int64_t sharedExpertNum;
82	- int64_t sharedExpertRankNum;
83	- int64_t globalBs;
84	- int64_t commQuantMode;
85	- int64_t commType;
86	- char* commAlg;
87	-
88	- // dtype
89	- aclDataType expandXDtype;
90	- aclDataType expertIdsDtype;
91	- aclDataType assistInfoForCombineDtype;
92	- aclDataType quantExpandXOutDtype;
93	- aclDataType commCmdInfoOutDtype;
94	-
95	- // format
96	- aclFormat expandXFormat;
97	- aclFormat expertIdsFormat;
98	- aclFormat assistInfoForCombineFormat;
99	- aclFormat quantExpandXOutFormat;
100	- aclFormat commCmdInfoOutFormat;

Mmc2/moe_distribute_combine_teardown/README.md+1-1

	@@ -1,3 +1,3 @@
1	# MoeDistributeCombineTeardown	1	# MoeDistributeCombineTeardown
2		2
3	-本目录仅包含MoeDistributeCombineTeardown算子对应的aclnn接口；如您想要贡献该算子的Ascend C实现，请参考[贡献流程](../../CONTRIBUTING.md)。	3	+本目录包含MoeDistributeCombineTeardown算子的实现，不提供公开的aclnn接口；如您想要为该算子贡献代码，请参考[贡献流程](../../CONTRIBUTING.md)。

Dmc2/moe_distribute_combine_teardown/docs/aclnnMoeDistributeCombineTeardown.md+0-476

