@qvac/tts-ggml 0.9.2 → 0.10.1

package/test/integration/cosyvoice3-clone.test.js

@@ -18,9 +18,12 @@ const {
   cosyvoiceBaseFileNames
 } = require('../utils/downloadModel')
 const { resolveRefWavPath } = require('../utils/runChatterboxTTS')
+const { TTS_TEST_THREADS } = require('../utils/testThreads')
+const { NO_GPU, GPU_ONLY_USE_GPU } = require('../utils/gpuOnly')
 
 const platform = os.platform()
 const isMobile = platform === 'ios' || platform === 'android'
+const useGPU = GPU_ONLY_USE_GPU
 
 function getBaseDir() {
   return isMobile && global.testDir ? global.testDir : '.'
@@ -66,7 +69,12 @@ function samplesDiffer(a, b) {
 }
 
 async function synthOnce(t, loadParams, label) {
-  const model = await loadCosyvoiceTTS({ ...loadParams, seed: CLONE_SEED })
+  const model = await loadCosyvoiceTTS({
+    threads: TTS_TEST_THREADS,
+    ...loadParams,
+    seed: CLONE_SEED,
+    useGPU
+  })
   try {
     const result = await runCosyvoiceTTS(
       model,
@@ -76,6 +84,11 @@ async function synthOnce(t, loadParams, label) {
     console.log(result.output)
     t.ok(result.passed, `${label} synth passes expectations`)
     t.ok(result.data.sampleCount > 0, `${label} produced audio`)
+    t.is(
+      result.data.stats?.backendDevice,
+      useGPU ? 1 : 0,
+      `${label} runs on the ${useGPU ? 'GPU' : 'CPU'} selected by the CI lane`
+    )
     return result.data
   } finally {
     try {
@@ -86,7 +99,7 @@ async function synthOnce(t, loadParams, label) {
 
 test(
   'CosyVoice3 cloning (ggml): baked, zero-shot and cross-lingual all differ at one seed',
-  { timeout: 1800000 },
+  { timeout: 1800000, skip: NO_GPU },
   async (t) => {
     const modelDir = await ensureCloneReadyModelDir(t)
     if (!modelDir) return
@@ -127,6 +140,7 @@ test(
     t.exception(
       () =>
         new TTSGgml({
+          threads: TTS_TEST_THREADS,
           engine: TTSGgml.ENGINE_COSYVOICE3,
           referenceAudio: resolveRefWavPath({}),
           files: {
@@ -167,6 +181,7 @@ test(
     }
 
     const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_COSYVOICE3,
       files: { cosyvoiceModelDir: baseOnlyDir },
       referenceAudio: resolveRefWavPath({})

package/test/integration/cosyvoice3-lavasr.test.js

@@ -7,7 +7,9 @@
 // CosyVoice3's ~2.4 GB model dir, while this one is listed on the dedicated
 // 'cosyvoice' row that pre-stages it (see test/mobile/test-groups.json).
 //
-// Text is kept short to bound CPU LM-decode time in CI, as in cosyvoice3.test.js.
+// As in cosyvoice3.test.js, the model-backed tests run on the GPU only and are
+// skipped on CPU-only runners (NO_GPU=true): even one short CosyVoice3 CPU
+// synthesis costs minutes there.
 //
 // Stage the enhancer GGUF via scripts/convert-lavasr-enhancer-to-gguf.py (from
 // the public LavaSRcpp ONNX release) into models/lavasr/lavasr-enhancer.gguf,
@@ -24,6 +26,8 @@ const {
   ensureLavaSRDenoiserGguf,
   ensureCosyvoiceModel
 } = require('../utils/downloadModel')
+const { TTS_TEST_THREADS } = require('../utils/testThreads')
+const { NO_GPU, GPU_ONLY_USE_GPU } = require('../utils/gpuOnly')
 
 const platform = os.platform()
 const isMobile = platform === 'ios' || platform === 'android'
@@ -117,7 +121,7 @@ async function stageCosyvoice(t, baseDir) {
 
