Everything is driven from the GitHub UI — no command line needed. Releases are semantic-version and tag-driven: there is no auto-release on pushes to main. You release by tagging a version; the built binaries are then attached to that version's Release.

1. Open the repository → Releases → Draft a new release.
2. Choose a tag → create a new v* tag, e.g. v1.2.0 (or v1.2.0-rc1).
3. Add a title/notes; check "Set as a pre-release" if it's a candidate.
4. Publish release.

Publishing pushes the tag, which starts the build. When it finishes (~1–2 h; the Windows CUDA jobs dominate), the prebuilt binaries are attached to that Release.

1. Actions tab → Release → Run workflow.
2. Version: enter 1.2.0 (or v1.2.0).
3. Leave the Publish a GitHub Release box unchecked.
4. Run workflow → builds all backends and uploads them as workflow artifacts, but creates no tag or Release. Use this to validate before cutting a release.

Notes:

• CUDA builds use GGML_BACKEND_DL, so CUDA kernels ship once as ggml-cuda.dll; the heavy CUDA runtime (cudart/cuBLAS/cuBLASLt/cuFFT) is bundled in a separate cudart-...zip.
• CUDA architectures are pinned (all-real) so binaries are portable across the supported NVIDIA GPUs instead of being tied to the (GPU-less) CI host.
• Windows packages bundle the MSVC runtime app-locally (vcruntime140*.dll, msvcp140*.dll, vcomp140.dll) so they run on clean Windows without a separately installed VC++ Redistributable. Total added size is ~1 MB compressed (negligible vs. the CPU/Vulkan/CUDA package sizes).

• Pushing a v* tag (e.g. by drafting + Publishing a Release) → builds and attaches the binaries to that tag's Release.
• Manual run from the Actions UI (with a version) → builds; publishes only if Publish is checked.
• main pushes do not trigger a release.

1. Confirm the expected assets exist on the Release page.
2. On a real NVIDIA GPU machine, run e.g. audiocpp_cli.exe --backend cuda ... and confirm the log reports a CUDA device (ggml_cuda_init / "found N CUDA devices"). GitHub CI runners have no GPU, so a green build does not prove GPU runtime attach — this check is required.