Updates

daemon/qvac-master.js

@@ -16,6 +16,8 @@ let loadPromise = null;
 let ownerCount = 0;
 let operationTail = Promise.resolve();
 const auxiliaryModels = new Map();
+const auxiliaryLoads = new Map();
+const auxiliaryOwners = new Map();
 
 export const QVAC_MASTER = Object.freeze({
   configPath: process.env.QVAC_CONFIG_PATH,
@@ -92,36 +94,51 @@ function resolveModelConfigAssets(sdk, config) {
   return copy;
 }
 
+function retainAuxiliaryModel(modelId) {
+  auxiliaryOwners.set(modelId, (auxiliaryOwners.get(modelId) || 0) + 1);
+  return modelId;
+}
+
 /** 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 = {}, modelType = undefined) {
   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 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: auxiliaryConfig,
-  };
-  if (modelType) loadOptions.modelType = modelType;
-  const modelId = await withQvacMaster(() => sdk.loadModel(loadOptions));
-  auxiliaryModels.set(String(name), modelId);
-  return modelId;
+  if (existing) return retainAuxiliaryModel(existing);
+  const key = String(name);
+  if (auxiliaryLoads.has(key)) return retainAuxiliaryModel(await auxiliaryLoads.get(key));
+  const pending = Promise.resolve().then(async () => {
+    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: auxiliaryConfig,
+    };
+    if (modelType) loadOptions.modelType = modelType;
+    const modelId = await withQvacMaster(() => sdk.loadModel(loadOptions));
+    auxiliaryModels.set(String(name), modelId);
+    return modelId;
+  });
+  auxiliaryLoads.set(key, pending);
+  try { return retainAuxiliaryModel(await pending); }
+  finally { if (auxiliaryLoads.get(key) === pending) auxiliaryLoads.delete(key); }
 }
 
-export async function unloadAuxiliaryModel(modelId) {
+export async function unloadAuxiliaryModel(modelId, { force = false } = {}) {
   if (!modelId) return;
+  const owners = auxiliaryOwners.get(modelId) || 0;
+  if (!force && owners > 1) { auxiliaryOwners.set(modelId, owners - 1); return; }
   const sdk = await qvacSdk();
   if (typeof sdk.unloadModel === 'function') await withQvacMaster(() => sdk.unloadModel({ modelId }));
+  auxiliaryOwners.delete(modelId);
   for (const [name, id] of auxiliaryModels) if (id === modelId) auxiliaryModels.delete(name);
 }
 
@@ -138,8 +155,9 @@ export function qvacBusy() {
 
 export async function closeQvac() {
   if (ownerCount > 0) return;
-  for (const modelId of auxiliaryModels.values()) await unloadAuxiliaryModel(modelId).catch(() => {});
+  for (const modelId of auxiliaryModels.values()) await unloadAuxiliaryModel(modelId, { force: true }).catch(() => {});
   auxiliaryModels.clear();
+  auxiliaryOwners.clear();
   loadPromise = null;
   await Agent.engine.close();
 }