 test(
   'CosyVoice3 + LavaSR enhancer (batch) reports 48 kHz enhanced output',
-  { timeout: 900000, skip: SKIP_COSYVOICE },
+  { timeout: 900000, skip: SKIP_COSYVOICE || NO_GPU },
   async (t) => {
     const baseDir = getBaseDir()
     const enh = await stageEnhancer(t, baseDir)
@@ -126,9 +130,10 @@ test(
     if (!dl) return
 
     const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_COSYVOICE3,
       files: { cosyvoiceModelDir: dl.modelDir, lavasrEnhancer: enh.path },
-      config: { language: 'en', useGPU: false },
+      config: { language: 'en', useGPU: GPU_ONLY_USE_GPU },
       opts: { stats: true }
     })
     await model.load()
@@ -139,8 +144,8 @@ test(
       t.ok(r.stats, 'runtimeStats returned (constructed with stats:true)')
       t.is(
         r.stats.enhancerBackendDevice,
-        0,
-        'enhancer loaded and ran on CPU (enhancerBackendDevice=0, not -1 = never loaded)'
+        r.stats.backendDevice,
+        'enhancer loaded and ran on the engine device (not -1 = never loaded)'
       )
     } finally {
       try {
@@ -152,7 +157,7 @@ test(
 
 test(
   'CosyVoice3 + LavaSR enhancer + native chunk streaming emits 48 kHz chunks',
-  { timeout: 900000, skip: SKIP_COSYVOICE },
+  { timeout: 900000, skip: SKIP_COSYVOICE || NO_GPU },
   async (t) => {
     const baseDir = getBaseDir()
     const enh = await stageEnhancer(t, baseDir)
@@ -161,10 +166,11 @@ test(
     if (!dl) return
 
     const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_COSYVOICE3,
       files: { cosyvoiceModelDir: dl.modelDir, lavasrEnhancer: enh.path },
       streamChunkTokens: 25,
-      config: { language: 'en', useGPU: false },
+      config: { language: 'en', useGPU: GPU_ONLY_USE_GPU },
       opts: { stats: true }
     })
     await model.load()
@@ -188,7 +194,7 @@ test(
 
 test(
   'CosyVoice3 + enhancer + streaming resamples every chunk to outputSampleRate',
-  { timeout: 900000, skip: SKIP_COSYVOICE },
+  { timeout: 900000, skip: SKIP_COSYVOICE || NO_GPU },
   async (t) => {
     // A non-native rate while streaming is only legal because the enhancer's
     // overlap-reprocess window folds the resample in seam-free. Without this the
@@ -201,10 +207,11 @@ test(
     if (!dl) return
 
     const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_COSYVOICE3,
       files: { cosyvoiceModelDir: dl.modelDir, lavasrEnhancer: enh.path },
       streamChunkTokens: 25,
-      config: { language: 'en', useGPU: false, outputSampleRate: 16000 },
+      config: { language: 'en', useGPU: GPU_ONLY_USE_GPU, outputSampleRate: 16000 },
       opts: { stats: true }
     })
     await model.load()
@@ -224,7 +231,7 @@ test(
 
 test(
   'CosyVoice3 + LavaSR denoiser changes the emitted PCM',
-  { timeout: 900000, skip: SKIP_COSYVOICE },
+  { timeout: 900000, skip: SKIP_COSYVOICE || NO_GPU },
   async (t) => {
     const baseDir = getBaseDir()
     const dl = await stageCosyvoice(t, baseDir)
@@ -240,9 +247,10 @@ test(
 
     const synth = async (files) => {
       const model = new TTSGgml({
+        threads: TTS_TEST_THREADS,
         engine: TTSGgml.ENGINE_COSYVOICE3,
         files: Object.assign({ cosyvoiceModelDir: dl.modelDir }, files),
-        config: { language: 'en', useGPU: false, seed: SEED },
+        config: { language: 'en', useGPU: GPU_ONLY_USE_GPU, seed: SEED },
         opts: { stats: true }
       })
       await model.load()
@@ -289,9 +297,11 @@ test(
     if (!dl) return
 
