P4

daemon/qvac-master.js

@@ -8,6 +8,8 @@ const Agent = require(path.join(harnessPath, 'index.js'));
 
 let loadPromise = null;
 let ownerCount = 0;
+let operationTail = Promise.resolve();
+const auxiliaryModels = new Map();
 
 export const QVAC_MASTER = Object.freeze({
   configPath: process.env.QVAC_CONFIG_PATH,
@@ -55,10 +57,58 @@ export async function acquireQvac() {
   return loadPromise;
 }
 
+export async function qvacSdk() {
+  return Agent.engine.ensureInit();
+}
+
+export function withQvacMaster(task) {
+  const operation = operationTail.then(task, task);
+  operationTail = operation.catch(() => {});
+  return operation;
+}
+
+function resolveSdkAsset(sdk, name) {
+  if (name && typeof name !== 'string') return name;
+  return sdk[name] || sdk.models?.[name] || name;
+}
+
+function resolveModelConfigAssets(sdk, config) {
+  const copy = { ...config };
+  for (const key of ['vadModelSrc', 'projectionModelSrc', 'vocabModelSrc']) {
+    if (typeof copy[key] === 'string') copy[key] = resolveSdkAsset(sdk, copy[key]);
+  }
+  return copy;
+}
+
+/** Load ASR/TTS models in the same SDK worker and under the same master lock.
+ * These are auxiliary model IDs; they do not create another QVAC runtime. */
+export async function loadAuxiliaryModel(name, modelConfig = {}) {
+  if (!name) throw new Error('auxiliary QVAC model name is required');
+  const existing = auxiliaryModels.get(String(name));
+  if (existing) return existing;
+  const sdk = await qvacSdk();
+  if (typeof sdk.loadModel !== 'function') throw new Error('QVAC SDK does not expose loadModel()');
+  const modelId = await withQvacMaster(() => sdk.loadModel({
+    modelSrc: resolveSdkAsset(sdk, name),
+    modelConfig: { ...resolveModelConfigAssets(sdk, modelConfig), device: 'gpu', gpu_layers: QVAC_MASTER.gpuLayers, 'mmproj-use-gpu': true },
+  }));
+  auxiliaryModels.set(String(name), modelId);
+  return modelId;
+}
+
+export async function unloadAuxiliaryModel(modelId) {
+  if (!modelId) return;
+  const sdk = await qvacSdk();
+  if (typeof sdk.unloadModel === 'function') await withQvacMaster(() => sdk.unloadModel({ modelId }));
+  for (const [name, id] of auxiliaryModels) if (id === modelId) auxiliaryModels.delete(name);
+}
+
 export function releaseQvac() { ownerCount = Math.max(0, ownerCount - 1); }
 
 export async function closeQvac() {
   if (ownerCount > 0) return;
+  for (const modelId of auxiliaryModels.values()) await unloadAuxiliaryModel(modelId).catch(() => {});
+  auxiliaryModels.clear();
   loadPromise = null;
   await Agent.engine.close();
 }
@@ -101,7 +151,7 @@ export async function callQvac(method, input) {
 }
 
 export function qvacStatus() {
-  return { ...QVAC_MASTER, owners: ownerCount, loaded: Agent.engine.getLoaded() };
+  return { ...QVAC_MASTER, owners: ownerCount, loaded: Agent.engine.getLoaded(), auxiliaryModels: auxiliaryModels.size };
 }
 
 export { Agent };