Star1574
3231
代码
代码
Issues495
Pull Requests844
Actions
项目讨论
Wiki
分析
Star1574
3231
已关闭
Fix CleanCode #11170
yayahello创建于 9月3日关闭于 9月4日
Fix CleanCode #11170
已关闭
yayahello创建于 9月3日关闭于 9月4日
从yayahello/ops-transformermaster合入到cann/ops-transformermaster
共 10 个文件变更+59-94
Mmc2/engram_fetch/op_api/aclnn_engram_fetch.cpp+4-12
@@ -24,9 +24,7 @@
24 24 
25using namespace op;25using namespace op;
26 26 
27-static bool CheckNotNull(const aclTensor *commContext, const aclTensor *indices, const aclTensor *localStorageAddr,27+static bool CheckNotNull(const aclTensor *commContext, const aclTensor *indices, const aclTensor *fetched)
28- aclTensor *fetched, aclTensor *permOut, aclTensor *sendCountsOut, aclTensor *recvCountsOut,
29- aclTensor *recvLocalEntryOut, aclTensor *numRecvOut)
30{28{
31 OP_CHECK_NULL(commContext, return false);29 OP_CHECK_NULL(commContext, return false);
32 OP_CHECK_NULL(indices, return false);30 OP_CHECK_NULL(indices, return false);
@@ -34,14 +32,9 @@ static bool CheckNotNull(const aclTensor *commContext, const aclTensor *indices,
34 return true;32 return true;
35}33}
36 34 
37-static aclnnStatus CheckParams(const aclTensor *commContext, const aclTensor *indices,35+static aclnnStatus CheckParams(const aclTensor *commContext, const aclTensor *indices, const aclTensor *fetched)
38- const aclTensor *localStorageAddr, aclTensor *fetched, aclTensor *permOut,
39- aclTensor *sendCountsOut, aclTensor *recvCountsOut, aclTensor *recvLocalEntryOut,
40- aclTensor *numRecvOut)
41{36{
42- CHECK_RET(CheckNotNull(commContext, indices, localStorageAddr, fetched, permOut, sendCountsOut, recvCountsOut,37+ CHECK_RET(CheckNotNull(commContext, indices, fetched), ACLNN_ERR_PARAM_NULLPTR);
43- recvLocalEntryOut, numRecvOut),
44- ACLNN_ERR_PARAM_NULLPTR);
45 return ACLNN_SUCCESS;38 return ACLNN_SUCCESS;
46}39}
47 40 
@@ -57,8 +50,7 @@ aclnnStatus aclnnEngramFetchGetWorkspaceSize(const aclTensor *commContext, const
57 int64_t commBufferSize, int64_t withGrad, uint64_t *workspaceSize,50 int64_t commBufferSize, int64_t withGrad, uint64_t *workspaceSize,
58 aclOpExecutor **executor)51 aclOpExecutor **executor)
59{52{
60- auto retParam = CheckParams(commContext, indices, localStorageAddr, fetched, permOut, sendCountsOut, recvCountsOut,53+ auto retParam = CheckParams(commContext, indices, fetched);
61- recvLocalEntryOut, numRecvOut);
62 CHECK_RET(retParam == ACLNN_SUCCESS, retParam);54 CHECK_RET(retParam == ACLNN_SUCCESS, retParam);
63 aclnnStatus ret = aclnnInnerEngramFetchGetWorkspaceSize(commContext, indices, localStorageAddr, hiddenSize,55 aclnnStatus ret = aclnnInnerEngramFetchGetWorkspaceSize(commContext, indices, localStorageAddr, hiddenSize,
64 numEntriesPerRank, numMaxTokensPerRank, commBufferSize,56 numEntriesPerRank, numMaxTokensPerRank, commBufferSize,
Mmc2/moe_distribute_combine_setup/op_api/aclnn_moe_distribute_combine_setup.h+4-4
@@ -35,7 +35,7 @@ extern "C" {
35 * @param [in] epWorldSize: 计算输入，int。ep通信域size。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。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。37 * @param [in] moeExpertNum: 计算输入，int。MOE专家数量。
38- * @param [in] expertSharedType: 计算可选输入，int。共享专家卡分布类型。当前仅支持传0。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。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。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。41 * @param [in] globalBs: 计算可选输入，int。
@@ -51,7 +51,7 @@ extern "C" {
51 */51 */
52ACLNN_API aclnnStatus aclnnMoeDistributeCombineSetupGetWorkspaceSize(52ACLNN_API aclnnStatus aclnnMoeDistributeCombineSetupGetWorkspaceSize(
