vllmstat 0.2.2

nvtop for vLLM — an interactive terminal dashboard for vLLM serving performance

vllmstat

nvtop for vLLM — a zero-infrastructure interactive terminal dashboard for vLLM serving performance.

---

Why vllmstat?

The standard observability stack for vLLM is Prometheus + Grafana: powerful, but heavyweight. You need a running Prometheus instance, a Grafana server, a dashboard JSON import, and a browser tab — all just to see whether your inference server is busy.

vllmstat replaces that for day-to-day monitoring. One command, no infrastructure. It scrapes the vLLM server's built-in /metrics endpoint directly and renders everything in your terminal, refreshing every second.

There is one other terminal tool (vllm-top on PyPI), but it is a basic watch-style metrics printer: no interactivity, no GPU panel, no latency percentiles, no speculative-decoding acceptance, no KV-compression ratio. vllmstat fills that gap — it is closer to nvtop than to watch.

---

Install

pip install vllmstat

Or with pipx (isolated install, globally available):

pipx install vllmstat

Or run it ephemerally without installing:

uvx vllmstat

---

Usage

Point it at your vLLM server and it starts immediately:

vllmstat

# Different host / port
vllmstat --url http://my-gpu-host:8000

# Try the dashboard without a real server (uses synthetic data)
vllmstat --mock

# Print a single snapshot as JSON and exit — useful for scripting / alerting
vllmstat --once --json

Key bindings

Key	Action
q	Quit
p	Pause / resume polling
g	Toggle GPU panel on/off
+ / =	Halve the refresh interval (faster)
-	Double the refresh interval (slower)

Flags

Flag	Default	Description
-u / --url	http://localhost:8000	vLLM server base URL
--metrics-path	/metrics	Prometheus metrics path
-i / --interval	1.0	Refresh interval in seconds
--api-key	—	Bearer token (VLLM_API_KEY env var also accepted)
--no-gpu	—	Disable the GPU panel entirely
--mock	—	Use synthetic data — no server required
--once --json	—	Print one snapshot as JSON and exit
--version	—	Print version and exit

---

What it shows

• Concurrency — running requests, waiting queue depth, preemption rate, with mini sparklines.
• Throughput — generation tok/s, prompt tok/s, tokens per iteration, requests per second.
• Cache & KV memory — prefix-cache hit rate (windowed and lifetime), token-source breakdown (compute vs. cache-hit vs. external KV transfer), KV-cache utilisation percentage, KV-cache capacity in tokens, and — when a quantised KV dtype is detected — the dtype (fp8_e4m3, turboquant_k3v4_nc, …), effective compression ratio vs. fp16, and how much fp16 memory the model's full context would require. For example, a turboquant k3v4 cache shows ~4.6× compression and a note that the full context would need 25.8 GB in fp16.
• Latency percentiles — TTFT, TPOT, end-to-end, and queue-wait time, each at p50 / p90 / p99, computed over a rolling window so recent spikes are visible immediately.
• Speculative decoding — acceptance rate, accepted tokens per draft, per-position acceptance (when the server reports it). The panel is hidden when spec-decode is not active.
• Per-GPU stats — utilisation %, VRAM used / total, temperature, power draw vs. limit, clocks, fan. Works on NVIDIA, AMD, and Intel GPUs (see GPU support for what each vendor reports). Multi-GPU and mixed-vendor hosts show every GPU.

---

GPU support

vllmstat detects each GPU's vendor from its DRM device and reads stats from the best source available. Every field degrades to — when its source is unavailable, and a missing driver, tool, or sysfs file never crashes the dashboard — it just shows less.

Vendor	What works	Prerequisite
NVIDIA	Full: util %, VRAM used/total, temperature, power draw/limit, SM & memory clocks, fan %.	NVIDIA driver. The bundled nvidia-ml-py uses NVML; nvidia-smi on PATH is used as a fallback.
AMD	Full: util %, VRAM used/total, temperature, power draw/limit, fan RPM, clock — via the amdgpu kernel driver's sysfs.	amdgpu kernel driver (in-tree on modern Linux). Install ROCm's amd-smi (or rocm-smi) for richer data; it's used automatically when on PATH.
Intel	Utilisation %, temperature, power draw/limit, clock, and fan RPM out of the box via the xe/i915 sysfs — no root. VRAM used via DRM fdinfo — see the note below for the root requirement.	xe or i915 kernel driver. No extra tools needed; util/temp/power/clock/fan work as a normal user. Root (or matching UID) is only needed for VRAM.