	@@ -1,476 +0,0 @@
1	-# aclnnMoeDistributeCombineTeardown
2	-
3	-## 产品支持情况
4	-
5	-<!-- npu="950" id1 -->
6	-- <term>Ascend 950DT系列产品</term>：支持
7	-<!-- end id1 -->
8	-<!-- npu="A3" id2 -->
9	-- <term>Atlas A3系列产品</term>：不支持
10	-<!-- end id2 -->
11	-<!-- npu="910b" id3 -->
12	-- <term>Atlas A2系列产品</term>：不支持
13	-<!-- end id3 -->
14	-<!-- npu="310b" id4 -->
15	-- <term>Atlas 200I/500 A2推理产品</term>：不支持
16	-<!-- end id4 -->
17	-<!-- npu="310p" id5 -->
18	-- <term>Atlas推理系列产品</term>：不支持
19	-<!-- end id5 -->
20	-<!-- npu="910" id6 -->
21	-- <term>Atlas训练系列产品</term>：不支持
22	-<!-- end id6 -->
23	-
24	-## 功能说明
25	-
26	-- 接口功能：接收aclnnMoeDistributeCombineSetup发来的数据并整合数据（乘权重再相加）。
27	-- 计算公式：
28	-
29	- $$
30	- ataOut = AllToAllV(expandX)\\
31	- xOut = Sum(expertScales * ataOut + expertScales * sharedExpertX)
32	- $$
33	-
34	- 按MoeDistributeDispatchSetup和MoeDistributeDispatchTeardown算子收集数据的路径原路返还，本算子只做通信状态确认和数据整理，对aclnnMoeDistributeCombineSetup接口发送的数据做通信状态确认，并按照专家的scale权重做加权求和。
35	-
36	-- 注意：该接口必须与aclnnMoeDistributeDispatchSetup、aclnnMoeDistributeDispatchTeardown及aclnnMoeDistributeCombineSetup配套使用。
37	-
38	-## 函数原型
39	-
40	-每个算子分为两段式接口，必须先调用“`aclnnMoeDistributeCombineTeardownGetWorkspaceSize`”接口获取入参并根据计算流程获取计算所需workspace大小以及包含了算子计算流程的执行器，再调用“`aclnnMoeDistributeCombineTeardown`”接口执行计算。
41	-
42	-```cpp
43	-aclnnStatus aclnnMoeDistributeCombineTeardownGetWorkspaceSize(
44	- const aclTensor* expandX,
45	- const aclTensor* quantExpandX,
46	- const aclTensor* expertIds,
47	- const aclTensor* expandIdx,
48	- const aclTensor* expertScales,
49	- const aclTensor* commCmdInfo,
50	- const aclTensor* xActiveMaskOptional,
51	- const aclTensor* sharedExpertXOptional,
52	- const char* groupEp,
53	- int64_t epWorldSize,
54	- int64_t epRankId,
55	- int64_t moeExpertNum,
56	- int64_t expertShardType,
57	- int64_t sharedExpertNum,
58	- int64_t sharedExpertRankNum,
59	- int64_t globalBs,
60	- int64_t commQuantMode,
61	- int64_t commType,
62	- const char* commAlg,
63	- aclTensor* xOut,
64	- uint64_t* workspaceSize,
65	- aclOpExecutor** executor)
66	-```
67	-
68	-```cpp
69	-aclnnStatus aclnnMoeDistributeCombineTeardown(
70	- void* workspace,
71	- uint64_t workspaceSize,
72	- aclOpExecutor* executor,
73	- aclrtStream stream)
74	-```
75	-
76	-## aclnnMoeDistributeCombineTeardownGetWorkspaceSize
77	-
78	-- 参数说明
79	-
80	- <table style="undefined;table-layout: fixed; width: 1548px"><colgroup>
81	- <col style="width: 162px">
82	- <col style="width: 94px">
83	- <col style="width: 485px">
84	- <col style="width: 458px">
85	- <col style="width: 84px">
86	- <col style="width: 78px">
87	- <col style="width: 76px">
88	- <col style="width: 111px">
89	- </colgroup>
90	- <thead>
91	- <tr>
92	- <th>参数名</th>
93	- <th>输入/输出</th>
94	- <th>描述</th>
95	- <th>使用说明</th>
96	- <th>数据类型</th>
97	- <th>数据格式</th>
98	- <th>维度</th>
99	- <th>非连续Tensor</th>
100	- </tr></thead>
101	- <tbody>
102	- <tr>
103	- <td>expandX（aclTensor*）</td>
104	- <td>输入</td>
105	- <td>根据expertIds进行扩展过的token特征</td>
106	- <td>不支持空Tensor。</td>
107	- <td>FLOAT16、BFLOAT16</td>
108	- <td>ND</td>
109	- <td>(A, H)</td>
110	- <td>√</td>
111	- </tr>
112	- <tr>
113	- <td>quantExpandX（aclTensor*）</td>
114	- <td>输入</td>
115	- <td>对应aclnnMoeDistributeCombineSetup中的quantExpandX输出</td>
116	- <td>不支持空Tensor。</td>
117	- <td>INT8</td>
118	- <td>ND</td>
119	- <td>(A, tokenMsgSize)</td>
120	- <td>√</td>
121	- </tr>
122	- <tr>
123	- <td>expertIds（aclTensor*）</td>
124	- <td>输入</td>
125	- <td>每个token的topK个专家索引</td>
126	- <td>不支持空Tensor。</td>
127	- <td>INT32</td>
128	- <td>ND</td>
129	- <td>(Bs, K)</td>
130	- <td>√</td>
131	- </tr>
132	- <tr>
133	- <td>expandIdx（aclTensor*）</td>
134	- <td>输入</td>
135	- <td>对应aclnnMoeDistributeDispatchSetup中的expandIdx输出</td>
136	- <td>不支持空Tensor。</td>
137	- <td>INT32</td>
138	- <td>ND</td>
139	- <td>(BS * K, )</td>
140	- <td>√</td>
141	- </tr>
142	- <tr>
143	- <td>expertScales（aclTensor*）</td>
144	- <td>输入</td>
145	- <td>每个token的topK个专家的权重</td>
146	- <td>不支持空Tensor。</td>
147	- <td>FLOAT32</td>
148	- <td>ND</td>
149	- <td>(BS, K)</td>
150	- <td>√</td>
151	- </tr>
152	- <tr>
153	- <td>commCmdInfo（aclTensor*）</td>
154	- <td>输入</td>
155	- <td>aclnnMoeDistributeCombineSetup的输出，通信的cmd信息</td>
156	- <td>不支持空Tensor。</td>
157	- <td>INT32</td>
158	- <td>ND</td>
159	- <td>((A + epWorldSize) * 16, )</td>
160	- <td>√</td>
161	- </tr>
162	- <tr>
163	- <td>xActiveMaskOptional（aclTensor*）</td>
164	- <td>可选输入</td>
165	- <td>表示token是否参与通信</td>
166	- <td>可选择传入有效数据或填空指针，传空指针时表示所有token都参与通信。</td>
167	- <td>BOOL</td>
168	- <td>ND</td>
169	- <td>(BS, )</td>
170	- <td>√</td>
171	- </tr>
172	- <tr>
173	- <td>sharedExpertXOptional（aclTensor*）</td>
174	- <td>可选输入</td>
175	- <td>表示共享专家计算后的Token</td>
176	- <td>不支持空Tensor。要求是一个2D或3D的Tensor，当Tensor为2D时，shape为(BS,H)；当Tensor为3D时，前两位的乘积需等于BS，第三维需等于H。数据类型需跟expandX保持一致。可选择传入有效数据或填空指针，传入有效数据时，sharedExpertNum需为0。</td>
177	- <td>FLOAT16、BFLOAT16</td>
178	- <td>ND</td>
179	- <td>(BS, H)或(a, b, H)</td>
180	- <td>√</td>
181	- </tr>
182	- <tr>
183	- <td>groupEp</td>
184	- <td>输入</td>
185	- <td>EP通信域名称</td>
186	- <td>字符串长度范围为[1, 128)</td>
187	- <td>-</td>
188	- <td>-</td>
189	- <td>-</td>
190	- <td>-</td>
191	- </tr>
192	- <tr>
193	- <td>epWorldSize</td>
194	- <td>输入</td>
195	- <td>EP通信域size</td>
196	- <td>取值支持[2, 384]</td>
197	- <td>-</td>
198	- <td>-</td>
199	- <td>-</td>
200	- <td>-</td>
201	- </tr>
202	- <tr>
203	- <td>epRankId</td>
204	- <td>输入</td>
205	- <td>EP域本卡Id</td>
206	- <td>取值范围[0, epWorldSize)。<br>同一个EP通信域中各卡的epRankId不重复。</td>
207	- <td>-</td>
208	- <td>-</td>
209	- <td>-</td>
210	- <td>-</td>
211	- </tr>
212	- <tr>
213	- <td>moeExpertNum</td>
214	- <td>输入</td>
215	- <td>MoE专家数量</td>
216	- <td>取值范围(0, 512]。<br>满足moeExpertNum % (epWorldSize - sharedExpertRankNum) = 0。</td>
217	- <td>-</td>
218	- <td>-</td>
219	- <td>-</td>
220	- <td>-</td>
221	- </tr>
222	- <tr>
223	- <td>expertShardType</td>
224	- <td>输入</td>
225	- <td>共享专家卡分布类型</td>
226	- <td>当前仅支持传0，表示共享专家卡排在MoE专家卡前面。</td>
227	- <td>-</td>
228	- <td>-</td>
229	- <td>-</td>
230	- <td>-</td>
231	- </tr>
232	- <tr>
233	- <td>sharedExpertNum</td>
234	- <td>输入</td>
235	- <td>共享专家数量</td>
236	- <td>当前取值范围[0, 4]。0表示无共享专家。当前仅支持无共享专家。</td>
237	- <td>-</td>
238	- <td>-</td>
239	- <td>-</td>
240	- <td>-</td>
241	- </tr>
242	- <tr>
243	- <td>sharedExpertRankNum</td>
244	- <td>输入</td>
245	- <td>共享专家卡数量</td>
246	- <td>当前取值范围[0, epWorldSize / 2]。</td>
247	- <td>-</td>
248	- <td>-</td>
249	- <td>-</td>
250	- <td>-</td>
251	- </tr>
252	- <tr>
253	- <td>globalBs</td>
254	- <td>输入</td>
255	- <td>EP域全局的batch size大小</td>
256	- <td>当每个rank的Bs数一致场景下，globalBs = Bs * epWorldSize或globalBs = 0；当每个rank的Bs数不一致场景下，globalBs = maxBs * epWorldSize，其中maxBs表示单卡Bs最大值。</td>
257	- <td>-</td>
258	- <td>-</td>
259	- <td>-</td>
260	- <td>-</td>
261	- </tr>
262	- <tr>
263	- <td>commQuantMode</td>
264	- <td>输入</td>
265	- <td>通信量化类型</td>
266	- <td>取值范围[0, 2]，0表示通信时不进行量化，当前仅支持0。</td>
267	- <td>-</td>
268	- <td>-</td>
269	- <td>-</td>
270	- <td>-</td>
271	- </tr>
272	- <tr>
273	- <td>commType</td>
274	- <td>输入</td>
275	- <td>通信方案选择</td>
