You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Harvey Chua edited this page Oct 12, 2026
·
3 revisions
Building the Image
Clone the repository and initialize its submodules (this also fetches the bundled MMseqs2 fork):
git clone https://github.com/bioinfodlsu/dummest.git
cd dummest
git submodule update --init --recursive
Both DUMMEST (portable AVX2) and the custom GPU MMseqs2 fork are built on the host and copied in, so the image carries no compiler and no CUDA toolkit:
./build_apptainer.sh
Host build tools required: apptainer, ninja, ccache (sudo apt install ninja-build ccache), plus cmake, the CUDA toolkit, and cargo (Rust).
build_apptainer.sh builds the fork with the host CUDA toolkit. The host must match the image base (Ubuntu noble, amd64), since the binaries are compiled on the host and run in the image. The Ubuntu version is checked at build time (SKIP_DISTRO_CHECK=1 to skip it), but the architecture is not — so if you skip the check, confirm your host is amd64 yourself.
Validate a build with the in-container suite:
apptainer test dummest.sif # if CPU only
apptainer test --nv dummest.sif # if you have a GPU