search

PaddlePaddle / Paddle

PArallel Distributed Deep LEarning: Machine Learning Framework from Industrial Practice （『飞桨』核心框架，深度学习&机器学习高性能单机、分布式训练和跨平台部署）
http://www.paddlepaddle.org/
Apache License 2.0
22.28k stars 5.61k forks source link

[CINN]fix layer_norm bug in combinatorial operator #69553

Closed zhanghonggeng closed 20 hours ago

zhanghonggeng commented 2 days ago

PR Category

Performance Optimization

PR Types

Performance

Description

在cinn前执行AutoMixedPrecisionPass后layer_norm的三个输入会添加cast算子转换成fp16，同时三个输出也变成了fp16。而组合算子拆分后的逻辑中后两个输出是fp32，导致check_decomp_outputs中检查错误 PreconditionNotMetError: [Prim] For op pd_op.layer_norm, its origin 1-index output dtype float16 is not equal to decomp output dtype float32 Pcard-67164

paddle-bot[bot] commented 2 days ago

你的PR提交成功，感谢你对开源项目的贡献! 请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。 Your PR has been submitted. Thanks for your contribution! Please wait for the result of CI firstly. See Paddle CI Manual for details.