Issues / #2921
#2921 [Bug]: Intel Arc telemetry keeps an idle-unloaded server GPU awake
open · @pankleks · 0 commentaires · Sur GitHub
Server & APIModels & quantsLinux
Description
### What happened On Intel Arc / xe, an idle-unloaded Strata server can keep the GPU awake even though there are no inference requests and the engine has exited. The Monitor displayed about 41 W, but that is not an independent power measurement. The server starts `Telemetry._loop` for its lifetime, sampling once per second. `_XeGpu.read()` continues reading hwmon temperature/energy/power-cap and PCIe sysfs attributes after the model unloads. Closing the browser is not sufficient to stop this background sampler. Reproduction: 1. Serve a model using the SYCL wrapper with `idle_unload_s=600` and Linux GPU `power/control=auto`. 2. Wait for `loaded:false`, no in-flight requests, and no DRM clients (also reproduced after explicit `/v1/unload`). 3. Observe only the driver's software PM counters, without sampling GPU sensors. Expected: the unloaded server does not poll hardware and prevent runtime suspend; CPU/RAM/server metrics remain available. Observed before the patch: `runtime_status=active`, `runtime_usage=0`, with +15,002 ms active and +0 ms suspended over a quiet 15-second interval. Stopping only Strata for 120 seconds yielded +14,422 ms active / +105,582 ms suspended, ending `suspended`. ### Strata version, GPU, OS Strata 0.1.45 / commit `477ac13cbe8e`, Intel Arc Pro B70 32 GB (ASRock, GSC 31.1062), Linux `7.0.14-24-pve`, unprivileged Linux container, IQ3_XXS at 262,144 context. No driver, firmware, model or power-policy change was needed for the reproduction. ### Engine log The server logged `model unloaded after 600 s idle; the next request loads it again`. At a later check, request count was unchanged, `loaded:false`, `in_flight:0`, and the last inference was about 79 minutes earlier. No raw configuration or private logs are attached. Prepared a small fix to pause SYCL GPU telemetry while `Service.loaded()` is false, keep non-GPU telemetry running, and reset the Intel reader's energy baseline before resuming. A check based only on transient Linux runtime-PM state was not sufficient; it is not part of the proposed final patch.
Sur le site
Liens install, modèles, releases.