compile

zzx-study · zzx-study · commit 31b4061e6357 · 2026-02-09T16:54:22.000+08:00
diff --git a/csrc/deepep/ops2/op_kernel/a2/moe_distribute_dispatch_a2_pipeline.h b/csrc/deepep/ops2/op_kernel/a2/moe_distribute_dispatch_a2_pipeline.h
@@ -503,7 +503,7 @@ __aicore__ inline void MoeDistributeDispatchA2Pipeline<TemplateMC2TypeA2Pipeline
                     continue;
                 }
                 GlobalTensor<uint8_t> dstRankRecvRingU8Tensor;
-                dstRankRecvRingU8Tensor.SetGlobalBuffer((__gm__ uint8_t *) (hccl_.GetWindowsInAddr(rankId_)) + halfWinSize_ / 2);
+                dstRankRecvRingU8Tensor.SetGlobalBuffer((__gm__ uint8_t *) (hccl_.GetWindowsInAddr(localDstRank)) + halfWinSize_ / 2);
                 LocalTensor<uint32_t> localHccsHeadTailTensor;
                 GlobalTensor<uint32_t> globalHccsHeadTailTensor;
                 globalHccsHeadTailTensor.SetGlobalBuffer((__gm__ uint32_t *)hccsHeadTailGM[localDstRank]);
@@ -525,7 +525,7 @@ __aicore__ inline void MoeDistributeDispatchA2Pipeline<TemplateMC2TypeA2Pipeline
                     }
                 }
                 SyncFunc<AscendC::HardEvent::S_MTE3>();
-                DataCopyPad(dstRankRecvRingU8Tensor[tokenStructLen_ * index], tokenStructInRdmaTensor_, 
+                DataCopyPad(dstRankRecvRingU8Tensor[hccsItemNum * localDstRank + tokenStructLen_ * index], tokenStructInRdmaTensor_, 
                 tokenStructParams);
                 DataCopyPad(rdmaRecvRingU8Tensor_[(i * rdmaItemNum + rdmaHead) * tokenStructLen_], tokenStructInRdmaTensor_, 
                 tokenStructParams);

Original file line number	Diff line number	Diff line change
`@@ -503,7 +503,7 @@ __aicore__ inline void MoeDistributeDispatchA2Pipeline<TemplateMC2TypeA2Pipeline`
`503`	`503`	`continue;`
`504`	`504`	`}`
`505`	`505`	`GlobalTensor<uint8_t> dstRankRecvRingU8Tensor;`
`506`		`- dstRankRecvRingU8Tensor.SetGlobalBuffer((__gm__ uint8_t *) (hccl_.GetWindowsInAddr(rankId_)) + halfWinSize_ / 2);`
	`506`	`+ dstRankRecvRingU8Tensor.SetGlobalBuffer((__gm__ uint8_t *) (hccl_.GetWindowsInAddr(localDstRank)) + halfWinSize_ / 2);`
`507`	`507`	`LocalTensor<uint32_t> localHccsHeadTailTensor;`
`508`	`508`	`GlobalTensor<uint32_t> globalHccsHeadTailTensor;`
`509`	`509`	`globalHccsHeadTailTensor.SetGlobalBuffer((__gm__ uint32_t *)hccsHeadTailGM[localDstRank]);`
`@@ -525,7 +525,7 @@ __aicore__ inline void MoeDistributeDispatchA2Pipeline<TemplateMC2TypeA2Pipeline`
`525`	`525`	`}`
`526`	`526`	`}`
`527`	`527`	`SyncFunc<AscendC::HardEvent::S_MTE3>();`
`528`		`- DataCopyPad(dstRankRecvRingU8Tensor[tokenStructLen_ * index], tokenStructInRdmaTensor_,`
	`528`	`+ DataCopyPad(dstRankRecvRingU8Tensor[hccsItemNum * localDstRank + tokenStructLen_ * index], tokenStructInRdmaTensor_,`
`529`	`529`	`tokenStructParams);`
`530`	`530`	`DataCopyPad(rdmaRecvRingU8Tensor_[(i * rdmaItemNum + rdmaHead) * tokenStructLen_], tokenStructInRdmaTensor_,`
`531`	`531`	`tokenStructParams);`