53 const aclTensor *expandX, const aclTensor *expertIds, const aclTensor *assistInfoForCombine, const char *groupEp,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 expertSharedType, int64_t sharedExpertNum,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,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);56 aclTensor *quantExpandXOut, aclTensor *commCmdInfoOut, uint64_t *workspaceSize, aclOpExecutor **executor);
57 57 
@@ -66,7 +66,7 @@ ACLNN_API aclnnStatus aclnnMoeDistributeCombineSetupGetWorkspaceSize(
66 * @param [in] epWorldSize: 计算输入，int。ep通信域size。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。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。68 * @param [in] moeExpertNum: 计算输入，int。MOE专家数量。
69- * @param [in] expertSharedType: 计算可选输入，int。共享专家卡分布类型。当前仅支持传0。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。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。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。72 * @param [in] globalBs: 计算可选输入，int。
@@ -80,7 +80,7 @@ ACLNN_API aclnnStatus aclnnMoeDistributeCombineSetupGetWorkspaceSize(
80 */80 */
81ACLNN_API aclnnStatus aclnnMoeDistributeCombineSetupTeardownCalcOutputSize(81ACLNN_API aclnnStatus aclnnMoeDistributeCombineSetupTeardownCalcOutputSize(
82 const aclTensor *expandX, const aclTensor *expertIds, const aclTensor *assistInfoForCombine, const char *groupEp,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 expertSharedType, int64_t sharedExpertNum,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,84 int64_t sharedExpertRankNum, int64_t globalBs, int64_t commQuantMode, int64_t commType, const char *commAlg,
85 uint64_t &tokenMsgSize, uint64_t &commCmdInfoOutSize);85 uint64_t &tokenMsgSize, uint64_t &commCmdInfoOutSize);
86 86 
Mmc2/moe_distribute_combine_teardown/op_api/aclnn_moe_distribute_combine_teardown.cpp+3-12
@@ -36,7 +36,6 @@ enum class NnopbaseHcclServerType : uint32_t {
36 36 
37static bool CheckNotNull(const aclTensor *expandX, const aclTensor *quantExpandX, const aclTensor *expertIds,37static bool CheckNotNull(const aclTensor *expandX, const aclTensor *quantExpandX, const aclTensor *expertIds,
38 const aclTensor *expandIdx, const aclTensor *expertScales, const aclTensor *commCmdInfo,38 const aclTensor *expandIdx, const aclTensor *expertScales, const aclTensor *commCmdInfo,
39- const aclTensor *xActiveMaskOptional, const aclTensor *sharedExpertXOptional,
40 const char *groupEp, aclTensor *xOut)39 const char *groupEp, aclTensor *xOut)
41{40{
42 OP_LOGD("aclnn_moe_distribute_combine_teardown CheckNotNull start");41 OP_LOGD("aclnn_moe_distribute_combine_teardown CheckNotNull start");
@@ -57,15 +56,10 @@ static bool CheckNotNull(const aclTensor *expandX, const aclTensor *quantExpandX
57 56 
58static aclnnStatus CheckParams(const aclTensor *expandX, const aclTensor *quantExpandX, const aclTensor *expertIds,57static aclnnStatus CheckParams(const aclTensor *expandX, const aclTensor *quantExpandX, const aclTensor *expertIds,
59 const aclTensor *expandIdx, const aclTensor *expertScales, const aclTensor *commCmdInfo,58 const aclTensor *expandIdx, const aclTensor *expertScales, const aclTensor *commCmdInfo,
60- const aclTensor *xActiveMaskOptional, const aclTensor *sharedExpertXOptional,59+ const char *groupEp, aclTensor *xOut)
61- const char *groupEp, int64_t epWorldSize, int64_t epRankId, int64_t moeExpertNum,
62- int64_t expertShardType, int64_t sharedExpertNum, int64_t sharedExpertRankNum,
63- int64_t globalBs, int64_t commQuantMode, int64_t commType, const char *commAlg,
64- aclTensor *xOut)
65{60{
66 OP_LOGD("aclnn_moe_distribute_combine_teardown checkparams start");61 OP_LOGD("aclnn_moe_distribute_combine_teardown checkparams start");
