rtx3080 RuntimeError: CUDA error: no kernel image is available for execution on the device

[qq260612718]

My experimental environment is cuda11.0 RTX3080 pytorch1.7.1
When I try to install apex,, the following problems occur:
nvcc fatal: Unsupported gpu architecture ‘compute_86’
I took this method of reducing the computing power, which showed that apex was successfully installed.
export TORCH_CUDA_ARCH_LIST="7.5"
But when I run the code, the following situation appears:
RuntimeError: CUDA error: no kernel image is available for execution on the device.
Can someone help me?I tried cuda10.0 and 10.2 lower versions and still have problems.

[ptrblck]

You would have to use TORCH_CUDA_ARCH_LIST="8.0" or ="8.6" to run apex on the 3080.
These architectures are supported in CUDA11, so your local 10.x toolkits won't work.

That being said, we recommend to use the native mixed-precision implementation as described here.