[planetf1]

Until now the release image was amd64-only, decisions-only, and could not take speech input without extra setup. The audio libraries were missing, so a server started with speech on would pass /health and then fail every audio request with HTTP 500. After this PR each release publishes two multi-arch images (amd64 and arm64), the audio libraries are in both, and the README is written for people deploying the plugin rather than for us.

What changes for a user

• ghcr.io/generative-computing/granite-switch-systemone:X.Y.Z (and latest) is the decisions image. X.Y.Z-asr (and latest-asr) also has the speech model baked in. Each tag is one manifest for linux/amd64 and linux/arm64.
• With SYSTEMONE_SPEECH_INPUT=1, a missing audio library or speech model now stops the container at startup with a message that names the cause and the fix, instead of failing later.
• The README is restructured: container, existing vLLM install and Hugging Face Inference Endpoints first, then request, security, speech, configuration, limits and test status. Everything in it was checked against the code.

How

• image.yml builds each variant on a native runner per architecture (ubuntu-latest, ubuntu-24.04-arm), pushes by digest, and a second job joins the digests into the tags. A branch run pushes only <branch> and <branch>-asr.
• Dockerfile installs granite-switch[audio] in every variant. ASR_REVISION is now the multilingual release of the speech model, pinned by commit; the earlier English-only revision is one build argument away.
• preflight.py checks for av, scipy, soundfile and soxr as well as the speech model. scripts/speech_check.py is a standard-library check for speech input (clips are made at run time, none are committed).

Checked

• Unit tests: 5 new preflight and serve.sh tests fail without the change and pass with it. actionlint, ruff and the SPDX header check pass.
• Workflow dispatched from this branch: build-only run 40013172836 and push run 40014366911 both succeeded. The workflow's own imagetools inspect shows linux/amd64 and linux/arm64 in both tags.
• Served from the pushed arm64 -asr image with docker run on an NVIDIA GB10 (public 8b, speech on): speech_check.py passed. The arm64 default image passed e2e_check, probe_check and accuracy_check, with accuracy matching amd64 on the same checkpoint.
• speech_check.py also passed on amd64 (H100) with a mounted speech-model cache.
• Not checked: the amd64 -asr image was not pulled and served, Hugging Face Inference Endpoints was not tried, and nothing was run under Docker on amd64.

For a reviewer

• The ghcr package is private (an anonymous pull returns 401), so the README tells people to log in or build locally. Can an org admin make it public? If so, that sentence should be dropped before release.
• The two branch tags feat-image-variants-and-build-args and feat-image-variants-and-build-args-asr are on ghcr and can be deleted after review.
• Draft. I have not merged it.