67- CHECK_RET(CheckNotNull(expandX, quantExpandX, expertIds, expandIdx, expertScales, commCmdInfo, xActiveMaskOptional,62+ CHECK_RET(CheckNotNull(expandX, quantExpandX, expertIds, expandIdx, expertScales, commCmdInfo, groupEp, xOut),
68- sharedExpertXOptional, groupEp, xOut),
69 ACLNN_ERR_PARAM_NULLPTR);63 ACLNN_ERR_PARAM_NULLPTR);
70 if (strnlen(groupEp, HCCL_GROUP_NAME_MAX) >= HCCL_GROUP_NAME_MAX) {64 if (strnlen(groupEp, HCCL_GROUP_NAME_MAX) >= HCCL_GROUP_NAME_MAX) {
71 OP_LOGE_FOR_INVALID_VALUE_WITH_REASON(65 OP_LOGE_FOR_INVALID_VALUE_WITH_REASON(
@@ -105,10 +99,7 @@ extern "C" aclnnStatus aclnnMoeDistributeCombineTeardownGetWorkspaceSize(
105 return ACLNN_ERR_PARAM_INVALID;99 return ACLNN_ERR_PARAM_INVALID;
106 }100 }
107 101 
108- auto ret_param =102+ auto ret_param = CheckParams(expandX, quantExpandX, expertIds, expandIdx, expertScales, commCmdInfo, groupEp, xOut);
109- CheckParams(expandX, quantExpandX, expertIds, expandIdx, expertScales, commCmdInfo, xActiveMaskOptional,
110- sharedExpertXOptional, groupEp, epWorldSize, epRankId, moeExpertNum, expertShardType,
111- sharedExpertNum, sharedExpertRankNum, globalBs, commQuantMode, commType, commAlg, xOut);
112 CHECK_RET(ret_param == ACLNN_SUCCESS, ret_param);103 CHECK_RET(ret_param == ACLNN_SUCCESS, ret_param);
113 104 
114 aclnnStatus ret = aclnnInnerMoeDistributeCombineTeardownGetWorkspaceSize(105 aclnnStatus ret = aclnnInnerMoeDistributeCombineTeardownGetWorkspaceSize(
Mmc2/moe_distribute_dispatch_setup/op_api/aclnn_moe_distribute_dispatch_setup.cpp+3-8
@@ -65,11 +65,8 @@ static bool CheckNotNull(const aclTensor *x, const aclTensor *expertIds, const c
65}65}
66 66 
67// 入参校验67// 入参校验
68-static aclnnStatus CheckParams(const aclTensor *x, const aclTensor *expertIds, const aclTensor *scalesOptional,68+static aclnnStatus CheckParams(const aclTensor *x, const aclTensor *expertIds, const char *groupEp, aclTensor *yOut,
69- const char *groupEp, int64_t epWorldSize, int64_t epRankId, int64_t expertShardType,69+ aclTensor *expandIdxOut, aclTensor *commCmdInfoOut)
70- int64_t shareExpertRankNum, int64_t moeExpertNum, int64_t quantMode, int64_t globalBs,
71- int64_t commType, const char *commAlg, aclTensor *yOut, aclTensor *expandIdxOut,
72- aclTensor *commCmdInfoOut)
73{70{
74 CHECK_RET(CheckNotNull(x, expertIds, groupEp, yOut, expandIdxOut, commCmdInfoOut), ACLNN_ERR_PARAM_NULLPTR);71 CHECK_RET(CheckNotNull(x, expertIds, groupEp, yOut, expandIdxOut, commCmdInfoOut), ACLNN_ERR_PARAM_NULLPTR);
75 if (strnlen(groupEp, HCCL_GROUP_NAME_MAX) >= HCCL_GROUP_NAME_MAX) {72 if (strnlen(groupEp, HCCL_GROUP_NAME_MAX) >= HCCL_GROUP_NAME_MAX) {
@@ -89,9 +86,7 @@ aclnnStatus aclnnMoeDistributeDispatchSetupGetWorkspaceSize(
89 aclTensor *expandIdxOut, aclTensor *commCmdInfoOut, uint64_t *workspaceSize, aclOpExecutor **executor)86 aclTensor *expandIdxOut, aclTensor *commCmdInfoOut, uint64_t *workspaceSize, aclOpExecutor **executor)
90{87{
91 OP_LOGD("aclnnMoeDistributeDispatchSetupGetWorkspaceSize start");88 OP_LOGD("aclnnMoeDistributeDispatchSetupGetWorkspaceSize start");
92- auto ret_param =89+ auto ret_param = CheckParams(x, expertIds, groupEp, yOut, expandIdxOut, commCmdInfoOut);
93- CheckParams(x, expertIds, scalesOptional, groupEp, epWorldSize, epRankId, expertShardType, shareExpertRankNum,
94- moeExpertNum, quantMode, globalBs, commType, commAlg, yOut, expandIdxOut, commCmdInfoOut);
95 CHECK_RET(ret_param == ACLNN_SUCCESS, ret_param);90 CHECK_RET(ret_param == ACLNN_SUCCESS, ret_param);
96 91 
97 aclnnStatus ret = aclnnInnerMoeDistributeDispatchSetupGetWorkspaceSize(92 aclnnStatus ret = aclnnInnerMoeDistributeDispatchSetupGetWorkspaceSize(
Mmc2/moe_distribute_dispatch_teardown/op_api/aclnn_moe_distribute_dispatch_teardown.cpp+4-6
