EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[Bugfix][DispatchFFNCombine] resolve vec error caused by unaligned UB access (#6707)

Browse Source 
### What this PR does / why we need it?
1. Fix a vec error caused by unaligned UB accesss in the
DispatchFFNCombine;
2. Fix expert_token_nums tensor defined on host instead of NPU in
moe_comm_method.py
3. Fix multi-core copy issue of expert_token_nums in dispatchffnCombine
op (single aiv copy is sufficient)

### Does this PR introduce _any_ user-facing change?

No, this PR does not introduce any user-facing changes. The fix only
addresses internal memory access logic and does not modify any public
APIs, interfaces, or user-visible behaviors.

### How was this patch tested?

`export VLLM_ASCEND_ENABLE_FUSED_MC2=1`

vLLM version: v0.15.0

- vLLM version: v0.15.0
- vLLM main:
9562912cea

Signed-off-by: xulei_ict <xulei292@huawei.com>
Co-authored-by: xulei_ict <xulei292@huawei.com>
This commit is contained in:
xulei
2026-02-14 10:32:50 +08:00
committed by GitHub
parent e2175d9c7e
commit 1e77077788
3 changed files with 19 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
csrc/dispatch_ffn_combine/op_kernel/dispatch_ffn_combine_kernel.hpp
View File

@@ -779,10 +779,10 @@ private:
AscendC::GlobalTensor<int32_t> ExpertTokenNums;
ExpertTokenNums.SetGlobalBuffer(reinterpret_cast<__gm__ int32_t*>(params.ptrExpertTokenNums));
AscendC::GlobalTensor<int32_t> LcalCumsumMM;
LcalCumsumMM.SetGlobalBuffer(reinterpret_cast<__gm__ int32_t*>(workspaceInfo.ptrcumsumMM + (params.EP - 1) * params.expertPerRank * sizeof(int32_t)));
CopyGMToGM(ExpertTokenNums, LcalCumsumMM, params.expertPerRank, params.ubMoveNum);
AscendC::SyncAll<true>();
if(coreIdx == 0)
{
CopyGMToGM(ExpertTokenNums, cumsumMM[(params.EP - 1) * params.expertPerRank], params.expertPerRank, params.ubMoveNum);
}
uint16_t syncgmm1Idx = 0;
AscendC::CrossCoreSetFlag<0x2, PIPE_MTE3>(syncgmm1Idx / CROSS_CORE_FLAG_MAX_SET_COUNT);
syncgmm1Idx++;
@@ -921,11 +921,11 @@ private:
AscendC::LocalTensor<float> statusTensor = resource.ubBuf.template GetBufferByByte<float>(uboffset);
uboffset += sendRankNum_ * UB_ALIGN;
AscendC::LocalTensor<float> gatherMaskOutTensor = resource.ubBuf.template GetBufferByByte<float>(uboffset);
uboffset += params.EP * sizeof(float);
uboffset += AlignUp(params.EP * sizeof(float), 32);
AscendC::LocalTensor<uint32_t> gatherTmpTensor = resource.ubBuf.template GetBufferByByte<uint32_t>(uboffset);
uboffset += sizeof(uint32_t);
uboffset += AlignUp(sizeof(uint32_t), 32);
AscendC::LocalTensor<float> statusSumOutTensor = resource.ubBuf.template GetBufferByByte<float>(uboffset);
uboffset += sizeof(float);
uboffset += AlignUp(sizeof(float), 32);
shmem.CrossRankSyncV2Wait(statusTensor, gatherMaskOutTensor, gatherTmpTensor, statusSumOutTensor);
MoeTokenUnpermuteTilingData tilingData;
MoeTokenUnpermuteTiling(params.problemShape.m() * params.topK, n2, params.topK, tilingData, coreNum / 2);