fixing experimental option and ReduceScatter BIRS-hccl-AtomGit

MCMakeLists.txt+1-1

	@@ -182,7 +182,7 @@ elseif(BUILD_OPEN_PROJECT)
182		182
183	# Device 侧编译构建	183	# Device 侧编译构建
184	if(ENABLE_DEVICE)	184	if(ENABLE_DEVICE)
185	- add_cann_device_project(hccl)	185	+ add_cann_device_project(hccl EXT_ARGS -D ENABLE_EXPERIMENTAL=${ENABLE_EXPERIMENTAL})
186	endif()	186	endif()
187		187
188	set(HCCL_HEAD	188	set(HCCL_HEAD

Mbuild.sh+9-0文件内容审核中，请稍后刷新重试

Mcmake/device/CMakeLists.txt+1-0

	@@ -14,6 +14,7 @@ set(PRODUCT_SIDE "device")
14	set(HCCL_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../../" CACHE PATH "Project top directory")	14	set(HCCL_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../../" CACHE PATH "Project top directory")
15	include(${HCCL_DIR}/cmake/fetch_cann_cmake.cmake)	15	include(${HCCL_DIR}/cmake/fetch_cann_cmake.cmake)
16	project(hccl_device)	16	project(hccl_device)
		17	+option(ENABLE_EXPERIMENTAL "Enable experimental features" OFF)
17		18
18	# 初始化 CANN 项目配置	19	# 初始化 CANN 项目配置
19	init_cann_project(PREPEND_MODULE_PATH)	20	init_cann_project(PREPEND_MODULE_PATH)

Mexperimental/ops/all_gather/all_gather.cc+6-0

	@@ -40,7 +40,13 @@ static HcclAlgorithm MakeAicpuAllGatherSequenceMeshNHRNHRMesh()
40	algo.hcclCmdType = HcclCMDType::HCCL_CMD_ALLGATHER;	40	algo.hcclCmdType = HcclCMDType::HCCL_CMD_ALLGATHER;
41	algo.engineType = HcclAlgEngineType::COMM_ENGINE_AICPU;	41	algo.engineType = HcclAlgEngineType::COMM_ENGINE_AICPU;
42	algo.topoMatch = std::make_shared<TopoMatchFourLevel>();	42	algo.topoMatch = std::make_shared<TopoMatchFourLevel>();
		43	+#ifndef AICPU_COMPILE
43	AlgAttrsRegistry::ParseAlgName("AicpuAllGatherSequenceMeshNHRNHRMesh", algo.algAttrs);	44	AlgAttrsRegistry::ParseAlgName("AicpuAllGatherSequenceMeshNHRNHRMesh", algo.algAttrs);
		45	+#else
		46	+ // AICPU 独立核库(scatter_aicpu_kernel.so)不链接 host-only 的 AlgAttrsRegistry,
		47	+ // 直接引用会导致 dlopen 阶段 undefined symbol, 与 REGISTER_ALG_ATTRS 在 AICPU_COMPILE 下置空同理。
		48	+ algo.algAttrs.name = "AicpuAllGatherSequenceMeshNHRNHRMesh";
		49	+#endif
44	algo.algoExecDesc = MakeAllGather4LevelAlgoExecDesc();	50	algo.algoExecDesc = MakeAllGather4LevelAlgoExecDesc();
45	algo.algName = "AicpuAllGatherSequenceMeshNHRNHRMesh";	51	algo.algName = "AicpuAllGatherSequenceMeshNHRNHRMesh";
46	return algo;	52	return algo;

Mexperimental/ops/all_reduce/example/executor/ins_v2_all_reduce_experimental_sole_executor.cc+3-1

