Pytorch1.7适配cuda篇

Pytorch1.7适配cuda篇

由于3090、A10以后的显卡使用安培Ampere架构，不再支持cuda11以下版本，导致cuda、pytorch必须升级，进而导致过去开的的模型需要重新适配。根据适配过程中遇见的问题，做优化记录（持续更新）…

服务器环境

cuda: 11.2
cudnn: 8.0
pytorch: 1.7.1

适配记录

1、报错：error: identifier “AT_CHECK” is undefined

 error: identifier “AT_CHECK” is undefined

解决方法：找到报错的.cpp、.cu、.h文件，搜索并替换

“AT_CHECK”修改为“TORCH_CHECK”

2、报错：type().is_cuda()

 error: identifier “type().is_cuda()” is undefined

解决方法：找到报错的.cpp、.cu、.h文件，搜索并替换

“type().is_cuda()”修改为“is_cuda()”

3、报错： error: ‘THCState_getCurrentStream’ was not declared in this scope

  error: ‘THCState_getCurrentStream’ was not declared in this scope

解决方法：找到报错的.cpp、.cu、.h文件，搜索并替换

 cudaStream_t stream = THCState_getCurrentStream(state);修改为“cudaStream_t stream = c10::cuda::getCurrentCUDAStream();”

4、报错：error: ‘getCurrentCUDAStream’ is not a member of ‘c10::cuda’

error: ‘getCurrentCUDAStream’ is not a member of ‘c10::cuda’

解决方法：#include <c10/cuda/CUDACachingAllocator.h>

#include <c10/cuda/CUDACachingAllocator.h>