276	- <td>取值范围[0, 2]，0表示AICPU-SDMA方案，1表示CCU方案，2表示URMA方案，当前仅支持2。</td>
277	- <td>-</td>
278	- <td>-</td>
279	- <td>-</td>
280	- <td>-</td>
281	- </tr>
282	- <tr>
283	- <td>commAlg</td>
284	- <td>输入</td>
285	- <td>通信算法选择</td>
286	- <td>仅支持传入空指针或空字符串</td>
287	- <td>-</td>
288	- <td>-</td>
289	- <td>-</td>
290	- <td>-</td>
291	- </tr>
292	- <tr>
293	- <td>xOut（aclTensor*）</td>
294	- <td>输出</td>
295	- <td>表示处理后的token</td>
296	- <td>不支持空Tensor。数据类型与expandX保持一致。</td>
297	- <td>FLOAT16、BFLOAT16</td>
298	- <td>ND</td>
299	- <td>(BS, H)</td>
300	- <td>√</td>
301	- </tr>
302	- <tr>
303	- <td>workspaceSize</td>
304	- <td>输出</td>
305	- <td>返回需要在Device侧申请的workspace大小。</td>
306	- <td>-</td>
307	- <td>-</td>
308	- <td>-</td>
309	- <td>-</td>
310	- <td>-</td>
311	- </tr>
312	- <tr>
313	- <td>executor</td>
314	- <td>输出</td>
315	- <td>返回op执行器，包含了算子计算流程。</td>
316	- <td>-</td>
317	- <td>-</td>
318	- <td>-</td>
319	- <td>-</td>
320	- <td>-</td>
321	- </tr>
322	- <tr>
323	- <td>stream</td>
324	- <td>输入</td>
325	- <td>指定执行任务的stream流。</td>
326	- </tr>
327	- </tbody></table>
328	-
329	- <!-- npu="950" id7 -->
330	- - Ascend 950DT系列产品：
331	- - 不支持共享专家场景。
332	- - epWorldSize当前取值仅支持2、8。
333	- - moeExpertNum表示MoE专家数量，当前仅能传入32。
334	- - expertShardType当前仅支持传0，表示共享专家卡排在MoE专家卡前面。
335	- - sharedExpertNum表示共享专家数量，当前不支持共享专家，仅能传入0。
336	- - sharedExpertRankNum表示共享专家卡数，当前不支持共享专家，仅能传入0。
337	- - commQuantMode当前仅支持传入0，表示不进行量化。
338	- - commType取值范围[0, 2]，当前仅支持2，表示URMA通路。
339	- - commAlg当前版本不支持，传空指针即可。
340	-
341	- <!-- end id7 -->
342	- - Atlas A3训练系列产品/Atlas A3推理系列产品：
343	- - 不支持共享专家场景。
344	- - epWorldSize当前取值仅支持2、8。
345	- - moeExpertNum表示MoE专家数量，当前仅能传入32。
346	- - expertShardType当前仅支持传0，表示共享专家卡排在MoE专家卡前面。
347	- - sharedExpertNum表示共享专家数量，当前不支持共享专家，仅能传入0。
348	- - sharedExpertRankNum表示共享专家卡数，当前不支持共享专家，仅能传入0。
349	- - commQuantMode当前仅支持传入0，表示不进行量化。
350	- - commType取值范围[0, 2]，当前仅支持2，表示URMA通路。
351	- - commAlg当前版本不支持，传空指针即可。
352	-
353	-- 返回值
354	-
355	- aclnnStatus：返回状态码，具体参见[aclnn返回码](../../../docs/zh/context/aclnn_return_code.md)。
356	-
357	- 第一段接口完成入参校验，出现以下场景时报错：
358	-
359	- <table style="undefined;table-layout: fixed; width: 1000px"><colgroup>
360	- <col style="width: 300px">
361	- <col style="width: 150px">
362	- <col style="width: 550px">
363	- </colgroup>
364	- <thead>
365	- <tr>
366	- <th>返回值</th>
367	- <th>错误码</th>
368	- <th>描述</th>
369	- </tr></thead>
370	- <tbody>
371	- <tr>
372	- <td>ACLNN_ERR_PARAM_NULLPTR </td>
373	- <td>161001</td>
374	- <td>输入和输出的必选参数Tensor是空指针。</td>
375	- </tr>
376	- <tr>
377	- <td>ACLNN_ERR_PARAM_INVALID </td>
378	- <td>161002</td>
379	- <td>输入和输出的数据类型不在支持的范围内。</td>
380	- </tr>
381	- <tr>
382	- <td rowspan="2">ACLNN_ERR_INNER_TILING_ERROR</td>
383	- <td rowspan="2">561002</td>
384	- <td>输入和输出的shape不在支持的范围内。</td>
385	- </tr>
386	- <tr>
387	- <td>参数的取值不在支持的范围内。</td>
388	- </tr>
389	- </tbody>
390	- </table>
391	-
392	-## aclnnMoeDistributeCombineTeardown
393	-
394	-- 参数说明：
395	-
396	- <table style="undefined;table-layout: fixed; width: 1000px"><colgroup>
397	- <col style="width: 200px">
398	- <col style="width: 130px">
399	- <col style="width: 670px">
400	- </colgroup>
401	- <thead>
402	- <tr>
403	- <th>参数名</th>
404	- <th>输入/输出</th>
405	- <th>描述</th>
406	- </tr></thead>
407	- <tbody>
408	- <tr>
409	- <td>workspace</td>
410	- <td>输入</td>
411	- <td>在Device侧申请的workspace内存地址。</td>
412	- </tr>
413	- <tr>
414	- <td>workspaceSize</td>
415	- <td>输入</td>
416	- <td>在Device侧申请的workspace大小，由第一段接口aclnnMoeDistributeCombineTeardownGetWorkspaceSize获取。</td>
417	- </tr>
418	- <tr>
419	- <td rowspan="2">executor</td>
420	- <td rowspan="2">输入</td>
421	- <td>op执行器，包含了算子计算流程。</td>
422	- </tr>
423	- <tr>
424	- <td>指定执行任务的stream流。</td>
425	- </tr>
426	- </tbody>
427	- </table>
428	-
429	-- 返回值：
430	-
431	- 返回aclnnStatus状态码，具体参见[aclnn返回码](../../../docs/zh/context/aclnn_return_code.md)。
432	-
433	-## 约束说明
434	-
435	-- 确定性计算：
436	- - aclnnMoeDistributeCombineTeardown默认确定性实现。
437	-- aclnnMoeDistributeDispatchSetup接口，aclnnMoeDistributeDispatchTeardown接口，aclnnMoeDistributeCombineSetup接口，aclnnMoeDistributeCombineTeardown接口必须配套使用。
438	-- 调用接口过程中使用的`groupEp`、`epWorldSize`、`moeExpertNum`、`expertShardType`、`sharedExpertNum`、`sharedExpertRankNum`、`globalBs`、`commQuantMode`、`commType`、`commAlg`参数取值所有卡需保持一致，`groupEp`、`epWorldSize`、`expertShardType`、`sharedExpertNum`、`sharedExpertRankNum`、`globalBs`、`commQuantMode`、`commType`、`commAlg`参数取值在网络中不同层中也需保持一致，且和aclnnMoeDistributeDispatchSetup接口，aclnnMoeDistributeDispatchTeardown接口，aclnnMoeDistributeCombineSetup接口对应参数也保持一致。
439	-- 参数说明里shape格式说明：
440	-
441	- - A：表示本卡需要分发的最大token数量，取值范围如下：
442	- - 对于共享专家，要满足A = BS \* epWorldSize \* sharedExpertNum / sharedExpertRankNum；当globalBs非0时，要满足A = globalBs\* sharedExpertNum / sharedExpertRankNum。
443	- - 对于MoE专家，当globalBs为0时，要满足A >= BS \* epWorldSize \* min(localExpertNum, K)；当globalBs非0时，要满足A >= globalBs \* min(localExpertNum, K)。
444	- - H：表示hidden size隐藏层大小。取值为[1024, 8192]。当前仅支持4096、7168。
445	- - HS：表示通信时的数据大小，Align512( Align32(H) + Align8(H) / 8 sizeof(float) )，其中Align512(x) = ((x + 512 - 1) / 512) \ 512，Align32(x) = ((x + 32 - 1) / 32) \* 32，Align8(x) = ((x + 8 - 1) / 8) \* 8。
446	- - BS：表示batch sequence size，即本卡最终输出的token数量。取值范围为0 < BS ≤ 512。当前仅支持8、16、256。
447	- - K：表示选取topK个专家，取值范围为0 < K ≤ 16同时满足0 < K ≤ moeExpertNum。当前仅支持6、8。
448	- - localExpertNum：表示本卡专家数量。
449	- - 对于共享专家卡，localExpertNum = 1
450	- - 对于MoE专家卡，localExpertNum = moeExpertNum / (epWorldSize - sharedExpertRankNum)。moeExpertNum当前仅支持32。
451	- - tokenMsgSize：表示每个token在数据通信时的维度信息，计算公式为Align512(Align32(H)+Align8(H)/8\sizeof(float))，其中AlignN(x)=((x+N-1)/NN)。
452	- - 当前不支持共享专家。sharedExpertNum和sharedExpertRankNum当前仅支持0。
453	-- HCCL_BUFFSIZE：
454	- 调用本接口前需检查HCCL_BUFFSIZE环境变量取值是否合理，该环境变量表示单个通信域占用内存大小，单位MB，不配置时默认为200MB。
455	-
456	- <!-- npu="950" id8 -->
457	- - Ascend 950DT系列产品：
458	- - 要求 >= 2且满足>= 4 \* (localExpertNum \* maxBs \* epWorldSize \* Align512(Align32(2 \* H) + 44) + (K + sharedExpertNum) \* maxBs \* Align512(2 \* H))，localExpertNum需使用MoE专家卡的本卡专家数，其中Align512(x) = ((x + 512 - 1) / 512) \* 512，Align32(x) = ((x + 32 - 1) / 32) \* 32
459	- <!-- end id8 -->
460	- - Atlas A3训练系列产品/Atlas A3推理系列产品：
461	- - 要求 >= 2且满足>= 2 \* (localExpertNum \* maxBs \* epWorldSize \* Align512(Align32(2 \* H) + 44) + (K + sharedExpertNum) \* maxBs \* Align512(2 \* H))，localExpertNum需使用MoE专家卡的本卡专家数，其中Align512(x) = ((x + 512 - 1) / 512) \* 512，Align32(x) = ((x + 32 - 1) / 32) \* 32
462	-- 通信域使用约束：
463	- - 一个模型中的aclnnMoeDistributeDispatchSetup接口，aclnnMoeDistributeDispatchTeardown接口，aclnnMoeDistributeCombineSetup接口，aclnnMoeDistributeCombineTeardown接口仅支持相同EP通信域，且该通信域中不允许有其他算子。
464	-- 通信方式约束：
465	-
466	- <!-- npu="950" id9 -->
467	- - <term>Ascend 950DT系列产品</term>：仅支持URMA通信。
468	-
469	- <!-- end id9 -->
470	-
471	-## 调用示例
472	-
473	-<!-- npu="950" id10 -->
474	-- Ascend 950DT系列产品：aclnnMoeDistributeCombineTeardown必须在aclnnMoeDistributeCombineSetup调用之后调用，请参考[aclnnMoeDistributeCombineSetup](../../moe_distribute_combine_setup/docs/aclnnMoeDistributeCombineSetup.md)中的调用示例。
475	-
476	-<!-- end id10 -->