	@@ -42,9 +42,11 @@ std::vector<CostModelParam> InsV2AllReduceExperimentalSoleExecutor<AlgTopoMatch,
42		42
43	template <typename AlgTopoMatch, typename InsAlgTemplate>	43	template <typename AlgTopoMatch, typename InsAlgTemplate>
44	AlgNetMeta InsV2AllReduceExperimentalSoleExecutor<AlgTopoMatch, InsAlgTemplate>::GetAlgNetMeta(	44	AlgNetMeta InsV2AllReduceExperimentalSoleExecutor<AlgTopoMatch, InsAlgTemplate>::GetAlgNetMeta(
45	- const TopoInfoWithNetLayerDetails* topoInfo) const	45	+ const TopoInfoWithNetLayerDetails* topoInfo, const OpParam& param, const char* algName) const
46	{	46	{
47	(void)topoInfo;	47	(void)topoInfo;
		48	+ (void)param;
		49	+ (void)algName;
48	AlgNetMeta meta;	50	AlgNetMeta meta;
49	meta.netTypes.push_back(CommTopo::COMM_TOPO_1DMESH);	51	meta.netTypes.push_back(CommTopo::COMM_TOPO_1DMESH);
50	meta.intraGroupMode = CostAggMode::SUM;	52	meta.intraGroupMode = CostAggMode::SUM;

Mexperimental/ops/all_reduce/example/executor/ins_v2_all_reduce_experimental_sole_executor.h+2-1

	@@ -37,7 +37,8 @@ public:
37		37
38	std::vector<CostModelParam> CalcCostCoeff(	38	std::vector<CostModelParam> CalcCostCoeff(
39	HcclComm comm, TopoInfoWithNetLayerDetails* topoInfo, const char* algName, const OpParam& param) override;	39	HcclComm comm, TopoInfoWithNetLayerDetails* topoInfo, const char* algName, const OpParam& param) override;
40	- AlgNetMeta GetAlgNetMeta(const TopoInfoWithNetLayerDetails* topoInfo) const override;	40	+ AlgNetMeta GetAlgNetMeta(
		41	+ const TopoInfoWithNetLayerDetails* topoInfo, const OpParam& param, const char* algName) const override;
41		42
42	#ifndef AICPU_COMPILE	43	#ifndef AICPU_COMPILE
43	HcclResult	44	HcclResult

Mexperimental/ops/op_common/CMakeLists.txt+0-5

	@@ -7,11 +7,6 @@
7	# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	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.	8	# See LICENSE in the root of the software repository for the full text of the License.
9	# -----------------------------------------------------------------------------------------------------------	9	# -----------------------------------------------------------------------------------------------------------
10	-
11	-if(NOT ENABLE_EXPERIMENTAL)
12	- return()
13	-endif()
14	-
15	add_subdirectory(template)	10	add_subdirectory(template)
16	add_subdirectory(topo)	11	add_subdirectory(topo)
17	add_subdirectory(recursive_executor)	12	add_subdirectory(recursive_executor)

Mexperimental/ops/op_common/recursive_executor/template/aicpu/allgather_mesh_template.cc+1-0

	@@ -9,6 +9,7 @@
9	*/	9	*/
10		10
11	#include "allgather_mesh_template.h"	11	#include "allgather_mesh_template.h"
		12	+#include "template_factory.h"
12		13
13	namespace ops_hccl {	14	namespace ops_hccl {
14		15

Mexperimental/ops/op_common/recursive_executor/template/aicpu/allgather_nhr_template.cc+1-0

	@@ -9,6 +9,7 @@
9	*/	9	*/
10		10
11	#include "allgather_nhr_template.h"	11	#include "allgather_nhr_template.h"
		12	+#include "template_factory.h"
12		13
13	namespace ops_hccl {	14	namespace ops_hccl {
14		15

Mexperimental/ops/op_common/template/CMakeLists.txt+0-5

	@@ -7,11 +7,6 @@
7	# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	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.	8	# See LICENSE in the root of the software repository for the full text of the License.
9	# -----------------------------------------------------------------------------------------------------------	9	# -----------------------------------------------------------------------------------------------------------
10	-
11	-if(NOT ENABLE_EXPERIMENTAL)
12	- return()
13	-endif()
14	-
15	set(src_list	10	set(src_list
16	${CMAKE_CURRENT_SOURCE_DIR}/alg_template_base_experimental.cc	11	${CMAKE_CURRENT_SOURCE_DIR}/alg_template_base_experimental.cc
17	)	12	)

