llamaMan

English · 简体中文 · 日本語 · Deutsch · Français · Español · Português · Русский

What it is. A browser UI + API front-end for running multiple llama.cpp server instances, across one machine or a small cluster of them. Ships an Ollama-compatible proxy so it drops in for Open WebUI.

What's different. llamaMan itself has no llama.cpp code and no GPU dependency - it spawns ghcr.io/ggml-org/llama.cpp:server-* containers as siblings over the Docker socket. Update llama.cpp without touching llamaMan: docker pull a newer server-* image. The same spawn-and-proxy machinery is engine-agnostic: Strata instances run alongside llama.cpp ones, launched, monitored, proxied, and slept the same way.

llama-server flag semantics are the source of truth on llama.cpp's side: server CLI reference.

Features

• Universal GPU image - one flow for NVIDIA, AMD (ROCm), Intel Arc, CPU; vendor auto-detected (Vulkan is opt-in via GPU_TYPE=vulkan)
• Model library + downloader - HuggingFace pulls with speed limits, resume, auto-retry, republish detection with atomic re-pull
• One-click launch + presets - every llama.cpp knob (spec-decoding, MoE offload, KV quant, mmproj, PDF input…) is a first-class UI field, with live preset updates for fields that don't need a relaunch
• Ollama + OpenAI proxy on :42069 - Open WebUI drops in; models auto-start on demand and LRU-evict once LLAMAMAN_MAX_MODELS is hit
• Image & PDF input - vision models with --mmproj; PDFs rasterized page-by-page
• Instance monitoring - per-GPU VRAM, container CPU/RAM, per-instance TTFT / throughput / latency
• Auth - user accounts, API keys (bearer tokens), per-endpoint auth toggle
• Multi-node clustering (opt-in) - heterogeneous nodes as one logical cluster; aggregated dashboard; shared-queue load balancing
• 🆕 Strata engine (opt-in, NVIDIA) - launch Strata instances next to llama.cpp ones
• 🆕 Knowledge base + MCP server (opt-in, MariaDB backend) - store documents as topics, search them by meaning with your own embedding model, and expose the whole thing to any MCP client — all local

Requirements

• Docker with access to /var/run/docker.sock
• One of:
  • NVIDIA Container Toolkit for NVIDIA GPUs
  • ROCm-compatible setup for AMD GPUs
  • Intel Arc with /dev/dri access
  • CPU-only

Quick Start

Edit docker-compose.yml first:

• Set HOST_MODELS_DIR / HOST_LOGS_DIR to the real host paths that back your volume mounts (llamaMan passes them to the Docker daemon when spawning sibling containers).
• LLAMAMAN_NODE_NAME (srv1 in the sample compose; there is no built-in default - the app refuses to start without it) is required for every install - it's the partition key for this deployment's stored state. Pick once, keep it (changing later orphans stored state).

Pull the llama.cpp image for your GPU and start:

docker pull <LLAMA_IMAGE>          # from the table above
# For non-CUDA: also set LLAMA_IMAGE=... in docker-compose.yml
docker compose up --build

For NVIDIA native VRAM monitoring, also uncomment deploy.resources.reservations in docker-compose.yml.

• Management UI: http://localhost:5000 (create admin on /setup)
• Ollama-compatible API: http://localhost:42069
• llama-server instance ports: 8000-8020

Security note: the mounted /var/run/docker.sock gives the container full control of Docker on the host. Standard Docker-socket-access implications apply.

Running bare-metal (dev, WSL)? See Running bare-metal.

Your first model

1. Get a model. Either drop a .gguf into your models directory, or open the UI, pick a HuggingFace repo under Downloads, and let llamaMan pull it.
2. Launch it. Select the model in the sidebar, click Launch. llamaMan spawns a llama-server container and the instance appears with a status badge. Defaults (4096 ctx, auto GPU layers) work for most models; Save Preset remembers your settings. Every launch field is documented in Launch settings.
3. Create an API key. Under API Keys, create one (llm-xxxxxxxxxx). Model endpoints require a bearer token by default.
4. Point your client at it. Any OpenAI or Ollama client works against :42069. For Open WebUI:

open-webui:
  environment:
    - OLLAMA_BASE_URL=http://llamaman:42069
    - OPENAI_API_BASE_URLS=http://llamaman:42069/v1
    - OPENAI_API_KEYS=llm-your-api-key-here

Open WebUI lists your models via /api/tags (by GGUF filename stem unless you set a Display Name); picking one auto-starts it, waits until healthy, then proxies. Set a per-model Display Name to show a friendly name instead of the raw quant filename.

