llama.cpp on StartOS

Upstream repo: https://github.com/ggml-org/llama.cpp

Upstream llama-server docs: https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md

Everything not listed here behaves the same as upstream llama-server. If a flag, endpoint, or behavior is not mentioned in this document, upstream documentation is accurate and fully applicable.

llama.cpp is a high-performance C/C++ runtime for large language models in GGUF format. This package wraps its built-in HTTP server (llama-server), which exposes an OpenAI-compatible API and a small in-browser chat UI on the same port.

---

Table of Contents

• Image and Container Runtime
• Volume and Data Layout
• Installation and First-Run Flow
• Configuration Management
• Network Access and Interfaces
• Actions (StartOS UI)
• Dependencies
• Backups and Restore
• Health Checks
• Limitations and Differences
• What Is Unchanged from Upstream
• Contributing
• Quick Reference for AI Consumers

---

Image and Container Runtime

The package ships four variants, selected at build time via the VARIANT env var (driven by the Makefile):

Variant	Image	Arches	Accelerator	Offered to GPU driver
generic	ghcr.io/ggml-org/llama.cpp:server	x86_64, aarch64	CPU only	— (universal fallback)
nvidia	ghcr.io/ggml-org/llama.cpp:server-cuda	x86_64, aarch64	CUDA (NVIDIA)	nvidia
rocm	ghcr.io/ggml-org/llama.cpp:server-rocm	x86_64	ROCm (AMD)	amdgpu
vulkan	ghcr.io/ggml-org/llama.cpp:server-vulkan	x86_64, aarch64	Vulkan	i915 (Intel)

All four variants publish under a single package version. Each declares a distinct hardwareRequirements.device, so StartOS serves each host the most specific variant its detected hardware satisfies — nvidia/rocm/vulkan for matching GPUs, and generic as the universal CPU fallback for everything else. Note that vulkan matches only Intel GPUs on the i915 driver; newer Intel GPUs on the xe driver (and non-Intel Vulkan-only setups) fall back to generic. rocm matches the amdgpu driver but is narrowed by GPU product name to discrete AMD GPUs (Navi / Radeon RX / Instinct); integrated Radeon graphics (e.g. the Radeon 680M in Ryzen APUs), where ROCm is unreliable, fall back to generic. nvidia matches the nvidia driver, which is present only when StartOS is installed from a -nvidia platform flavor (x86_64-nvidia / aarch64-nvidia, bundling the NVIDIA driver and container toolkit); on the standard or -nonfree flavors an NVIDIA card isn't detected and falls back to generic (CPU), even with the card physically present.

Property	Value
Entrypoint	/app/llama-server
Working dir	/app
Default port	8080

---

Volume and Data Layout

Volume	Mount Point	Purpose
main	/data	store.json (serve args) and models/ (GGUF cache)

The container runs with LLAMA_CACHE=/data/models and HF_HOME=/data/huggingface, so all -hf <repo> downloads land on the persistent volume.

---

Installation and First-Run Flow

Step	StartOS
Install	Marketplace install or sideload .s9pk
First-run tasks	Two critical tasks: Set UI Password (created whenever no password is set) and Set Model (created whenever no model is selected). Both are created on install and re-surface if the underlying value is later cleared.
Start service	After Set Model has been run; until then the daemon idles
Pull the model	Automatic on first start (cached on the main volume)

Until Set Model has been run, the daemon stays in an idle (sleep infinity) state and the API port is closed — the health check reports "No model selected." Once a model is selected, llama-server is restarted with the chosen serve arguments.

---

Configuration Management

Serve configuration is stored at /data/store.json and managed via the Set Model action:

{
  "serveArgs": [
    "-hf",
    "unsloth/Qwen2.5-7B-Instruct-GGUF:Q4_K_M",
    "-c",
    "8192",
    "-ngl",
    "999"
  ]
}

serveArgs is the exact list of arguments appended after /app/llama-server. The daemon adds --host 0.0.0.0 and --port 8080 at runtime.

llama-server itself runs keyless — no --api-key. Access is instead gated by HTTP basic auth enforced at the StartOS reverse proxy (addSsl.auth): the OS validates credentials before any request reaches the container. The username is hard-coded to admin; the password is generated by the Set UI Password action and stored as uiPassword in store.json. setupInterfaces reads it reactively, so rotating it via the action takes effect without a manual restart. Set UI Password is a critical task, which blocks the service from starting until a password is set — so the service never runs (and the gate never serves) without one.

Dependent StartOS services reach llama.cpp over the internal service mesh (http://llama-cpp.startos:8080), which is not behind the proxy gate, so they connect keyless.

Curated presets: the Set Model action surfaces a hardware-tier-aware list of GGUF presets and disables ones too large for the detected memory:

Preset	Repo (-hf)	Min memory
Llama 3.2 1B Instruct	unsloth/Llama-3.2-1B-Instruct-GGUF:Q4_K_M	2 GB
Llama 3.2 3B Instruct	unsloth/Llama-3.2-3B-Instruct-GGUF:Q4_K_M	4 GB
Qwen2.5 7B Instruct	unsloth/Qwen2.5-7B-Instruct-GGUF:Q4_K_M	6 GB
Llama 3.1 8B Instruct	unsloth/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M	8 GB
Qwen2.5 14B Instruct	unsloth/Qwen2.5-14B-Instruct-GGUF:Q4_K_M	12 GB
Mistral Small 3.2 24B Instruct	unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF:Q4_K_M	18 GB
Qwen3 30B-A3B Instruct	unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF:Q4_K_M	22 GB
Qwen2.5 32B Instruct	unsloth/Qwen2.5-32B-Instruct-GGUF:Q4_K_M	24 GB
Llama 3.3 70B Instruct	unsloth/Llama-3.3-70B-Instruct-GGUF:Q4_K_M	48 GB

The Custom variant accepts a HuggingFace repo, optional filename, context size, GPU layer count, and extra llama-server flags. For settings that can't be expressed cleanly via the form (quoted JSON, multi-word strings), edit store.json directly.