Mexperimental/ops/op_common/topo/CMakeLists.txt+0-5

	@@ -7,11 +7,6 @@
7	# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	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.	8	# See LICENSE in the root of the software repository for the full text of the License.
9	# -----------------------------------------------------------------------------------------------------------	9	# -----------------------------------------------------------------------------------------------------------
10	-
11	-if(NOT ENABLE_EXPERIMENTAL)
12	- return()
13	-endif()
14	-
15	set(src_list	10	set(src_list
16	${CMAKE_CURRENT_SOURCE_DIR}/topo_experimental.cc	11	${CMAKE_CURRENT_SOURCE_DIR}/topo_experimental.cc
17	)	12	)

Mexperimental/ops/reduce_scatter/CMakeLists.txt+0-4

	@@ -8,10 +8,6 @@
8	# See LICENSE in the root of the software repository for the full text of the License.	8	# See LICENSE in the root of the software repository for the full text of the License.
9	# -----------------------------------------------------------------------------------------------------------	9	# -----------------------------------------------------------------------------------------------------------
10		10
11	-if(NOT ENABLE_EXPERIMENTAL)
12	- return()
13	-endif()
14	-
15	add_subdirectory(birs)	11	add_subdirectory(birs)
16		12
17	set(src_list	13	set(src_list

Mexperimental/ops/reduce_scatter/birs/CMakeLists.txt+0-4

	@@ -8,10 +8,6 @@
8	# See LICENSE in the root of the software repository for the full text of the License.	8	# See LICENSE in the root of the software repository for the full text of the License.
9	# -----------------------------------------------------------------------------------------------------------	9	# -----------------------------------------------------------------------------------------------------------
10		10
11	-if(NOT ENABLE_EXPERIMENTAL)
12	- return()
13	-endif()
14	-
15	add_subdirectory(template)	11	add_subdirectory(template)
16		12
17	set(src_list	13	set(src_list

Mexperimental/ops/reduce_scatter/birs/template/CMakeLists.txt+0-4

	@@ -8,10 +8,6 @@
8	# See LICENSE in the root of the software repository for the full text of the License.	8	# See LICENSE in the root of the software repository for the full text of the License.
9	# -----------------------------------------------------------------------------------------------------------	9	# -----------------------------------------------------------------------------------------------------------
10		10
11	-if(NOT ENABLE_EXPERIMENTAL)
12	- return()
13	-endif()
14	-
15	set(src_list	11	set(src_list
16	${CMAKE_CURRENT_SOURCE_DIR}/reduce_scatter_birs_inter.cc	12	${CMAKE_CURRENT_SOURCE_DIR}/reduce_scatter_birs_inter.cc
17	${CMAKE_CURRENT_SOURCE_DIR}/reduce_scatter_birs.cc	13	${CMAKE_CURRENT_SOURCE_DIR}/reduce_scatter_birs.cc

