dsh-remote-gpu

dsh-remote-gpu

A DSH plugin that shows live GPU utilization of the server dsh-remote is connected to.

It piggybacks on dsh-remote: the host half calls dsh-remote's own rw_exec tool through the host tools service, so no second SSH connection, no duplicated credential handling, and no extra dependency.

What you get

Seat	What it shows
Conversation header chip	空闲 {idle}/{total} — only how many remote GPUs are free (colour-coded: green = all free, amber = some free, red = none). Click it to open the panel for the details. The tooltip also lists the users currently on the GPU box (占用 <user>×<n>).
Right sidebar tab 远程显卡	Every GPU of the remote host in detail: name, utilization %, memory used/total, temperature, the user who owns the task running on it (占用 <user>×<n>), a 空闲 tag on the free cards, plus a sparkline of recent samples.
Settings → 远程显卡	Probe status, remote host, backend, idle count, refresh interval, the exact probe command, a "refresh now" button, the per-GPU process list (user · command · pid · memory) and an "active users" summary.
Host tool remote_gpu	The same snapshot as text, so the agent can read it directly.

What counts as a free GPU

A card is idle when it is barely drawing power and barely holding memory:

• utilization < idleUtilPercent (default 10 %), and
• memory used < idleMemoryPercent (default 5 %) of the card.

Both must hold: a model that is loaded but idle still owns its VRAM, so that card is not free for a new job. The host computes the verdict and ships it with every sample, so the chip, the panel and the remote_gpu tool always agree on the same number.

Remote host <host> — <backend>
Idle GPUs: <idle>/<total> (util < 10%, memory < 5%)
GPU <index> <name> · util <n>% · memory <used>/<total> MiB · temp <n>°C · idle
GPU <index> <name> · util <n>% · memory <used>/<total> MiB · temp <n>°C · users <user>×<n>

Who is using each GPU

Every GPU card names the owner of the task that runs on it:

GPU <index> · <name>
<util>%   <used>/<total>   <mem%>   <temp> °C   占用 <user>×<n>

The panel header adds the overall count (<total> 张显卡 · 空闲 <idle> 张 · <users> 个用户), and the chip tooltip summarises the same thing. On NVIDIA servers the probe adds three extra sections to the command:

nvidia-smi --query-gpu=index,uuid --format=csv,noheader,nounits   # ==DSHGPU uuid==
nvidia-smi --query-compute-apps=pid,gpu_uuid,used_memory …        # ==DSHGPU procs==
ps -o pid=,user=,comm= -p <pids>                                  # ==DSHGPU ps==

lib/gpu-parse.js maps the GPU UUIDs to indices and the PIDs to their owner (user + command), so each GPU ends up with processes[] and a per-GPU users[] summary. On ROCm / gpustat servers the process list stays empty — those tools do not expose per-process GPU ownership.

Motion

Every reading eases into place instead of snapping:

• one shared requestAnimationFrame loop drives all tweens, so a panel with many GPUs still costs a single frame callback;
• utilization, memory and temperature glide to their new value over ~520 ms (TWEEN_MS in client.js);
• the utilization and memory bars, the sparkline line and the chip dot morph continuously — the sparkline even slides one sample to the left;
• the panel header carries a slow breathing dot, a GPU card fades in when it appears, and the loading state uses a shimmer instead of static text;
• prefers-reduced-motion: reduce switches every animation off.

How the probe works

The host runs one POSIX shell line on the remote machine. It picks the first available backend and wraps every part in ==DSHGPU …== markers:

Section	Content
host	hostname of the remote machine
backend	nvidia-smi, rocm-smi, gpustat or none
gpus	one CSV row per GPU (index,name,utilization.gpu,memory.used,memory.total)
temp	GPU temperature (NVIDIA only)
uuid	GPU index → UUID (NVIDIA only, links processes to GPUs)
procs	running compute apps: pid,gpu_uuid,used_memory (NVIDIA only)
ps	owner of each PID: pid,user,command (NVIDIA only)

lib/gpu-parse.js parses those sections into the snapshot the UI renders.

Requirements on the remote server: nvidia-smi (or rocm-smi / gpustat) on PATH, and a dsh-remote active remote workspace (the probe uses the machine dsh-remote is currently connected to).

Configuration

cordis.patch.yml carries the tunables:

- insert:
    - id: dsh-remote-gpu
      name: '@local/dsh-remote-gpu'
      config:
        intervalMs: 3000        # refresh cadence (500 – 300000)
        commandTimeoutMs: 15000 # remote command timeout
        idleStopMs: 120000      # stop probing when no panel is open
        historyLength: 90       # samples kept for the sparkline
        idleUtilPercent: 10     # a card is free below this utilization …
        idleMemoryPercent: 5    # … and below this memory usage
        # command: '…'          # override the probe command entirely

Install

plugin_manager install_bundle target=<abs path to this directory>
plugin_manager set_bundle target=@local/dsh-remote-gpu enabled=true

install_bundle only runs pnpm; the second command is what puts the plugin's row into the profile's bundle layer (dsh.profile.bundles in the profile package.json). Without it the plugin works until the next DSH restart and then silently disappears — the symptom is a vanished chip and a "removed tool" notice in the session.

The plugin is offline-safe: it has zero runtime dependencies and imports nothing from @deepseek-ai/*.

Layout

One card is ~79 px tall (compact paddings, a single-line footer, a 5 px utilization bar and a 3 px memory bar), so a full node fits into the right sidebar without scrolling; the panel scrolls when there are more GPUs than fit.