Open
[确定性性能] sparse_lightning_indexer_kl_loss_grad:网络用例确定性/非确定性性能对比 #566
Alice_zhouAdd on 2 days ago
2 days ago Modify description of issue
2 days ago Modify description of issue
2 days ago Modify description of issue
1 day ago Modify description of issue
1 day ago Modify description of issue
1 day ago Modify description of issue
1 day ago Modify description of issue
背景
sparse_lightning_indexer_kl_loss_grad。sparse_lightning_indexer_kl_loss_grad_network_9582.xlsx。关键指标
deterministic_check=Fail可稳定复现。网络用例结果(有效对 Top 10)
确定性与 AICORE 问题(已拆分为独立 issue)
执行步骤
cd <WORKDIR>/TuringBigKernel/FusionOp_Test export PATH=<CONDA_BIN>:$PATH source activate <CONDA_ENV> source <CANN_TOOLKIT>/set_env.sh python3 dailyauto_fusion_david_fast.py \ --timeout=1800 \ --casefile=<CASE_DIR>/sparse_lightning_indexer_kl_loss_grad_network_9582.xlsx \ --runtype=2 --deviceid=<DEVICE_ID> --need_lock_freq=False \ --sheetname=Sheet1 --ti=0-26附件
备注/限制
Actual_backward_kernel_time;路径使用占位符。