@@ -56,10 +56,9 @@ static bool CheckNotNull(const aclTensor *x, const aclTensor *y, const aclTensor
56 56 
57// check invalid parameters57// check invalid parameters
58static aclnnStatus CheckParams(const aclTensor *x, const aclTensor *y, const aclTensor *expertIds,58static aclnnStatus CheckParams(const aclTensor *x, const aclTensor *y, const aclTensor *expertIds,
59- const aclTensor *commCmdInfo, const char *groupEp, int64_t expertShardType,59+ const aclTensor *commCmdInfo, const char *groupEp, int64_t quantMode,
60- int64_t sharedExpertRankNum, int64_t moeExpertNum, int64_t quantMode, int64_t globalBs,60+ aclTensor *expandXOut, aclTensor *dynamicScalesOut, aclTensor *assistInfoForCombineOut,
61- int64_t commType, char *commAlg, aclTensor *expandXOut, aclTensor *dynamicScalesOut,61+ aclTensor *expertTokenNumsOut)
62- aclTensor *assistInfoForCombineOut, aclTensor *expertTokenNumsOut)
63{62{
64 CHECK_RET(CheckNotNull(x, y, expertIds, commCmdInfo, groupEp, expandXOut, dynamicScalesOut, assistInfoForCombineOut,63 CHECK_RET(CheckNotNull(x, y, expertIds, commCmdInfo, groupEp, expandXOut, dynamicScalesOut, assistInfoForCombineOut,
65 expertTokenNumsOut),64 expertTokenNumsOut),
@@ -94,8 +93,7 @@ aclnnStatus aclnnMoeDistributeDispatchTeardownGetWorkspaceSize(
94 "unsupported npuArch");93 "unsupported npuArch");
95 return ACLNN_ERR_PARAM_INVALID;94 return ACLNN_ERR_PARAM_INVALID;
96 }95 }
97- auto ret_param = CheckParams(x, y, expertIds, commCmdInfo, groupEp, expertShardType, sharedExpertRankNum,96+ auto ret_param = CheckParams(x, y, expertIds, commCmdInfo, groupEp, quantMode, expandXOut, dynamicScalesOut,
98- moeExpertNum, quantMode, globalBs, commType, commAlg, expandXOut, dynamicScalesOut,
99 assistInfoForCombineOut, expertTokenNumsOut);97 assistInfoForCombineOut, expertTokenNumsOut);
100 CHECK_RET(ret_param == ACLNN_SUCCESS, ret_param);98 CHECK_RET(ret_param == ACLNN_SUCCESS, ret_param);
101 99 
Mmc2/moe_distribute_dispatch_v2/op_api/moe_distribute_dispatch_v2_base.cpp+4-5
@@ -86,7 +86,7 @@ bool DispatchCheckNotNull(const aclTensor *x, const aclTensor *expertIds, const
86aclnnStatus DispatchCheckParams(const aclTensor *x, const aclTensor *expertIds, const aclTensor *scalesOptional,86aclnnStatus DispatchCheckParams(const aclTensor *x, const aclTensor *expertIds, const aclTensor *scalesOptional,
87 const char *groupEp, const char *groupTp, int64_t quantMode, aclTensor *expandX,87 const char *groupEp, const char *groupTp, int64_t quantMode, aclTensor *expandX,
88 aclTensor *dynamicScales, aclTensor *assistInfoForCombine, aclTensor *expertTokensNums,88 aclTensor *dynamicScales, aclTensor *assistInfoForCombine, aclTensor *expertTokensNums,
89- aclTensor *epRecvCounts, aclTensor *tpRecvCounts)89+ aclTensor *epRecvCounts)
90{90{
91 CHECK_RET(DispatchCheckNotNull(x, expertIds, groupEp, groupTp, expandX, dynamicScales, assistInfoForCombine,91 CHECK_RET(DispatchCheckNotNull(x, expertIds, groupEp, groupTp, expandX, dynamicScales, assistInfoForCombine,
92 expertTokensNums, epRecvCounts),92 expertTokensNums, epRecvCounts),
@@ -100,7 +100,7 @@ aclnnStatus DispatchCheckParams(const aclTensor *x, const aclTensor *expertIds,
100 OP_LOGD("quantMode = 2, dynamicScales can't be null");100 OP_LOGD("quantMode = 2, dynamicScales can't be null");
101 CHECK_RET(dynamicScales != nullptr, ACLNN_ERR_PARAM_NULLPTR);101 CHECK_RET(dynamicScales != nullptr, ACLNN_ERR_PARAM_NULLPTR);
102 }102 }
103- if (x->GetDataType() == ACL_INT32) {103+ if (static_cast<int32_t>(x->GetDataType()) == static_cast<int32_t>(ACL_INT32)) {