Intel utilisation (no root): the xe driver exposes no gpu_busy_percent, but it does expose a world-readable, cumulative GT-idle counter at …/device/tile*/gt*/gtidle/idle_residency_ms. vllmstat reads it each refresh and derives util % as 100 × (1 − Δidle_ms / Δwall_ms), taking the busiest GT (a card can have a render/compute gt0 and a media gt1). No root, no extra tools. Utilisation needs two refreshes to produce its first delta; Intel power is derived from the energy1_input counter, so it likewise appears one refresh after the panel opens.

Intel VRAM (DRM fdinfo, root-gated): the xe driver exposes no mem_info_vram_* in sysfs, so vllmstat reads VRAM the way nvtop does — by summing each GPU client's drm-resident-vram0 from /proc/<pid>/fdinfo/<fd>. Reading another process's fdinfo requires a matching UID or root, so VRAM appears only when vllmstat can read the vLLM worker processes (see Getting GPU stats below). Without that access VRAM shows — with a (VRAM needs root) hint; total VRAM capacity isn't reliably exposed on xe yet, so VRAM is shown as used/—.

---

Getting GPU stats

The GPU panel works with no configuration on all three vendors — but each vendor sources its data differently, and one case (Intel VRAM) can need elevated permissions. Here's how to get the full set.

NVIDIA

Install the NVIDIA driver. Utilisation, VRAM used/total, temperature, power draw/limit, and SM/memory clocks all come from NVML via the bundled nvidia-ml-py; if NVML isn't importable, vllmstat falls back to nvidia-smi on your PATH. No root required.

AMD

The in-tree amdgpu kernel driver (present on modern Linux) exposes utilisation, VRAM used/total, temperature, power, and fan via sysfs out of the box — no root, no extra tools. For richer data, install ROCm's amd-smi (or the older rocm-smi); vllmstat uses whichever is on your PATH automatically.

Intel (Arc / xe or i915)

Utilisation, temperature, power, clocks, and fan work out of the box, no root — they come from world-readable sysfs (utilisation from the GT idle-residency counter; see GPU support above for details).

VRAM is the one exception. It's read per-process from DRM fdinfo, so it only appears when vllmstat can read the GPU process. If your vLLM runs as root (e.g. inside Docker) while you run vllmstat as a normal user, VRAM shows — with a (VRAM needs root) hint. To get VRAM, either:

• Run vllmstat as the same user as vLLM (simplest if you launched vLLM yourself), or

• Run vllmstat as root to match a root-owned vLLM:

  sudo $(which vllmstat)
  # for a pipx install:
  sudo ~/.local/bin/vllmstat
  

Note: kernel.yama.ptrace_scope does not help here. Reading another user's fdinfo is blocked by a cross-UID ptrace_may_access check that requires a matching UID or root — relaxing ptrace_scope does not change it.

Keeping vllmstat current

pipx upgrade vllmstat

---

Remote and containerised setups

vllmstat does not need to run on the GPU machine. If no GPU is reachable from the machine you run it on — no NVML/nvidia-smi, no amdgpu/xe sysfs — for example when monitoring a remote server or when vLLM is isolated in its own GPU container, the GPU panel shows "unavailable" and all the vLLM telemetry panels (concurrency, throughput, cache, latency, spec-decode) continue to work normally. Pass --no-gpu to suppress the panel entirely.

---

Requirements

• Python ≥ 3.10
• A running vLLM server that exposes its Prometheus /metrics endpoint (all vLLM ≥ 0.4 deployments do this by default)
• A GPU driver — optional, only needed for the GPU panel. NVIDIA (NVML/nvidia-smi), AMD (amdgpu), or Intel (xe/i915); see GPU support.

---

Development

See CONTRIBUTING.md.

---

License

Apache-2.0. See LICENSE.