fix(aclnn): RT2执行前设置确定性配置-ge-AtomGit

Mbase/common/op_tiling/op_tiling_rt2.cc+3-17

	@@ -1061,24 +1061,10 @@ ge::graphStatus GetDeterministicLevel(int32_t &deterministic_level, bool &has_de
1061	return ge::GRAPH_SUCCESS;	1061	return ge::GRAPH_SUCCESS;
1062	}	1062	}
1063		1063
1064	-ge::graphStatus CheckDeterministicConfig(const int32_t deterministic, const int32_t deterministic_level) {
1065	- if ((deterministic == 0) != (deterministic_level == 0)) {
1066	- const std::string value = std::to_string(deterministic) + " / " + std::to_string(deterministic_level);
1067	- const char *reason = "must both be 0 or both non-zero.";
1068	- REPORT_PREDEFINED_ERR_MSG(
1069	- "E10001", std::vector<const char *>({"parameter", "value", "reason"}),
1070	- std::vector<const char *>({"deterministic / deterministic_level", value.c_str(), reason}));
1071	- GELOGE(ge::PARAM_INVALID, "Deterministic[%d] and deterministic level[%d] are inconsistent.", deterministic,
1072	- deterministic_level);
1073	- return ge::PARAM_INVALID;
1074	- }
1075	- return ge::GRAPH_SUCCESS;
1076	-}
1077	-
1078	ge::graphStatus SetDeterministicConfig(const int32_t deterministic, const int32_t deterministic_level) {	1064	ge::graphStatus SetDeterministicConfig(const int32_t deterministic, const int32_t deterministic_level) {
1079	- GE_ASSERT_SUCCESS(CheckDeterministicConfig(deterministic, deterministic_level));	1065	+ const int32_t effective_level = (deterministic != 0 && deterministic_level == 0) ? 1 : deterministic_level;
1080	- GE_CHK_ACL_RET(aclrtSetSysParamOpt(ACL_OPT_DETERMINISTIC, deterministic_level));	1066	+ GE_CHK_ACL_RET(aclrtSetSysParamOpt(ACL_OPT_DETERMINISTIC, effective_level));
1081	- GE_CHK_ACL_RET(aclrtCtxSetSysParamOpt(ACL_OPT_DETERMINISTIC, deterministic_level));	1067	+ GE_CHK_ACL_RET(aclrtCtxSetSysParamOpt(ACL_OPT_DETERMINISTIC, effective_level));
1082	return ge::GRAPH_SUCCESS;	1068	return ge::GRAPH_SUCCESS;
1083	}	1069	}
1084		1070

Mdocs/zh/api/graph_engine_api/cpp/ge/aclgrphBuildInitialize_config_params/basic_functions.md+4-0

	@@ -109,6 +109,10 @@ cluster\_config.json配置示例如下：
109	{ge::ir_option::DETERMINISTIC, "1"}	109	{ge::ir_option::DETERMINISTIC, "1"}
110	```	110	```
111		111
		112	+使用约束：
		113	+
		114	+通过GE图引擎接口构建离线模型时，建议在模型构建阶段通过`ge::ir_option::DETERMINISTIC`和`ge::ir_option::DETERMINISTIC_LEVEL`设置确定性及一致性配置。通过`aclmdl`接口加载、执行离线模型时，不建议调用`aclrtSetSysParamOpt`或`aclrtCtxSetSysParamOpt`，通过`ACL_OPT_DETERMINISTIC`修改该配置，否则可能导致运行时配置与模型构建时的配置不一致，无法保证预期的确定性及一致性效果。如需调整该配置，建议重新构建离线模型。
		115	+
112	产品支持情况：	116	产品支持情况：
113		117
114	<!-- npu="950" id8 -->	118	<!-- npu="950" id8 -->

Mdocs/zh/api/graph_engine_api/cpp/ge/aclgrphBuildInitialize_config_params/experimental_parameters.md+4-0

