Container Releases

Each published image ghcr.io/enzymead/reactantserver:X.Y.Z is built from the tagged ReactantServer-vX.Y.Z tree, its committed deploy/Manifest.toml, and the base image and Julia distribution pinned in MODULE.bazel. Each release has a page listing what it ships, what it requires of the host, its digests, and what changed from the release before it.

Julia, the base image, and the compiler and runtime packages move only through bazel run //deploy:relock or a new base image digest, never as a side effect of a release, so consecutive releases often share them. Changes from previous names what moved. A release that changes only ReactantServer's own packages keeps the numerics stack; any other change is a change to revalidate (see Reproducibility & Regulated Use).

Accelerators

Each release page has a section for every accelerator its image supports, listing what the image bundles for it and what it requires of the host. Released images support CUDA only.

CUDA

The CUDA that runs a model comes from Reactant_jll, not from the base image. The image ships the Reactant_jll artifact selected for REACTANT_GPU=cuda and the release's REACTANT_GPU_VERSION, which carries ptxas, fatbinary, libdevice and libnccl.so.2 and links cuDNN, cuBLAS, cuBLASLt and NVSHMEM into libReactantExtra.so. A release page's CUDA rows are the versions that artifact was built against. Reactant loads libnvJitLink.so.13, libnvrtc.so.13 and libcupti.so at run time, and those come from the base image's CUDA. The base also carries its own cuDNN and cuBLAS packages, which Reactant does not load.

NVIDIA driver. The base image sets NVIDIA_REQUIRE_CUDA, which the NVIDIA Container Toolkit checks when the container starts, so the host driver must support the base image's CUDA release. The toolkit itself is required as described in Deployment.

GPU. The minimum compute capability is the oldest target the Reactant_jll artifact is built for.