[kennuardo]

Issue: none (related: #2001 is about the Intel reader picking the wrong card; this is about the charts on a layer split)

Summary

On a two-card layer split the Monitor tab's sparklines show one series per tile (mean load, total VRAM and power, the hottest temperature, the first card's PCIe traffic), so the charts cannot say which card is working. With this change the GPU load, VRAM, GPU temp, Power and PCIe tiles draw GPU 0 as today and GPU 1 as a dashed second line on the same scale, and the tiles' subtext names the cards ("RTX 5070 Ti 72% · RTX 5060 Ti 26%"). One card: nothing changes, on the server or on the page.

What changed

• serve/telemetry.py: with more than one card, _loop also records each card's own gpu<N>_util, gpu<N>_mem_used, gpu<N>_temp, gpu<N>_power, gpu<N>_pcie_rx_mb in the same 60-sample history, and the gpus entries carry pcie_rx_mb / pcie_tx_mb (the reader already read them; they were only summed). Single card: no new keys.
• serve/web/app.js: sparkPair() draws GPU 0 (solid line + area) and GPU 1 (dashed) on one scale when the per-card series exist, else falls back to spark() exactly as before. The per-card subtext uses short card names split from hardware_static.gpu_name (vendor prefix dropped, falls back to GPU <n> when the names and the card count disagree), one card per side of the tile with the value in bold; the Power and PCIe tiles get the per-card subtext the other tiles already had.
• serve/web/app.css: the second line in --st-info; the per-card subtext as a flex row.

Extra Notes

• Measured on an RTX 5070 Ti + RTX 5060 Ti layer split, Linux, engine 0.1.43.1: /metrics grows by five 60-sample series per card (about 3 KB); the sampler does no new reads.
• Default path: with one card the history keys, the gpus entries and the rendered tiles are the same as before (sparkPair returns to spark() when gpu1_* is absent).
• Touches the same PCIe tile as Monitor: the PCIe traffic from the bytes the engine moved, not from a driver counter #2272; both apply, the conflict is a few lines in render().
• Checked: the serve test suite, python -m unittest discover -s serve -p "test_*.py" -t . on this branch and on main at a0116b2: Ran 725 tests ... OK (skipped=11) on both (Linux, Python 3.14). The page was checked by hand on the two-card split (screenshot) and reads the same as before with one card (the per-card series are absent, sparkPair falls back to spark).

[qiuzan2001]

One case the second line misses: serve/telemetry.py stores each card's series under its real index (gpu{g['index']}_util, from the config's GPU list, idx = list(gpu_indices)), but sparkPair() in serve/web/app.js always reads h.gpu0_* and h.gpu1_*. With a split over GPUs 0 and 2 (--gpus 0,2, as in docs/INSTALL.md) or 1 and 2, gpu1_* does not exist, so every tile falls back to the single aggregate line while the subtext names both cards. Taking the two series keys from hw.gpus[0].index and hw.gpus[1].index would cover it. (Read on the branch, not run in a browser; no multi-GPU box here.) Also, the new code comments are marked local:, which reads like a note from a private fork.

[Niko1221]

Thanks for the pull request.

We're resetting the tracker for 0.2.0-beta. There are more than 1,300 open issues and pull requests, too many for us to review properly, so we're closing everything that isn't being worked on right now. Closed here does not mean rejected.

To reopen it takes about two minutes:

1. Rebase on main if it no longer applies, and re-run your checks on 0.2.0-beta.
2. Start the title with its area: [Engine], [GUI], [TUI], [Benchmark], [Docs] or [Other].
3. Fill in the pull request template: the problem and result, and the validation (what you ran, on which revision, what you saw).
4. Comment /reopen. A check runs and either reopens the pull request or lists what is still missing.

The rules are in CONTRIBUTING.md.