     const badEnhancer = path.join(baseDir, 'models', 'lavasr', 'invalid-enhancer.gguf')
+    fs.mkdirSync(path.dirname(badEnhancer), { recursive: true })
     fs.writeFileSync(badEnhancer, 'not a gguf')
 
     const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_COSYVOICE3,
       files: { cosyvoiceModelDir: dl.modelDir, lavasrEnhancer: badEnhancer },
       config: { language: 'en', useGPU: false },

package/test/integration/cosyvoice3.test.js

@@ -3,8 +3,10 @@
 // CosyVoice3 engine integration smoke: real synthesis in the desktop
 // `run-integration-tests` lane.  Mirrors parler.test.js (download model → build
 // TTSGgml → run → assert) but for the directory-consuming CosyVoice3 engine
-// (Qwen2 speech LM + DiT flow + CausalHiFT vocoder; native 24 kHz, CPU-only).
-// Text is kept SHORT to bound CPU LM-decode time in CI.
+// (Qwen2 speech LM + DiT flow + CausalHiFT vocoder; native 24 kHz).
+// Even one short CPU synthesis costs minutes on the CPU-only runners, so these
+// tests run on the GPU only and are skipped there (NO_GPU=true); the CPU crash
+// check is the CosyVoice3 CPU smoke in gpu-smoke.test.js.
 
 const os = require('bare-os')
 const path = require('bare-path')
@@ -12,6 +14,8 @@ const test = require('brittle')
 
 const { loadCosyvoiceTTS, runCosyvoiceTTS } = require('../utils/runCosyvoiceTTS')
 const { ensureCosyvoiceModel } = require('../utils/downloadModel')
+const { TTS_TEST_THREADS } = require('../utils/testThreads')
+const { NO_GPU, GPU_ONLY_USE_GPU } = require('../utils/gpuOnly')
 
 const platform = os.platform()
 const isMobile = platform === 'ios' || platform === 'android'
@@ -26,7 +30,7 @@ const MODEL_MISSING =
 
 test(
   'CosyVoice3 TTS (ggml): outputSampleRate=16000 resamples and reports 16 kHz',
-  { timeout: 600000 },
+  { timeout: 600000, skip: NO_GPU },
   async (t) => {
     const baseDir = getBaseDir()
     const download = await ensureCosyvoiceModel({
@@ -40,8 +44,10 @@ test(
     // Second model with outputSampleRate wired through the config surface, so the
     // whole resample path (engine → addon → reported chunk) is exercised end-to-end.
     const model = await loadCosyvoiceTTS({
+      threads: TTS_TEST_THREADS,
       cosyvoiceModelDir: download.modelDir,
-      outputSampleRate: 16000
+      outputSampleRate: 16000,
+      useGPU: GPU_ONLY_USE_GPU
     })
     try {
       const text = 'Hello from CosyVoice.'
@@ -68,7 +74,7 @@ function samplesEqual(a, b) {
 
 test(
   'CosyVoice3 TTS (ggml): emotion changes the audio; per-call switch needs no reload',
-  { timeout: 900000 },
+  { timeout: 900000, skip: NO_GPU },
   async (t) => {
     const baseDir = getBaseDir()
     const download = await ensureCosyvoiceModel({
@@ -83,8 +89,10 @@ test(
     // logits, and cosyvoice reseeds per synthesis, so pinning it is the
     // difference between measuring the conditioning and measuring the sampler.
     const model = await loadCosyvoiceTTS({
+      threads: TTS_TEST_THREADS,
       cosyvoiceModelDir: download.modelDir,
-      seed: 42
+      seed: 42,
+      useGPU: GPU_ONLY_USE_GPU
     })
     try {
       const text = 'Hello from CosyVoice.'
@@ -119,6 +127,16 @@ test(
         'plain cosyvoice reports 24 kHz native sample rate'
       )
       t.ok(plain.data.durationMs > 0, 'plain cosyvoice audio duration is > 0 ms')
+      // Engine StageTimings: each stage did work, the zero-shot baked voice
+      // prompts the LM with its speech tokens, and no denoiser is loaded.
+      const stats = plain.data.stats
+      t.ok(stats.speechTokens > 0, 'stats report the speech tokens generated')
+      t.ok(stats.decodeSteps > 0, 'stats report the LM decode steps')
+      t.ok(stats.promptSpeechTokens > 0, 'zero-shot prompts the LM with speech tokens')
+      for (const key of ['lmDecodeMs', 'ditEulerMs', 'hiftDecodeMs', 'stageTotalMs']) {
+        t.ok(stats[key] > 0, `stats report ${key}`)
+      }
+      t.is(stats.denoiserBackendDevice, -1, 'no denoiser loaded -> denoiserBackendDevice=-1')
       const neutral = await runCosyvoiceTTS(
         model,
         { text, perCallEmotion: 'neutral' },
@@ -159,7 +177,7 @@ test(
 