Dmc2/moe_distribute_combine_teardown/examples/test_aclnn_moe_distribute_combine_teardown.cpp+0-356

	@@ -1,356 +0,0 @@
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 test_aclnn_moe_distribute_combine_teardown.cpp
13	- * \brief aclnn直调测试脚本
14	- */
15	-
16	-#include <thread>
17	-#include <iostream>
18	-#include <string>
19	-#include <vector>
20	-#include <cstring>
21	-#include "acl/acl.h"
22	-#include "hccl/hccl.h"
23	-#include "aclnnop/aclnn_moe_distribute_combine_teardown.h"
24	-#include "aclnnop/aclnn_moe_distribute_combine_setup.h"
25	-
26	-#define CHECK_RET(cond, return_expr) \
27	- do { \
28	- if (!(cond)) { \
29	- return_expr; \
30	- } \
31	- } while (0)
32	-
33	-#define LOG_PRINT(message, ...) \
34	- do { \
35	- printf(message, ##__VA_ARGS__); \
36	- } while (0)
37	-
38	-constexpr int DEV_NUM = 2;
39	-
40	-template <typename Func>
41	-class Guard {
42	-public:
43	- explicit Guard(Func& func)
44	- : func_(func)
45	- {}
46	- ~Guard()
47	- {
48	- func_();
49	- }
50	-
51	-private:
52	- Func& func_;
53	-};
54	-
55	-int64_t AlignN(int64_t x, int64_t n)
56	-{
57	- return (x + n - 1) / n * n;
58	-}
59	-
60	-int64_t GetShapeSize(const std::vector<int64_t>& shape)
61	-{
62	- int64_t shape_size = 1;
63	- for (auto i : shape) {
64	- shape_size *= i;
65	- }
66	- return shape_size;
67	-}
68	-
69	-struct Args {
70	- uint32_t rankId;
71	- uint32_t epRankId;
72	- HcclComm hcclEpComm;
73	- aclrtStream combinesetupstream;
74	- aclrtStream combineteardownstream;
75	- aclrtContext context;
76	-};
77	-
78	-template <typename T>
79	-int CreateAclTensor(const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr,
80	- aclDataType dataType, aclTensor** tensor)
81	-{
82	- auto size = GetShapeSize(shape) * sizeof(T);
83	- auto ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
84	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclrtMalloc failed. ret: %d\n", ret); return ret);
85	-
86	- ret = aclrtMemcpy(*deviceAddr, size, hostData.data(), size, ACL_MEMCPY_HOST_TO_DEVICE);
87	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclrtMemcpy failed. ret: %d\n", ret); return ret);
88	- std::vector<int64_t> strides(shape.size(), 1);
89	- for (int64_t i = shape.size() - 2; i >= 0; i--) {
90	- strides[i] = shape[i + 1] * strides[i + 1];
91	- }
92	- *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND,
93	- shape.data(), shape.size(), *deviceAddr);
94	- return 0;
95	-}
96	-
97	-void DestroyTensor(aclTensor* tensor)
98	-{
99	- if (tensor != nullptr) {
100	- aclDestroyTensor(tensor);
101	- }
102	-}
103	-
104	-void FreeDeviceAddr(void* deviceAddr)
105	-{
106	- if (deviceAddr != nullptr) {
107	- aclrtFree(deviceAddr);
108	- }
109	-}
110	-
111	-int LaunchOneProcess(Args& args)
112	-{
113	- int ret = aclrtSetCurrentContext(args.context);
114	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclrtSetCurrentContext failed. ret: %d\n", ret); return ret);
115	-
116	- char hcomEpName[128] = {0};
117	- ret = HcclGetCommName(args.hcclEpComm, hcomEpName);
118	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] HcclGetCommName failed. ret: %d\n", ret); return ret);
119	-
120	- auto destroyFunc = [&args]() {
121	- std::cout << "== begin to destroy " << std::endl;
122	- HcclCommDestroy(args.hcclEpComm);
123	- aclrtDestroyStream(args.combinesetupstream);
124	- aclrtDestroyStream(args.combineteardownstream);
125	- aclrtDestroyContext(args.context);
126	- aclrtResetDevice(args.rankId);
127	- };
128	- auto guard = Guard<decltype(destroyFunc)>(destroyFunc);
129	-
130	- // 场景参数
131	- int64_t bs = 16;
132	- int64_t h = 4096;
133	- int64_t k = 6;
134	- int64_t expertSharedType = 0;
135	- int64_t sharedExpertNum = 0;
136	- int64_t sharedExpertRankNum = 0;
137	- int64_t moeExpertNum = 32;
138	- int64_t commQuantMode = 0;
139	- int64_t tpWorldSize = 1;
140	- int64_t epWorldSize = DEV_NUM;
141	- int64_t commType = 2;
142	- int64_t timeOut = 100000000;
143	-
144	- int64_t globalBS = bs * epWorldSize;
145	- int64_t localExpertNum;
146	- int64_t localToken;
147	-
148	- if (args.epRankId < sharedExpertRankNum) {
149	- localExpertNum = 1;
150	- localToken = globalBS / sharedExpertRankNum;
151	- } else {
152	- localExpertNum = moeExpertNum / (epWorldSize - sharedExpertRankNum);
153	- localToken = globalBS * (localExpertNum < k ? localExpertNum : k);
154	- }
155	-
156	- int64_t assistInfoForCombineOutSize = localToken * 128;
157	-
158	- // 定义 shape
159	- std::vector<int64_t> expandXShape{tpWorldSize * localToken, h};
160	- std::vector<int64_t> expertIdsShape{bs, k};
161	- std::vector<int64_t> expertScalesShape{bs, k};
162	- std::vector<int64_t> expandIdxShape{bs * k};
163	- std::vector<int64_t> assistInfoForCombineOutShape{assistInfoForCombineOutSize};
164	- std::vector<int64_t> quantExpandXOutShape{tpWorldSize * localToken,
165	- AlignN(AlignN(h, 32) + AlignN(h, 8) / 8 * sizeof(float), 512)};
166	- std::vector<int64_t> commCmdInfoOutShapeforcombine{(localToken + epWorldSize) * 16};
167	- std::vector<int64_t> xOutShape{bs, h};
168	-
169	- int64_t expandXShapeSize = GetShapeSize(expandXShape);
170	- int64_t expertIdsShapeSize = GetShapeSize(expertIdsShape);
171	- int64_t expertScalesShapeSize = GetShapeSize(expertScalesShape);
172	- int64_t expandIdxShapeSize = GetShapeSize(expandIdxShape);
173	- int64_t assistInfoForCombineOutShapeSize = GetShapeSize(assistInfoForCombineOutShape);
174	- int64_t quantExpandXOutShapeSize = GetShapeSize(quantExpandXOutShape);
175	- int64_t commCmdInfoOutShapeSizeforcombine = GetShapeSize(commCmdInfoOutShapeforcombine);
176	- int64_t xOutShapeSize = GetShapeSize(xOutShape);
177	-
178	- // 构造 host 数据
179	- std::vector<int16_t> expandXHostData(expandXShapeSize, 0);
180	- std::vector<int32_t> expertIdsHostData;
181	- for (int32_t token_id = 0; token_id < expertIdsShape[0]; token_id++) {
182	- for (int32_t k_id = 0; k_id < expertIdsShape[1]; k_id++) {
183	- expertIdsHostData.push_back(k_id);
184	- }
185	- }
186	- std::vector<float> expertScalesHostData(expertScalesShapeSize, 0);
187	- std::vector<int32_t> expandIdxHostData(expandIdxShapeSize, 0);
188	- std::vector<int32_t> assistInfoForCombineOutHostData(assistInfoForCombineOutShapeSize, 0);
189	- std::vector<int8_t> quantExpandXOutHostData(quantExpandXOutShapeSize, 0);
190	- std::vector<int32_t> commCmdInfoOutforCombineHostData(commCmdInfoOutShapeSizeforcombine, 0);
191	- std::vector<float> xOutHostData(xOutShapeSize, 0);
192	-
193	- // 声明 device 地址和 tensor
194	- void* expandXDeviceAddr = nullptr;
195	- void* expertIdsDeviceAddr = nullptr;
196	- void* expertScalesDeviceAddr = nullptr;
197	- void* expandIdxDeviceAddr = nullptr;
198	- void* assistInfoForCombineOutDeviceAddr = nullptr;
199	- void* quantExpandXOutDeviceAddr = nullptr;
200	- void* commCmdInfoOutforCombineDeviceAddr = nullptr;
201	- void* xOutDeviceAddr = nullptr;
202	-
203	- aclTensor* expandX = nullptr;
204	- aclTensor* expertIds = nullptr;
205	- aclTensor* expertScales = nullptr;
206	- aclTensor* expandIdx = nullptr;
207	- aclTensor* assistInfoForCombineOut = nullptr;
208	- aclTensor* quantExpandXOut = nullptr;
209	- aclTensor* commCmdInfoOutforCombine = nullptr;
210	- aclTensor* xOut = nullptr;
211	-
212	- // 创建 tensor
213	- ret = CreateAclTensor(expandXHostData, expandXShape, &expandXDeviceAddr, aclDataType::ACL_FLOAT16, &expandX);
214	- CHECK_RET(ret == ACL_SUCCESS, return ret);
215	- ret = CreateAclTensor(expertIdsHostData, expertIdsShape, &expertIdsDeviceAddr, aclDataType::ACL_INT32, &expertIds);
216	- CHECK_RET(ret == ACL_SUCCESS, return ret);
217	- ret = CreateAclTensor(expertScalesHostData, expertScalesShape, &expertScalesDeviceAddr, aclDataType::ACL_FLOAT,
218	- &expertScales);
219	- CHECK_RET(ret == ACL_SUCCESS, return ret);
220	- ret = CreateAclTensor(expandIdxHostData, expandIdxShape, &expandIdxDeviceAddr, aclDataType::ACL_INT32, &expandIdx);
221	- CHECK_RET(ret == ACL_SUCCESS, return ret);
222	- ret = CreateAclTensor(assistInfoForCombineOutHostData, assistInfoForCombineOutShape,
223	- &assistInfoForCombineOutDeviceAddr, aclDataType::ACL_INT32, &assistInfoForCombineOut);
224	- CHECK_RET(ret == ACL_SUCCESS, return ret);
225	- ret = CreateAclTensor(quantExpandXOutHostData, quantExpandXOutShape, &quantExpandXOutDeviceAddr,
226	- aclDataType::ACL_INT8, &quantExpandXOut);
227	- CHECK_RET(ret == ACL_SUCCESS, return ret);
228	- ret = CreateAclTensor(commCmdInfoOutforCombineHostData, commCmdInfoOutShapeforcombine,
229	- &commCmdInfoOutforCombineDeviceAddr, aclDataType::ACL_INT32, &commCmdInfoOutforCombine);
230	- CHECK_RET(ret == ACL_SUCCESS, return ret);
231	- ret = CreateAclTensor(xOutHostData, xOutShape, &xOutDeviceAddr, aclDataType::ACL_FLOAT16, &xOut);
232	- CHECK_RET(ret == ACL_SUCCESS, return ret);
233	-
234	- // 声明算子执行必需变量
235	- uint64_t workspaceSize = 0;
236	- aclOpExecutor* executor = nullptr;
237	- void* workspaceAddr = nullptr;
238	-
239	- /****************************调用combine_setup******************************************/
240	- ret = aclnnMoeDistributeCombineSetupGetWorkspaceSize(
241	- expandX, expertIds, assistInfoForCombineOut, hcomEpName, epWorldSize, args.epRankId, moeExpertNum,
242	- expertSharedType, sharedExpertNum, sharedExpertRankNum, globalBS, commQuantMode, commType, nullptr,
243	- quantExpandXOut, commCmdInfoOutforCombine, &workspaceSize, &executor);
244	- CHECK_RET(ret == ACL_SUCCESS,
245	- LOG_PRINT("[ERROR] aclnnMoeDistributeCombineSetupGetWorkspaceSize failed. ret = %d \n", ret);
246	- return ret);
247	-
248	- if (workspaceSize > 0) {
249	- ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
250	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] CombineSetup aclrtMalloc failed. ret = %d\n", ret);
251	- return ret);
252	- }
253	-
254	- ret = aclnnMoeDistributeCombineSetup(workspaceAddr, workspaceSize, executor, args.combinesetupstream);
255	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclnnMoeDistributeCombineSetup failed. ret = %d \n", ret);
256	- return ret);
257	-
258	- ret = aclrtSynchronizeStreamWithTimeout(args.combinesetupstream, timeOut);
259	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclrtSynchronizeStreamWithTimeout failed. ret = %d \n", ret);
260	- return ret);
261	- LOG_PRINT("[INFO] device_%d aclnnMoeDistributeCombineSetup execute successfully.\n", args.rankId);
262	-
263	- /****************************调用combine_teardown******************************************/
264	- ret = aclnnMoeDistributeCombineTeardownGetWorkspaceSize(
265	- expandX, quantExpandXOut, expertIds, expandIdx, expertScales, commCmdInfoOutforCombine, nullptr, nullptr,
266	- hcomEpName, epWorldSize, args.epRankId, moeExpertNum, expertSharedType, sharedExpertNum, sharedExpertRankNum,
267	- globalBS, commQuantMode, commType, nullptr, xOut, &workspaceSize, &executor);
268	- CHECK_RET(ret == ACL_SUCCESS,
269	- LOG_PRINT("[ERROR] aclnnMoeDistributeCombineTeardownGetWorkspaceSize failed. ret = %d \n", ret);
270	- return ret);
271	-
272	- if (workspaceSize > 0) {
273	- ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
274	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] CombineTeardown aclrtMalloc failed. ret = %d\n", ret);
275	- return ret);
276	- }
277	-
278	- ret = aclnnMoeDistributeCombineTeardown(workspaceAddr, workspaceSize, executor, args.combineteardownstream);
279	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclnnMoeDistributeCombineTeardown failed. ret = %d \n", ret);
280	- return ret);
281	-
282	- ret = aclrtSynchronizeStreamWithTimeout(args.combineteardownstream, timeOut);
283	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclrtSynchronizeStreamWithTimeout failed. ret = %d \n", ret);
284	- return ret);
285	- LOG_PRINT("[INFO] device_%d aclnnMoeDistributeCombineTeardown execute successfully.\n", args.rankId);
286	-
287	- // 释放资源
288	- if (workspaceSize > 0) {
289	- aclrtFree(workspaceAddr);
290	- }
291	- DestroyTensor(expandX);
292	- DestroyTensor(expertIds);
293	- DestroyTensor(expertScales);
294	- DestroyTensor(expandIdx);
295	- DestroyTensor(assistInfoForCombineOut);
296	- DestroyTensor(quantExpandXOut);
297	- DestroyTensor(commCmdInfoOutforCombine);
298	- DestroyTensor(xOut);
299	-
300	- FreeDeviceAddr(expandXDeviceAddr);
301	- FreeDeviceAddr(expertIdsDeviceAddr);
302	- FreeDeviceAddr(expertScalesDeviceAddr);
303	- FreeDeviceAddr(expandIdxDeviceAddr);
304	- FreeDeviceAddr(assistInfoForCombineOutDeviceAddr);
305	- FreeDeviceAddr(quantExpandXOutDeviceAddr);
306	- FreeDeviceAddr(commCmdInfoOutforCombineDeviceAddr);
307	- FreeDeviceAddr(xOutDeviceAddr);
308	-
309	- return 0;
310	-}
311	-
312	-int main(int argc, char* argv[])
313	-{
314	- int ret = aclInit(nullptr);
315	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclInit failed. ret = %d \n", ret); return ret);
316	-
317	- aclrtStream combineSetupStream[DEV_NUM];
318	- aclrtStream combineTeardownStream[DEV_NUM];
319	- aclrtContext context[DEV_NUM];
320	- for (uint32_t rankId = 0; rankId < DEV_NUM; rankId++) {
321	- ret = aclrtSetDevice(rankId);
322	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclrtSetDevice failed. ret = %d \n", ret); return ret);
323	- ret = aclrtCreateContext(&context[rankId], rankId);
324	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclrtCreateContext failed. ret = %d \n", ret); return ret);
325	- ret = aclrtCreateStream(&combineSetupStream[rankId]);
326	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclrtCreateStream failed. ret = %d \n", ret); return ret);
327	- ret = aclrtCreateStream(&combineTeardownStream[rankId]);
328	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] aclrtCreateStream failed. ret = %d \n", ret); return ret);
329	- }
330	-
331	- int32_t devices[DEV_NUM];
332	- for (int i = 0; i < DEV_NUM; i++) {
333	- devices[i] = i;
334	- }
335	-
336	- HcclComm comms[DEV_NUM];
337	- ret = HcclCommInitAll(DEV_NUM, devices, comms);
338	- CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] HcclCommInitAll failed. ret = %d \n", ret); return ret);
339	-
340	- Args args[DEV_NUM];
341	- std::vector<std::unique_ptr<std::thread>> threads(DEV_NUM);
342	- for (uint32_t rankId = 0; rankId < DEV_NUM; rankId++) {
343	- args[rankId].rankId = rankId;
344	- args[rankId].epRankId = rankId;
345	- args[rankId].hcclEpComm = comms[rankId];
346	- args[rankId].combinesetupstream = combineSetupStream[rankId];
347	- args[rankId].combineteardownstream = combineTeardownStream[rankId];
348	- args[rankId].context = context[rankId];
349	- threads[rankId].reset(new (std::nothrow) std::thread(&LaunchOneProcess, std::ref(args[rankId])));
350	- }
351	- for (uint32_t rankId = 0; rankId < DEV_NUM; rankId++) {
352	- threads[rankId]->join();
353	- }
354	- aclFinalize();
355	- return 0;
356	-}

