v1.5.2

who-gpu v1.5.2

A round of small fixes, plus a smoke test.

• uninstall.sh removes only the files the dashboard wrote. It no longer deletes a custom WHO_GPU_OUT directory that holds your own files.
• --setup on macOS's stock bash 3.2 no longer aborts when you only turn hosts on, or only off.
• GPU process names containing a quote are no longer dropped.
• The full view prints "(none)" instead of a blank line when nobody is logged in.
• A host's GPU users show "?" for a process whose owner is unknown, matching the per-GPU lists.
• When there is no new release, the update notice links to a compare view of exactly what --update would pull.
• New test.sh: a smoke test that runs every mode locally against a fake ssh and nvidia-smi, needing only bash.
• README: fix the connection-reuse note that pointed at an install-only flag.

v1.5.1

who-gpu v1.5.1

Shows which GPU each process holds, and who holds each GPU.

• The --full view names the card next to every GPU process.
• Each GPU on the dashboard lists its users under its label, and --json gains a per-GPU users field.

v1.5.0

who-gpu v1.5.0

This release makes the web dashboard show more information, and tidies up the code underneath.

Dashboard

• Each card names the GPU models and current users under the hostname and shows total VRAM. New sort columns: GPU model, memory used, memory total.
• Every GPU bar is split in two: utilization on top, memory used below, with both numbers alongside.
• Unified-memory machines (GB10, Jetson) report system RAM as their GPU memory, are marked as such, and are counted once.
• Starting a second --web over the same directory exits with a hint to use the one already running, instead of two engines fighting over the same page.

Minor changes:

• A GPU with a process attached now counts as busy even at 0% utilization.
• Hosts behind a jump host get a longer connection timeout, and a host that is merely slow to answer is retried once before being called unreachable.
• --update is more robust when you pull by hand.
• --help prints the whole header again (the last line had been cut off) and now lists --json and --update.
• --full output blocks no longer interleave between parallel workers.
• A missing or non-numeric value for -t, -n, -p or WHO_GPU_INTERVAL gives a plain message instead of a bash error.

Under the hood:

• The dashboard page now lives in fleet.html in the repo instead of an 800-line heredoc, and the main script's flow is gathered into one main(). Behaviour remains unchanged.

v1.4.0

who-gpu v1.4.0

This release upgrades the web dashboard. New features include a list view, ability to see nvidia-smi output directly, and controls for sorting and grouping.

Dashboard

• Click anywhere on a machine's card to open or close its details. The details now have two tabs: You can toggle between who-gpu's full output, and also the plain nvidia-smi output.
• A new, more compact list view. switch with the "List view" button in the top bar.
• Sort by any column, ascending or descending, from either list view or card view.
• Option to enable/disable grouping by activity.

Minor changes:

• Opening and closing details, switching tabs and widening a card are animated. The animations are switched off if your system asks for reduced motion.
• Empty user fields show a dash instead of "nobody".
• The update notice in the corner is now a clickable link to the release it is announcing.
• Your view, grouping and sort choices are remembered by the browser.
• Dashboard files now live in ~/.local/share/who-gpu (or $XDG_DATA_HOME/who-gpu) instead of a separate who-gpu-web folder in your home directory. An existing folder is moved there the first time you run --web, and uninstall.sh removes the generated files.
• --json output includes the nvidia-smi table as "smi" and the update link as "update_url". Terminal output is unchanged.

v1.3.3

who-gpu v1.3.3

Adds a default mode setting and on-demand probing.

• DEFAULT_MODE=terminal|web in the config decides what a plain who-gpu opens. Set it with --setup, install.sh --default-mode, or by editing the config. Explicit flags still win, and the desktop icon keeps following ICON_MODE.
• While --web runs you can probe before the interval is up: press Enter in the terminal, or click the new "Probe now" button on the page.

v1.3.2

who-gpu v1.3.2

The dashboard opens immediately and fills in as hosts answer.

• The page used to appear only after the slowest host had replied. Now it opens at once with every machine marked "probing", and cards fill in one by one. On later cycles a host keeps its previous numbers until new ones arrive.
• README: add screenshots.

v1.3.1

who-gpu v1.3.1

Fixes the dashboard never opening in the browser on Windows.

• Git Bash rewrote the /c argument to cmd.exe as a drive path, so the browser was silently never launched. That call is now protected, with PowerShell and explorer.exe as fallbacks.
• README: trim the connection-reuse explanation.

v1.3

who-gpu v1.3

Slows the dashboard refresh automatically when SSH connections cannot be reused.

• When connection reuse turns out to be unavailable (mainly Git Bash), the refresh interval moves to 60s on its own instead of only printing advice. An interval you chose yourself is used exactly as given.
• The effective rate is announced once, and the dashboard's staleness warning scales with it.
• README: explain the one-time manual upgrade (git pull && ./install.sh) for installs older than v1.1, which predate --update.

v1.2

who-gpu v1.2

Windows fixes for the v1.1 dashboard release.

• The installer puts ~/.local/bin on PATH in Git Bash so who-gpu --update is runnable. It only appends one line, backs up the profile first, and --no-path opts out.
• SSH connection reuse is now verified after the first probe cycle and switched off where multiplexing is unsupported, such as Git Bash, instead of reporting every host as unreachable.
• install.sh --help no longer prints a stray line of shell code.

v1.1

who-gpu v1.1

Adds --web: a live fleet dashboard in the browser, with no new dependencies. Still just bash and ssh, on all three platforms.

• --web opens a live dashboard that refreshes in place, shows how old its data is, and warns loudly if the probe loop stops. Files go to ~/who-gpu-web.
• One SSH connection is reused per host while the dashboard runs, instead of a fresh login on every refresh.
• --update fast-forwards the checkout and --version reports the version. A once-a-day check mentions new versions but never installs them.
• The desktop icon now opens the dashboard by default. install.sh asks which mode you prefer and saves the choice in ~/.config/who-gpu/config.
• Fix --setup on macOS, where BSD mktemp rejected the bare call.