	@@ -90,6 +90,10 @@ HF32是昇腾推出的专门用于算子内部计算的单精度浮点类型，
90	{ge::ir_option::DETERMINISTIC_LEVEL, "2"}	90	{ge::ir_option::DETERMINISTIC_LEVEL, "2"}
91	```	91	```
92		92
		93	+使用约束：
		94	+
		95	+通过GE图引擎接口构建离线模型时，建议在模型构建阶段通过`ge::ir_option::DETERMINISTIC`和`ge::ir_option::DETERMINISTIC_LEVEL`设置确定性及一致性配置。通过`aclmdl`接口加载、执行离线模型时，不建议调用`aclrtSetSysParamOpt`或`aclrtCtxSetSysParamOpt`，通过`ACL_OPT_DETERMINISTIC`修改该配置，否则可能导致运行时配置与模型构建时的配置不一致，无法保证预期的确定性及一致性效果。如需调整该配置，建议重新构建离线模型。
		96	+
93	产品支持情况：	97	产品支持情况：
94		98
95	<!-- npu="950" id15 -->	99	<!-- npu="950" id15 -->

Mdocs/zh/user_guides/atc_tools/CLI_options/--deterministic.md+1-1

	@@ -50,4 +50,4 @@ atc --deterministic=0 ...
50		50
51	## 依赖约束	51	## 依赖约束
52		52
53	-无	53	+离线推理场景下，建议在模型转换阶段通过`--deterministic`和`--deterministic_level`设置确定性及一致性配置。通过`aclmdl`接口加载、执行离线模型时，不建议调用`aclrtSetSysParamOpt`或`aclrtCtxSetSysParamOpt`，通过`ACL_OPT_DETERMINISTIC`修改该配置，否则可能导致运行时配置与模型转换时的配置不一致，无法保证预期的确定性及一致性效果。如需调整该配置，建议使用目标配置重新转换模型。

Mdocs/zh/user_guides/atc_tools/CLI_options/--deterministic_level.md+1-1

	@@ -64,4 +64,4 @@ atc --deterministic=0 --deterministic_level=0 ...
64		64
65	## 依赖约束	65	## 依赖约束
66		66
67	-无	67	+离线推理场景下，建议在模型转换阶段通过`--deterministic`和`--deterministic_level`设置确定性及一致性配置。通过`aclmdl`接口加载、执行离线模型时，不建议调用`aclrtSetSysParamOpt`或`aclrtCtxSetSysParamOpt`，通过`ACL_OPT_DETERMINISTIC`修改该配置，否则可能导致运行时配置与模型转换时的配置不一致，无法保证预期的确定性及一致性效果。如需调整该配置，建议使用目标配置重新转换模型。

Minc/framework/runtime/model_v2_executor.h+3-0

	@@ -243,6 +243,9 @@ class VISIBILITY_EXPORT ModelV2Executor {
243	rtStream_t default_stream_ = nullptr;	243	rtStream_t default_stream_ = nullptr;
244	ExecutorSubscribersScheduler subscribers_;	244	ExecutorSubscribersScheduler subscribers_;
245	ExecutorState state_ = ExecutorState::kInit;	245	ExecutorState state_ = ExecutorState::kInit;
		246	+ bool need_set_deterministic_config_ = false;
		247	+ int32_t deterministic_ = 0;
		248	+ int32_t deterministic_level_ = 0;
246	std::string file_constant_weight_dir_;	249	std::string file_constant_weight_dir_;
247	// 自动多流寻优标识，空表示不打点；本执行器无 model_id，由打点模块分配	250	// 自动多流寻优标识，空表示不打点；本执行器无 model_id，由打点模块分配
248	std::string auto_multistream_tuning_mode_;	251	std::string auto_multistream_tuning_mode_;

