已合并
修复WeightQuantBatchMatmulV2 W4A16 PerGroup-ND场景OOM问题 #3895
sunduiyang创建于 4月16日
修复WeightQuantBatchMatmulV2 W4A16 PerGroup-ND场景OOM问题 #3895
已合并
共 1 个文件变更+1-1
Mmatmul/weight_quant_batch_matmul_v2/op_kernel/arch35/weight_quant_batch_matmul_v2_reg_base_common.h+1-1
| @@ -1704,7 +1704,7 @@ __aicore__ inline void WeightQuantBatchMatmulV2RegBaseCommonKernel< | |||
| 1704 | 1704 | ||
| 1705 | WaitFlag<HardEvent::V_MTE2>(ubInBufIdx_); | 1705 | WaitFlag<HardEvent::V_MTE2>(ubInBufIdx_); |
| 1706 | 1706 | ||
| 1707 | - int32_t bubNLen = CeilAlign(Min(vecNBL1Len_, static_cast<int64_t>(tiling_->nBubSize)), BLOCK_CUBE); | 1707 | + int32_t bubNLen = Min(vecNBL1Len_, static_cast<int64_t>(tiling_->nBubSize)); |
| 1708 | CopyInWeight(kBL1Offset, nBL1Offset, bubKLen, bubNLen); | 1708 | CopyInWeight(kBL1Offset, nBL1Offset, bubKLen, bubNLen); |
| 1709 | CopyInScaleOffset(nBL1Offset, bubNLen, 0, 0, kBL1Offset, 1); | 1709 | CopyInScaleOffset(nBL1Offset, bubNLen, 0, 0, kBL1Offset, 1); |
| 1710 | SetFlag<HardEvent::MTE2_V>(ubInBufIdx_); | 1710 | SetFlag<HardEvent::MTE2_V>(ubInBufIdx_); |