Hi, I’m using the nvidia docker runtime to launch containers from the nvidia docker base images.
Despite the fact that the applications are compiled with the appropriate fat binaries for the compute capability of the GPUs, all CUDA code is still JIT compiling using CICC.
Has anyone seen a similar issue? Why is it not detecting that the GPU the container uses is compatible with the binaries included? Is there a special location for the JIT cache for the nvidia docker runtime?