 test(
   'CosyVoice3 TTS (ggml): instruct conditioning produces audio',
-  { timeout: 600000 },
+  { timeout: 600000, skip: NO_GPU },
   async (t) => {
     const baseDir = getBaseDir()
     const download = await ensureCosyvoiceModel({
@@ -171,8 +189,10 @@ test(
     }
 
     const model = await loadCosyvoiceTTS({
+      threads: TTS_TEST_THREADS,
       cosyvoiceModelDir: download.modelDir,
-      emotion: 'happy'
+      emotion: 'happy',
+      useGPU: GPU_ONLY_USE_GPU
     })
     try {
       const text = 'Hello from CosyVoice.'
@@ -182,6 +202,11 @@ test(
       t.ok(result.passed, 'cosyvoice instruct synth passes expectations')
       t.ok(result.data.sampleCount > 0, 'cosyvoice instruct produced audio')
       t.is(result.data.reportedSampleRate, 24000, 'cosyvoice instruct reports 24 kHz')
+      t.is(
+        result.data.stats.promptSpeechTokens,
+        0,
+        'instruct mode drops the prompt speech tokens (StageTimings)'
+      )
     } finally {
       try {
         await model.unload()

package/test/integration/fit.test.js

@@ -0,0 +1,174 @@
+'use strict'
+
+const fs = require('bare-fs')
+const os = require('bare-os')
+const path = require('bare-path')
+const test = require('brittle')
+
+const TTSGgml = require('@qvac/tts-ggml')
+const { ensureSupertonicModel } = require('../utils/downloadModel')
+
+const TEST_TIMEOUT_MS = 900000
+
+let modelPath = null
+
+async function supertonicPath(t) {
+  if (modelPath === null) {
+    const resolved = await ensureSupertonicModel()
+    modelPath = resolved.success ? resolved.path : ''
+  }
+  if (modelPath === '') {
+    t.pass('Supertonic GGUF unavailable on this runner')
+    return null
+  }
+  return modelPath
+}
+
+/**
+ * The supertonic fitter covers the fused graph path: a validated GPU, or a CPU
+ * without the Accelerate pointwise kernels. A runner outside that answers
+ * `compute-path-not-supported`. That is a refusal, and it carries no figures.
+ */
+function projected(t, fit) {
+  if (fit.reason === 'compute-path-not-supported') {
+    t.pass('this runner’s compute path is not modelled')
+    return false
+  }
+  return true
+}
+
+function tempDir(t) {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'tts-fit-'))
+  t.teardown(() => fs.rmSync(dir, { recursive: true, force: true }))
+  return dir
+}
+
+/**
+ * The registry's weightless copy of a GGUF: the header and tensor table as
+ * written, and a hole where the tensor data was. The file keeps its length so
+ * every offset in the header still resolves.
+ */
+function weightlessCopy(dir, sourcePath, headerBytes) {
+  const target = path.join(dir, 'weightless.gguf')
+  const source = fs.openSync(sourcePath, 'r')
+  const header = Buffer.alloc(headerBytes)
+  fs.readSync(source, header, 0, headerBytes, 0)
+  fs.closeSync(source)
+
+  fs.writeFileSync(target, header)
+  fs.truncateSync(target, fs.statSync(sourcePath).size)
+  return target
+}
+
+test('a projection is internally consistent', { timeout: TEST_TIMEOUT_MS }, async (t) => {
+  const model = await supertonicPath(t)
+  if (model === null) return
+
+  const fit = TTSGgml.assessFit({
+    engineType: 'supertonic',
+    supertonicModelPath: model,
+    useGPU: true
+  })
+
+  t.comment(`status=${fit.status} reason=${fit.reason} device=${fit.deviceName}`)