Mexperimental/ops/reduce_scatter/birs/template/reduce_scatter_birs.cc+13-9

	@@ -114,7 +114,7 @@ HcclResult ReduceScatterBIRS::Preprocess(const u32 rank, const u32 rankSize, std
114	return HCCL_E_INTERNAL;	114	return HCCL_E_INTERNAL;
115	}	115	}
116	if (slices_.size() == 0) {	116	if (slices_.size() == 0) {
117	- PrepareSlicesData(unitSize, count_, rankSize);	117	+ CHK_RET(PrepareSlicesData(unitSize, count_, rankSize));
118	}	118	}
119	return HCCL_SUCCESS;	119	return HCCL_SUCCESS;
120	}	120	}
	@@ -132,7 +132,8 @@ HcclResult ReduceScatterBIRS::HCCSProcessMainLoop(
132	void* src = static_cast<void>(static_cast<u8>(scratchMem_.addr) + localOffsetByte);	132	void* src = static_cast<void>(static_cast<u8>(scratchMem_.addr) + localOffsetByte);
133	void* dst = static_cast<void>(static_cast<u8>(hccs_links[round - 1].remoteOutput.addr) + remoteOffsetByte);	133	void* dst = static_cast<void>(static_cast<u8>(hccs_links[round - 1].remoteOutput.addr) + remoteOffsetByte);
134		134
135	- HcommWriteOnThread(subThreads[0], hccs_links[round - 1].handle, dst, src, sliceSize);	135	+ CHK_RET(static_cast<HcclResult>(
		136	+ HcommWriteOnThread(subThreads[0], hccs_links[round - 1].handle, dst, src, sliceSize)));
136		137
137	CHK_RET(static_cast<HcclResult>(	138	CHK_RET(static_cast<HcclResult>(
138	HcommChannelNotifyRecordOnThread(subThreads[0], hccs_links[round - 1].handle, NOTIFY_IDX_DATA_SIGNAL)));	139	HcommChannelNotifyRecordOnThread(subThreads[0], hccs_links[round - 1].handle, NOTIFY_IDX_DATA_SIGNAL)));
	@@ -156,9 +157,9 @@ HcclResult ReduceScatterBIRS::SIOProcessMainLoop(
156	void* src = static_cast<void>(static_cast<u8>(inputMem_.addr) + localOffsetByte);	157	void* src = static_cast<void>(static_cast<u8>(inputMem_.addr) + localOffsetByte);
157	void* dst = static_cast<void>(static_cast<u8>(sio_link.remoteOutput.addr) + remoteOffsetByte);	158	void* dst = static_cast<void>(static_cast<u8>(sio_link.remoteOutput.addr) + remoteOffsetByte);
158		159
159	- HcommWriteReduceOnThread(	160	+ CHK_RET(static_cast<HcclResult>(HcommWriteReduceOnThread(
160	mainThread, sio_link.handle, dst, src, sliceSize / unitSize, static_cast<HcommDataType>(dataType_),	161	mainThread, sio_link.handle, dst, src, sliceSize / unitSize, static_cast<HcommDataType>(dataType_),
161	- static_cast<HcommReduceOp>(reductionOp_));	162	+ static_cast<HcommReduceOp>(reductionOp_))));
162		163
163	CHK_RET(	164	CHK_RET(
164	static_cast<HcclResult>(HcommChannelNotifyRecordOnThread(mainThread, sio_link.handle, NOTIFY_IDX_DATA_SIGNAL)));	165	static_cast<HcclResult>(HcommChannelNotifyRecordOnThread(mainThread, sio_link.handle, NOTIFY_IDX_DATA_SIGNAL)));
	@@ -207,7 +208,7 @@ HcclResult ReduceScatterBIRS::RunAsync(const u32 rank, const u32 rankSize, std::
207	"ReduceScatterBIRS run: rank[%u] rankSize[%u] inputMem[%p] to outputMem[%p] count[%llu]", rank, rankSize,	208	"ReduceScatterBIRS run: rank[%u] rankSize[%u] inputMem[%p] to outputMem[%p] count[%llu]", rank, rankSize,
208	inputMem_.addr, outputMem_.addr, count_);	209	inputMem_.addr, outputMem_.addr, count_);
209		210