Dmc2/moe_distribute_combine_teardown/op_api/aclnn_moe_distribute_combine_teardown.cpp+0-135

	@@ -1,135 +0,0 @@
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 aclnn_moe_distribute_combine_teardown.cpp
13	- * \brief
14	- */
15	-#include "aclnn_moe_distribute_combine_teardown.h"
16	-#include <algorithm>
17	-#include "common/utils/op_mc2.h"
18	-#include "common/utils/matmul_util.h"
19	-#include "common/utils/op_mc2_def.h"
20	-#include "aclnn_kernels/common/op_error_check.h"
21	-#include "opdev/op_log.h"
22	-#include "mc2_log_compat.h"
23	-#include "opdev/common_types.h"
24	-#include "opdev/platform.h"
25	-#include "op_host/util/op_const_def.h"
26	-#include "common/op_api/mc2_context.h"
27	-
28	-namespace {
29	-
30	-using namespace Mc2Aclnn;
31	-using namespace op;
32	-
33	-enum class NnopbaseHcclServerType : uint32_t {
34	- NNOPBASE_HCCL_SERVER_TYPE_AICPU = 0,
35	- NNOPBASE_HCCL_SERVER_TYPE_MTE,
36	- NNOPBASE_HCCL_SERVER_TYPE_CCU,
37	- NNOPBASE_HCCL_SERVER_TYPE_END
38	-};
39	-
40	-static bool CheckNotNull(const aclTensor* expandX, const aclTensor* quantExpandX, const aclTensor* expertIds,
41	- const aclTensor* expandIdx, const aclTensor* expertScales, const aclTensor* commCmdInfo,
42	- const char* groupEp, aclTensor* xOut)
43	-{
44	- OP_LOGD("aclnn_moe_distribute_combine_teardown CheckNotNull start");
45	- OP_CHECK_NULL(expandX, return false);
46	- OP_CHECK_NULL(quantExpandX, return false);
47	- OP_CHECK_NULL(expertIds, return false);
48	- OP_CHECK_NULL(expandIdx, return false);
49	- OP_CHECK_NULL(expertScales, return false);
50	- OP_CHECK_NULL(commCmdInfo, return false);
51	- OP_CHECK_NULL(xOut, return false);
52	- if ((groupEp == nullptr) \|\| (strnlen(groupEp, HCCL_GROUP_NAME_MAX) == 0)) {
53	- OP_LOGE_WITH_INVALID_INPUT("aclnnMoeDistributeCombineTeardown", "groupEp");
54	- return false;
55	- }
56	- OP_LOGD("aclnn_moe_distribute_combine_teardown CheckNotNull success");
57	- return true;
58	-}
59	-
60	-static aclnnStatus CheckParams(const aclTensor* expandX, const aclTensor* quantExpandX, const aclTensor* expertIds,
61	- const aclTensor* expandIdx, const aclTensor* expertScales, const aclTensor* commCmdInfo,
62	- const char* groupEp, aclTensor* xOut)
63	-{
64	- OP_LOGD("aclnn_moe_distribute_combine_teardown checkparams start");
65	- CHECK_RET(CheckNotNull(expandX, quantExpandX, expertIds, expandIdx, expertScales, commCmdInfo, groupEp, xOut),
66	- ACLNN_ERR_PARAM_NULLPTR);
67	- if (strnlen(groupEp, HCCL_GROUP_NAME_MAX) >= HCCL_GROUP_NAME_MAX) {
68	- OP_LOGE_FOR_INVALID_VALUE_WITH_REASON(
69	- "aclnnMoeDistributeCombineTeardown", "groupEp",
70	- std::to_string(strnlen(groupEp, HCCL_GROUP_NAME_MAX)).c_str(),
71	- std::string("groupEp name exceeds ").append(std::to_string(HCCL_GROUP_NAME_MAX)).c_str());
72	- return ACLNN_ERR_PARAM_INVALID;
73	- }
74	- OP_LOGD("aclnn_moe_distribute_combine_teardown checkparams success");
75	- return ACLNN_SUCCESS;
76	-}
77	-
78	-} // namespace
79	-
80	-extern "C" aclnnStatus aclnnInnerMoeDistributeCombineTeardownGetWorkspaceSize(
81	- aclTensor* mc2Context, const aclTensor* expandX, const aclTensor* quantExpandX, const aclTensor* expertIds,
82	- const aclTensor* expandIdx, const aclTensor* expertScales, const aclTensor* commCmdInfo,
83	- const aclTensor* xActiveMaskOptional, const aclTensor* sharedExpertXOptional, int64_t epWorldSize, int64_t epRankId,
84	- int64_t moeExpertNum, int64_t cclBuffSize, int64_t expertShardType, int64_t sharedExpertNum,
85	- int64_t sharedExpertRankNum, int64_t globalBs, int64_t commQuantMode, int64_t commType, const char* commAlg,
86	- aclTensor* xOut, uint64_t* workspaceSize, aclOpExecutor** executor);
87	-
88	-extern "C" aclnnStatus aclnnInnerMoeDistributeCombineTeardown(void* workspace, uint64_t workspaceSize,
89	- aclOpExecutor* executor, aclrtStream stream);
90	-extern "C" void __attribute__((weak)) NnopbaseSetHcclServerType(void* executor, NnopbaseHcclServerType sType);
91	-
92	-extern "C" aclnnStatus aclnnMoeDistributeCombineTeardownGetWorkspaceSize(
93	- const aclTensor* expandX, const aclTensor* quantExpandX, const aclTensor* expertIds, const aclTensor* expandIdx,
94	- const aclTensor* expertScales, const aclTensor* commCmdInfo, const aclTensor* xActiveMaskOptional,
95	- const aclTensor* sharedExpertXOptional, const char* groupEp, int64_t epWorldSize, int64_t epRankId,
96	- int64_t moeExpertNum, int64_t expertShardType, int64_t sharedExpertNum, int64_t sharedExpertRankNum,
97	- int64_t globalBs, int64_t commQuantMode, int64_t commType, const char* commAlg, aclTensor* xOut,
98	- uint64_t* workspaceSize, aclOpExecutor** executor)
99	-{
100	- OP_LOGD("aclnn_moe_distribute_combine_teardown get_workspace_size start");
101	- if (GetCurrentPlatformInfo().GetCurNpuArch() != Ops::Base::DAV_3510) {
102	- OP_LOGE_LIBOPAPI_REPORT("aclnnMoeDistributeCombineTeardown", "Unsupported npuArch");
103	- return ACLNN_ERR_PARAM_INVALID;
104	- }
105	-
106	- auto ret_param = CheckParams(expandX, quantExpandX, expertIds, expandIdx, expertScales, commCmdInfo, groupEp, xOut);
107	- CHECK_RET(ret_param == ACLNN_SUCCESS, ret_param);
108	-
109	- aclnnStatus retStatus;
110	- uint64_t hcclBuffSize = 0;
111	- aclTensor* mc2Context = nullptr;
112	- const char* opName = "moe_distribute_a2av_setup_teardown";
113	- retStatus = Mc2Aclnn::Mc2Context::GetMc2ContextTensor(groupEp, opName, hcclBuffSize, mc2Context,
114	- CommProtocol::COMM_PROTOCOL_UBC_CTP);
115	- CHECK_RET(retStatus == ACLNN_SUCCESS, retStatus);
116	- retStatus = aclnnInnerMoeDistributeCombineTeardownGetWorkspaceSize(
117	- mc2Context, expandX, quantExpandX, expertIds, expandIdx, expertScales, commCmdInfo, xActiveMaskOptional,
118	- sharedExpertXOptional, epWorldSize, epRankId, moeExpertNum, hcclBuffSize, expertShardType, sharedExpertNum,
119	- sharedExpertRankNum, globalBs, commQuantMode, commType, const_cast<char*>(commAlg), xOut, workspaceSize,
120	- executor);
121	- OP_LOGD("aclnn_moe_distribute_combine_teardown get_workspace_size success");
122	- return retStatus;
123	-}
124	-
125	-extern "C" aclnnStatus aclnnMoeDistributeCombineTeardown(void* workspace, uint64_t workspaceSize,
126	- aclOpExecutor* executor, aclrtStream stream)
127	-{
128	- OP_LOGD("aclnnMoeDistributeCombineTeardown start");
129	- if (NnopbaseSetHcclServerType) {
130	- NnopbaseSetHcclServerType(executor, NnopbaseHcclServerType::NNOPBASE_HCCL_SERVER_TYPE_MTE);
131	- }
132	- aclnnStatus ret = aclnnInnerMoeDistributeCombineTeardown(workspace, workspaceSize, executor, stream);
133	- OP_LOGD("aclnnMoeDistributeCombineTeardown success");
134	- return ret;
135	-}

Dmc2/moe_distribute_combine_teardown/op_api/aclnn_moe_distribute_combine_teardown.h+0-82

	@@ -1,82 +0,0 @@
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 aclnn_moe_distribute_combine_teardown.h
13	- * \brief
14	- */
15	-#ifndef OP_API_INC_MOE_DISTRIBUTE_COMBINE_TEARDOWN_
16	-#define OP_API_INC_MOE_DISTRIBUTE_COMBINE_TEARDOWN_
17	-
18	-#include <string>
19	-
20	-#include "aclnn/aclnn_base.h"
21	-#include "aclnn_util.h"
22	-
23	-#ifdef __cplusplus
24	-extern "C" {
25	-#endif
26	-
27	-/**
28	- * 算子功能：实现MoeDistributeCombineTeardown功能
29	- * @brief aclnnMoeDistributeCombineTeardown的第一段接口，根据具体的计算流程，计算workspace大小。
30	- * @domain aclnn_ops_infer
31	- * @param [in] expandX: 计算输入，Tensor，数据类型float16, bfloat16，必须为2维，数据格式支持ND。
32	- * @param [in] quantExpandX: 计算输入，Tensor，数据类型int8，必须为2维，数据格式支持ND。
33	- * @param [in] expertIds: 计算输入，Tensor，数据类型int32，必须为2维，数据格式支持ND。
34	- * @param [in] expandIdx: 计算输入，Tensor，数据类型int32，必须为2维，数据格式支持ND。
35	- * @param [in] expertScales: 计算输入，Tensor，数据类型float32，必须为2维，数据格式支持ND。
36	- * @param [in] commCmdInfo:
37	- * 计算输入，Tensor，数据类型int32，必须为1维，数据格式支持ND，aclnnMoeDistributeCombineSetUp的输出。
38	- * @param [in] xActiveMaskOptional:
39	- * 计算输入，Tensor，数据类型bool，必须为1维，数据格式支持ND。预留参数，暂未使用，传空即可。
40	- * @param [in] sharedExpertXOptional:
41	- * 计算输入，Tensor，数据类型与expandX保持一致，必须为2维，数据格式支持ND。预留参数，暂未使用，传空即可。
42	- * @param [in] groupEp: 计算输入，str。ep通信域名称，专家并行的通信域。不能和groupTp相同。
43	- * @param [in] epWorldSize: 计算输入，int。ep通信域size。
44	- * @param [in] epRankId: 计算输入，int。ep本卡Id。同一个EP通信域中各卡的epRankId不能重复。
45	- * @param [in] moeExpertNum: 计算输入，int。MOE专家数量。
46	- * @param [in] expertShardType: 计算可选输入，int。共享专家卡分布类型。当前仅支持传0。
47	- * @param [in] sharedExpertNum: 计算可选输入，int。共享专家数量。当前仅支持传0。
48	- * @param [in] sharedExpertRankNum: 计算可选输入，int。共享专家卡数量。当前仅支持传0。
49	- * @param [in] globalBs: 计算可选输入，int。
50	- * @param [in] commQuantMode: 计算可选输入，int。通信量化类型。预留参数，暂未使用，传0即可。
51	- * @param [in] commType: 计算可选输入，int。通信方案选择。预留参数，暂未使用，传0即可。
52	- * @param [in] commAlg: 计算可选输入，str。通信算法类型。预留参数，暂未使用。
53	- * @param [out] xOut: 计算输出，Tensor，必选输出，数据类型支持float16，bfloat16，仅支持2维，数据格式支持ND。
54	- * @param [out] workspaceSize: 出参，返回需要在npu device侧申请的workspace大小。
55	- * @param [out] executor: 出参，返回op执行器，包含了算子计算流程。
56	- * @return aclnnStatus: 返回值，返回状态码。
57	- */
58	-ACLNN_API aclnnStatus aclnnMoeDistributeCombineTeardownGetWorkspaceSize(
59	- const aclTensor* expandX, const aclTensor* quantExpandX, const aclTensor* expertIds, const aclTensor* expandIdx,
60	- const aclTensor* expertScales, const aclTensor* commCmdInfo, const aclTensor* xActiveMaskOptional,
61	- const aclTensor* sharedExpertXOptional, const char* groupEp, int64_t epWorldSize, int64_t epRankId,
62	- int64_t moeExpertNum, int64_t expertShardType, int64_t sharedExpertNum, int64_t sharedExpertRankNum,
63	- int64_t globalBs, int64_t commQuantMode, int64_t commType, const char* commAlg, aclTensor* xOut,
64	- uint64_t* workspaceSize, aclOpExecutor** executor);
65	-
66	-/**
67	- * @brief aclnnMoeDistributeCombineTeardown的第二段接口，用于执行计算。
68	- * @param [in] workspace: 在npu device侧申请的workspace内存地址。
69	- * @param [in] workspace_size: 在npu device侧申请的workspace大小，
70	- * 由第一段接口aclnnMoeDistributeCombineTeardownGetWorkspaceSize获取。
71	- * @param [in] executor: op执行器，包含了算子计算流程。
72	- * @param [in] stream: acl stream流。
73	- * @return aclnnStatus: 返回状态码。
74	- */
75	-ACLNN_API aclnnStatus aclnnMoeDistributeCombineTeardown(void* workspace, uint64_t workspaceSize,
76	- aclOpExecutor* executor, aclrtStream stream);
77	-
78	-#ifdef __cplusplus
79	-}
80	-#endif
81	-
82	-#endif // OP_API_INC_MOE_DISTRIBUTE_COMBINE_TEARDOWN_

