ci : ROCm + Vulkan release builds per merge to master (lemonade-sdk/llamacpp-rocm layout)

[dzannotti]

Overview

Adds ROCm and Vulkan release builds, one release per merge to master, Linux zips by default, in the same layout as lemonade-sdk/llamacpp-rocm: same pipeline, same assets, same release format. Tools that consume their releases (e.g. Lemonade) can point at this repo.

The files are lemonade's own, copied as-is (MIT, attribution in each file):

file	origin
.github/workflows/build-llamacpp-rocm.yml	build-llamacpp-rocm.yml
.github/workflows/test-llamacpp-rocm.yml	test-llamacpp-rocm.yml
.github/workflows/nightly-failure-alert.yml	nightly-failure-alert.yml
.github/actions/test-llamacpp-build/action.yml	.github/actions/test-llamacpp-build/action.yml
scripts/rocm-release/gather_required_libs.py	utils/gather_required_libs.py

Kept identical to lemonade:

• Build and packaging:
  • Ubuntu 22.04 builds. The Windows (windows-2022, VS2022) job is still there, selectable in a manual run.
  • Targets gfx1151,gfx1150,gfx120X,gfx110X,gfx103X,gfx90a,gfx908, including the multi-arch target mapping.
  • Newest TheRock nightly tarball from nightly.repo.amd.com, or a pinned version.
  • Identical cmake flags.
  • The same set of bundled ROCm DLLs / .sos plus the rocblas/library and hipblaslt/library kernel folders.
  • $ORIGIN RPATH on Linux.
• Releases:
  • Assets named llama-bNNNN-ubuntu-rocm-{target}-x64.zip.
  • Sequential bNNNN tags starting at b1000.
  • The same release notes block.
  • An issue opened when a release build fails.

Changed, because this repo is not a wrapper around upstream:

1. It clones this repository (${{ github.repository }}) instead of ggerganov/llama.cpp. The llamacpp_version input now means a branch or tag here, and latest means master.
2. secrets.GH_TOKEN || github.token, so it works without adding a secret.
3. The self-hosted hardware tests (test-stx-halo, test-stx, test-llamacpp-rocm.yml) only run when the repo variables STX_HALO_RUNNERS / STX_RUNNERS are true. When skipped they still allow the release, exactly as in lemonade's create-release condition. Without runners they would sit in the queue until they time out, as the self-hosted jobs did on Merge halo-box/llama.cpp master into strix (2026-09-15) #64.
4. The PR trigger is limited to changes to the pipeline files. Lemonade builds the full 14-zip matrix on every PR; here that would be every PR.
5. The alert goes to @dzannotti.
6. It releases on every push to master (the merged commit, checked out by SHA) instead of lemonade's nightly cron. A newer merge cancels a release that is still building (concurrency).
7. It adds a Linux Vulkan zip (llama-bNNNN-ubuntu-vulkan-x64.zip) to each release. It is built from the same commit on ubuntu-22.04 with the LunarG Vulkan SDK, GGML_VULKAN=ON, GGML_BACKEND_DL + GGML_CPU_ALL_VARIANTS for portable CPU code, and $ORIGIN RPATH, using the same flags as upstream's release.yml Vulkan job. Its failure blocks the release, like a ROCm failure does. The vulkan dispatch input turns it off.
8. It builds Linux only by default. workflow_dispatch with operating_systems: windows,ubuntu still builds Windows.

The bNNNN pattern only matches four digits, so it does not collide with the b1xxxx tags carried over from upstream.

Measurements

Not applicable: CI configuration only, no runtime code changes.

Additional information

• First run: this PR's own run (the build matrix, no release) is the first real test of the pipeline.
• Windows: not built by default. The first run's Windows gfx120X job failed on a network timeout while downloading, not on our code. The Windows test build is also broken on master by test-qsa-fast.cpp (fixed in tests : build test-qsa-fast on Windows #113); release builds use LLAMA_BUILD_TESTS=OFF and are not affected.
• Other targets: the CDNA/RDNA2-4 targets are built as lemonade does. Strix's HIP paths are gated on RDNA3.5, so they should compile, but that is unverified until the run finishes.
• Existing failure: upstream's release.yml fails on every push to master here (the last 3 runs). This PR does not touch it.

Requirements

• I have read and agree with the contributing guidelines
• This change is Strix Halo specific, or justified by measurements on Strix Halo: release packaging for this fork, not an upstream change.
• AI usage disclosure: AGENT-AUTHORED. Claude (Opus 5.5) ported the workflows and wrote this description, and @dzannotti requested it.
• What was NOT verified:
  • YAML parses, and actionlint reports only the custom stx-halo runner label.
  • No build has run yet. This PR's CI run is the first.
  • The release step (tagging and upload) cannot run from a PR. It runs on the first merge to master after this lands, which includes this PR's own merge.

🤖 Generated with Claude Code

[iggy]

file is called rocm-releases.md but it talks about rocm and vulkan... but I also don't see anything that actually builds vulkan or is it just using the upstream vulkan build action?

[iggy]

nevermind on not seeing the vulkan action. It was furled in the github UI (and in a file called build-llamacpp-rocm.yml, which I didn't think to look in at first)