210	- Preprocess(rank, rankSize, channels);	211	+ CHK_RET(Preprocess(rank, rankSize, channels));
211		212
212	u32 rankSizeX_ = 2;	213	u32 rankSizeX_ = 2;
213	if (rankSize % rankSizeX_ != 0) {	214	if (rankSize % rankSizeX_ != 0) {
	@@ -243,15 +244,18 @@ HcclResult ReduceScatterBIRS::RunAsync(const u32 rank, const u32 rankSize, std::
243		244
244	GetNotifyIdxSubToMain(notifyIdxSubToMain_);	245	GetNotifyIdxSubToMain(notifyIdxSubToMain_);
245	CHK_RET(PostSyncInterThreads(mainThread, subThreads, notifyIdxSubToMain_));	246	CHK_RET(PostSyncInterThreads(mainThread, subThreads, notifyIdxSubToMain_));
		247	+ // 说明：Hcomm*OnThread 失败视为致命内部错误，提前返回依赖上层 comm 销毁统一清理线程/通道，
		248	+ // 此处不做子线程兜底同步。子线程侧等待均受 CUSTOM_TIMEOUT/ExecTimeout 约束，不会永久挂起；
		249	+ // PostSync 的 wait+record 成对发射，整体跳过不会造成线程 notify 计数失衡。
246	for (u32 round = 0; round < hccs_ranks.size() + 1; round++) {	250	for (u32 round = 0; round < hccs_ranks.size() + 1; round++) {
247	// MainRecordSub + SubWaitMain	251	// MainRecordSub + SubWaitMain
248	CHK_RET(PreSyncInterThreads(mainThread, subThreads, notifyIdxMainToSub_));	252	CHK_RET(PreSyncInterThreads(mainThread, subThreads, notifyIdxMainToSub_));
249		253
250	- HCCSProcessMainLoop(round, rank, rankSize, rankSizeX_, sliceSize, localStrideSize);	254	+ CHK_RET(HCCSProcessMainLoop(round, rank, rankSize, rankSizeX_, sliceSize, localStrideSize));
		W W wenxuemin9月15日 ⚠️ [中等] [代码正确性] experimental/ops/reduce_scatter/birs/template/reduce_scatter_birs.cc:251 给 HCCSProcessMainLoop / SIOProcessMainLoop / LocalCopyMainLoop 包上 CHK_RET 后，若失败会提前 return 跳过 PostSyncInterThreads。此时 subThreads[0]/[1] 可能仍阻塞在等待主线程 NotifyRecord，导致子线程挂起、channel notify 状态不一致。建议：在注释中明确「Hcomm*OnThread 失败即致命，提前返回依赖上层 comm 销毁清理，不再做子线程同步」；或在 RunAsync 出口增加错误清理路径兜底线程同步。 D Davydov_Danil 9月15日评论： 9月15日评论：
251		255
252	- SIOProcessMainLoop(round, rank, rankSize, rankSizeX_, sliceSize, localStrideSize);	256	+ CHK_RET(SIOProcessMainLoop(round, rank, rankSize, rankSizeX_, sliceSize, localStrideSize));
253		257
254	- LocalCopyMainLoop(round, rank, rankSize, rankSizeX_, sliceSize, localStrideSize);	258	+ CHK_RET(LocalCopyMainLoop(round, rank, rankSize, rankSizeX_, sliceSize, localStrideSize));
255		259
256	// SubRecordMain + MainWaitSub	260	// SubRecordMain + MainWaitSub
257	CHK_RET(PostSyncInterThreads(mainThread, subThreads, notifyIdxSubToMain_));	261	CHK_RET(PostSyncInterThreads(mainThread, subThreads, notifyIdxSubToMain_));
	@@ -260,7 +264,7 @@ HcclResult ReduceScatterBIRS::RunAsync(const u32 rank, const u32 rankSize, std::
260	// MainRecordSub + SubWaitMain	264	// MainRecordSub + SubWaitMain
261	CHK_RET(PreSyncInterThreads(mainThread, subThreads, notifyIdxMainToSub_));	265	CHK_RET(PreSyncInterThreads(mainThread, subThreads, notifyIdxMainToSub_));
262		266
263	- FinalStep(rank, rankSize, rankSizeX_, sliceSize, localStrideSize);	267	+ CHK_RET(FinalStep(rank, rankSize, rankSizeX_, sliceSize, localStrideSize));
264		268
265	CHK_RET(PostSyncInterThreads(mainThread, subThreads, notifyIdxSubToMain_));	269	CHK_RET(PostSyncInterThreads(mainThread, subThreads, notifyIdxSubToMain_));
266		270