104 OP_LOGD("x dtype is INT32, scales and dynamicScales can't be null");104 OP_LOGD("x dtype is INT32, scales and dynamicScales can't be null");
105 CHECK_RET(scalesOptional != nullptr, ACLNN_ERR_PARAM_NULLPTR);105 CHECK_RET(scalesOptional != nullptr, ACLNN_ERR_PARAM_NULLPTR);
106 CHECK_RET(dynamicScales != nullptr, ACLNN_ERR_PARAM_NULLPTR);106 CHECK_RET(dynamicScales != nullptr, ACLNN_ERR_PARAM_NULLPTR);
@@ -164,9 +164,8 @@ aclnnStatus aclnnMoeDistributeDispatchGetWorkspaceSizeBase(
164 OP_LOGD("enter to the aclnnMoeDistributeDispatchGetWorkspaceSizeBase\n");164 OP_LOGD("enter to the aclnnMoeDistributeDispatchGetWorkspaceSizeBase\n");
165 const static bool is910B = GetCurrentPlatformInfo().GetSocVersion() == SocVersion::ASCEND910B;165 const static bool is910B = GetCurrentPlatformInfo().GetSocVersion() == SocVersion::ASCEND910B;
166 const static bool is950 = GetCurrentPlatformInfo().GetCurNpuArch() == NpuArch::DAV_3510;166 const static bool is950 = GetCurrentPlatformInfo().GetCurNpuArch() == NpuArch::DAV_3510;
167- auto retParam =167+ auto retParam = DispatchCheckParams(x, expertIds, scalesOptional, groupEp, groupTp, quantMode, expandXOut,
168- DispatchCheckParams(x, expertIds, scalesOptional, groupEp, groupTp, quantMode, expandXOut, dynamicScalesOut,168+ dynamicScalesOut, assistInfoForCombineOut, expertTokenNumsOut, epRecvCountsOut);
169- assistInfoForCombineOut, expertTokenNumsOut, epRecvCountsOut, tpRecvCountsOut);
170 CHECK_RET(retParam == ACLNN_SUCCESS, retParam);169 CHECK_RET(retParam == ACLNN_SUCCESS, retParam);
171 170 
172 aclTensor *mc2Context = nullptr;171 aclTensor *mc2Context = nullptr;
Mmc2/moe_distribute_dispatch_v2/op_api/moe_distribute_dispatch_v2_base.h+1-1
@@ -45,7 +45,7 @@ ACLNN_API aclnnStatus DispatchCheckParams(const aclTensor *x, const aclTensor *e
45 const aclTensor *scalesOptional, const char *groupEp, const char *groupTp,45 const aclTensor *scalesOptional, const char *groupEp, const char *groupTp,
46 int64_t quantMode, aclTensor *expandX, aclTensor *dynamicScales,46 int64_t quantMode, aclTensor *expandX, aclTensor *dynamicScales,
47 aclTensor *assistInfoForCombine, aclTensor *expertTokensNums,47 aclTensor *assistInfoForCombine, aclTensor *expertTokensNums,
48- aclTensor *epRecvCounts, aclTensor *tpRecvCounts);48+ aclTensor *epRecvCounts);
49 49 
50ACLNN_API aclnnStatus aclnnMoeDistributeDispatchBase(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor,50ACLNN_API aclnnStatus aclnnMoeDistributeDispatchBase(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor,
51 aclrtStream stream);51 aclrtStream stream);
Mmc2/moe_ep_combine_epilogue/op_api/aclnn_moe_ep_combine_epilogue.cpp+1-1
@@ -22,7 +22,7 @@ constexpr int64_t NETWORK_DIRECT = 0;
22constexpr int64_t NETWORK_HYBRID = 1;22constexpr int64_t NETWORK_HYBRID = 1;
23} // namespace23} // namespace
24 24 
25-static aclnnStatus CheckNotNull(const aclTensor *context, const aclTensor *topkIdx, aclTensor *combinedX)25+static aclnnStatus CheckNotNull(const aclTensor *context, const aclTensor *topkIdx, const aclTensor *combinedX)
26{26{
27 CHECK_RET(context != nullptr, ACLNN_ERR_PARAM_NULLPTR);27 CHECK_RET(context != nullptr, ACLNN_ERR_PARAM_NULLPTR);
28 CHECK_RET(topkIdx != nullptr, ACLNN_ERR_PARAM_NULLPTR);28 CHECK_RET(topkIdx != nullptr, ACLNN_ERR_PARAM_NULLPTR);
Mmc2/moe_ep_dispatch/op_api/aclnn_moe_ep_dispatch.cpp+26-33
@@ -23,9 +23,9 @@ constexpr int64_t NETWORK_HYBRID = 1;
23} // namespace23} // namespace
24 24 
25static aclnnStatus CheckNotNull(const aclTensor *context, const aclTensor *x, const aclTensor *topkIdx,25static aclnnStatus CheckNotNull(const aclTensor *context, const aclTensor *x, const aclTensor *topkIdx,