+  if (!projected(t, fit)) return
+  t.ok(fit.status === 'fits' || fit.status === 'does-not-fit', 'a verdict, not an error')
+  t.ok(fit.deviceTotalBytes > 0, 'the device reported its capacity')
+  t.ok(fit.deviceFreeBytes <= fit.deviceTotalBytes, 'free never exceeds total')
+  t.ok(fit.weightsBytes > 0, 'the weights were measured')
+  t.ok(fit.deviceBytes >= fit.weightsBytes, 'weights are part of the device demand')
+  t.is(fit.lavasrFileBytes, 0, 'no LavaSR stage was named')
+})
+
+test('a weightless copy projects the same', { timeout: TEST_TIMEOUT_MS }, async (t) => {
+  const model = await supertonicPath(t)
+  if (model === null) return
+
+  const full = TTSGgml.assessFit({
+    engineType: 'supertonic',
+    supertonicModelPath: model,
+    useGPU: true
+  })
+  if (!projected(t, full)) return
+  const stub = TTSGgml.assessFit({
+    engineType: 'supertonic',
+    supertonicModelPath: weightlessCopy(tempDir(t), model, 1024 * 1024),
+    useGPU: true
+  })
+
+  t.is(stub.status, full.status, 'the same verdict')
+  t.is(stub.weightsBytes, full.weightsBytes, 'the same weight footprint')
+  t.is(stub.deviceBytes, full.deviceBytes, 'the same device demand')
+  t.is(stub.hostBytes, full.hostBytes, 'the same host demand')
+})
+
+test('the load’s offload keys reach the projection', { timeout: TEST_TIMEOUT_MS }, async (t) => {
+  const model = await supertonicPath(t)
+  if (model === null) return
+
+  const cpu = TTSGgml.assessFit({
+    engineType: 'supertonic',
+    supertonicModelPath: model,
+    useGPU: false
+  })
+  const pinned = TTSGgml.assessFit({
+    engineType: 'supertonic',
+    supertonicModelPath: model,
+    nGpuLayers: 0
+  })
+
+  t.ok(cpu.deviceIsCpu, 'useGPU false projects on the CPU device')
+  t.is(pinned.deviceIsCpu, cpu.deviceIsCpu, 'nGpuLayers 0 agrees with useGPU false')
+})
+
+test('a LavaSR stage is sized beside the projection', { timeout: TEST_TIMEOUT_MS }, async (t) => {
+  const model = await supertonicPath(t)
+  if (model === null) return
+
+  const dir = tempDir(t)
+  const enhancer = path.join(dir, 'enhancer.gguf')
+  fs.writeFileSync(enhancer, Buffer.alloc(4096, 1))
+
+  const fit = TTSGgml.assessFit({
+    engineType: 'supertonic',
+    supertonicModelPath: model,
+    useGPU: true,
+    lavasrEnhancerPath: enhancer
+  })
+
+  t.is(fit.lavasrFileBytes, 4096, 'the enhancer is reported at its size on disk')
+  if (!projected(t, fit)) return
+  t.ok(fit.lavasrFileBytes < fit.deviceBytes, 'reported beside the projection')
+})
+
+test('a voice with no fitter is an outcome, not a throw', (t) => {
+  const fit = TTSGgml.assessFit({
+    engineType: 'moss',
+    mossBackbonePath: '/models/moss-tts-delay.gguf'
+  })
+
+  t.is(fit.status, 'error')
+  t.is(fit.reason, 'unsupported-engine')
+  t.is(fit.modelVariant, 'moss')
+})
+
+test('a model that cannot be read is an outcome, not a throw', (t) => {
+  const fit = TTSGgml.assessFit({
+    engineType: 'supertonic',
+    supertonicModelPath: '/nonexistent/model.gguf'
+  })
+
+  t.is(fit.status, 'error')
+})
+
+test('a count that is not a count is refused', (t) => {
+  t.exception(
+    () =>
+      TTSGgml.assessFit({
+        engineType: 'supertonic',
+        supertonicModelPath: '/models/supertonic.gguf',
+        textTokens: -1
+      }),
+    /non-negative count/
+  )
+})