Dmc2/moe_distribute_combine_teardown/op_graph/CMakeLists.txt+0-13

	@@ -1,13 +0,0 @@
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	-if (BUILD_OPEN_PROJECT)
12	- add_graph_plugin_sources()
13	-endif()

Dmc2/moe_distribute_combine_teardown/op_graph/fusion_pass/.gitkeep+0-0

The file is empty

Dmc2/moe_distribute_combine_teardown/op_graph/moe_distribute_combine_teardown_proto.h+0-86

	@@ -1,86 +0,0 @@
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 moe_distribute_combine_teardown_proto.h
13	- * \brief 图模式原型定义
14	- */
15	-#ifndef MOE_DISTRIBUTE_COMBINE_TEARDOWN_PROTO_H
16	-#define MOE_DISTRIBUTE_COMBINE_TEARDOWN_PROTO_H
17	-
18	-#include <graph/operator_reg.h>
19	-
20	-namespace ge {
21	-/**
22	- * @brief MoeDistributeCombineTeardown operator interface implementation.
23	-
24	- * @par Inputs
25	- * Eight inputs, including:
26	- * @li context: A tensor. Support dtype: int32, dimension must be 1. Shape supports (2052, ), support format: ND.
27	- * @li expand_x: A tensor. Support dtype: float16, bfloat16, dimension must be 2, Support Shape: (A, H),
28	- * support format: ND.
29	- * @li quant_expand_x: A tensor. Support dtype: int8, dimension must be 2, Support Shape: (A, tokenMsgSize), support
30	- * format: ND.
31	- * @li expert_ids: A tensor. Support dtype: int32, dimension must be 2, Support Shape: (BS, K), support format: ND.
32	- * @li expand_idx: A tensor. Support dtype: int32, dimension must be 1, Support Shape: (BS * K, ), support format: ND.
33	- * @li expert_scales: A tensor. Support dtype: float32, dimension must be 2, Support Shape: (BS, K),
34	- * support format: ND.
35	- * @li comm_cmd_info: A tensor. Support dtype: int32, dimension must be 1, Call the
36	- * `aclnnMoeDistributeCombineSetupTeardownCalcOutputSize` API to obtain the Shape, support format: ND.
37	- * @li x_active_mask: An optional tensor. Support dtype: bool, dimension must be 1, Support Shape: (BS, ) support
38	- * format: ND.
39	- * @li shared_expert_x: An optional tensor. Dtype should be same with expand_x, dimension must be 2 or 3,
40	- * Support Shape: (BS, H) or (a, b, H), where a * b = BS, support format: ND.
41	-
42	- * @par Attributes
43	- * @li ep_world_size: Input ep comm world size, dtype: Int64.
44	- * @li ep_rank_id: Input ep comm rank Id, dtype: Int64.
45	- * @li moe_expert_num: Input moe expert num, dtype: Int64.
46	- * @li ccl_buffer_size: Required, Input ccl buffer size, Support Range: [0, MAX_INT32)，MAX_INT32 = 2^31 - 1,
47	- * dtype: Int64.
48	- * @li expert_shard_type: Input moe shard type, dtype: Int64. Default: 0.
49	- * @li shared_expert_num: Input shared expert num, dtype: Int64. Default: 1.
50	- * @li shared_expert_rank_num: Input shared expert rank num, dtype: Int64. Default: 0.
51	- * @li global_bs: Input global batch size, dtype: Int64. Default: 0.
52	- * @li comm_quant_mode: communication quantization mode, 1 for enable, 0 for disable, dtype: Int64. Default: 0.
53	- * @li comm_type: type of communication, dtype: Int64. Default: 0.
54	- * @li comm_alg: Input comm alg type, dtype: String. Default: "".
55	-
56	- * @par Outputs
57	- * One outputs, including:
58	- * @li x_out: A tensor. Result of combine. Dtype should be same with expand_x, Support Shape: (BS, H), support format:
59	- * ND.
60	- */
61	-REG_OP(MoeDistributeCombineTeardown)
62	- .INPUT(context, TensorType({DT_INT32}))
63	- .INPUT(expand_x, TensorType({DT_FLOAT16, DT_BF16}))
64	- .INPUT(quant_expand_x, TensorType({DT_INT8}))
65	- .INPUT(expert_ids, TensorType({DT_INT32}))
66	- .INPUT(expand_idx, TensorType({DT_INT32}))
67	- .INPUT(expert_scales, TensorType({DT_FLOAT}))
68	- .INPUT(comm_cmd_info, TensorType({DT_INT32}))
69	- .OPTIONAL_INPUT(x_active_mask, TensorType({DT_BOOL}))
70	- .OPTIONAL_INPUT(shared_expert_x, TensorType({DT_FLOAT16, DT_BF16}))
71	- .OUTPUT(x_out, TensorType({DT_FLOAT16, DT_BF16}))
72	- .REQUIRED_ATTR(ep_world_size, Int)
73	- .REQUIRED_ATTR(ep_rank_id, Int)
74	- .REQUIRED_ATTR(moe_expert_num, Int)
75	- .REQUIRED_ATTR(ccl_buffer_size, Int)
76	- .ATTR(expert_shard_type, Int, 0)
77	- .ATTR(shared_expert_num, Int, 1)
78	- .ATTR(shared_expert_rank_num, Int, 0)
79	- .ATTR(global_bs, Int, 0)
80	- .ATTR(comm_quant_mode, Int, 0)
81	- .ATTR(comm_type, Int, 0)
82	- .ATTR(comm_alg, String, "")
83	- .OP_END_FACTORY_REG(MoeDistributeCombineTeardown)
84	-} // namespace ge
85	-
86	-#endif // MOE_DISTRIBUTE_COMBINE_TEARDOWN_PROTO_H

Mmc2/moe_distribute_combine_teardown/op_host/CMakeLists.txt+2-4

	@@ -13,14 +13,12 @@ if (BUILD_OPEN_PROJECT) # custom
13	moe_distribute_combine_teardown_def.cpp	13	moe_distribute_combine_teardown_def.cpp
14	)	14	)
15	add_modules_sources_with_soc(	15	add_modules_sources_with_soc(
16	- OP_API_INDEPENDENT ON
17	- OP_API_DIR ${CMAKE_CURRENT_SOURCE_DIR}/../op_api
18	OP_MC2_ENABLE ON	16	OP_MC2_ENABLE ON
19	- OPTYPE moe_distribute_combine_teardown moe_distribute_combine_teardown ACLNNTYPE aclnn aclnn_inner)	17	+ OPTYPE moe_distribute_combine_teardown ACLNNTYPE aclnn_inner)
20	set(SUB_MC2_COMPILE TRUE PARENT_SCOPE)	18	set(SUB_MC2_COMPILE TRUE PARENT_SCOPE)
21	set(MC2_OPT ON PARENT_SCOPE)	19	set(MC2_OPT ON PARENT_SCOPE)
22	set(moe_distribute_combine_teardown_depends mc2/common mc2/3rd mc2/moe_distribute_combine_setup PARENT_SCOPE)	20	set(moe_distribute_combine_teardown_depends mc2/common mc2/3rd mc2/moe_distribute_combine_setup PARENT_SCOPE)
23	else() # 回黄host	21	else() # 回黄host
24	- add_mc2_modules_sources(OPTYPE moe_distribute_combine_teardown moe_distribute_combine_teardown ACLNNTYPE aclnn aclnn_inner)	22	+ add_mc2_modules_sources(OPTYPE moe_distribute_combine_teardown ACLNNTYPE aclnn_inner)
25	set(SUB_MC2_COMPILE TRUE PARENT_SCOPE)	23	set(SUB_MC2_COMPILE TRUE PARENT_SCOPE)
26	endif()	24	endif()

Dmc2/moe_distribute_combine_teardown/tests/ut/op_api/CMakeLists.txt+0-13

	@@ -1,13 +0,0 @@
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	-if(UT_TEST_ALL OR OP_API_UT)
12	- add_modules_ut_sources(UT_NAME ${OP_API_MODULE_NAME} MODE PRIVATE DIR ${CMAKE_CURRENT_SOURCE_DIR})
13	-endif()

Dmc2/moe_distribute_combine_teardown/tests/ut/op_api/test_aclnn_moe_distribute_combine_teardown.cpp+0-516

Mmc2/moe_distribute_dispatch_setup/README.md+1-1

Dmc2/moe_distribute_dispatch_setup/docs/aclnnMoeDistributeDispatchSetup.md+0-1047

Dmc2/moe_distribute_dispatch_setup/examples/test_aclnn_moe_distribute_dispatch_setup.cpp+0-337

Dmc2/moe_distribute_dispatch_setup/op_api/aclnn_moe_distribute_dispatch_setup.cpp+0-196

Dmc2/moe_distribute_dispatch_setup/op_api/aclnn_moe_distribute_dispatch_setup.h+0-115

Dmc2/moe_distribute_dispatch_setup/op_graph/CMakeLists.txt+0-13

Dmc2/moe_distribute_dispatch_setup/op_graph/fusion_pass/.gitkeep+0-0

Dmc2/moe_distribute_dispatch_setup/op_graph/moe_distribute_dispatch_setup_graph_infer.cpp+0-53

Dmc2/moe_distribute_dispatch_setup/op_graph/moe_distribute_dispatch_setup_proto.h+0-88

Mmc2/moe_distribute_dispatch_setup/op_host/CMakeLists.txt+2-4

Dmc2/moe_distribute_dispatch_setup/tests/ut/op_api/CMakeLists.txt+0-13

Dmc2/moe_distribute_dispatch_setup/tests/ut/op_api/test_aclnn_moe_distribute_dispatch_setup.cpp+0-581

Mmc2/moe_distribute_dispatch_teardown/README.md+1-1

Dmc2/moe_distribute_dispatch_teardown/docs/aclnnMoeDistributeDispatchTeardown.md+0-491

Dmc2/moe_distribute_dispatch_teardown/examples/test_aclnn_moe_distribute_dispatch_teardown.cpp+0-339

Dmc2/moe_distribute_dispatch_teardown/op_api/aclnn_moe_distribute_dispatch_teardown.cpp+0-130

Dmc2/moe_distribute_dispatch_teardown/op_api/aclnn_moe_distribute_dispatch_teardown.h+0-80

Dmc2/moe_distribute_dispatch_teardown/op_graph/CMakeLists.txt+0-13

Dmc2/moe_distribute_dispatch_teardown/op_graph/fusion_pass/.gitkeep+0-0

Dmc2/moe_distribute_dispatch_teardown/op_graph/moe_distribute_dispatch_teardown_graph_infer.cpp+0-55

Dmc2/moe_distribute_dispatch_teardown/op_graph/moe_distribute_dispatch_teardown_proto.h+0-96

Mmc2/moe_distribute_dispatch_teardown/op_host/CMakeLists.txt+2-4

Dmc2/moe_distribute_dispatch_teardown/tests/ut/op_api/CMakeLists.txt+0-13

Dmc2/moe_distribute_dispatch_teardown/tests/ut/op_api/test_aclnn_moe_distribute_dispatch_teardown.cpp+0-14