Mexperimental/ops/reduce_scatter/birs/template/reduce_scatter_birs_inter.cc+17-13

	@@ -74,7 +74,8 @@ HcclResult ReduceScatterBIRSInter::HCCSIntraStep(
74	void* src = static_cast<void>(static_cast<u8>(scratchMem_.addr) + localOffsetByte);	74	void* src = static_cast<void>(static_cast<u8>(scratchMem_.addr) + localOffsetByte);
75	void* dst = static_cast<void>(static_cast<u8>(hccs_links[round - 1].remoteOutput.addr) + remoteOffsetByte);	75	void* dst = static_cast<void>(static_cast<u8>(hccs_links[round - 1].remoteOutput.addr) + remoteOffsetByte);
76		76
77	- HcommWriteOnThread(subThreads[0], hccs_links[round - 1].handle, dst, src, localStrideSize * serverNum_);	77	+ CHK_RET(static_cast<HcclResult>(
		78	+ HcommWriteOnThread(subThreads[0], hccs_links[round - 1].handle, dst, src, localStrideSize * serverNum_)));
78		79
79	CHK_RET(static_cast<HcclResult>(	80	CHK_RET(static_cast<HcclResult>(
80	HcommChannelNotifyRecordOnThread(subThreads[0], hccs_links[round - 1].handle, NOTIFY_IDX_DATA_SIGNAL)));	81	HcommChannelNotifyRecordOnThread(subThreads[0], hccs_links[round - 1].handle, NOTIFY_IDX_DATA_SIGNAL)));
	@@ -99,9 +100,9 @@ HcclResult ReduceScatterBIRSInter::SIOIntraStep(
99	void* src = static_cast<void>(static_cast<u8>(scratchMem_.addr) + localOffsetByte);	100	void* src = static_cast<void>(static_cast<u8>(scratchMem_.addr) + localOffsetByte);
100	void* dst = static_cast<void>(static_cast<u8>(sio_link.remoteOutput.addr) + remoteOffsetByte);	101	void* dst = static_cast<void>(static_cast<u8>(sio_link.remoteOutput.addr) + remoteOffsetByte);
101		102
102	- HcommWriteReduceOnThread(	103	+ CHK_RET(static_cast<HcclResult>(HcommWriteReduceOnThread(
103	mainThread, sio_link.handle, dst, src, localStrideSize * serverNum_ / unitSize,	104	mainThread, sio_link.handle, dst, src, localStrideSize * serverNum_ / unitSize,
104	- static_cast<HcommDataType>(dataType_), static_cast<HcommReduceOp>(reductionOp_));	105	+ static_cast<HcommDataType>(dataType_), static_cast<HcommReduceOp>(reductionOp_))));
105		106
106	CHK_RET(	107	CHK_RET(
107	static_cast<HcclResult>(HcommChannelNotifyRecordOnThread(mainThread, sio_link.handle, NOTIFY_IDX_DATA_SIGNAL)));	108	static_cast<HcclResult>(HcommChannelNotifyRecordOnThread(mainThread, sio_link.handle, NOTIFY_IDX_DATA_SIGNAL)));
	@@ -175,8 +176,8 @@ HcclResult ReduceScatterBIRSInter::InterServer(
175	void* dst = static_cast<void*>(	176	void* dst = static_cast<void*>(
176	static_cast<u8>(channels[((round)intraRankSize_ + rank) % rankSize].remoteOutput.addr)	177	static_cast<u8>(channels[((round)intraRankSize_ + rank) % rankSize].remoteOutput.addr)
177	+ remoteOffsetByte);	178	+ remoteOffsetByte);
178	- HcommWriteOnThread(	179	+ CHK_RET(static_cast<HcclResult>(HcommWriteOnThread(
179	- mainThread, channels[((round)*intraRankSize_ + rank) % rankSize].handle, dst, src, sliceSize);	180	+ mainThread, channels[((round)*intraRankSize_ + rank) % rankSize].handle, dst, src, sliceSize)));
180	CHK_RET(static_cast<HcclResult>(HcommChannelNotifyRecordOnThread(	181	CHK_RET(static_cast<HcclResult>(HcommChannelNotifyRecordOnThread(