Mruntime/v2/core/builder/model_v2_executor_builder.cc+4-0

	@@ -149,6 +149,10 @@ std::unique_ptr<ModelV2Executor> ModelV2ExecutorBuilder::Build(const ExecutorOpt
149		149
150	ge::ComputeGraphPtr root_graph = root_model_->GetRootGraph();	150	ge::ComputeGraphPtr root_graph = root_model_->GetRootGraph();
151	GE_ASSERT_NOTNULL(root_graph);	151	GE_ASSERT_NOTNULL(root_graph);
		152	+ (void)ge::AttrUtils::GetInt(root_graph, ge::DETERMINISTIC, executor->deterministic_);
		153	+ (void)ge::AttrUtils::GetInt(root_graph, ge::DETERMINISTIC_LEVEL, executor->deterministic_level_);
		154	+ executor->need_set_deterministic_config_ =
		155	+ (executor->deterministic_ != 0) \|\| ge::AttrUtils::HasAttr(root_graph, ge::DETERMINISTIC_LEVEL);
152	GE_ASSERT_GRAPH_SUCCESS(RestoreDeviceVarMem(*executor));	156	GE_ASSERT_GRAPH_SUCCESS(RestoreDeviceVarMem(*executor));
153	uint32_t cur_model_id = root_model_->GetCurModelId();	157	uint32_t cur_model_id = root_model_->GetCurModelId();
154	std::string model_name = root_model_->GetModelName();	158	std::string model_name = root_model_->GetModelName();

Mruntime/v2/core/model_v2_executor.cc+5-0

	@@ -31,6 +31,7 @@
31	#include "graph/manager/session_id_manager.h"	31	#include "graph/manager/session_id_manager.h"
32	#include "acl/acl_rt.h"	32	#include "acl/acl_rt.h"
33	#include "common/ge_rts_decl.h"	33	#include "common/ge_rts_decl.h"
		34	+#include "common/op_tiling/op_tiling_rt2.h"
34		35
35	namespace gert {	36	namespace gert {
36	namespace {	37	namespace {
	@@ -244,6 +245,10 @@ ge::graphStatus ModelV2Executor::Execute(const ModelExecuteArg &arg, Tensor **in
244		245
245	GE_RETURN_IF_ERROR(CheckIoReuseAddrs(inputs, input_num, outputs, output_num));	246	GE_RETURN_IF_ERROR(CheckIoReuseAddrs(inputs, input_num, outputs, output_num));
246		247
		248	+ if (need_set_deterministic_config_) {
		249	+ GE_RETURN_IF_ERROR(optiling::SetGlobalDeterministicConfig(deterministic_, deterministic_level_));
		250	+ }
		251	+
247	ge::multistream_tune::StepScope step(ge::multistream_tune::kSiteModelV2Executor, auto_multistream_tuning_mode_,	252	ge::multistream_tune::StepScope step(ge::multistream_tune::kSiteModelV2Executor, auto_multistream_tuning_mode_,
248	auto_multistream_tuning_id_, arg.stream);	253	auto_multistream_tuning_id_, arg.stream);
249	ge::graphStatus ret = ge::GRAPH_FAILED;	254	ge::graphStatus ret = ge::GRAPH_FAILED;

Mruntime/v2/engine/aicore/converter/aclnn_node_converter.cc+10-4

	@@ -158,8 +158,15 @@ ge::graphStatus CreateAclnnDeterministicConfig(const ge::NodePtr &node, const Lo
158	std::to_string(config.op_deterministic) + "_" +	158	std::to_string(config.op_deterministic) + "_" +
159	std::to_string(static_cast<int32_t>(config.has_op_deterministic_level)) + "_" +	159	std::to_string(static_cast<int32_t>(config.has_op_deterministic_level)) + "_" +
160	std::to_string(config.op_deterministic_level);	160	std::to_string(config.op_deterministic_level);
161	- config_holder = lower_input.global_data->GetOrCreateUniqueValueHolder(	161	+ const auto holders =
162	- key, [&config]() { return bg::ValueHolder::CreateConst(&config, sizeof(config)); });	162	+ lower_input.global_data->GetOrCreateUniqueValueHolder(key, [&config]() -> std::vector<bg::ValueHolderPtr> {
		163	+ return bg::FrameSelector::OnInitRoot([&config]() -> std::vector<bg::ValueHolderPtr> {
		164	+ return {bg::ValueHolder::CreateConst(&config, sizeof(config))};
		165	+ });
		166	+ });
		167	+ GE_ASSERT_TRUE(!holders.empty());
		168	+ GE_ASSERT_NOTNULL(holders[0]);
		169	+ config_holder = holders[0];
163	return ge::GRAPH_SUCCESS;	170	return ge::GRAPH_SUCCESS;
164	}	171	}
165		172
	@@ -172,8 +179,7 @@ ge::graphStatus BuildAclnnOriginalDeterministicConfig(const ge::NodePtr &node,
172		179
173	GE_ASSERT_SUCCESS(	180	GE_ASSERT_SUCCESS(
174	optiling::GetGraphDeterministicConfig(root_compute_graph, config.deterministic, config.deterministic_level));	181	optiling::GetGraphDeterministicConfig(root_compute_graph, config.deterministic, config.deterministic_level));
175	-	182	+ return ge::GRAPH_SUCCESS;
176	- return optiling::SetGlobalDeterministicConfig(config.deterministic, config.deterministic_level);
177	}	183	}
178		184
179	ge::graphStatus CreateAclnnOriginalDeterministicConfig(const ge::NodePtr &node, const LowerInput &lower_input,	185	ge::graphStatus CreateAclnnOriginalDeterministicConfig(const ge::NodePtr &node, const LowerInput &lower_input,

