Skip to content

cstream-gpu-va-tooling

Version 2026.243.0429
Repo box/cachyos

BED FIXTURE ONLY – never composed into a shipped box.

Installs libva-utils so the GPU bed can ANSWER its own question.

The bed exists to measure one thing: whether a passed-through NVIDIA card offers VA H.264 EncSlice. The plan asserts it does not, and three later encoder gates are built on that premise – encode: va must fail loudly on this venue, encode: auto must land on software.

Without vainfo the guest cannot be asked. layer-nvidia installs the VA driver (libva-nvidia-driver) only on its FEDORA arm; the arch/cachyos arm installs nvidia-utils, which carries no VA driver and no vainfo. So on this guest the measurement had nothing to run, and the first version of the check “recorded” that fact and PASSED – reporting success having determined nothing, which is the failure shape this bed exists to catch.

Deliberately NOT libva-nvidia-driver. That package is decode-only (it maps VA-API onto NVDEC and exposes no EncSlice entrypoint at all), so installing it would not create an encode path – it would only change which of two no-encode answers the guest gives. The production cstream image does not install it either, so measuring without it is measuring the real venue.

Installed on every distro:

  • libva-utils

This candy’s plan: — the runnable spec charly check executes against a live deployment. check: steps are idempotent probes; run: steps change state.

Intent Step
check vainfo is installed, so the VA inventory can actually be read