How lemonade-sdk/llamacpp-rocm works
This repository delivers nightly llama.cpp builds accelerated by ROCm, rather than hosting the inference engine itself. Its documented user path is to download a platform/GPU-targeted release, extract it, and run llama-server with a GGUF model. The tree identifies build, smoke-test, and failure-alert workflows; README describes the published build matrix and bundled runtime. The sampled utility gathers missing ROCm libraries for llama-server, but its wiring to the workflows is not established by the available source.
Main entry points: build-llamacpp-rocm.yml Build workflow · gather_required_libs.py Library gatherer · test-llamacpp-rocm.yml Build smoke test · nightly-failure-alert.yml Nightly failure alert
Components 14
Build automation
- Build workflow
.github/workflows/build-llamacpp-rocm.yml - Build smoke test
.github/workflows/test-llamacpp-rocm.yml - Test action
.github/actions/test-llamacpp-build/action.yml - Nightly failure alert
.github/workflows/nightly-failure-alert.yml
Build inputs
- Library gatherer
utils/gather_required_libs.py - llama.cpp
- ROCm TheRock
- HIP
Download and use
- Release builds
- llama-server
Other components
- Maintainers
- Build user
- GGUF model
- Supported AMD GPU
Connections 13
- Build workflow → llama.cpp: builds with
- Build workflow → ROCm TheRock: includes runtime
- Build workflow → HIP: targets
- Build workflow → Release builds: publishes
- Build smoke test → Test action: uses
- Nightly failure alert → Maintainers: alerts
- Library gatherer → llama-server: runs
- Library gatherer → ROCm TheRock: searches libraries
- Library gatherer → llama-server: copies libraries beside
- Build user → Release builds: downloads
- Build user → llama-server: runs
- Build user → GGUF model: supplies
- llama-server → Supported AMD GPU: accelerates inference