skills.rest
yolanother
yolanotherCOMMUNITY
Explore all skills by this author →
WebsiteGitHub

build-llama-cpp

Builds and updates the ROCm llama.cpp llama-server engine for AMD Strix Halo iGPU inference.

4|2|Updated Feb 2, 2026
Search Tags:#build#amd gpu#llama.cpp#distrobox#inference engine#rocm#llama-server
One-click install
npx skills add https://github.com/yolanother/WebLlamaManager --skill build-llama-cpp-yolanother
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: build-llama-cpp
Source: https://github.com/yolanother/WebLlamaManager/tree/main/.claude/skills/build-llama-cpp
Command: npx skills add https://github.com/yolanother/WebLlamaManager --skill build-llama-cpp-yolanother

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Keeping the llama.cpp inference engine current on an AMD Strix Halo (gfx1151) box is error-prone: new model architectures require newer upstream commits, the ROCm build must run inside a specific distrobox toolchain, and mistakes (wrong flags, deleted build trees, missing UMA env vars) cause hard freezes or silent CPU fallback. This Skill encodes the pinned-version build workflow and the hard-won gotchas so engine updates succeed on the first try. ## Core Features & Use Cases - Pinned version builds: Sets LLAMA_CPP_REF in .llama-cpp-version and compiles via scripts/build-llama-cpp.sh inside the ROCm 7.2.4 distrobox with the known-good gfx1151 flags. - Safe deploy path: Stops the manager service before building, checks thermals before the 30-minute all-core compile, and deploys through install.sh which prefers ~/.local/bin/llama-server. - Failure-avoidance guidance: Documents RUNPATH-into-build-tree behavior, Makefile-vs-Ninja toolchain limits, required UMA runtime env vars, and the --models-preset router requirement for multimodal models. - Use Case: A new Gemma or Qwen architecture needs upstream support newer than the installed b10752 build — bump the pinned commit, rebuild in the distrobox, pass the validation gate, and restart the service. ## Quick Start Ask the assistant to update llama.cpp to the latest upstream commit and rebuild the llama-server binary for the Strix Halo iGPU.

Frequently Asked Questions about build-llama-cpp

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I build llama.cpp with ROCm support for AMD Strix Halo?▼

Pin the desired commit in `.llama-cpp-version`, then run `scripts/build-llama-cpp.sh` inside the ROCm distrobox. It compiles with `GGML_HIP=ON`, `AMDGPU_TARGETS=gfx1151`, and `GGML_HIP_ROCWMMA_FATTN=ON`, then installs the binary to `~/.local/bin/llama-server`.

How do I update llama-server to support a new model architecture?▼

Bump `LLAMA_CPP_REF` in `.llama-cpp-version` to a commit new enough for the architecture, stop the llama-manager service, run the build script, and pass the validation gate before restarting. Use `LLAMA_CPP_CLEAN=1` after a large version jump.

Why does llama-server fall back to CPU instead of using the iGPU?▼

The Strix Halo needs UMA runtime environment variables: `GGML_HIP_UMA=1`, `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1`, and `HSA_OVERRIDE_GFX_VERSION=11.5.1`. These are set by `container-start.sh` at runtime, not during the build; without them models run on CPU.

Can the runtime-only ROCm toolbox compile llama.cpp?▼

No, the runtime-only `llama-rocm-7.2.4` toolbox has no compiler. For custom builds, point `DISTROBOX_CONTAINER` at a toolchain image such as a devel or rocwmma toolbox that includes a working compiler.

Why does the installed llama-server break after deleting the build directory?▼

The installed binary finds its shared libraries via RUNPATH pointing into `~/llama.cpp/build/bin`. The install step only copies the binary, so deleting the build tree removes the libraries it depends on.

Why do multimodal vision models fail with --models-dir in llama.cpp?▼

The `--models-dir` router mode cannot attach a per-model vision projector (mmproj). Multimodal models like Gemma 4 require the `--models-preset` router with an INI configuration in `container-start.sh` to serve vision on the GPU.

Related Skills

Explore More Software Engineering Skills→
crycriMcrycriM

turboquant-vulkan-toolbox

Build and configure TurboQuant Vulkan/HIP toolboxes for AMD Strix Halo GPUs.

Community
Advanced
jayll1303jayll1303
17

llama-cpp-inference

Run local GGUF model inference and OpenAI-compatible serving with llama.cpp.

Community
Advanced
crycriMcrycriM

strix-halo-monitoring

Monitor CPU/GPU temps, power draw, and clock speeds on Strix Halo during inference workloads.

Community
Intermediate
Mesh-LLMMesh-LLM
3.1k

llama-stage-patch-changes

Regenerate the llama-stage.cpp patch queue and update the upstream SHA pin.

Official
Advanced