LMStudio: select GPU runtime in CI, harden ping test

[lucbruni-amd]

Addresses: #419, #457

Summary

• lmstudio-ping-endpoint timed out on runners where LM Studio ran on the CPU (avx2) runtime.
• Add a hidden Linux CI step that selects an installed Vulkan runtime before model load.
• Harden the ping test: max_tokens 500 to 64, socket timeout 60s to 120s.

Why

Nothing pinned the LM Studio runtime, so an app update could silently select the CPU engine, and a 500-token reasoning generation exceeded the hardcoded 60s client timeout. All 6 Linux runners were switched to Vulkan; the fix passed in the latest scheduled run:

https://github.com/amd/playbooks/actions/runs/29484903834

Test plan

• Latest scheduled run: all Linux lmstudio jobs (krk, halo, stx, rx7900xt, rx9070xt, r9700) passed.

[lucbruni-amd]

Getting another recurrent failure in the failing check for this PR: https://github.com/amd/playbooks/actions/runs/29502189131/job/87634358274?pr=549

No models are currently loaded.

Tackling these as well.

[sreeram-11]

Getting another recurrent failure in the failing check for this PR: https://github.com/amd/playbooks/actions/runs/29502189131/job/87634358274?pr=549

No models are currently loaded.

Tackling these as well.

---

The issue appears to be that the model fails to load initially. As a result, when the next command tries to chat with the model, it reports that no model is available because the model was never loaded successfully.

Model loading failure

The model load step fails with the following error:

Error: Failed to load model

   (X) CAUSE

Failed to load model

This also explains why lms ps shows the following immediately after the model load command:

No models are currently loaded.

---

Follow-up chat failure

After that, when we try to chat with the model, it fails because the model was not loaded:

Error: Model "gpt-oss-120b-29502189131" not found, load with:
       lms load gpt-oss-120b-29502189131

So the root problem is the initial model loading failure. The later “model not found” error is a downstream issue caused by the model never being loaded successfully.