26- aclTensor *numRecvTokensPerRank, aclTensor *numRecvTokensPerExpert,26+ const aclTensor *numRecvTokensPerRank, const aclTensor *numRecvTokensPerExpert,
27- aclTensor *dstBufferSlotIdx, aclTensor *routeCount, aclTensor *routeDstScaleout,27+ const aclTensor *dstBufferSlotIdx, const aclTensor *routeCount,
28- aclTensor *routeScaleoutSlot)28+ const aclTensor *routeDstScaleout, const aclTensor *routeScaleoutSlot)
29{29{
30 CHECK_RET(context != nullptr, ACLNN_ERR_PARAM_NULLPTR);30 CHECK_RET(context != nullptr, ACLNN_ERR_PARAM_NULLPTR);
31 CHECK_RET(x != nullptr, ACLNN_ERR_PARAM_NULLPTR);31 CHECK_RET(x != nullptr, ACLNN_ERR_PARAM_NULLPTR);
@@ -40,10 +40,9 @@ static aclnnStatus CheckNotNull(const aclTensor *context, const aclTensor *x, co
40}40}
41 41 
42static aclnnStatus CheckParams(int64_t epWorldSize, int64_t epRankId, int64_t numExperts, int64_t numMaxTokensPerRank,42static aclnnStatus CheckParams(int64_t epWorldSize, int64_t epRankId, int64_t numExperts, int64_t numMaxTokensPerRank,
43- int64_t cclBufferSize, int64_t expertAlignment, bool doCpuSync, int64_t topoType,43+ int64_t cclBufferSize, bool doCpuSync, int64_t topoType, int64_t rankNumPerServer,
44- int64_t rankNumPerServer, const aclTensor *cachedDstSlotIdx,44+ const aclTensor *cachedDstSlotIdx, const aclTensor *cachedRouteCount,
45- const aclTensor *cachedRouteCount, const aclTensor *cachedRouteDstScaleout,45+ const aclTensor *cachedRouteDstScaleout, const aclTensor *cachedRouteScaleoutSlot)
46- const aclTensor *cachedRouteScaleoutSlot)
47{46{
48 CHECK_RET(epWorldSize > 1, ACLNN_ERR_PARAM_INVALID);47 CHECK_RET(epWorldSize > 1, ACLNN_ERR_PARAM_INVALID);
49 CHECK_RET(epRankId >= 0 && epRankId < epWorldSize, ACLNN_ERR_PARAM_INVALID);48 CHECK_RET(epRankId >= 0 && epRankId < epWorldSize, ACLNN_ERR_PARAM_INVALID);
@@ -56,13 +55,12 @@ static aclnnStatus CheckParams(int64_t epWorldSize, int64_t epRankId, int64_t nu
56 55 
57 bool cachedMode = (cachedDstSlotIdx != nullptr);56 bool cachedMode = (cachedDstSlotIdx != nullptr);
58 CHECK_RET(!(cachedMode && doCpuSync), ACLNN_ERR_PARAM_INVALID);57 CHECK_RET(!(cachedMode && doCpuSync), ACLNN_ERR_PARAM_INVALID);
59- bool anyCachedRoute = cachedRouteCount != nullptr || cachedRouteDstScaleout != nullptr ||58+ bool anyCachedRoute =
60- cachedRouteScaleoutSlot != nullptr;59+ cachedRouteCount != nullptr || cachedRouteDstScaleout != nullptr || cachedRouteScaleoutSlot != nullptr;
61- bool allCachedRoute = cachedRouteCount != nullptr && cachedRouteDstScaleout != nullptr &&60+ bool allCachedRoute =
62- cachedRouteScaleoutSlot != nullptr;61+ cachedRouteCount != nullptr && cachedRouteDstScaleout != nullptr && cachedRouteScaleoutSlot != nullptr;
63 CHECK_RET(!anyCachedRoute || allCachedRoute, ACLNN_ERR_PARAM_INVALID);62 CHECK_RET(!anyCachedRoute || allCachedRoute, ACLNN_ERR_PARAM_INVALID);
64- bool hybridCachedMode =63+ bool hybridCachedMode = cachedMode && topoType == NETWORK_HYBRID && epWorldSize / rankNumPerServer > 1;
65- cachedMode && topoType == NETWORK_HYBRID && epWorldSize / rankNumPerServer > 1;
66 CHECK_RET(!hybridCachedMode || allCachedRoute, ACLNN_ERR_PARAM_INVALID);64 CHECK_RET(!hybridCachedMode || allCachedRoute, ACLNN_ERR_PARAM_INVALID);
67 65 
68 return ACLNN_SUCCESS;66 return ACLNN_SUCCESS;
@@ -81,19 +79,15 @@ enum NnopbaseHcclServerType {
81 79 
82extern "C" void __attribute__((weak)) NnopbaseSetHcclServerType(void *executor, NnopbaseHcclServerType sType);80extern "C" void __attribute__((weak)) NnopbaseSetHcclServerType(void *executor, NnopbaseHcclServerType sType);
83 81 
84-aclnnStatus aclnnMoeEpDispatchGetWorkspaceSize(const aclTensor *context, const aclTensor *x, const aclTensor *topkIdx,82+aclnnStatus aclnnMoeEpDispatchGetWorkspaceSize(
