[Testing] Migrate accelerator-backed tests to vLLM CPU servers in CI

[Bslabe123]

The tests that needed accelerators never ran in CI, and most of what they check does not depend on GPU speed, so they move to real vLLM CPU servers inside the existing E2E Test on change job. This is the #703 pattern (a CPU vLLM as the real-server oracle, implemented for token accounting and /metrics in #697) applied to the rest of the accelerator-backed surface.

What was accelerator-backed

tests/optional, the hand-run @pytest.mark.live tier: eight multimodal cases (images, video, mixed, frame sampling, two shared-prefix variants, VisionArena, audio) and one text chat case, each deploying Qwen3-VL-8B or Qwen2-Audio-7B on an H100 node through deploy/manifests.yaml and asserting that every request completes. With no accelerator capacity in CI (#641), the tier ran only when someone ran it by hand, so changes to multimodal payload building shipped with no real-server check.

The rule for what moves

Fake the conditions, never the oracle (#606): a case moves when its oracle is the server accepting a payload, or the server's own numbers, because neither depends on model size or speed, and the sim cannot stand in (it matches multimodal blocks by type string and never decodes bytes). A case stays on a cluster only when the cluster is the oracle.

Work

• test: run the optional tier's payload cases against vLLM CPU servers #745: the nine payload cases become e2e/tests/test_vllm_cpu_multimodal.py against OpenGVLab/InternVL3-1B-hf (image + video) and ibm-granite/granite-4.0-1b-speech (audio), run by a multimodal-e2e job on ARM runners per release in e2e/vllm_multimodal_releases.txt, merge-blocking from its first run. tests/optional keeps text/chat only.

Stays accelerator-backed, on purpose

• tests/optional/text/chat: the deploy path itself (manifests, published image, ConfigMap, Job on a GPU node) is the oracle, and nothing else exercises it. Hand-run with --kubeconfigs.
• [Testing] Periodic metric-name drift detection for vLLM, SGLang, and TGI #669's SGLang scrape (test: catch SGLang and TGI metric-name drift #737): SGLang's CPU backend requires AMX, which hosted runners do not have, and test: catch SGLang and TGI metric-name drift #737 carries its own runner gate. TGI ships an -intel-cpu image and can follow later. Neither is vLLM, so both are out of scope here.

Part of the v0.7.0 release (#606), Test infrastructure and gating.

[Bslabe123]

#745 merged, which was the only item on this issue's Work list. The nine payload cases now run as a multimodal-e2e job inside E2E Test on change on ARM runners, merge-blocking, against OpenGVLab/InternVL3-1B-hf for image and video and ibm-granite/granite-4.0-1b-speech for audio. tests/optional keeps text/chat only.

Two surfaces stay accelerator-backed on purpose, as scoped above: tests/optional/text/chat, where the deploy path itself is the oracle and nothing else exercises it, and #669's SGLang scrape in #737, whose CPU backend requires AMX that hosted runners do not have. Neither is a regression from this migration.

Closing as complete.