feat(model-serving): the muse-glimmer-30b preset on vLLM's own image

[teemow]

Problem

The signed model image for Muse Glimmer 30B exists (giantswarm/models/muse-glimmer-30b-fp8:1deb4641ff84). Its architecture (MuseGlimmerForConditionalGeneration) and its muse_glimmer tool and reasoning parsers first ship in vLLM 0.28.0. The platform's llm-d runtime carries vLLM 0.23, and no llm-d release carries 0.28 yet (v0.9.0 ships 0.26), so a preset on the well-known runtime would never start (#597).

Change

• A new preset, muse-glimmer-30b: Red Hat AI's FP8 block quantization, 32.03 GiB of weights, on one 48 GB GPU (g6e.2xlarge/g6e.4xlarge). It serves 32k context, four concurrent requests and one image per prompt, with --reasoning-parser=muse_glimmer --enable-auto-tool-choice --tool-call-parser=muse_glimmer under vLLM v0.30.0's examples/tool_chat_template_muse_glimmer.jinja (as chat-templates/muse-glimmer-30b.jinja).
• The runtime: the preset's spec.template.containers[main].image names gsoci.azurecr.io/giantswarm/vllm-openai:v0.30.0. That's the copy of vllm/vllm-openai:v0.30.0 that retagger pins by index digest sha256:8a69ffad… (chore: retag vllm/vllm-openai v0.30.0 pinned by digest retagger#1245), and it verifies against the Giant Swarm CircleCI identity. The well-known template's command runs it unchanged: bash, vllm serve /mnt/models, and version-gated flags read from vllm --version. It's the one shipped preset that names an image, until an llm-d release carries vLLM ≥ 0.28.
• Sizing: 52 layers × 2 KV heads × 128 dimensions is 52 KiB of bf16 KV per token, so four 32k sequences take 6.5 GiB. The fit is 33 + 11 GiB of 48 at 0.92.
• tests/verify-gpu-pool.py: the preset is in OCI_PRESETS, and a golden hold leaves its ConfigMaps and discovery name out until origin/main carries them. README and CHANGELOG are updated.

Checks

make verify-presets verify-gpu-pool verify-serving-slice verify-model-images verify-images verify-preset-weights pass. 14 shipped presets' arguments survive the template's eval, and the weights are declared at 33 GiB against the Hub's 32.03 (+3.0%).

Proof

On gazelle, 2026-09-26, one L40S (g6e.4xlarge), this preset rendered as an LLMInferenceService: vLLM 0.30.0 resolved MuseGlimmerForConditionalGeneration and answered a tool-calling request with HTTP 200 in 3.5 s (finish_reason: tool_calls, reasoning via muse_glimmer). Details are on #597. The branch also adds the family's row to model-families.yaml (runtimeImage: vLLM 0.30.0's parsers) and moves the verify-wiring preset count to 14.

Closes #597.

[circleci-architect]

Published Helm chart agent-platform

4.81.3-rb7c6d220t20260926143123h5c1bc38

Chart	agent-platform
Version	4.81.3-rb7c6d220t20260926143123h5c1bc38
OCI reference	oci://gsoci.azurecr.io/charts/giantswarm/agent-platform:4.81.3-rb7c6d220t20260926143123h5c1bc38
Digest	sha256:a9d3387cbed0f73b289b0e6723b4d21d9beb373bbbdc4d99d5c7c3de31cd22b9
Registry	public — gsoci.azurecr.io
Git catalog	giantswarm-test-catalog (index)

Pull this chart

helm pull oci://gsoci.azurecr.io/charts/giantswarm/agent-platform --version 4.81.3-rb7c6d220t20260926143123h5c1bc38

_{Posted by architect-orb · build 14809 · commit 5c1bc38 · updated in place on every push}

[circleci-architect]

Published Helm chart agent-platform-connectivity

4.81.3-rb7c6d220t20260926143123h5c1bc38

Chart	agent-platform-connectivity
Version	4.81.3-rb7c6d220t20260926143123h5c1bc38
OCI reference	oci://gsoci.azurecr.io/charts/giantswarm/agent-platform-connectivity:4.81.3-rb7c6d220t20260926143123h5c1bc38
Digest	sha256:fb27bae6397db24f61763848fa3d05715fe139e3646470e0cde212873de1d0d4
Registry	public — gsoci.azurecr.io
Git catalog	giantswarm-test-catalog (index)

Pull this chart

helm pull oci://gsoci.azurecr.io/charts/giantswarm/agent-platform-connectivity --version 4.81.3-rb7c6d220t20260926143123h5c1bc38

_{Posted by architect-orb · build 14807 · commit 5c1bc38 · updated in place on every push}