Skip to content

release-process.md

Latest commit

 

History

History
69 lines (57 loc) · 4.51 KB

release-process.md

File metadata and controls

69 lines (57 loc) · 4.51 KB
Raw

Cutting a release

This repository publishes its binaries as the Docker image, which CI builds from every master commit that passes it and changes a file the image is built from (README); the image is tagged latest, sha-<commit> and the contents of VERSION. Release archives for Windows and Linux are an additional, manual path that no release has used yet. This page is that path.

The flow

  1. Prepare. Branch from master, set VERSION to the full tag (for example 0.13.0-rtx3090), write RELEASE_NOTES_<X.Y.Z>.md for it, and merge. The packagers derive everything version-specific from VERSION and refuse to run without the matching release notes.
  2. Build both platforms from the same commit. The binaries do not embed VERSION, so they can be built before the release PR merges.
    • Windows: scripts\build.ps1 -Package builds ninfer, ninfer-serve and ninfer_bench (Windows build) and packages them.
    • Linux: a fresh git archive of the commit, -DCMAKE_CUDA_ARCHITECTURES=86, -DBUILD_TESTING=OFF, -DNINFER_BUILD_BENCHMARKS=ON, targets ninfer ninfer-serve ninfer_bench, then scripts/package-release.sh (or scripts/build.sh --package).
  3. Package. scripts/package-release.ps1 writes the Windows zip and scripts/package-release.sh the Linux tar.gz, each with SHA256SUMS-v<X.Y.Z>-{windows,linux}.txt, under dist/. The archive README is release-archive-windows.md or release-archive-linux.md; check that it still describes the archive's requirements and launcher profiles.
  4. Smoke-test on a real card (below), not just the unit tests.
  5. Publish. Tag the merge commit (annotated), push the tag, then gh release create vX.Y.Z-rtx3090 <zip> <tar.gz> <both SHA256SUMS files> -R iamwavecut/ninfer-all --notes-file RELEASE_NOTES_X.Y.Z.md --title "..." --latest.

Gotchas

Always pass -R iamwavecut/ninfer-all to gh. The checkout also has the upstream and ashalliants remotes, and gh without -R may resolve to another fork.

Windows executables import cuBLAS, and the archive must carry it. The cuBLAS prefill route makes every executable import cublas64_12.dll (which imports cublasLt64_12.dll), so a machine without the CUDA Toolkit cannot start any of them (0xC0000135), even when --prefill-cublas is never used. package-release.ps1 copies both DLLs from %CUDA_PATH%\bin and refuses to produce an archive whose executables do not start with only Windows on PATH. Do not weaken that check.

Redistributing NVIDIA's DLLs has conditions. The CUDA Toolkit EULA (Attachment A) lists the cuBLAS DLLs as redistributable. The grant requires that the application has material functionality of its own, that the DLLs are used only by it, and that the distribution terms are consistent with the EULA and protect NVIDIA's rights. The packager therefore ships the toolkit's EULA.txt as NVIDIA-CUDA-EULA.txt, and the archive README names the two DLLs as NVIDIA's, outside the project's Apache-2.0 licence. Re-read the EULA of the toolkit that builds a release.

Linux links cuBLAS, FFmpeg and glibc dynamically. The Linux archive does not ship libcublas (the Lt library alone is about 500 MB), so release-archive-linux.md states the requirements: the glibc of the build host, the CUDA runtime libraries including cuBLAS, FFmpeg and libcurl. Check them with ldd on the new binaries and update the archive README when they change.

Asset sizes. Every executable links every kernel and weighs about half a gigabyte; the Windows zip with the cuBLAS DLLs is about 1.5 GB. GitHub rejects release assets over 2 GiB.

Smoke-test the launcher default on the real card. A desktop holding 2-3 GiB of a 24 GB card refuses the default DFlash2 profile, both on the engine's runtime reservation and, on Windows, on pinned host state. run.{sh,bat} step down by themselves when startup is refused for lack of GPU memory (launcher profiles). Start the launcher with nothing overridden and confirm it serves a request.

Checks before publishing

  • scripts/check-linux-scripts.sh passes.
  • Both archives: every file matches its inner SHA256SUMS.txt, and the outer checksum file verifies.
  • Windows: all three executables start with only Windows on PATH.
  • Linux: a fresh unpack runs ./ninfer-serve --help and ./run.sh --help.
  • The default launcher profile starts and serves a chat completion on a real 24 GB card.