Mtests/ge/st/testcase/fast_runtime_v2/ai_core/aclnn_node_converter_stest.cc+14-0

	@@ -110,6 +110,13 @@ TEST_F(AclnnNodeConverterST, TestHgl) {
110	global_data.SetSpaceRegistriesV2(*space_registry_array);	110	global_data.SetSpaceRegistriesV2(*space_registry_array);
111	auto add_ret = LoweringAclnnNode(add_node, add_input);	111	auto add_ret = LoweringAclnnNode(add_node, add_input);
112	ASSERT_TRUE(add_ret.result.IsSuccess());	112	ASSERT_TRUE(add_ret.result.IsSuccess());
		113	+ size_t set_config_count = 0U;
		114	+ for (const auto &node : init_frame_->GetExecuteGraph()->GetAllNodes()) {
		115	+ if (node->GetType() == "SetAclnnGlobalDeterministicConfig") {
		116	+ ++set_config_count;
		117	+ }
		118	+ }
		119	+ EXPECT_EQ(set_config_count, 0U);
113	}	120	}
114		121
115	TEST_F(AclnnNodeConverterST, LoweringTwoStagesWithCombinedDeterministicAttrs) {	122	TEST_F(AclnnNodeConverterST, LoweringTwoStagesWithCombinedDeterministicAttrs) {
	@@ -157,6 +164,13 @@ TEST_F(AclnnNodeConverterST, LoweringTwoStagesWithCombinedDeterministicAttrs) {
157	global_data.SetSpaceRegistriesV2(*space_registry_array);	164	global_data.SetSpaceRegistriesV2(*space_registry_array);
158	auto add_ret = LoweringAclnnNode(add_node, add_input);	165	auto add_ret = LoweringAclnnNode(add_node, add_input);
159	ASSERT_TRUE(add_ret.result.IsSuccess());	166	ASSERT_TRUE(add_ret.result.IsSuccess());
		167	+ size_t set_config_count = 0U;
		168	+ for (const auto &node : init_frame_->GetExecuteGraph()->GetAllNodes()) {
		169	+ if (node->GetType() == "SetAclnnGlobalDeterministicConfig") {
		170	+ ++set_config_count;
		171	+ }
		172	+ }
		173	+ EXPECT_EQ(set_config_count, 0U);
160	}	174	}
161		175
162	TEST_F(AclnnNodeConverterST, LoweringWithDeterministicAttr) {	176	TEST_F(AclnnNodeConverterST, LoweringWithDeterministicAttr) {

