Realtime troubleshooting

Qwen3-Omni Realtime operation

Open Models → Create → Location: Local → Inference Engine: (Experimental) vLLM-Omni. Location is not an engine selector and contains only Local and External. Regular engines appear alphabetically above the experimental FreeToken and vLLM-Omni options; the labels do not change saved model configurations.

Omni supports Hugging Face search and Direct reference just like the regular model discovery workflow. Magic Stick does not reject quantized, unknown-architecture or incomplete checkpoints and does not fetch config.json to approve a direct reference. The selected runtime performs loading and reports its own errors; inspect Logs, not an old compatibility allowlist. A valid direct reference remains usable when Hub discovery is unavailable.

See Realtime for the new Models create/edit form, immutable Omni image, GPU allocation requirements and LiteLLM Playground test. Back up LiteLLM's PostgreSQL database before the bundled 1.84.0 → 1.101.0 upgrade; verify database migration, existing routes/keys, OIDC and Private Mesh on the test appliance. Realtime Ready checks deployment generation and /health, not actual audio quality. Live acceptance must include input/output audio, a second turn, interruptions, authenticated Gateway WebSockets and lifecycle recovery. Stop/Remove preserve the shared downloaded cache and other models.

The experimental AMD profile is a separate ROCm build of the same Omni revision. It is no longer tied to Strix Halo, a specific GPU generation, driver inventory or confirmed memory telemetry. CPU and Intel are also selectable experimental targets, with a matching image supplied under Advanced Settings; an empty catalog image is not evidence of hardware incompatibility. Default images are digest-pinned, while per-model runtime overrides may use tags for experiments. Actual runtime support and upstream parameter validation still apply. One time-slicing/DRA slot is not two physical GPUs. Coordinate GPU budgets with other models; RAM/headroom estimates are warnings rather than start blockers. Do not stop other models or change sharing automatically. See experimental boundaries.

If a Realtime activation exists but no Pod appears, inspect its status and the magicstick-operator logs. RealtimePermissionDenied (HTTP 401/403) means the runtime resources cannot be managed by the operator. In particular, verify the ai/magicstick-realtime-runtime Role/RoleBinding grants ConfigMap get/create/patch/delete to ai-system/magicstick-operator; apply the matching operator RBAC and controller together. The generated stage ConfigMap requires server-side-apply patch permission even on first creation. API failures are retried automatically; recreating the activation is not necessary. A custom target namespace needs an equivalent namespaced binding, not cluster-wide ConfigMap write access.

no structured config owner: cpu_offload_gb is a pinned Omni configuration projection error, not an out-of-memory error. The operator's generated bootstrap now omits zero offload and applies a source-hash-guarded compatibility repair for positive offload. See the repair and verification contract. Roll out the matching controller; its bootstrap hash recreates the model Pod. Do not disable upstream argument validation or increase RAM merely to hide this parser error. Unknown source hashes require reviewing the pinned runtime.