Complete local Bare smoke path

daemon/dbus-service.js

@@ -22,8 +22,8 @@ export async function serveOnSessionBus(daemon) {
     Sleep() { daemon.sleep(); }
     Shutdown() { daemon.close(); }
     PushToTalk(pressed) { daemon.setPushToTalk?.(Boolean(pressed)); }
-    Say(text) { return daemon.say?.(text); }
-    Ask(text) { return daemon.ask(text); }
+    async Say(text) { await daemon.say?.(text); }
+    async Ask(text) { await daemon.ask(text); }
     Cancel() { daemon.cancel(); }
     SetMode(mode) { daemon.mode = mode; daemon.emit('ModeChanged', mode); }
     GetState() { return daemon.state; }

daemon/qvac-master.js

@@ -105,9 +105,17 @@ export async function loadAuxiliaryModel(name, modelConfig = {}, modelType = und
   if (existing) return existing;
   const sdk = await qvacSdk();
   if (typeof sdk.loadModel !== 'function') throw new Error('QVAC SDK does not expose loadModel()');
+  const auxiliaryConfig = resolveModelConfigAssets(sdk, modelConfig);
+  // GPU placement belongs to the LLM model configuration. ASR and TTS have
+  // their own validated schemas and reject llama.cpp-only keys.
+  if (!modelType || modelType === 'llm') {
+    auxiliaryConfig.device = 'gpu';
+    auxiliaryConfig.gpu_layers = QVAC_MASTER.gpuLayers;
+    auxiliaryConfig['mmproj-use-gpu'] = true;
+  }
   const loadOptions = {
     modelSrc: resolveSdkAsset(sdk, name),
-    modelConfig: { ...resolveModelConfigAssets(sdk, modelConfig), device: 'gpu', gpu_layers: QVAC_MASTER.gpuLayers, 'mmproj-use-gpu': true },
+    modelConfig: auxiliaryConfig,
   };
   if (modelType) loadOptions.modelType = modelType;
   const modelId = await withQvacMaster(() => sdk.loadModel(loadOptions));