You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Shader wallpapers now render every vsync (60 fps; 30 fps after 4 s idle) with pure rAF pacing
instead of 18 fps / 6 fps idle through setTimeout. GPU cost is held constant by a pixel budget
(36 Mpx/s → ~0.6 Mpx canvas, CSS-upscaled); an adaptive governor measures real frame deltas and
steps resolution down 15% when frames are dropped, creeping back after 8 s of clean frames.
Per-theme perf.fps below 30 is clamped. Asset version bumped so cached runners refresh.
[2.1.0] — 2026-10-31
Hero throughput card
"Total tok/s" is now the deck's hero instrument: a large odometer that tweens toward the live value
(fast attack, slow decay) with glow and colour temperature scaling against the session peak; a token
particle stream over the chart whose density and speed follow throughput and converge on the chart
head; thicker glowing area chart with pulsing head dot and a dashed session-peak line;
↑out ↓in · pk · rec sub-line (record persisted per browser). The card takes 1.55× the width of the
cluster-util chart in fit mode.
Endpoint scraper: counters appearing for the first time (endpoint just turned healthy / server
replaying lifetime totals) no longer register as a throughput spike.
[2.0.0] — 2026-10-31
Layout — "fit" mode
The dashboard is now a single viewport-filling grid (header / deck / GPU grid / footer) with no page
scroll whenever the window is ≥ 760×540. All four GPUs are always on screen, whatever the tile size.
Toggle to the old scrolling page with L or the ⊞ button (persisted in localStorage).
GPU cards are CSS size containers: every dimension inside is em off a --fs derived from the card's
own height/width, so gauges, bars and text scale to the tile. Container queries re-arrange the card:
short cards keep hardware tags in the head row; tall cards drop them under the instruments; portrait
cards (4-across on ultrawide) stack instruments over telemetry.
GPU grid shape follows the window shape: 2×2 by default, 1×4 for portrait tiles, 4×1 on ultrawide.
Models and the two cluster charts share one horizontal "deck" band; the band grows to 50 vh while a
model boot is in progress.
Compact horizontal model cards in fit mode (identity · meta line · stats · spark); long-form blurbs
only in the scrolling layout.
GPU card — meter strips
Bars + sparklines merged into four meter strips (Memory, Power, Util, Temp): header with peak,
a thick glowing bar whose thickness scales with the strip, and the history graph filling the rest.
2×2 on short cards, 4 rows on tall; below ~58 px of strip height the graph yields and the bar thickens.
Hardware tags moved to pills: PCIe link, AER, SM/MEM clocks, P-state, fan, per-GPU TPS, model.
Fan % and TPS-share are no longer bars.
Processes rendered as one chip line in fit mode (heaviest first), full table in the scrolling layout.
GPU name shortened for the head row (RTX PRO 6000 Blackwell WS), full name and UUID in tooltips.
Model boot — launch sequence (new bootlog.py)
The serving engine's own stdout is tailed (docker container via /proc/<pid>/cgroup → docker logs,
or a log file on fd 1/2) and parsed into a structured boot: engine/version, architecture, TP,
max model len, KV dtype, expert dtype, checkpoint size, MoE backend, NCCL version.
Stage track INIT / WEIGHTS / WARMUP / API with live per-stage durations and the previous boot's
durations beside them; ETA from the last successful boot of the same model (boot_history.json).
Progress bar driven by whatever bar the engine is drawing: safetensors shards, CUDA-graph capture
(FULL / PIECEWISE), JIT kernel warmup, parallel compile, GEMM tuning.
Per-rank VRAM fill against the per-rank weight target; warmup checklist (model init, weights,
KV budget GiB/rank, KV cache tokens + max concurrency, compile, CUDA graphs, API server, TTFT probe);
host-side cells (disk read, VRAM fill rate, loader CPU, RAM staged, host RAM free at load).
One-shot TTFT probe (1-token streaming completion) when the API comes up; retried if refused.
Failure state: root-cause line surfaced (first <X>Error: after the traceback), bar frozen,
card and stage track in red; BOOT FAILED / ONLINE toasts and wallpaper pulses on milestones.
6-line engine log tail with level colouring; traceback frames and tqdm spam filtered.
Resident models show boot · ttft · N boots from history.
Handles docker logs spanning container restarts (a second engine banner resets the boot) and
ignores vLLM's early "Starting vLLM API server" line (readiness = "Application startup complete").
Telemetry
PCIe RX/TX per GPU (nvmlDeviceGetPcieThroughput) with link capacity at the current Gen × width:
shown in the PCIe pill as ↑/↓ GB/s with a bus-utilisation fill, cluster total in the hostline,
per-GPU pcie history series.
Energy: lifetime NVML energy counter → kWh since local midnight (baseline persisted in energy_day.json, survives restarts and driver counter resets) as a sixth header KPI;
tokens per kJ over a 30 s window in the hostline.
vLLM serving quality from /metrics histograms: TTFT p50/p95 and inter-token latency p50/p95
(windowed bucket deltas), prefix-cache hit rate, preemptions — as live cells on the model card.
Sniffer
Loading entries deduplicated by pid set (export/quant jobs touching many weight files produced
one card per file group); boots sort first.
NVTOP_DEMO_BOOT=/path/server.log replays a captured engine log as a synthetic loading model
(NVTOP_DEMO_LPS lines/s, NVTOP_PORT to run a second instance).