Mtests/test_config.yaml+0-10File content is under review, please try again later

	@@ -755,14 +755,10 @@ mc2_moe_advanced@ops-transformer:
755	- ops/ops-transformer/mc2/moe_distribute_combine_add_rms_norm/CMakeLists.txt	755	- ops/ops-transformer/mc2/moe_distribute_combine_add_rms_norm/CMakeLists.txt
756	- ops/ops-transformer/mc2/moe_distribute_combine_add_rms_norm/README.md	756	- ops/ops-transformer/mc2/moe_distribute_combine_add_rms_norm/README.md
757	- ops/ops-transformer/mc2/moe_distribute_combine_setup/op_host/op_tiling/arch35/	757	- ops/ops-transformer/mc2/moe_distribute_combine_setup/op_host/op_tiling/arch35/
758	- - ops/ops-transformer/mc2/moe_distribute_combine_setup/op_api/
759	- - ops/ops-transformer/mc2/moe_distribute_combine_setup/op_graph/
760	- ops/ops-transformer/mc2/moe_distribute_combine_setup/op_kernel/arch35/	758	- ops/ops-transformer/mc2/moe_distribute_combine_setup/op_kernel/arch35/
761	- ops/ops-transformer/mc2/moe_distribute_combine_setup/README.md	759	- ops/ops-transformer/mc2/moe_distribute_combine_setup/README.md
762	- ops/ops-transformer/mc2/moe_distribute_combine_setup/CMakeLists.txt	760	- ops/ops-transformer/mc2/moe_distribute_combine_setup/CMakeLists.txt
763	- ops/ops-transformer/mc2/moe_distribute_combine_teardown/op_host/op_tiling/arch35/	761	- ops/ops-transformer/mc2/moe_distribute_combine_teardown/op_host/op_tiling/arch35/
764	- - ops/ops-transformer/mc2/moe_distribute_combine_teardown/op_api/
765	- - ops/ops-transformer/mc2/moe_distribute_combine_teardown/op_graph/
766	- ops/ops-transformer/mc2/moe_distribute_combine_teardown/op_kernel/arch35/	762	- ops/ops-transformer/mc2/moe_distribute_combine_teardown/op_kernel/arch35/
767	- ops/ops-transformer/mc2/moe_distribute_combine_teardown/CMakeLists.txt	763	- ops/ops-transformer/mc2/moe_distribute_combine_teardown/CMakeLists.txt
768	- ops/ops-transformer/mc2/moe_distribute_combine_teardown/README.md	764	- ops/ops-transformer/mc2/moe_distribute_combine_teardown/README.md
	@@ -809,14 +805,10 @@ mc2_moe_advanced@ops-transformer:
809	- ops/ops-transformer/mc2/moe_distribute_dispatch/CMakeLists.txt	805	- ops/ops-transformer/mc2/moe_distribute_dispatch/CMakeLists.txt
810	- ops/ops-transformer/mc2/moe_distribute_dispatch/README.md	806	- ops/ops-transformer/mc2/moe_distribute_dispatch/README.md
811	- ops/ops-transformer/mc2/moe_distribute_dispatch_setup/op_host/op_tiling/arch35/	807	- ops/ops-transformer/mc2/moe_distribute_dispatch_setup/op_host/op_tiling/arch35/
812	- - ops/ops-transformer/mc2/moe_distribute_dispatch_setup/op_api/
813	- - ops/ops-transformer/mc2/moe_distribute_dispatch_setup/op_graph/
814	- ops/ops-transformer/mc2/moe_distribute_dispatch_setup/op_kernel/arch35/	808	- ops/ops-transformer/mc2/moe_distribute_dispatch_setup/op_kernel/arch35/
815	- ops/ops-transformer/mc2/moe_distribute_dispatch_setup/CMakeLists.txt	809	- ops/ops-transformer/mc2/moe_distribute_dispatch_setup/CMakeLists.txt
816	- ops/ops-transformer/mc2/moe_distribute_dispatch_setup/README.md	810	- ops/ops-transformer/mc2/moe_distribute_dispatch_setup/README.md
817	- ops/ops-transformer/mc2/moe_distribute_dispatch_teardown/op_host/op_tiling/arch35/	811	- ops/ops-transformer/mc2/moe_distribute_dispatch_teardown/op_host/op_tiling/arch35/
818	- - ops/ops-transformer/mc2/moe_distribute_dispatch_teardown/op_api/
819	- - ops/ops-transformer/mc2/moe_distribute_dispatch_teardown/op_graph/
820	- ops/ops-transformer/mc2/moe_distribute_dispatch_teardown/op_kernel/arch35/	812	- ops/ops-transformer/mc2/moe_distribute_dispatch_teardown/op_kernel/arch35/
821	- ops/ops-transformer/mc2/moe_distribute_dispatch_teardown/CMakeLists.txt	813	- ops/ops-transformer/mc2/moe_distribute_dispatch_teardown/CMakeLists.txt
822	- ops/ops-transformer/mc2/moe_distribute_dispatch_teardown/README.md	814	- ops/ops-transformer/mc2/moe_distribute_dispatch_teardown/README.md
	@@ -897,10 +889,7 @@ mc2_moe_advanced@ops-transformer:
897	- ops/ops-transformer/mc2/moe_distribute_combine_add_rms_norm/examples/	889	- ops/ops-transformer/mc2/moe_distribute_combine_add_rms_norm/examples/
898	- ops/ops-transformer/mc2/moe_distribute_combine_add_rms_norm/docs/	890	- ops/ops-transformer/mc2/moe_distribute_combine_add_rms_norm/docs/
899	- ops/ops-transformer/mc2/moe_distribute_combine_setup/tests/	891	- ops/ops-transformer/mc2/moe_distribute_combine_setup/tests/
		Z Z zhong-zixinSep 23 后面删除了combine_setup的tests，这里还需要保留吗 jiangxiuhan Sep 23 Comments： Sep 23 Comments：
900	- - ops/ops-transformer/mc2/moe_distribute_combine_setup/examples/
901	- - ops/ops-transformer/mc2/moe_distribute_combine_setup/docs/
902	- ops/ops-transformer/mc2/moe_distribute_combine_teardown/tests/	892	- ops/ops-transformer/mc2/moe_distribute_combine_teardown/tests/
903	- - ops/ops-transformer/mc2/moe_distribute_combine_teardown/docs/
904	- ops/ops-transformer/mc2/moe_distribute_combine_v2/tests/	893	- ops/ops-transformer/mc2/moe_distribute_combine_v2/tests/
905	- ops/ops-transformer/mc2/moe_distribute_combine_v2/examples/	894	- ops/ops-transformer/mc2/moe_distribute_combine_v2/examples/
906	- ops/ops-transformer/mc2/moe_distribute_combine_v2/docs/	895	- ops/ops-transformer/mc2/moe_distribute_combine_v2/docs/
	@@ -909,10 +898,7 @@ mc2_moe_advanced@ops-transformer:
909	- ops/ops-transformer/mc2/moe_distribute_dispatch/examples/	898	- ops/ops-transformer/mc2/moe_distribute_dispatch/examples/
910	- ops/ops-transformer/mc2/moe_distribute_dispatch/docs/	899	- ops/ops-transformer/mc2/moe_distribute_dispatch/docs/
911	- ops/ops-transformer/mc2/moe_distribute_dispatch_setup/tests/	900	- ops/ops-transformer/mc2/moe_distribute_dispatch_setup/tests/
912	- - ops/ops-transformer/mc2/moe_distribute_dispatch_setup/examples/
913	- - ops/ops-transformer/mc2/moe_distribute_dispatch_setup/docs/
914	- ops/ops-transformer/mc2/moe_distribute_dispatch_teardown/tests/	901	- ops/ops-transformer/mc2/moe_distribute_dispatch_teardown/tests/
915	- - ops/ops-transformer/mc2/moe_distribute_dispatch_teardown/docs/
916	- ops/ops-transformer/mc2/moe_distribute_dispatch_v2/tests/	902	- ops/ops-transformer/mc2/moe_distribute_dispatch_v2/tests/
917	- ops/ops-transformer/mc2/moe_distribute_dispatch_v2/examples/	903	- ops/ops-transformer/mc2/moe_distribute_dispatch_v2/examples/
918	- ops/ops-transformer/mc2/moe_distribute_dispatch_v2/docs/	904	- ops/ops-transformer/mc2/moe_distribute_dispatch_v2/docs/

	@@ -162,14 +162,10 @@
162	- [aclnnMoeDistributeCombine](../../mc2/moe_distribute_combine/docs/aclnnMoeDistributeCombine.md)	162	- [aclnnMoeDistributeCombine](../../mc2/moe_distribute_combine/docs/aclnnMoeDistributeCombine.md)
163	- [aclnnMoeDistributeCombineAddRmsNorm](../../mc2/moe_distribute_combine_add_rms_norm/docs/aclnnMoeDistributeCombineAddRmsNorm.md)	163	- [aclnnMoeDistributeCombineAddRmsNorm](../../mc2/moe_distribute_combine_add_rms_norm/docs/aclnnMoeDistributeCombineAddRmsNorm.md)
164	- [aclnnMoeDistributeCombineAddRmsNormV2](../../mc2/moe_distribute_combine_add_rms_norm/docs/aclnnMoeDistributeCombineAddRmsNormV2.md)	164	- [aclnnMoeDistributeCombineAddRmsNormV2](../../mc2/moe_distribute_combine_add_rms_norm/docs/aclnnMoeDistributeCombineAddRmsNormV2.md)
165	-- [aclnnMoeDistributeCombineSetup](../../mc2/moe_distribute_combine_setup/docs/aclnnMoeDistributeCombineSetup.md)
166	-- [aclnnMoeDistributeCombineTeardown](../../mc2/moe_distribute_combine_teardown/docs/aclnnMoeDistributeCombineTeardown.md)
167	- [aclnnMoeDistributeCombineV2](../../mc2/moe_distribute_combine_v2/docs/aclnnMoeDistributeCombineV2.md)	165	- [aclnnMoeDistributeCombineV2](../../mc2/moe_distribute_combine_v2/docs/aclnnMoeDistributeCombineV2.md)
168	- [aclnnMoeDistributeCombineV3](../../mc2/moe_distribute_combine_v2/docs/aclnnMoeDistributeCombineV3.md)	166	- [aclnnMoeDistributeCombineV3](../../mc2/moe_distribute_combine_v2/docs/aclnnMoeDistributeCombineV3.md)
169	- [aclnnMoeDistributeCombineV4](../../mc2/moe_distribute_combine_v2/docs/aclnnMoeDistributeCombineV4.md)	167	- [aclnnMoeDistributeCombineV4](../../mc2/moe_distribute_combine_v2/docs/aclnnMoeDistributeCombineV4.md)
170	- [aclnnMoeDistributeDispatch](../../mc2/moe_distribute_dispatch/docs/aclnnMoeDistributeDispatch.md)	168	- [aclnnMoeDistributeDispatch](../../mc2/moe_distribute_dispatch/docs/aclnnMoeDistributeDispatch.md)
171	-- [aclnnMoeDistributeDispatchSetup](../../mc2/moe_distribute_dispatch_setup/docs/aclnnMoeDistributeDispatchSetup.md)
172	-- [aclnnMoeDistributeDispatchTeardown](../../mc2/moe_distribute_dispatch_teardown/docs/aclnnMoeDistributeDispatchTeardown.md)
173	- [aclnnMoeDistributeDispatchV2](../../mc2/moe_distribute_dispatch_v2/docs/aclnnMoeDistributeDispatchV2.md)	169	- [aclnnMoeDistributeDispatchV2](../../mc2/moe_distribute_dispatch_v2/docs/aclnnMoeDistributeDispatchV2.md)
174	- [aclnnMoeDistributeDispatchV3](../../mc2/moe_distribute_dispatch_v2/docs/aclnnMoeDistributeDispatchV3.md)	170	- [aclnnMoeDistributeDispatchV3](../../mc2/moe_distribute_dispatch_v2/docs/aclnnMoeDistributeDispatchV3.md)
175	- [aclnnMoeDistributeDispatchV4](../../mc2/moe_distribute_dispatch_v2/docs/aclnnMoeDistributeDispatchV4.md)	171	- [aclnnMoeDistributeDispatchV4](../../mc2/moe_distribute_dispatch_v2/docs/aclnnMoeDistributeDispatchV4.md)