---

Network Access and Interfaces

Interface	Port	Protocol	Type	Purpose
llama.cpp Server	8080	HTTP	ui	Built-in chat UI + OpenAI-compatible API

The chat UI and the API share a single port, gated by basic auth (admin + the generated password) at the proxy. Access methods (StartOS 0.4.x): LAN IP, <hostname>.local, Tor .onion, and custom domains if configured. Browsers get a native login prompt. OpenAI-compatible clients hitting the public interface use base URL <interface-url>/v1 and must supply the basic-auth credentials (e.g. curl -u admin:<password>); other StartOS services use the keyless internal http://llama-cpp.startos:8080/v1.

Selected upstream endpoints:

Endpoint	Method	Purpose
/v1/chat/completions	POST	OpenAI-compatible chat
/v1/completions	POST	OpenAI-compatible text completion
/v1/embeddings	POST	Embeddings (when the loaded model supports them)
/health	GET	Health probe
/props	GET	Loaded model info

The full surface area is documented in upstream tools/server/README.md.

---

Actions (StartOS UI)

Action	Purpose
Set Model	Choose a curated preset (with hardware-tier-aware availability) or a custom HuggingFace GGUF. Writes serveArgs to store.json and restarts the daemon.
Set UI Password	Generate (or rotate) the web UI login password. Username is always admin. Returns the new credentials; the proxy gate picks them up automatically.
Delete Model Cache	Remove a specific filename from /data/models to reclaim disk space.

---

Dependencies

None.

---

Backups and Restore

Included in backup:

• main volume — store.json and all cached GGUF weights under models/.

Restore behavior:

• Serve args and any locally cached models are restored verbatim. No reconfiguration needed.

Backups can be very large depending on how many models you've cached — a single 70B Q4 file is ~40 GB.

---

Health Checks

Check	Method	Grace period	Messages
llama.cpp API	Port listening on 8080	60 minutes (cold-cache model downloads)	"The llama.cpp API is ready" / "The llama.cpp API is not ready" or "No model selected. Run the "Set Model" action."

---

Limitations and Differences

1. One model per process. llama-server holds a single GGUF in memory. To switch models, run Set Model again — the service restarts with the new weights.
2. Custom-action arg splitting. The Custom variant's Extra arguments field is split on whitespace, so JSON values with quoted spaces will not survive — edit store.json directly for those.
3. Hardware-tier detection is best-effort. GPU memory is read from nvidia-smi / rocm-smi; on Vulkan and unsupported topologies, the preset filter falls back to total system RAM as a memory budget.
4. Variants are independent installs. Switching from e.g. generic to nvidia is an uninstall + reinstall, not an in-place change; cached models on the main volume can be restored from backup.

---

What Is Unchanged from Upstream

• The full llama-server HTTP API and built-in chat UI.
• All llama-server CLI flags — anything not consumed by the package wrapper passes straight through (via the Custom variant's extra args).
• HuggingFace -hf model downloads and the LLAMA_CACHE layout.
• GGUF model support, embedding endpoints, OpenAI-compatible response shapes, and tool-call formats.

---

Contributing

See CONTRIBUTING.md for build instructions and development workflow, and UPDATING.md for the upstream-bump procedure.

---

Quick Reference for AI Consumers

package_id: llama-cpp
hardware_acceleration: true
variants: # all publish under one version; StartOS matches by detected GPU driver
  generic:
    image: ghcr.io/ggml-org/llama.cpp:server
    arch: [x86_64, aarch64]
    accel: cpu
    gpu_driver: null # universal CPU fallback
  nvidia:
    image: ghcr.io/ggml-org/llama.cpp:server-cuda
    arch: [x86_64, aarch64]
    accel: cuda
    gpu_driver: nvidia # present only on -nvidia StartOS flavors; CPU fallback otherwise
  rocm:
    image: ghcr.io/ggml-org/llama.cpp:server-rocm
    arch: [x86_64]
    accel: rocm
    gpu_driver: amdgpu # discrete AMD GPU only (integrated Radeon -> generic)
  vulkan:
    image: ghcr.io/ggml-org/llama.cpp:server-vulkan
    arch: [x86_64, aarch64]
    accel: vulkan
    gpu_driver: i915 # Intel GPUs only
volumes:
  main: /data
ports:
  api_and_ui: 8080
env:
  LLAMA_CACHE: /data/models
  HF_HOME: /data/huggingface
dependencies: none
auth: # llama-server runs keyless; basic auth enforced at the OS reverse proxy
  type: basic
  username: admin # hard-coded
  password: generated by set-ui-password, stored as store.json uiPassword
  internal_mesh: keyless # http://llama-cpp.startos:8080 bypasses the proxy gate
startos_managed_args: ['--host 0.0.0.0', '--port 8080']
actions:
  - set-model
  - set-ui-password
  - delete-model-cache