llamacpp_ci

This repository contains GitHub Actions and local build scripts for validating the HRX runtime build and the pinned llama.cpp HRX backend integration.

The CI job builds HRX from a pinned ROCm/hrx revision on the default ubuntu-latest GitHub runner, validates the installed HRX runtime, then builds pinned ROCm/llama.cpp with GGML_HRX=ON. ROCm toolchain and runtime dependencies are fetched from HRX/TheRock ROCm assets instead of installing full ROCm packages from apt.

What Gets Built

The active workflow builds HRX, validates the HRX install with hrx-info --device=cpu:0, then builds llama.cpp with the HRX backend enabled.

The HRX source revision and related repository pins live in:

scripts/hrx/versions.sh

Current pinned inputs are:

• ROCm/hrx main checksum for the HRX build.
• ROCm/llama.cpp hrx-integration checksum for scripts that need a matching llama.cpp checkout.

GitHub Actions

The workflow is defined in:

.github/workflows/build-hrx.yml

It performs these steps:

1. Check out this CI repository.
2. Install base build dependencies.
3. Restore/configure ccache.
4. Check out the pinned HRX revision.
5. Fetch ROCm assets using this repository's TheRock artifact fetcher.
6. Configure, build, and install HRX with amdclang and amdclang++ from those assets.
7. Validate the HRX install with hrx-info --device=cpu:0.
8. Check out pinned ROCm/llama.cpp.
9. Configure and build llama.cpp with GGML_HRX=ON.

Workflow dispatch inputs control the ROCm asset selection:

• hrx_release_type
• hrx_run_id
• hrx_artifact_set

The HRX repository and checkout revision are intentionally controlled by scripts/hrx/versions.sh, not workflow dispatch inputs.

Local Build

Run the same HRX build flow locally with:

scripts/hrx/local-build-hrx.sh

Build the HRX+llama.cpp HRX flow locally with:

scripts/hrx/local-build-llama.sh

Prepare source checkouts and ROCm assets locally with:

scripts/hrx/local-build-setup.sh

By default, local generated files are placed under ignored paths:

assets/hrx-src
assets/hrx-rocm-root
assets/hrx-rocm-downloads
build-hrx
build-hrx-install
assets/llama-src
assets/llama-src/build-hrx

For local-build-hrx.sh to run correctly, run local-build-setup.sh first (or otherwise ensure sources and assets are already prepared), then it runs:

scripts/hrx/build-hrx.sh
scripts/hrx/validate-hrx.sh

scripts/hrx/local-build-llama.sh runs:

scripts/hrx/build-llama-hrx.sh

scripts/hrx/local-build-setup.sh runs:

scripts/hrx/checkout-hrx.sh
scripts/hrx/checkout-llama.sh
scripts/hrx/fetch-rocm-assets.sh

Script Layout

• scripts/hrx/versions.sh: pinned upstream repository revisions.
• scripts/hrx/env.sh: shared default environment values.
• scripts/hrx/checkout-repo.sh: exact checkout helper used by HRX and llama.cpp checkout scripts.
• scripts/hrx/checkout-hrx.sh: checks out pinned HRX.
• scripts/hrx/fetch-rocm-assets.sh: fetches and exposes ROCm assets, including the default rocwmma_dev_gfx110X-all header artifact.
• scripts/hrx/build-hrx.sh: configures and builds HRX.
• scripts/hrx/validate-hrx.sh: validates the installed HRX runtime with a CPU smoke test.
• scripts/hrx/checkout-llama.sh: checks out pinned ROCm/llama.cpp.
• scripts/hrx/build-llama-hrx.sh: configures and builds llama.cpp with the HRX backend.
• scripts/hrx/run-build.sh: full HRX checkout, ROCm asset fetch, build, and validation flow.
• scripts/hrx/local-build-hrx.sh: local entrypoint using ignored repo-local output paths.
• scripts/hrx/local-build-llama.sh: local entrypoint for full HRX + llama.cpp HRX build flow.
• scripts/hrx/local-build-setup.sh: local setup entrypoint for repository checkouts and ROCm asset fetch.
• scripts/hrx/ci-build-hrx.sh: CI entrypoint.

Notes

Generated build outputs and downloaded assets are ignored by .gitignore.

AGENTS.md is local branch guidance for Codex work and should not be committed.