	@@ -158,14 +158,10 @@
158	\|[aclnnMoeDistributeCombineV4](../../mc2/moe_distribute_combine_v2/docs/aclnnMoeDistributeCombineV4.md)\|当存在TP域通信时，先进行ReduceScatterV通信，再进行AllToAllV通信，最后将接收的数据整合；当不存在TP域通信时，进行AllToAllV通信，最后将接收的数据整合。expertScales非空时乘权重再相加，为空Tensor时直接相加。\|默认确定性实现\|默认确定性实现\|默认确定性实现\|	158	\|[aclnnMoeDistributeCombineV4](../../mc2/moe_distribute_combine_v2/docs/aclnnMoeDistributeCombineV4.md)\|当存在TP域通信时，先进行ReduceScatterV通信，再进行AllToAllV通信，最后将接收的数据整合；当不存在TP域通信时，进行AllToAllV通信，最后将接收的数据整合。expertScales非空时乘权重再相加，为空Tensor时直接相加。\|默认确定性实现\|默认确定性实现\|默认确定性实现\|
159	\|[aclnnMoeDistributeCombineAddRmsNorm](../../mc2/moe_distribute_combine_add_rms_norm/docs/aclnnMoeDistributeCombineAddRmsNorm.md)\|当存在TP域通信时，先进行ReduceScatterV通信，再进行AlltoAllV通信，最后将接收的数据整合；当不存在TP域通信时，进行AlltoAllV通信，最后将接收的数据整合，之后完成Add + RmsNorm融合。\|-\|默认确定性实现\|默认确定性实现\|	159	\|[aclnnMoeDistributeCombineAddRmsNorm](../../mc2/moe_distribute_combine_add_rms_norm/docs/aclnnMoeDistributeCombineAddRmsNorm.md)\|当存在TP域通信时，先进行ReduceScatterV通信，再进行AlltoAllV通信，最后将接收的数据整合；当不存在TP域通信时，进行AlltoAllV通信，最后将接收的数据整合，之后完成Add + RmsNorm融合。\|-\|默认确定性实现\|默认确定性实现\|
160	\|[aclnnMoeDistributeCombineAddRmsNormV2](../../mc2/moe_distribute_combine_add_rms_norm/docs/aclnnMoeDistributeCombineAddRmsNormV2.md)\|当存在TP域通信时，先进行ReduceScatterV通信，再进行AlltoAllV通信，最后将接收的数据整合；当不存在TP域通信时，进行AlltoAllV通信，最后将接收的数据整合，之后完成Add + RmsNorm融合。\|-\|默认确定性实现\|默认确定性实现\|	160	\|[aclnnMoeDistributeCombineAddRmsNormV2](../../mc2/moe_distribute_combine_add_rms_norm/docs/aclnnMoeDistributeCombineAddRmsNormV2.md)\|当存在TP域通信时，先进行ReduceScatterV通信，再进行AlltoAllV通信，最后将接收的数据整合；当不存在TP域通信时，进行AlltoAllV通信，最后将接收的数据整合，之后完成Add + RmsNorm融合。\|-\|默认确定性实现\|默认确定性实现\|
161	-\|[aclnnMoeDistributeCombineSetup](../../mc2/moe_distribute_combine_setup/docs/aclnnMoeDistributeCombineSetup.md)\|进行AlltoAllV通信，将数据写入对端GM。\|-\|-\|默认确定性实现\|
162	-\|[aclnnMoeDistributeCombineTeardown](../../mc2/moe_distribute_combine_teardown/docs/aclnnMoeDistributeCombineTeardown.md)\|接收aclnnMoeDistributeCombineSetup发来的数据并整合数据（乘权重再相加）。\|-\|-\|默认确定性实现\|
163	\|[aclnnMoeDistributeDispatch](../../mc2/moe_distribute_dispatch/docs/aclnnMoeDistributeDispatch.md)\|对Token数据进行量化，当存在TP域通信时，先进行EP域的AllToAllV通信，再进行TP域的AllGatherV通信；当不存在TP域通信时，进行EP域的AllToAllV通信。\|默认确定性实现\|默认确定性实现\|默认确定性实现\|	161	\|[aclnnMoeDistributeDispatch](../../mc2/moe_distribute_dispatch/docs/aclnnMoeDistributeDispatch.md)\|对Token数据进行量化，当存在TP域通信时，先进行EP域的AllToAllV通信，再进行TP域的AllGatherV通信；当不存在TP域通信时，进行EP域的AllToAllV通信。\|默认确定性实现\|默认确定性实现\|默认确定性实现\|
164	\|[aclnnMoeDistributeDispatchV2](../../mc2/moe_distribute_dispatch_v2/docs/aclnnMoeDistributeDispatchV2.md)\|对token数据进行量化，当存在TP域通信时，先进行EP域的AllToAllV通信，再进行TP域的AllGatherV通信；当不存在TP域通信时，进行EP域的AllToAllV通信。\|默认确定性实现\|默认确定性实现\|默认确定性实现\|	162	\|[aclnnMoeDistributeDispatchV2](../../mc2/moe_distribute_dispatch_v2/docs/aclnnMoeDistributeDispatchV2.md)\|对token数据进行量化，当存在TP域通信时，先进行EP域的AllToAllV通信，再进行TP域的AllGatherV通信；当不存在TP域通信时，进行EP域的AllToAllV通信。\|默认确定性实现\|默认确定性实现\|默认确定性实现\|
165	\|[aclnnMoeDistributeDispatchV3](../../mc2/moe_distribute_dispatch_v2/docs/aclnnMoeDistributeDispatchV3.md)\|对token数据进行量化，当存在TP域通信时，先进行EP域的AllToAllV通信，再进行TP域的AllGatherV通信；当不存在TP域通信时，进行EP域的AllToAllV通信。\|默认确定性实现\|默认确定性实现\|默认确定性实现\|	163	\|[aclnnMoeDistributeDispatchV3](../../mc2/moe_distribute_dispatch_v2/docs/aclnnMoeDistributeDispatchV3.md)\|对token数据进行量化，当存在TP域通信时，先进行EP域的AllToAllV通信，再进行TP域的AllGatherV通信；当不存在TP域通信时，进行EP域的AllToAllV通信。\|默认确定性实现\|默认确定性实现\|默认确定性实现\|
166	\|[aclnnMoeDistributeDispatchV4](../../mc2/moe_distribute_dispatch_v2/docs/aclnnMoeDistributeDispatchV4.md)\|对token数据进行量化，当存在TP域通信时，先进行EP域的AllToAllV通信，再进行TP域的AllGatherV通信；当不存在TP域通信时，进行EP域的AllToAllV通信。\|默认确定性实现\|默认确定性实现\|默认确定性实现\|	164	\|[aclnnMoeDistributeDispatchV4](../../mc2/moe_distribute_dispatch_v2/docs/aclnnMoeDistributeDispatchV4.md)\|对token数据进行量化，当存在TP域通信时，先进行EP域的AllToAllV通信，再进行TP域的AllGatherV通信；当不存在TP域通信时，进行EP域的AllToAllV通信。\|默认确定性实现\|默认确定性实现\|默认确定性实现\|
167	-\|[aclnnMoeDistributeDispatchSetup](../../mc2/moe_distribute_dispatch_setup/docs/aclnnMoeDistributeDispatchSetup.md)\|对Token数据进行量化（可选），根据token选择的topK专家在EP（Expert Parallelism）域的AllToAllV通信，只进行数据发送和通信状态发送，通信指令发出后算子即刻退出，无需等待通信完成。数据的接收和后处理由aclnnMoeDistributeDispatchTeardown接口完成。\|-\|-\|默认确定性实现\|
168	-\|[aclnnMoeDistributeDispatchTeardown](../../mc2/moe_distribute_dispatch_teardown/docs/aclnnMoeDistributeDispatchTeardown.md)\|接收MoE层EP（Expert Parallelism）域的AllToAllV通信发过来的数据，数据发送端由aclnnMoeDistributeDispatchSetup完成，本接口内完成通信状态确认和数据整理。\|-\|-\|默认确定性实现\|
169	\|[aclnnMoeFinalizeRouting](../../moe/moe_finalize_routing/docs/aclnnMoeFinalizeRouting.md)\|MoE计算中，最后处理合并MoE FFN的输出结果。\|默认确定性实现\|默认确定性实现\|-\|	165	\|[aclnnMoeFinalizeRouting](../../moe/moe_finalize_routing/docs/aclnnMoeFinalizeRouting.md)\|MoE计算中，最后处理合并MoE FFN的输出结果。\|默认确定性实现\|默认确定性实现\|-\|
170	\|[aclnnMoeFinalizeRoutingV2](../../moe/moe_finalize_routing_v2/docs/aclnnMoeFinalizeRoutingV2.md)\|MoE计算中，最后处理合并MoE FFN的输出结果，支持配置dropPadMode。\|默认确定性实现\|默认确定性实现\|默认确定性实现\|	166	\|[aclnnMoeFinalizeRoutingV2](../../moe/moe_finalize_routing_v2/docs/aclnnMoeFinalizeRoutingV2.md)\|MoE计算中，最后处理合并MoE FFN的输出结果，支持配置dropPadMode。\|默认确定性实现\|默认确定性实现\|默认确定性实现\|
171	\|[aclnnMoeFinalizeRoutingV2Grad](../../moe/moe_finalize_routing_v2_grad/docs/aclnnMoeFinalizeRoutingV2Grad.md)\|aclnnMoeFinalizeRoutingV2的反向传播。\|默认确定性实现\|默认确定性实现\|默认确定性实现\|	167	\|[aclnnMoeFinalizeRoutingV2Grad](../../moe/moe_finalize_routing_v2_grad/docs/aclnnMoeFinalizeRoutingV2Grad.md)\|aclnnMoeFinalizeRoutingV2的反向传播。\|默认确定性实现\|默认确定性实现\|默认确定性实现\|