QVAC-18995 feat[api]: bump transcription-whispercpp to 0.9.0 and surface backend/GPU stats

packages/sdk/package.json

@@ -214,7 +214,7 @@
     "@qvac/rag": "^0.6.2",
     "@qvac/registry-client": "^0.6.0",
     "@qvac/transcription-parakeet": "^0.7.1",
-    "@qvac/transcription-whispercpp": "^0.8.0",
+    "@qvac/transcription-whispercpp": "^0.9.0",
     "@qvac/translation-nmtcpp": "^5.0.1",
     "@qvac/tts-ggml": "^0.2.0",
     "@qvac/vla-ggml": "^0.3.2",

packages/sdk/schemas/transcription.ts

@@ -37,6 +37,14 @@ export const transcribeStatsSchema = z.object({
   encoderTime: z.number().optional(),
   decoderTime: z.number().optional(),
   melSpecTime: z.number().optional(),
+  // Backend selection captured once at model load (whisper engine).
+  // `0` CPU / `1` GPU; `backendId` codes the GPU family (see the addon's
+  // BackendId enum). gpu memory fields are `-1` when the backend exposes
+  // no accounting, and absent entirely for the parakeet engine.
+  backendDevice: z.number().optional(),
+  backendId: z.number().optional(),
+  gpuMemTotalMb: z.number().optional(),
+  gpuMemFreeMb: z.number().optional(),
 });
 
 export const transcribeSegmentSchema = z.object({

packages/sdk/server/bare/ops/transcribe.ts

@@ -242,6 +242,10 @@ export async function* transcribe(
     ...(response.stats?.encoderMs !== undefined && { encoderTime: response.stats.encoderMs }),
     ...(response.stats?.decoderMs !== undefined && { decoderTime: response.stats.decoderMs }),
     ...(response.stats?.melSpecMs !== undefined && { melSpecTime: response.stats.melSpecMs }),
+    ...(response.stats?.backendDevice !== undefined && { backendDevice: response.stats.backendDevice }),
+    ...(response.stats?.backendId !== undefined && { backendId: response.stats.backendId }),
+    ...(response.stats?.gpuMemTotalMb !== undefined && { gpuMemTotalMb: response.stats.gpuMemTotalMb }),
+    ...(response.stats?.gpuMemFreeMb !== undefined && { gpuMemFreeMb: response.stats.gpuMemFreeMb }),
   };
 
   return buildStreamResult(modelExecutionMs, stats);

packages/sdk/server/bare/types/addon-responses.ts

@@ -60,6 +60,10 @@ export interface TranscribeStats {
   encoderMs?: number;
   decoderMs?: number;
   melSpecMs?: number;
+  backendDevice?: number;
+  backendId?: number;
+  gpuMemTotalMb?: number;
+  gpuMemFreeMb?: number;
 }
 
 export interface TranscribeAddonSegment {

packages/sdk/test/unit/transcription-stats-schema.test.ts

@@ -0,0 +1,42 @@
+import test from "brittle";
+import { transcribeStatsSchema } from "@/schemas/transcription";
+
+test("transcribeStatsSchema: round-trips whisper backend/GPU stats", (t) => {
+  const result = transcribeStatsSchema.parse({
+    realTimeFactor: 0.25,
+    tokensPerSecond: 120,
+    backendDevice: 1,
+    backendId: 3,
+    gpuMemTotalMb: 8192,
+    gpuMemFreeMb: 4096,
+  });
+  t.is(result.backendDevice, 1);
+  t.is(result.backendId, 3);
+  t.is(result.gpuMemTotalMb, 8192);
+  t.is(result.gpuMemFreeMb, 4096);
+});
+
+test("transcribeStatsSchema: backend/GPU fields are optional (CPU / parakeet path)", (t) => {
+  const result = transcribeStatsSchema.parse({ realTimeFactor: 1.5 });
+  t.absent(result.backendDevice);
+  t.absent(result.backendId);
+  t.absent(result.gpuMemTotalMb);
+  t.absent(result.gpuMemFreeMb);
+});
+
+test("transcribeStatsSchema: accepts the -1 no-accounting sentinel for gpu memory", (t) => {
+  const result = transcribeStatsSchema.parse({
+    backendDevice: 1,
+    backendId: 2,
+    gpuMemTotalMb: -1,
+    gpuMemFreeMb: -1,
+  });
+  t.is(result.gpuMemTotalMb, -1);
+  t.is(result.gpuMemFreeMb, -1);
+});
+
+test("transcribeStatsSchema: rejects non-numeric backend fields", (t) => {
+  t.exception(() =>
+    transcribeStatsSchema.parse({ backendId: "vulkan" }),
+  );
+});