llama.cpp rolling release (no numbered versions)

llama.cpp doesn't publish a single "supported CUDA version" the way PyTorch or vLLM do -- it's compiled from source (or from AMD/community Docker images) against whatever CUDA Toolkit is installed, and the project's own CI simply tracks a moving target rather than pinning one. As of build b10826 (2026-09-06): its official Docker images build against CUDA 12.8.1 (the "cuda12" tag) and CUDA 13.3.0 (the "cuda13" tag) per .github/workflows/docker.yml; its Ubuntu CUDA CI job builds inside an nvidia/cuda:12.6.2-devel container per .github/workflows/build-cuda-ubuntu.yml; and its Windows CUDA CI job tests 12.4, 13.3, and 13.4 per .github/workflows/build-cuda-windows.yml. 12.4 through 13.4 is therefore a reasonable "known-good" range rather than a hard minimum or maximum -- older CUDA Toolkits (down to roughly 11.x) are widely reported by users to still work but aren't exercised by current CI. ROCm: the same Ubuntu CI job's HIP build runs inside a rocm/dev-ubuntu-22.04:6.1.2 container (this project's own practical floor), while AMD's own ROCm documentation separately validates a specific llama.cpp build (b6652) against ROCm 7.0.0 for the Instinct MI325X/MI300X/MI210 -- see rocm.docs.amd.com/projects/llama-cpp/en/docs-26.02/install/llama-cpp-install.html. Both figures are included as rocmVersions since they represent two different (and both legitimate) "what's actually been tested" claims from two different sources, not a single official minimum. This entry is a point-in-time snapshot, not a version history -- unlike PyTorch/vLLM, llama.cpp has no stable version numbers to track over time. pythonVersions is omitted: llama.cpp itself is a C++ CLI/server project with no Python version dependency of its own, distinct from the separately versioned `llama-cpp-python` PyPI bindings package, which isn't covered by this entry. Source: github.com/ggml-org/llama.cpp workflow files (.github/workflows/build-cuda-ubuntu.yml, build-cuda-windows.yml, docker.yml) and docs/build.md at the current master commit, plus AMD's ROCm documentation above, verified 2026-09-06.