Cap caveat: the sample compose ships LLAMAMAN_MAX_MODELS=1. With a cap set, the Ollama connection evicts the LRU model when full, but the OpenAI connection (OPENAI_API_BASE_URLS) returns 503 by default - pick one path or raise the cap. Rules: Model Eviction.

If something goes wrong, Troubleshooting below covers the common first-run failures.

Authentication

Two credential types:

• Session cookies for the UI (created via /setup on first launch)
• API keys for external clients (Authorization: Bearer llm-xxxxxxxxxx), managed under API Keys in the UI

The "Require authentication for all endpoints" toggle (on by default) controls whether model-serving endpoints demand a bearer token. Details: Authentication.

Architecture

llamaMan is a small Python app that never touches llama.cpp code directly. It spawns each llama-server as a sibling container over the Docker socket, monitors the fleet, and hands out an OpenAI/Ollama front door.

%%{init: {'theme': 'base', 'themeVariables': {'fontFamily': 'ui-sans-serif, system-ui, sans-serif', 'fontSize': '14px', 'primaryColor': '#eef2ff', 'primaryBorderColor': '#6366f1', 'primaryTextColor': '#1e1b4b', 'lineColor': '#64748b', 'clusterBkg': '#f8fafc', 'clusterBorder': '#cbd5e1', 'edgeLabelBackground': '#f8fafc'}, 'flowchart': {'curve': 'basis', 'nodeSpacing': 40, 'rankSpacing': 60}}}%%
flowchart LR
  client[Client<br/>Open WebUI / any OpenAI or Ollama client]

  subgraph nodeA[Node A]
    lmA[llamaman<br/>:5000 UI · :42069 compat]
    subgraph workersA[llama-server siblings on host GPUs]
      wA1[qwen2.5-14b]
      wA2[gpt-oss-20b]
    end
  end

  subgraph nodeB[Node B]
    lmB[llamaman]
    subgraph workersB[llama-server siblings]
      wB1[qwen2.5-14b]
      wB2[deepseek-v3]
    end
  end

  db[(Shared storage<br/>presets · instance state · heartbeats)]

  client -->|prompt| lmA
  lmA -->|forward when a peer is less loaded| lmB
  lmA --> wA1
  lmA --> wA2
  lmB --> wB1
  lmB --> wB2
  lmA <-.-> db
  lmB <-.-> db

  classDef core fill:#4f46e5,stroke:#312e81,color:#fff
  classDef worker fill:#0d9488,stroke:#134e4a,color:#fff
  classDef ext fill:#64748b,stroke:#334155,color:#fff
  class lmA,lmB core
  class wA1,wA2,wB1,wB2 worker
  class client,db ext

The single-node picture (default) is just the top half - one llamaman container, N llama-server siblings on the local Docker daemon, no cluster peers, JSON files instead of the shared DB. Clustering is opt-in and additive.

How it works in depth - request flow, queue-aware dispatch, failure behavior, design trade-offs: Architecture.

Documentation

Environment Variables

The ones you'll touch on day one (full reference: Environment Variables):

Troubleshooting

Full table: Troubleshooting.

Credits

This work would not be possible without ggml-org/llama.cpp.

License

llamaMan is licensed under the Elastic License 2.0. You may use, copy, distribute, and modify the software, subject to:

• No providing the software to third parties as a hosted / managed service where the service gives users access to a substantial set of its features or functionality.
• No removing or obscuring licensing, copyright, or other notices of the licensor.

Third-party licenses

• Font Awesome Free 7.1.0 by Fonticons, Inc. - icons (CC BY 4.0), fonts (SIL OFL 1.1), and code (MIT). Full license text ships in static/fontawesome-free-7.1.0-web/LICENSE.txt.