Mtests/ge/ut/ge/common/op_tiling_rt2_unittest.cc+9-6

	@@ -1346,12 +1346,15 @@ TEST_F(RegisterOpTilingRT2UT, SetDeterministicConfigPassesThroughLevel) {
1346	}	1346	}
1347	}	1347	}
1348		1348
1349	-TEST_F(RegisterOpTilingRT2UT, SetDeterministicConfigRejectsInconsistentConfig) {	1349	+TEST_F(RegisterOpTilingRT2UT, SetDeterministicConfigAcceptsIndependentValues) {
1350	- EXPECT_NE(SetDeterministicConfig(1, 0), GRAPH_SUCCESS);	1350	+ EXPECT_EQ(SetDeterministicConfig(1, 0), GRAPH_SUCCESS);
1351	- EXPECT_TRUE(acl_runtime_stub_->GetSysParamSetRecords().empty());	1351	+ EXPECT_EQ(SetDeterministicConfig(0, 1), GRAPH_SUCCESS);
1352	-	1352	+ const auto &records = acl_runtime_stub_->GetSysParamSetRecords();
1353	- EXPECT_NE(SetDeterministicConfig(0, 1), GRAPH_SUCCESS);	1353	+ ASSERT_EQ(records.size(), 4U);
1354	- EXPECT_TRUE(acl_runtime_stub_->GetSysParamSetRecords().empty());	1354	+ EXPECT_EQ(records[0].value, 1);
		1355	+ EXPECT_EQ(records[1].value, 1);
		1356	+ EXPECT_EQ(records[2].value, 1);
		1357	+ EXPECT_EQ(records[3].value, 1);
1355	}	1358	}
1356		1359
1357	TEST_F(RegisterOpTilingRT2UT, SetDeterministicConfigPropagatesAclFailures) {	1360	TEST_F(RegisterOpTilingRT2UT, SetDeterministicConfigPropagatesAclFailures) {

Mtests/ge/ut/ge/runtime/fast_v2/engine/aicore/aclnn_node_converter_unitest.cc+20-13