package/test/integration/gpu-smoke.test.js

@@ -45,6 +45,7 @@ const {
   ensureCosyvoiceModel
 } = require('../utils/downloadModel')
 const { recordTtsStats } = require('../utils/perf-helper')
+const { TTS_TEST_THREADS } = require('../utils/testThreads')
 
 const platform = os.platform()
 const isMobile = platform === 'ios' || platform === 'android'
@@ -228,6 +229,7 @@ test(
     }
 
     const model = await loadChatterboxTTS({
+      threads: TTS_TEST_THREADS,
       modelDir: download.targetDir,
       refWavPath,
       language: 'en',
@@ -287,6 +289,7 @@ test(
     }
 
     const model = await loadChatterboxTTS({
+      threads: TTS_TEST_THREADS,
       modelDir: download.targetDir,
       t3ModelPath: path.join(download.targetDir, 'chatterbox-t3-mtl.gguf'),
       s3genModelPath: path.join(download.targetDir, 'chatterbox-s3gen-mtl.gguf'),
@@ -338,6 +341,7 @@ test(
     const supertonicPath = download.path || path.join(modelsDir, 'supertonic.gguf')
 
     const model = await loadSupertonicTTS({
+      threads: TTS_TEST_THREADS,
       supertonicModelPath: supertonicPath,
       language: 'en',
       voice: 'F1',
@@ -385,6 +389,7 @@ test(
     }
 
     const model = await loadSupertonicTTS({
+      threads: TTS_TEST_THREADS,
       supertonicModelPath: download.path,
       language: 'en',
       voice: 'F1',
@@ -431,6 +436,7 @@ test(
     }
 
     const model = await loadSupertonicTTS({
+      threads: TTS_TEST_THREADS,
       supertonicModelPath: download.path,
       language: 'en',
       voice: 'F1',
@@ -488,6 +494,7 @@ test(
     }
 
     const model = await loadChatterboxTTS({
+      threads: TTS_TEST_THREADS,
       modelDir: download.targetDir,
       refWavPath,
       language: 'en',
@@ -532,6 +539,7 @@ test(
     const supertonicPath = download.path || path.join(modelsDir, 'supertonic.gguf')
 
     const model = await loadSupertonicTTS({
+      threads: TTS_TEST_THREADS,
       supertonicModelPath: supertonicPath,
       language: 'en',
       voice: 'F1',
@@ -583,6 +591,7 @@ for (const v of [
         return
       }
       const model = await loadParlerTTS({
+        threads: TTS_TEST_THREADS,
         parlerModelPath: download.path,
         seed: 42,
         useGPU: true
@@ -625,6 +634,7 @@ for (const v of [
         return
       }
       const model = await loadParlerTTS({
+        threads: TTS_TEST_THREADS,
         parlerModelPath: download.path,
         seed: 42,
         useGPU: false
@@ -678,6 +688,7 @@ async function probeAndroidGpuVendor(t) {
   })
   if (!download || !download.success) return null
   const model = await loadSupertonicTTS({
+    threads: TTS_TEST_THREADS,
     supertonicModelPath: download.path,
     language: 'en',
     voice: 'F1',
@@ -730,6 +741,7 @@ test(
       }
     }
     const model = await loadCosyvoiceTTS({
+      threads: TTS_TEST_THREADS,
       cosyvoiceModelDir: download.modelDir,
       useGPU: true
     })
@@ -788,6 +800,7 @@ test(
       return
     }
     const model = await loadCosyvoiceTTS({
+      threads: TTS_TEST_THREADS,
       cosyvoiceModelDir: download.modelDir,
       useGPU: false
     })