Pull requests / #2429

#2429 Add Intel GPU monitoring on Windows via IGCL

open · @marcelloc · 0 comentarios · En GitHub

Setup & installServer & APIMulti-GPUAMD / HIPNVIDIA / CUDAModels & quantsDocumentationWindowsLinux

Descripción

# Add Intel GPU monitoring on Windows via IGCL

## Summary

**Compatible with Strata 0.1.44.1**, the upstream version used when preparing
this contribution (commit `296b59d2`). The IGCL monitor was validated with the
native Windows SYCL server after integrating that release.

Populate the SYCL monitor's missing Intel GPU readings on Windows using an
optional native IGCL helper. The user explicitly builds/installs the helper;
inference continues without it when it is unavailable.

Related: https://github.com/Niko1221/Strata/pull/1940 and
https://github.com/Niko1221/Strata/pull/1390. This contribution can complement
their native Windows work; it does not include engine portability changes.

## What changed

- Read supported activity, memory, temperatures, chip energy and PCIe link data.
- Derive rates safely; discard invalid, reset or stale readings.
- Clean up the helper on server/parent exit and launch it without a console.
- Label chip power and PCIe link capability accurately in the Windows monitor.
- Add collector and UI label tests, a manual build script and setup documentation.

The helper uses the driver-installed ControlLib.dll. No vendor binaries or
headers are bundled. Shared CUDA/HIP server files are unchanged.

## Validation

Windows 10 / Intel Arc Pro B70: helper build, actual API readings during
inference, parent-exit cleanup, collector and monitor unit tests, and JavaScript
label tests were exercised in the local implementation. The source files in
this contribution match that implementation. See docs/WINDOWS_INTEL_MONITOR.md.

Additional user-reported usage: code review with IQ3_XXS, MTP and a context
configured to 128K, with the server remaining up since the previous day.
This is practical usage evidence, not an instrumented soak test; it does not
establish that the conversation filled the entire 128K context window.

## Extra Notes

Ready for review based on the Windows 10 / B70 validation above. Windows 11
and multi-GPU mapping remain untested; results can be added when available.

The IGCL collector is Windows-specific, but the integration also touches
`sycl/serve/server_intel.py`, which is shared with Linux. Linux regression
validation has not been performed. Maintainer or community help checking that
path is welcome; no Linux compatibility validation is claimed here.

Power is GPU chip power, not total board power. PCIe throughput is
unavailable and is not inferred from generation/width. A failed helper requires
server restart after the underlying issue is resolved.

En el sitio

Enlaces a install, modelos, releases.