@qvac/tts-ggml 0.10.0 → 0.11.0

package/test/integration/parler.test.js

@@ -7,18 +7,17 @@
 
 const os = require('bare-os')
 const path = require('bare-path')
-const proc = require('bare-process')
 const test = require('brittle')
 
 const { loadParlerTTS, runParlerTTS } = require('../utils/runParlerTTS')
 const { ensureParlerModel } = require('../utils/downloadModel')
 const { recordTtsStats } = require('../utils/perf-helper')
 const { TTS_TEST_THREADS } = require('../utils/testThreads')
+const { NO_GPU, SKIP_PARLER_GPU_ONLY, PARLER_GPU_ONLY_USE_GPU } = require('../utils/gpuOnly')
 
 const platform = os.platform()
 const isMobile = platform === 'ios' || platform === 'android'
 const isApple = platform === 'darwin' || platform === 'ios'
-const NO_GPU = proc.env && proc.env.NO_GPU === 'true'
 
 function getBaseDir() {
   return isMobile && global.testDir ? global.testDir : '.'
@@ -137,7 +136,7 @@ test(
 
 test(
   'Parler TTS (ggml): description matrix — conflicts throw, per-call wins, defaults work',
-  { timeout: 600000 },
+  { timeout: 600000, skip: SKIP_PARLER_GPU_ONLY },
   async (t) => {
     const baseDir = getBaseDir()
     const download = await ensureParlerModel({ targetDir: path.join(baseDir, 'models') })
@@ -191,7 +190,8 @@ test(
       threads: TTS_TEST_THREADS,
       parlerModelPath: download.path,
       voice: 'Laura',
-      seed: 42
+      seed: 42,
+      useGPU: PARLER_GPU_ONLY_USE_GPU
     })
     try {
       // Per-call description conflict (same level) rejects.
@@ -234,7 +234,8 @@ test(
       threads: TTS_TEST_THREADS,
       parlerModelPath: download.path,
       description: 'A calm female voice, very clear audio.',
-      seed: 42
+      seed: 42,
+      useGPU: PARLER_GPU_ONLY_USE_GPU
     })
     try {
       await t.exception(
@@ -252,7 +253,7 @@ test(
 
 test(
   'Parler TTS (ggml): emotion flag changes the audio; per-call switch needs no reload',
-  { timeout: 600000 },
+  { timeout: 600000, skip: SKIP_PARLER_GPU_ONLY },
   async (t) => {
     const baseDir = getBaseDir()
     const download = await ensureParlerModel({ targetDir: path.join(baseDir, 'models') })
@@ -265,7 +266,8 @@ test(
       threads: TTS_TEST_THREADS,
       parlerModelPath: download.path,
       voice: 'Laura',
-      seed: 42
+      seed: 42,
+      useGPU: PARLER_GPU_ONLY_USE_GPU
     })
     try {
       const text = 'The weather is wonderful today.'
@@ -300,7 +302,7 @@ test(
 
 test(
   'Parler TTS (ggml): fixed seed is deterministic across runs and reload',
-  { timeout: 600000 },
+  { timeout: 600000, skip: SKIP_PARLER_GPU_ONLY },
   async (t) => {
     const baseDir = getBaseDir()
     const download = await ensureParlerModel({ targetDir: path.join(baseDir, 'models') })
@@ -314,7 +316,8 @@ test(
       parlerModelPath: download.path,
       voice: 'Laura',
       emotion: 'neutral',
-      seed: 42
+      seed: 42,
+      useGPU: PARLER_GPU_ONLY_USE_GPU
     })
     try {
       const text = 'Deterministic synthesis check.'
@@ -345,7 +348,7 @@ test(
 
 test(
   'Parler TTS (ggml): runStream emits per-sentence chunks with one pinned description',
-  { timeout: 600000 },
+  { timeout: 600000, skip: SKIP_PARLER_GPU_ONLY },
   async (t) => {
     const baseDir = getBaseDir()
     const download = await ensureParlerModel({ targetDir: path.join(baseDir, 'models') })
@@ -360,6 +363,7 @@ test(
       engine: TTSGgml.ENGINE_PARLER,
       files: { parlerModel: download.path },
       seed: 42,
+      config: { useGPU: PARLER_GPU_ONLY_USE_GPU },
       opts: { stats: true }
     })
     await model.load()
@@ -393,7 +397,7 @@ test(
 
 test(
   'Parler TTS (ggml): native streaming emits monotonic chunks equal to batch',
-  { timeout: 900000 },
+  { timeout: 900000, skip: SKIP_PARLER_GPU_ONLY },
   async (t) => {
     const baseDir = getBaseDir()
     const download = await ensureParlerModel({ targetDir: path.join(baseDir, 'models') })
@@ -526,22 +530,24 @@ test(
 
