[almshary]

On Windows with AMD GPUs, the web Monitor leaves GPU load, dedicated VRAM, temperature, power and PCIe link fields blank. Add an ADL reader from the installed display driver, including HIP-to-ADL device matching, unit conversion and cleanup. Unsupported readings remain unavailable.

Add an optional Arabic interface to Chat, Monitor and API Monitor. English stays the default; the browser remembers the language. Arabic uses RTL layout while code and protocol values retain their direction and values. Language switching preserves drafts and in-flight responses, and the current conversation-cache and model-settings controls are translated.

Replaces #712, which GitHub closed automatically after the upstream history rewrite. This branch starts from the new upstream main (82f46a8) and cherry-picks only the contribution and screenshot commits, preserving original authorship.

Validation:

• 30 tests passed: python -m unittest serve.test_amd_windows serve.test_server.AmdTelemetry serve.test_monitor serve.test_runconfig -q.
• JavaScript syntax checks passed for app.js, monitor.js and i18n.js.
• Playwright checks passed for English defaults, Arabic/RTL, persistence, active requests, API values, conversation cache, PCIe share, unsaved model settings, code direction, API Monitor, mobile layout and unavailable browser storage.

Arabic Monitor screenshot from the original contribution, captured with the model running on an RX 9070 XT:

[crkva]

Hardware validation on 2026-10-09: RX 7900 XTX 24 GB, Ryzen 7 7800X3D, 64 GB RAM, Windows 11 build 26200, AMD driver 32.0.31041.1004. The installed HIP engine was Strata 0.1.41, Coder IQ1_M, 131072 context.

Each PR was checked out separately at the commit below. The hardware reader was sampled 15 times at one-second intervals while the existing model server continued its ordinary workload. These were not controlled performance benchmarks. For HTTP integration, a separate loopback server used MockEngine plus the real GPU reader; /metrics contained the readings and histories, and / returned HTTP 200. No second model was loaded, no inference was submitted by these probes, and the existing server was not restarted or modified.

Tested head: 54f3e428dbc78a68fd120d8b2cf9ab3bd5217e94, Python 3.14.2.

The AMD reader works on this discrete RDNA3 card, including board power.

Reading	Observed range (15 samples, 1 s apart)
GPU activity	1–99 %
Dedicated VRAM used	21.5–21.7 GiB
Total VRAM	24 GiB
GPU temperature	51–61 degrees C
Power	78–362 W
Reported PCIe generation	3
Reported PCIe width	16 lanes

• python -m unittest serve.test_amd_windows serve.test_server.AmdTelemetry serve.test_monitor serve.test_runconfig -q: 30 tests passed.
• The reader identified the RX 7900 XTX; invalid index 99 returned ok() == False and unknown free VRAM, without a fault.
• Explicit double-close completed normally.
• After warm-up, 40 calls to free_vram_mib(0, amd=True) kept process handles at 269 → 269 (delta 0). This is a short cleanup check, not proof of long-term leak freedom.
• PMLog sensor 73 was supported and returned board power; sensors 23 and 30 were absent on this card. The existing (73, 23) preference therefore matters on this driver.

Repeatable admission-query check from the PR checkout:

import gc
import psutil
from serve.telemetry import free_vram_mib

for _ in range(3):
    free_vram_mib(0, amd=True)
gc.collect()
process = psutil.Process()
before = process.num_handles()
for _ in range(40):
    free_vram_mib(0, amd=True)
gc.collect()
print(before, process.num_handles())

Limitations: one discrete card only; no independent PCIe-link or temperature calibration, no browser rendering/Arabic UI test, and no continuous-hours soak. The telemetry unit tests and HTTP/JSON integration above are the scope of this report.

[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.