HERMETIC_CUDA_COMPUTE_CAPABILITIES semantics

[Joel Berkeley]

Hi,

How does HERMETIC_CUDA_COMPUTE_CAPABILITIES work? I'm trying to build target //xla/pjrt/c:pjrt_c_api_gpu_plugin.so for as broad a set of GPUs as possible, but for fixed CUDA and CUDNN versions. What should I specify for HERMETIC_CUDA_COMPUTE_CAPABILITIES? Would be ideal if I could support everything from compute capability 6.1 to 12.0. Is that HERMETIC_CUDA_COMPUTE_CAPABILITIES=6.1,6.2,7.0,7.2,8.0,8.6,8.7,8.9,9.0,10.0,10.3,11.0,12.0

Thanks,

Joel