 // Desktop quant + backend coverage: q8_0 + f16 for indic/mini, q8_0 for large.
 // Runs on the platform's GPU-if-available backend (Metal on Apple GPU runners,
-// CPU otherwise). large self-gates on RAM; any tier that isn't staged skips
-// (t.pass) so a not-yet-registered tier never fails CI. Registered per PR #3372.
-// Desktop-only: the mobile Parler surface is covered by the gpu-smoke legs.
+// CPU otherwise). mini f16 and large are the slowest tiers on CPU (roughly
+// 45-80 s per synthesis), so they run on Metal only. large self-gates on RAM; any tier that isn't
+// staged skips (t.pass) so a not-yet-registered tier never fails CI. Registered
+// per PR #3372. Desktop-only: the mobile Parler surface is covered by the
+// gpu-smoke legs.
 const PARLER_QUANT_MATRIX = [
-  { variant: 'indic', quant: 'q8_0' },
-  { variant: 'indic', quant: 'f16' },
-  { variant: 'mini', quant: 'q8_0' },
-  { variant: 'mini', quant: 'f16' },
-  { variant: 'large', quant: 'q8_0' }
+  { variant: 'indic', quant: 'q8_0', gpuOnly: false },
+  { variant: 'indic', quant: 'f16', gpuOnly: false },
+  { variant: 'mini', quant: 'q8_0', gpuOnly: false },
+  { variant: 'mini', quant: 'f16', gpuOnly: true },
+  { variant: 'large', quant: 'q8_0', gpuOnly: true }
 ]
 const LARGE_MIN_RAM_BYTES = 16 * 1024 ** 3
 
-for (const { variant, quant } of PARLER_QUANT_MATRIX) {
+for (const { variant, quant, gpuOnly } of PARLER_QUANT_MATRIX) {
   test(
     `Parler TTS (ggml): ${variant} ${quant} synthesizes 44.1 kHz audio on the target backend`,
-    { timeout: 900000, skip: isMobile },
+    { timeout: 900000, skip: isMobile || (gpuOnly && SKIP_PARLER_GPU_ONLY) },
     async (t) => {
       if (variant === 'large' && os.totalmem() < LARGE_MIN_RAM_BYTES) {
         t.pass(

package/test/mobile/integration.auto.cjs

@@ -50,6 +50,14 @@ async function runLavasrEnhancerTest (options = {}) { // eslint-disable-line no-
   return runIntegrationModule('../integration/lavasr-enhancer.test.js', options)
 }
 
+async function runMossSfxTest (options = {}) { // eslint-disable-line no-unused-vars
+  return runIntegrationModule('../integration/moss-sfx.test.js', options)
+}
+
+async function runMossSpeechTest (options = {}) { // eslint-disable-line no-unused-vars
+  return runIntegrationModule('../integration/moss-speech.test.js', options)
+}
+
 async function runMossTest (options = {}) { // eslint-disable-line no-unused-vars
   return runIntegrationModule('../integration/moss.test.js', options)
 }
@@ -70,6 +78,10 @@ async function runParlerTest (options = {}) { // eslint-disable-line no-unused-v
   return runIntegrationModule('../integration/parler.test.js', options)
 }
 
+async function runPocketIntegrationTest (options = {}) { // eslint-disable-line no-unused-vars
+  return runIntegrationModule('../integration/pocket.integration.test.js', options)
+}
+
 async function runRtfBenchmarkTest (options = {}) { // eslint-disable-line no-unused-vars
   return runIntegrationModule('../integration/rtf-benchmark.test.js', options)
 }
@@ -106,11 +118,14 @@ module.exports = {
   runFitTest,
   runGpuSmokeTest,
   runLavasrEnhancerTest,
+  runMossSfxTest,
+  runMossSpeechTest,
   runMossTest,
   runMultipleRunsTest,
   runOutputSampleRateTest,
   runParlerWerTest,
   runParlerTest,
+  runPocketIntegrationTest,
   runRtfBenchmarkTest,
   runStreamingBenchmarkTest,
   runSupertonicMtlTest,

package/tts.d.ts

@@ -17,6 +17,26 @@ export interface TTSJobData {
     instruct?: string;
     referenceAudio?: string;
     referenceText?: string;
+    seconds?: number;
+    negativePrompt?: string;
+    steps?: number;
+    guidance?: number;
+    shift?: number;
+    messages?: Array<{
+        role: string;
+        text?: string;
+        audio?: Int16Array | Float32Array;
+        sampleRate?: number;
+    }>;
+    replyVoice?: Int16Array | Float32Array;
+    replyVoiceSampleRate?: number;
+    textReply?: boolean;
+    maxReplySeconds?: number;
+    maxNewTokens?: number;
+    greedy?: boolean;
+    temperature?: number;
+    topP?: number;
+    topK?: number;
 }
 export interface TTSWeightData {
     filename: string;