generated: '2026-08-04'
method: searched
source: https://llmboost.mangoboost.io/docs/running/lbh
name: MangoBoost command-line interfaces
description: >-
  MangoBoost ships three first-party CLIs across two product lines: `lbh` (LLMBoost Hub)
  and `llmboost` for the LLM inference server, and `mango-ctl` / `mango-smi` for DPU
  device control and monitoring. Command surfaces below are transcribed from the vendor
  documentation; no command was invented.

clis:
  - name: lbh
    title: LLMBoost Hub
    package: llmboost-hub (PyPI)
    docs: https://llmboost.mangoboost.io/docs/running/lbh
    summary: >-
      One CLI that manages LLMBoost container images, model assets, the license and the
      server — on a single node or a Kubernetes cluster.
    install:
      - pip install llmboost_hub
      - uv tool install llmboost_hub
      - pip install --upgrade llmboost_hub
    help:
      - lbh -h
      - lbh <command> -h
      - lbh -v <command>
    commands:
      - command: lbh login
        does: Import and validate the license file (prompts for the EULA if needed).
      - command: lbh fetch [model]
        does: Refresh the supported-model lookup cache, filtered to your GPU.
      - command: lbh list [query] [--discover PATH]
        does: List prepared models; --discover scans a directory for existing model folders.
      - command: lbh prep <model> [--only-verify] [--fresh]
        does: Pull the image and download model assets; caches the resolved path.
      - command: lbh run <model> [opts] -- [docker flags]
        does: Start a detached container (mounts workspace, maps GPUs). Flags after `--` go to `docker run`.
      - command: lbh serve <model> [opts] -- [llmboost args]
        does: Start the inference server inside the container. Flags after `--` go to `llmboost serve`.
        options: ['--host', '--port', '--detached', '--force']
      - command: lbh test <model> [--query "…"] [-t N]
        does: Send a request to /v1/chat/completions as a smoke test.
      - command: lbh attach <model>
        does: Open an interactive shell in the running container.
      - command: lbh stop <model>
        does: Stop the running container.
      - command: lbh status [model]
        does: Concise status for prepared models.
      - command: lbh completions
        does: Print or install shell completions.
    environment:
      - name: LBH_HOME
        default: ~/.llmboost_hub
        purpose: Root data dir (config, cache, models, license).
      - name: LBH_MODELS
        default: $LBH_HOME/models
        purpose: Downloaded model assets.
      - name: LBH_MODEL_PATHS
        default: $LBH_HOME/model_paths.yaml
        purpose: Saved model-name to host-path map.
      - name: LBH_LICENSE_PATH
        default: $LBH_HOME/license.skm
        purpose: License file.
      - name: LBH_AUTO_PREP
        default: 'True'
        purpose: Auto-run prep when run/serve need missing assets.
      - name: HF_TOKEN
        default: unset
        purpose: Hugging Face token, forwarded into containers.
    config_resolution: environment variable -> $LBH_HOME/config.yaml -> built-in default

  - name: llmboost
    title: LLMBoost inference server CLI
    docs: https://llmboost.mangoboost.io/docs/running/configuration
    summary: >-
      The server CLI inside the LLMBoost container. `llmboost serve` exposes a
      vLLM-compatible serve interface — same flag names and semantics — so existing vLLM
      launch scripts run against it unchanged.
    help:
      - llmboost --help
      - llmboost <command> --help
      - llmboost serve --help   # authoritative, version-correct flag set
    commands:
      - command: llmboost serve <Repo/Model-Name> [flags]
        does: Start the OpenAI-compatible HTTP server.
    flags:
      - flag: --port N
        default: '8000'
        does: Port for the HTTP server.
      - flag: --tensor-parallel-size N
        default: auto
        does: Shard the model across N GPUs.
      - flag: --max-model-len N
        default: model max
        does: Cap context length (prompt + output).
      - flag: --gpu-memory-utilization 0-1
        default: tuned
        does: Fraction of VRAM for weights + KV cache.
      - flag: --max-num-seqs N
        default: tuned
        does: Max concurrent sequences per batch.
      - flag: --enforce-eager
        default: 'off'
        does: Skip graph capture (faster start, lower peak throughput).
      - flag: --scheduling-policy fcfs
        default: fcfs
        does: Request ordering policy.
      - flag: --disable-llmboost-opts
        default: 'off'
        does: Run without LLMBoost's licensed optimizations (diagnostics).
      - flag: --disable-auto-config
        does: Disable automatic configuration when user args conflict with it.
      - flag: --chat-template ./chat_template.jinja
        does: Supply a Jinja2 chat template for base/custom models.

  - name: mango-ctl
    title: MangoBoost DPU device control
    package: mango-cli
    docs: https://sdk.mangoboost.io/docs/guide/cli
    summary: Configure MangoBoost DPU devices. Supports bash tab-completion.
    commands:
      - command: mango-ctl dev list
        does: List the available MangoBoost PCIe devices by BDF.
      - command: mango-ctl dev show [nvme]
        does: Show device details (vendor/device ID, BARs, NUMA node, PCIe link, kernel module).
      - command: mango-ctl dev rescan
        does: Rescan the PCIe devices.
      - command: mango-ctl dev remove
        does: Remove the PCIe device.
      - command: mango-ctl dev enable
        does: Enable the MangoBoost project.
      - command: mango-ctl dev disable
        does: Disable the MangoBoost project.
      - command: mango-ctl dev sriov
        does: Enable SR-IOV for the PCIe device.
      - command: mango-ctl dev reset
        does: Perform a PCIe hot reset on the device.
      - command: mango-ctl ntt start
        does: Start the NVMe-oF Target (NTT) service.
      - command: mango-ctl nvme start
        does: Start the NVMe-oF Initiator (NTI) service on the DPU SoC.

  - name: mango-smi
    title: MangoBoost DPU monitoring
    package: mango-cli
    docs: https://sdk.mangoboost.io/docs/guide/cli
    summary: >-
      Monitor MangoBoost DPU devices — card information plus sensor monitoring including
      total power, FPGA temperature and SoC status.
    commands:
      - command: mango-smi
        does: Print the device table (BDF, vendor:device ID, serial, description, power, temperature, SoC status).

x-evidence:
  fetched: '2026-08-04'
  probes:
    - url: https://llmboost.mangoboost.io/docs/running/lbh
      http_status: 200
    - url: https://llmboost.mangoboost.io/docs/running/configuration
      http_status: 200
    - url: https://sdk.mangoboost.io/docs/guide/cli
      http_status: 200