Source files live under envs/<cpu>/<gpu>/, where each leaf directory owns a spack.yaml manifest and a Dockerfile. Current environments are envs/x86/rocm (AMD GPUs), envs/x86/cuda (NVIDIA GPUs, parameterised by CUDA_ARCH and CUDA_VERSION build args), and envs/x86/none (CPU-only). Keep CPU targets (x86, …) and accelerator targets (rocm, cuda, none) as separate directories so images stay purpose-built, and limit the root README.md to high-level context.
docker build -f envs/x86/rocm/Dockerfile --build-arg GPU_ARCH=gfx90a -t selfish:x86-rocm-gfx90a .— builds the ROCm image for a specific AMD GPU arch.docker build -f envs/x86/cuda/Dockerfile --build-arg CUDA_ARCH=70 --build-arg CUDA_VERSION=12.4 -t selfish:x86-cuda-sm70 .— builds the CUDA image for a specific NVIDIA GPU arch.docker build -f envs/x86/none/Dockerfile -t selfish:x86-none .— builds the CPU-only image.docker run --rm selfish:x86-none spack find hdf5— smoke-tests that the expected view was installed inside the image.
Spack YAML uses 2-space indentation, lowercase keys, and quoted constraint strings ("target=x86_64_v3"). Group specs alphabetically, keep packages overrides sorted by scope, and rely on multiline RUN blocks with trailing \ alignment plus brief comments for non-obvious workarounds. Name new environments after the hardware tuple (x86/cuda, x86/none) so downstream scripts can glob predictably.
For each environment change, run spack spec followed by spack install --fail-fast inside a disposable builder container to verify concretization. Container builds must pass docker build locally before review; capture the last ~20 lines for the PR description. When adding MPI/HDF5 variants, run docker run --rm <tag> mpichversion (or another representative binary) to prove runtime availability. There is no coverage gate, but every new spec should ship with at least one build log, and GitHub Actions now double-checks gfx90a builds and publishes them to higherordermethods/selfish.
Existing history uses short, imperative subject lines (“Initial commit”); follow the same format and include the touched environment in parentheses when practical, e.g., Add feq-parse 2.2.2 to gfx90a. One logical change per commit keeps bisects clean. PRs should describe the motivation, list updated directories, attach the relevant spack spec or docker build excerpt, and link any upstream SELF issues. Paste terminal snippets when reviewing GPU-specific behavior.
Pin base images (rockylinux:9) and Spack refs in manifests, and run dnf update -y at build time to pick up CVEs. Never embed registry credentials or cluster hostnames in spack.yaml; rely on build-time secrets where required. Before publishing, scan the resulting image with docker scout cves selfish:gfx90a (or equivalent) to catch dependency vulnerabilities.