resource_stats

Live readings for the dashboards' gauges: GPU utilization and memory, universal across vendors, the session's CPU and memory (SystemUsage), and the one sampler both transports run over them (ResourceMonitor).

Sources, best-first: NVIDIA reads NVML in-process via nvidia-ml-py (the API behind nvitop/nvtop; exact PCI identity, no subprocess per poll) with an nvidia-smi fallback, and Tegra's integrated GPU, which has neither, reports its load through devfreq. Every other card is read through DRM: amdgpu counts utilization and VRAM device-wide in sysfs, and the rest have the engine times in their clients' fdinfo summed, the kernel's vendor-neutral interface that i915, xe, Mali, Adreno and VideoCore all write. A card whose driver writes neither -- Apple's, on Asahi -- is listed with no utilization rather than a zero.

get_gpus(dri_node=...) keys the readings to the render node the pipeline captures/encodes on (PCI match when the source knows its address, else a vendor-unique match), so the monitored GPU is always the one doing the work. Objects expose .load as a 0..1 fraction, None where nothing on the host counts the card's utilization, and .memoryTotal / .memoryUsed in MiB, the units the stats collectors serialize.