	@@ -113,6 +113,13 @@ TEST_F(AclnnNodeConverterUT, TestHgl) {
113	global_data.SetSpaceRegistriesV2(*space_registry_array);	113	global_data.SetSpaceRegistriesV2(*space_registry_array);
114	auto add_ret = LoweringAclnnNode(add_node, add_input);	114	auto add_ret = LoweringAclnnNode(add_node, add_input);
115	ASSERT_TRUE(add_ret.result.IsSuccess());	115	ASSERT_TRUE(add_ret.result.IsSuccess());
		116	+ size_t set_config_count = 0U;
		117	+ for (const auto &node : init_frame_->GetExecuteGraph()->GetAllNodes()) {
		118	+ if (node->GetType() == "SetAclnnGlobalDeterministicConfig") {
		119	+ ++set_config_count;
		120	+ }
		121	+ }
		122	+ EXPECT_EQ(set_config_count, 0U);
116	}	123	}
117		124
118	TEST_F(AclnnNodeConverterUT, TestHgl_Twostages) {	125	TEST_F(AclnnNodeConverterUT, TestHgl_Twostages) {
	@@ -162,9 +169,9 @@ TEST_F(AclnnNodeConverterUT, TestHgl_Twostages) {
162	ge::DumpGraph(exe_graph, "AclnnNodeConverterUT");	169	ge::DumpGraph(exe_graph, "AclnnNodeConverterUT");
163		170
164	// check main 图中节点数量, 由于没走CEM, main图上有俩InnerData连给PrepareCacheableTilingFwkData	171	// check main 图中节点数量, 由于没走CEM, main图上有俩InnerData连给PrepareCacheableTilingFwkData
165	- // 确定性配置新增一个Const和一个InnerData输入	172	+ // 确定性配置的Const在Init图中，main图新增一个InnerData输入
166	ASSERT_EQ(ExeGraphSummaryChecker(exe_graph).StrictAllNodeTypes({{"Data", 4},	173	ASSERT_EQ(ExeGraphSummaryChecker(exe_graph).StrictAllNodeTypes({{"Data", 4},
167	- {"Const", 10},	174	+ {"Const", 9},
168	{"AllocBatchHbm", 1},	175	{"AllocBatchHbm", 1},
169	{"AllocMemHbm", 1},	176	{"AllocMemHbm", 1},
170	{"BuildDualStageAclnnOpFwkData", 1},	177	{"BuildDualStageAclnnOpFwkData", 1},
	@@ -177,13 +184,20 @@ TEST_F(AclnnNodeConverterUT, TestHgl_Twostages) {
177	{"FreeBatchHbm", 1},	184	{"FreeBatchHbm", 1},
178	{"FreeMemory", 3},	185	{"FreeMemory", 3},
179	{"InferShape", 1},	186	{"InferShape", 1},
180	- {"InnerData", 10},	187	+ {"InnerData", 11},
181	{"MakeSureTensorAtDevice", 2},	188	{"MakeSureTensorAtDevice", 2},
182	{"SelectL1Allocator", 1},	189	{"SelectL1Allocator", 1},
183	{"SelectL2Allocator", 1},	190	{"SelectL2Allocator", 1},
184	{"SplitDataTensor", 2},	191	{"SplitDataTensor", 2},
185	{"SplitRtStreams", 1}}),	192	{"SplitRtStreams", 1}}),
186	"success");	193	"success");
		194	+ size_t set_config_count = 0U;
		195	+ for (const auto &node : init_frame_->GetExecuteGraph()->GetAllNodes()) {
		196	+ if (node->GetType() == "SetAclnnGlobalDeterministicConfig") {
		197	+ ++set_config_count;
		198	+ }
		199	+ }
		200	+ EXPECT_EQ(set_config_count, 0U);
187	}	201	}
188		202
189	TEST_F(AclnnNodeConverterUT, TestHgl_Twostages_Failed) {	203	TEST_F(AclnnNodeConverterUT, TestHgl_Twostages_Failed) {
	@@ -263,10 +277,10 @@ TEST_F(AclnnNodeConverterUT, LoweringWithDeterministicAttr) {
263		277
264	auto add_ret = LoweringAclnnNode(add_node, add_input);	278	auto add_ret = LoweringAclnnNode(add_node, add_input);
265	ASSERT_TRUE(add_ret.result.IsSuccess());	279	ASSERT_TRUE(add_ret.result.IsSuccess());
266	- EXPECT_EQ(acl_runtime_stub_->GetSysParamSetRecords().size(), 2U);	280	+ EXPECT_TRUE(acl_runtime_stub_->GetSysParamSetRecords().empty());
267	}	281	}
268		282
269	-TEST_F(AclnnNodeConverterUT, LoweringUsesLatestDeterministicConfig) {	283	+TEST_F(AclnnNodeConverterUT, LoweringDoesNotSetGlobalDeterministicConfig) {
270	auto graph = ShareGraph::AicoreGraph();	284	auto graph = ShareGraph::AicoreGraph();
271	auto add_node = graph->FindNode("add1");	285	auto add_node = graph->FindNode("add1");
272	(void)ge::AttrUtils::SetStr(add_node->GetOpDesc(), "_deterministic", "1");	286	(void)ge::AttrUtils::SetStr(add_node->GetOpDesc(), "_deterministic", "1");
	@@ -301,14 +315,7 @@ TEST_F(AclnnNodeConverterUT, LoweringUsesLatestDeterministicConfig) {
301		315
302	auto add_ret = LoweringAclnnNode(add_node, add_input);	316	auto add_ret = LoweringAclnnNode(add_node, add_input);
303	ASSERT_TRUE(add_ret.result.IsSuccess());	317	ASSERT_TRUE(add_ret.result.IsSuccess());
304	- const auto &records = acl_runtime_stub_->GetSysParamSetRecords();	318	+ EXPECT_TRUE(acl_runtime_stub_->GetSysParamSetRecords().empty());
305	- ASSERT_EQ(records.size(), 2U);
306	- EXPECT_FALSE(records[0].is_context);
307	- EXPECT_EQ(records[0].opt, ACL_OPT_DETERMINISTIC);
308	- EXPECT_EQ(records[0].value, 0);
309	- EXPECT_TRUE(records[1].is_context);
310	- EXPECT_EQ(records[1].opt, ACL_OPT_DETERMINISTIC);
311	- EXPECT_EQ(records[1].value, 0);
312	}	319	}
313		320
314	TEST_F(AclnnNodeConverterUT, LoweringWithDeterministicLevelAttr) {	321	TEST_F(AclnnNodeConverterUT, LoweringWithDeterministicLevelAttr) {