swin的mmcv血和泪教训，8.6算力的3070不要装MMCV 1.4.0!

会报错：RuntimeError: CUDA error: no kernel image is available for execution on the device

GIT上的原贴在这里：

https://github.com/open-mmlab/mmcv/issues/1623

以下是swin的mvcc部分安装代码，3070卡：

pip install torch==1.10.0+cu113   torchvision==0.11.0+cu113 torchaudio==0.10.0+cu113  -f https://download.pytorch.org/whl/torch_stable.html

pip install mmcv-full==1.4.1 -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.10/index.html
 

pip install absl-py astunparse flatbuffers gast==0.4.0 google-pasta grpcio h5py keras libclang opt-einsum protobuf==3.19   tensorboard tensorflow-estimator==2.11 tensorflow-io-gcs-filesystem wrapt

pip3 install cython==0.29.33 -i http://mirrors.aliyun.com/pypi/simple/ --trusted-host mirrors.aliyun.com
pip install mmpycocotools

pip install -r requirements.txt
 

cd 刚刚mmcv目录位置 
//环境配置
$env:TORCH_CUDA_ARCH_LIST="8.6" 
$env:MMCV_WITH_OPS = 1
$env:MAX_JOBS = 6