generated: '2026-08-16'
method: searched
source: >-
  https://developer.furiosa.ai/latest/en/device_management/system_management_interface/furiosa_smi_cli.html,
  https://developer.furiosa.ai/latest/en/furiosa_llm/furiosa-llm-serve.html,
  https://developer.furiosa.ai/latest/en/get_started/prerequisites.html
summary: >-
  FuriosaAI ships two first-party command-line tools plus a kubectl plugin. `furiosa-smi` is the
  device-management CLI (the nvidia-smi analogue for RNGD) and `furiosa-llm` is the model
  lifecycle and serving CLI. Note the asymmetry: furiosa-smi has a dedicated reference page with
  worked output for every subcommand, while `furiosa-llm serve` is documented through prose and
  option tables scattered across the serving, responses and tool-calling pages - there is no
  single `furiosa-llm --help` reference in the developer center.
clis:
  - name: furiosa-smi
    description: Device management and telemetry for FuriosaAI NPUs.
    docs: https://developer.furiosa.ai/latest/en/device_management/system_management_interface/furiosa_smi_cli.html
    install:
      - method: apt
        command: sudo apt install -y furiosa-smi
        note: Installs furiosa-libsmi and furiosa-smi. Requires the FuriosaAI APT repo (asia-northeast3-apt.pkg.dev/projects/furiosa-ai).
      - method: dnf
        command: sudo dnf install furiosa-smi
        note: Rocky Linux 10 / RHEL 10 via asia-northeast3-yum.pkg.dev/projects/furiosa-ai/el10.
    requirements:
      - Ubuntu 22.04 LTS / Debian Bookworm or later, or Rocky Linux 10 / RHEL 10
      - Linux kernel 6.3 or later
      - root privileges
    synopsis: furiosa-smi <sub command> [option]..
    commands:
      - name: info
        description: Architecture, device, firmware version, temperature, power draw and PCI BDF per NPU.
        flags:
          - --full
          - --format full
        note: --full / --format full adds UUID, serial number, core clock, governor and PCI-DEV.
      - name: status
        description: Per-device liveness, memory used/total, and per-core utilization (shows "-" when the core is idle/unused).
      - name: ps
        description: Lists the OS processes currently occupying each NPU device.
    library:
      name: furiosa-smi library
      docs: https://developer.furiosa.ai/latest/en/device_management/system_management_interface/furiosa_smi_lib.html
      bindings:
        - python: furiosa-smi-py (PyPI)
        - go: github.com/furiosa-ai/furiosa-smi-go
  - name: furiosa-llm
    description: Build, prepare and serve models on RNGD.
    docs: https://developer.furiosa.ai/latest/en/furiosa_llm/furiosa-llm-serve.html
    install:
      - method: pip
        command: pip install furiosa-llm
      - method: docker
        command: docker pull furiosaai/furiosa-llm:latest
    commands:
      - name: serve
        synopsis: furiosa-llm serve [ARTIFACT_PATH] [options]
        description: Start the OpenAI-compatible HTTP server hosting one model artifact.
        flags:
          - flag: --chat-template
            description: Path to a Jinja2 chat template, overriding the tokenizer's built-in one.
          - flag: --enable-auto-tool-choice
            description: Enable automatic tool-call detection in model output.
          - flag: --tool-call-parser
            description: Tool-call parser to use (e.g. hermes for EXAONE-4.0).
          - flag: --reasoning-parser
            description: Reasoning parser to use (e.g. qwen3). --help lists supported parsers.
          - flag: --enable-responses-api-store
            description: Enable the in-memory Responses API store (default off).
          - flag: --responses-api-store-max-entries
            description: Max stored responses; default 10000, oldest evicted first.
          - flag: --responses-api-store-ttl
            description: TTL for stored responses in seconds; default 3600.
        example: furiosa-llm serve furiosa-ai/Qwen3-32B-FP8 --reasoning-parser qwen3
  - name: kubectl view-rngd
    description: kubectl plugin for viewing RNGD device allocation across a cluster.
    source: https://github.com/furiosa-ai/kubectl-view-rngd
    install:
      - method: krew
        command: null
        note: >-
          Published to the krew-index (furiosa-ai maintains a krew-index fork). The developer
          center does not document an install command, so none is asserted here.
    status: published-undocumented
firmware_tools:
  - name: furiosa_rngd_updater_all
    description: Upgrade firmware on all RNGD devices. As of SDK 2026.3.0 this must be run manually after installing the firmware image package, followed by a cold reboot.
  - name: furiosa_rngd_updater
    synopsis: furiosa_rngd_updater -b <BDF> -f <firmware image>
    description: Upgrade firmware on a single device by PCI BDF.
cli_count: 3