85- const aclTensor *topkWeights, const aclTensor *scales,83+ const aclTensor *context, const aclTensor *x, const aclTensor *topkIdx, const aclTensor *topkWeights,
86- const aclTensor *cachedDstSlotIdx, const aclTensor *cachedRouteCount,84+ const aclTensor *scales, const aclTensor *cachedDstSlotIdx, const aclTensor *cachedRouteCount,
87- const aclTensor *cachedRouteDstScaleout,85+ const aclTensor *cachedRouteDstScaleout, const aclTensor *cachedRouteScaleoutSlot, int64_t epWorldSize,
88- const aclTensor *cachedRouteScaleoutSlot, int64_t epWorldSize,86+ int64_t epRankId, int64_t numExperts, int64_t numMaxTokensPerRank, int64_t cclBufferSize, int64_t expertAlignment,
89- int64_t epRankId, int64_t numExperts, int64_t numMaxTokensPerRank,87+ bool doCpuSync, int64_t hostPinnedCounterAddr, int64_t topoType, int64_t rankNumPerServer,
90- int64_t cclBufferSize, int64_t expertAlignment, bool doCpuSync,88+ aclTensor *numRecvTokensPerRank, aclTensor *numRecvTokensPerExpert, aclTensor *dstBufferSlotIdx,
91- int64_t hostPinnedCounterAddr, int64_t topoType,89+ aclTensor *routeCount, aclTensor *routeDstScaleout, aclTensor *routeScaleoutSlot, uint64_t *workspaceSize,
92- int64_t rankNumPerServer, aclTensor *numRecvTokensPerRank,90+ aclOpExecutor **executor)
93- aclTensor *numRecvTokensPerExpert, aclTensor *dstBufferSlotIdx,
94- aclTensor *routeCount, aclTensor *routeDstScaleout,
95- aclTensor *routeScaleoutSlot, uint64_t *workspaceSize,
96- aclOpExecutor **executor)
97{91{
98 OP_LOGD("aclnnMoeEpDispatch WorkspaceSize start");92 OP_LOGD("aclnnMoeEpDispatch WorkspaceSize start");
99 93 
@@ -101,17 +95,16 @@ aclnnStatus aclnnMoeEpDispatchGetWorkspaceSize(const aclTensor *context, const a
101 routeCount, routeDstScaleout, routeScaleoutSlot);95 routeCount, routeDstScaleout, routeScaleoutSlot);
102 CHECK_RET(retNotNull == ACLNN_SUCCESS, retNotNull);96 CHECK_RET(retNotNull == ACLNN_SUCCESS, retNotNull);
103 97 
104- auto retParams = CheckParams(epWorldSize, epRankId, numExperts, numMaxTokensPerRank, cclBufferSize,98+ auto retParams = CheckParams(epWorldSize, epRankId, numExperts, numMaxTokensPerRank, cclBufferSize, doCpuSync,
105- expertAlignment, doCpuSync, topoType, rankNumPerServer, cachedDstSlotIdx,99+ topoType, rankNumPerServer, cachedDstSlotIdx, cachedRouteCount, cachedRouteDstScaleout,
106- cachedRouteCount, cachedRouteDstScaleout, cachedRouteScaleoutSlot);100+ cachedRouteScaleoutSlot);
107 CHECK_RET(retParams == ACLNN_SUCCESS, retParams);101 CHECK_RET(retParams == ACLNN_SUCCESS, retParams);
108 102 
109 aclnnStatus getWorkspaceSizesRes = aclnnInnerMoeEpDispatchGetWorkspaceSize(103 aclnnStatus getWorkspaceSizesRes = aclnnInnerMoeEpDispatchGetWorkspaceSize(
110 context, x, topkIdx, topkWeights, scales, cachedDstSlotIdx, cachedRouteCount, cachedRouteDstScaleout,104 context, x, topkIdx, topkWeights, scales, cachedDstSlotIdx, cachedRouteCount, cachedRouteDstScaleout,
111- cachedRouteScaleoutSlot, epWorldSize, epRankId, numExperts, numMaxTokensPerRank, cclBufferSize,105+ cachedRouteScaleoutSlot, epWorldSize, epRankId, numExperts, numMaxTokensPerRank, cclBufferSize, expertAlignment,
112- expertAlignment, doCpuSync, hostPinnedCounterAddr, topoType, rankNumPerServer, numRecvTokensPerRank,106+ doCpuSync, hostPinnedCounterAddr, topoType, rankNumPerServer, numRecvTokensPerRank, numRecvTokensPerExpert,
113- numRecvTokensPerExpert, dstBufferSlotIdx, routeCount, routeDstScaleout, routeScaleoutSlot, workspaceSize,107+ dstBufferSlotIdx, routeCount, routeDstScaleout, routeScaleoutSlot, workspaceSize, executor);
114- executor);
115 108 
116 return getWorkspaceSizesRes;109 return getWorkspaceSizesRes;
117}110}