181	mainThread, channels[((round)*intraRankSize_ + rank) % rankSize].handle, NOTIFY_IDX_DATA_SIGNAL)));	182	mainThread, channels[((round)*intraRankSize_ + rank) % rankSize].handle, NOTIFY_IDX_DATA_SIGNAL)));
182	CHK_RET(static_cast<HcclResult>(HcommChannelNotifyWaitOnThread(	183	CHK_RET(static_cast<HcclResult>(HcommChannelNotifyWaitOnThread(
	@@ -190,7 +191,7 @@ HcclResult ReduceScatterBIRSInter::InterServer(
190	CHK_RET(PreSyncInterThreads(mainThread, subThreads, notifyIdxMainToSub_));	191	CHK_RET(PreSyncInterThreads(mainThread, subThreads, notifyIdxMainToSub_));
191	for (u32 step = 1; step < serverNum_; step *= 2) {	192	for (u32 step = 1; step < serverNum_; step *= 2) {
192	for (u32 i = 0; i + step < serverNum_; i += 2 * step) {	193	for (u32 i = 0; i + step < serverNum_; i += 2 * step) {
193	- LocalReduceCCLToCCL(ptrSz + sliceSize * (i + step), ptrSz + sliceSize * i, sliceSize, mainThread);	194	+ CHK_RET(LocalReduceCCLToCCL(ptrSz + sliceSize * (i + step), ptrSz + sliceSize * i, sliceSize, mainThread));
194	}	195	}
195	}	196	}
196	void* srcSlice = static_cast<void>(static_cast<u8>(scratchMem_.addr) + ptrSz);	197	void* srcSlice = static_cast<void>(static_cast<u8>(scratchMem_.addr) + ptrSz);
	@@ -202,7 +203,7 @@ HcclResult ReduceScatterBIRSInter::InterServer(
202	// scatter的入口函数	203	// scatter的入口函数
203	HcclResult ReduceScatterBIRSInter::RunAsync(const u32 rank, const u32 rankSize, std::vector<ChannelInfo>& channels)	204	HcclResult ReduceScatterBIRSInter::RunAsync(const u32 rank, const u32 rankSize, std::vector<ChannelInfo>& channels)
204	{	205	{
205	- Preprocess(rank, rankSize, channels);	206	+ CHK_RET(Preprocess(rank, rankSize, channels));
		严正行9月4日 [LOW] 描述测试仅覆盖单机场景，本次修改的 ReduceScatterBIRSInter（跨服务器路径）错误处理行为变更无验证说明第 206 行：`CHK_RET(Preprocess(rank, rankSize, channels));` 问题分析：本 PR 对 reduce_scatter_birs_inter.cc 的 RunAsync/HCCSIntraStep/SIOIntraStep/PreprocInterServer/InterServer 等函数系统性补 CHK_RET（含第 206 行 RunAsync 入口的 Preprocess、InterServer 内 HcommWriteOnThread 与 LocalReduceCCLToCCL），改变了失败语义：原先这些通信调用失败后继续执行，现在即时向上返回。 PR 描述 Testing 一节仅说明 "Single-node: ReduceScatterBirs was deployed on one Atlas A3 node with RankSize set between 4 and 16"。单机（serverNum==1）走的是 ReduceScatterBIRS（reduce_scatter_birs.cc，SelectAndPrepareBirsTemplate 按 isSingleServer 选 TEMPLATE_REDUCE_SCATTER_BIRS）；ReduceScatterBIRSInter 的跨服务器路径（含 InterServer 的 NOTIFY_IDX_ACK/DATA_SIGNAL 等待序列）无任何测试覆盖说明。集合通信中错误传播提前返回可能导致对端 rank 在 notify 等待处超时，该行为变更至少需要在描述中说明验证方式。建议修复：在 PR 描述 Testing 一节补充 ReduceScatterBIRSInter（多服务器）路径的测试情况；若暂无条件覆盖双机环境，明确说明该路径当前无 ST 覆盖并给出风险评估，便于合入决策。 D Davydov_Danil 9月7日评论： 9月7日评论：严正行 9月8日评论： 9月8日评论：
206	u32 rankSizeX_ = 2;	207	u32 rankSizeX_ = 2;
207	if (intraRankSize_ % rankSizeX_ != 0) {	208	if (intraRankSize_ % rankSizeX_ != 0) {
208	HCCL_ERROR(	209	HCCL_ERROR(
	@@ -231,17 +232,20 @@ HcclResult ReduceScatterBIRSInter::RunAsync(const u32 rank, const u32 rankSize,
231	GetNotifyIdxMainToSub(notifyIdxMainToSub_);	232	GetNotifyIdxMainToSub(notifyIdxMainToSub_);
