nagadomi/torch7 repository overview

Torch7 built with CUDA10.2, cuDNN v7.5, Ubuntu 18.04 (based on nvidia/cuda:10.2-cudnn7-devel-ubuntu18.04).

The image is not official Torch7, it is created from my personal repository maintained for CUDA10. https://github.com/nagadomi/distro⁠

TORCH_CUDA_ARCH_LIST: Kepler, Kepler+Tegra, Kepler+Tesla, Maxwell, Maxwell+Tegra, Pascal, Volta, 7.5+PTX(Turing)

For modern devices, it works with JIT compile from PTX.

Note that running this container in without JIT caching⁠ is very slow, which is what would happen if you use docker. For a workaround, you can mount a host volume to the CUDA_CACHE_PATH, for instance,

docker run --gpus all -v $PWD/ComputeCache:/root/.nv/ComputeCache -it nagadomi/torch7:latest