Mmc2/moe_ep_dispatch_epilogue/op_api/aclnn_moe_ep_dispatch_epilogue.cpp+9-12
@@ -24,8 +24,8 @@ constexpr int64_t NETWORK_HYBRID = 1;
24 24 
25static aclnnStatus CheckNotNull(const aclTensor *context, const aclTensor *dstBufferSlotIdx,25static aclnnStatus CheckNotNull(const aclTensor *context, const aclTensor *dstBufferSlotIdx,
26 const aclTensor *numRecvTokensPerRank, const aclTensor *numRecvTokensPerExpert,26 const aclTensor *numRecvTokensPerRank, const aclTensor *numRecvTokensPerExpert,
27- aclTensor *recvX, aclTensor *recvSrcMetadata, aclTensor *recvTopkWeights,27+ const aclTensor *recvX, const aclTensor *recvSrcMetadata,
28- aclTensor *recvScales)28+ const aclTensor *recvTopkWeights, const aclTensor *recvScales)
29{29{
30 CHECK_RET(context != nullptr, ACLNN_ERR_PARAM_NULLPTR);30 CHECK_RET(context != nullptr, ACLNN_ERR_PARAM_NULLPTR);
31 CHECK_RET(dstBufferSlotIdx != nullptr, ACLNN_ERR_PARAM_NULLPTR);31 CHECK_RET(dstBufferSlotIdx != nullptr, ACLNN_ERR_PARAM_NULLPTR);
@@ -62,14 +62,12 @@ aclnnStatus MoeEpDispatchEpilogueGetWorkspaceSize(
62 aclTensor *recvTopkWeights, aclTensor *recvScales, uint64_t *workspaceSize, aclOpExecutor **executor)62 aclTensor *recvTopkWeights, aclTensor *recvScales, uint64_t *workspaceSize, aclOpExecutor **executor)
63{63{
64 OP_LOGD("MoeEpDispatchEpilogue", "Begin to do MoeEpDispatchEpilogueGetWorkspaceSize");64 OP_LOGD("MoeEpDispatchEpilogue", "Begin to do MoeEpDispatchEpilogueGetWorkspaceSize");
65- auto retNotNull =65+ auto retNotNull = CheckNotNull(context, dstBufferSlotIdx, numRecvTokensPerRank, numRecvTokensPerExpert, recvX,
66- CheckNotNull(context, dstBufferSlotIdx, numRecvTokensPerRank, numRecvTokensPerExpert, recvX, recvSrcMetadata,66+ recvSrcMetadata, recvTopkWeights, recvScales);
67- recvTopkWeights, recvScales);
68 CHECK_RET(retNotNull == ACLNN_SUCCESS, retNotNull);67 CHECK_RET(retNotNull == ACLNN_SUCCESS, retNotNull);
69 68 
70 auto retParams =69 auto retParams =
71- CheckParams(epWorldSize, epRankId, numExperts, numMaxTokensPerRank, cclBufferSize, topoType,70+ CheckParams(epWorldSize, epRankId, numExperts, numMaxTokensPerRank, cclBufferSize, topoType, rankNumPerServer);
72- rankNumPerServer);
73 CHECK_RET(retParams == ACLNN_SUCCESS, retParams);71 CHECK_RET(retParams == ACLNN_SUCCESS, retParams);
74 72 
75 return aclnnInnerMoeEpDispatchEpilogueGetWorkspaceSize(73 return aclnnInnerMoeEpDispatchEpilogueGetWorkspaceSize(
@@ -85,11 +83,10 @@ aclnnStatus aclnnMoeEpDispatchEpilogueGetWorkspaceSize(
85 int64_t topoType, int64_t rankNumPerServer, aclTensor *recvX, aclTensor *recvSrcMetadata,83 int64_t topoType, int64_t rankNumPerServer, aclTensor *recvX, aclTensor *recvSrcMetadata,
86 aclTensor *recvTopkWeights, aclTensor *recvScales, uint64_t *workspaceSize, aclOpExecutor **executor)84 aclTensor *recvTopkWeights, aclTensor *recvScales, uint64_t *workspaceSize, aclOpExecutor **executor)
87{85{
88- return MoeEpDispatchEpilogueGetWorkspaceSize(context, dstBufferSlotIdx, numRecvTokensPerRank,86+ return MoeEpDispatchEpilogueGetWorkspaceSize(
89- numRecvTokensPerExpert, cachedRecvSrcMetadata, epWorldSize, epRankId,87+ context, dstBufferSlotIdx, numRecvTokensPerRank, numRecvTokensPerExpert, cachedRecvSrcMetadata, epWorldSize,
90- numExperts, numMaxTokensPerRank, cclBufferSize, hasTopkWeights,88+ epRankId, numExperts, numMaxTokensPerRank, cclBufferSize, hasTopkWeights, topoType, rankNumPerServer, recvX,
91- topoType, rankNumPerServer, recvX, recvSrcMetadata, recvTopkWeights,89+ recvSrcMetadata, recvTopkWeights, recvScales, workspaceSize, executor);
92- recvScales, workspaceSize, executor);
93}90}
94 91 
95enum NnopbaseHcclServerType {92enum NnopbaseHcclServerType {