232	CHK_RET(PreSyncInterThreads(mainThread, subThreads, notifyIdxMainToSub_));	233	CHK_RET(PreSyncInterThreads(mainThread, subThreads, notifyIdxMainToSub_));
233		234
234	- LocalCopyPreproc(mainThread, rank, sliceSize, localStrideSize);	235	+ CHK_RET(LocalCopyPreproc(mainThread, rank, sliceSize, localStrideSize));
235		236
236	GetNotifyIdxSubToMain(notifyIdxSubToMain_);	237	GetNotifyIdxSubToMain(notifyIdxSubToMain_);
237	CHK_RET(PostSyncInterThreads(mainThread, subThreads, notifyIdxSubToMain_));	238	CHK_RET(PostSyncInterThreads(mainThread, subThreads, notifyIdxSubToMain_));
		239	+ // 说明：Hcomm*OnThread 失败视为致命内部错误，提前返回依赖上层 comm 销毁统一清理线程/通道，
		240	+ // 此处不做子线程兜底同步。子线程侧等待均受 CUSTOM_TIMEOUT/ExecTimeout 约束，不会永久挂起；
		241	+ // PostSync 的 wait+record 成对发射，整体跳过不会造成线程 notify 计数失衡。
238	for (u32 round = 0; round < hccs_ranks.size() + 1; round++) {	242	for (u32 round = 0; round < hccs_ranks.size() + 1; round++) {
239	// MainRecordSub + SubWaitMain	243	// MainRecordSub + SubWaitMain
240	CHK_RET(PreSyncInterThreads(mainThread, subThreads, notifyIdxMainToSub_));	244	CHK_RET(PreSyncInterThreads(mainThread, subThreads, notifyIdxMainToSub_));
241		245
242	- HCCSIntraStep(round, rank, rankSize, rankSizeX_, sliceSize, localStrideSize);	246	+ CHK_RET(HCCSIntraStep(round, rank, rankSize, rankSizeX_, sliceSize, localStrideSize));
243	- SIOIntraStep(round, rank, rankSize, rankSizeX_, sliceSize, localStrideSize);	247	+ CHK_RET(SIOIntraStep(round, rank, rankSize, rankSizeX_, sliceSize, localStrideSize));
244	- LocalCopyIntraStep(round, rank, rankSize, rankSizeX_, sliceSize, localStrideSize);	248	+ CHK_RET(LocalCopyIntraStep(round, rank, rankSize, rankSizeX_, sliceSize, localStrideSize));
245		249
246	// SubRecordMain + MainWaitSub	250	// SubRecordMain + MainWaitSub
247	CHK_RET(PostSyncInterThreads(mainThread, subThreads, notifyIdxSubToMain_));	251	CHK_RET(PostSyncInterThreads(mainThread, subThreads, notifyIdxSubToMain_));
	@@ -250,11 +254,11 @@ HcclResult ReduceScatterBIRSInter::RunAsync(const u32 rank, const u32 rankSize,
250	// MainRecordSub + SubWaitMain	254	// MainRecordSub + SubWaitMain
251	CHK_RET(PreSyncInterThreads(mainThread, subThreads, notifyIdxMainToSub_));	255	CHK_RET(PreSyncInterThreads(mainThread, subThreads, notifyIdxMainToSub_));
252		256
253	- PreprocInterServer(rank, rankSize, rankSizeX_, sliceSize, localStrideSize, channels);	257	+ CHK_RET(PreprocInterServer(rank, rankSize, rankSizeX_, sliceSize, localStrideSize, channels));
254		258
255	CHK_RET(PostSyncInterThreads(mainThread, subThreads, notifyIdxSubToMain_));	259	CHK_RET(PostSyncInterThreads(mainThread, subThreads, notifyIdxSubToMain_));
256		260
257	- InterServer(rank, rankSize, rankSizeX_, sliceSize, localStrideSize, channels);	261	+ CHK_RET(InterServer(rank, rankSize, rankSizeX_, sliceSize, localStrideSize, channels));
258		262
259	CHK_RET(PostSyncInterThreads(mainThread, subThreads, notifyIdxSubToMain_));	263	CHK_RET(PostSyncInterThreads(mainThread, subThreads, notifyIdxSubToMain_));
260	HCCL_INFO("ReduceScatterBIRS finished: rank[%u]", rank);	264	HCCL_INFO("ReduceScatterBIRS finished: rank[%u]", rank);