@qvac/tts-ggml 0.9.1 → 0.10.0

package/test/integration/cosyvoice3.test.js

@@ -12,6 +12,7 @@ const test = require('brittle')
 
 const { loadCosyvoiceTTS, runCosyvoiceTTS } = require('../utils/runCosyvoiceTTS')
 const { ensureCosyvoiceModel } = require('../utils/downloadModel')
+const { TTS_TEST_THREADS } = require('../utils/testThreads')
 
 const platform = os.platform()
 const isMobile = platform === 'ios' || platform === 'android'
@@ -40,6 +41,7 @@ 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
     })
@@ -83,6 +85,7 @@ 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
     })
@@ -119,6 +122,16 @@ 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' },
@@ -171,6 +184,7 @@ test(
     }
 
     const model = await loadCosyvoiceTTS({
+      threads: TTS_TEST_THREADS,
       cosyvoiceModelDir: download.modelDir,
       emotion: 'happy'
     })
@@ -182,6 +196,11 @@ 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
     })

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

@@ -32,6 +32,7 @@ const {
   ensureParlerModel
 } = require('../utils/downloadModel')
 const { resolveRefWavPath } = require('../utils/runChatterboxTTS')
+const { TTS_TEST_THREADS } = require('../utils/testThreads')
 
 const platform = os.platform()
 const isMobile = platform === 'ios' || platform === 'android'
@@ -159,6 +160,13 @@ async function runAndCollect(model, text) {
   return { samples, sampleRate, stats: response.stats || null }
 }
 
+// A CPU engine keeps the denoiser on its scalar core, which runtimeStats
+// reports as CPU (denoiserBackendDevice 0, denoiserBackendId 0).
+function assertDenoiserOnCpu(t, stats, label) {
+  t.is(stats.denoiserBackendDevice, 0, `${label}: useGPU:false -> denoiser on CPU`)
+  t.is(stats.denoiserBackendId, 0, `${label}: denoiser backend reported as CPU`)
+}
+
 // Every chunk that carries audio must be tagged at the enhanced 48 kHz rate
 // rather than the engine's native rate — the mislabel this feature prevents.
 function assertStreamedChunksReportEnhancedRate(t, updates) {
@@ -202,6 +210,7 @@ function assertStreamTerminatesOnce(t, updates, terminal) {
 
 async function runParlerBatch(files) {
   const model = new TTSGgml({
+    threads: TTS_TEST_THREADS,
     engine: TTSGgml.ENGINE_PARLER,
     files,
     voice: 'Laura',
@@ -225,6 +234,7 @@ test('Chatterbox: enhancer + streamChunkTokens constructs and forwards both', (t
   // construct and forward both knobs to the addon (which runs the streaming
   // enhancer per chunk).
   const model = new TTSGgml({
+    threads: TTS_TEST_THREADS,
     engine: TTSGgml.ENGINE_CHATTERBOX,
     files: {
       t3Model: './models/chatterbox-t3-turbo.gguf',
@@ -247,6 +257,7 @@ test('Parler: enhancer + streamChunkTokens constructs and forwards both', (t) =>
   // Parler used to reject the enhancer outright; it now enhances both the batch
   // path and native chunk streaming (seam-free via the streaming enhancer).
   const model = new TTSGgml({
+    threads: TTS_TEST_THREADS,
     engine: TTSGgml.ENGINE_PARLER,
     files: {
       parlerModel: './models/parler-mini-v1-q8_0.gguf',
@@ -266,6 +277,7 @@ test('Parler: enhancer + streamChunkTokens constructs and forwards both', (t) =>
 
 test('Parler: denoiser forwards on the batch path and is rejected while streaming', (t) => {
   const model = new TTSGgml({
+    threads: TTS_TEST_THREADS,
     engine: TTSGgml.ENGINE_PARLER,
     files: {
       parlerModel: './models/parler-mini-v1-q8_0.gguf',
@@ -281,6 +293,7 @@ test('Parler: denoiser forwards on the batch path and is rejected while streamin
   t.exception(
     () =>
       new TTSGgml({
+        threads: TTS_TEST_THREADS,
         engine: TTSGgml.ENGINE_PARLER,
         files: {
           parlerModel: './models/parler-mini-v1-q8_0.gguf',
@@ -299,6 +312,7 @@ test('CosyVoice3: enhancer + streamChunkTokens constructs and forwards both', (t
   // batch path and on native chunk streaming, so both knobs must reach the
   // addon together.
   const model = new TTSGgml({
+    threads: TTS_TEST_THREADS,
     engine: TTSGgml.ENGINE_COSYVOICE3,
     files: {
       cosyvoiceModelDir: './models/cosyvoice3',
@@ -318,6 +332,7 @@ test('CosyVoice3: enhancer + streamChunkTokens constructs and forwards both', (t
 
 test('CosyVoice3: denoiser is forwarded for batch but rejected with streaming', (t) => {
   const batch = new TTSGgml({
+    threads: TTS_TEST_THREADS,
     engine: TTSGgml.ENGINE_COSYVOICE3,
     files: {
       cosyvoiceModelDir: './models/cosyvoice3',
@@ -334,6 +349,7 @@ test('CosyVoice3: denoiser is forwarded for batch but rejected with streaming',
   t.exception(
     () =>
       new TTSGgml({
+        threads: TTS_TEST_THREADS,
         engine: TTSGgml.ENGINE_COSYVOICE3,
         files: {
           cosyvoiceModelDir: './models/cosyvoice3',
@@ -351,6 +367,7 @@ test('enhancer with an unknown type is rejected at construction', (t) => {
   t.exception(
     () =>
       new TTSGgml({
+        threads: TTS_TEST_THREADS,
         engine: TTSGgml.ENGINE_SUPERTONIC,
         files: {
           supertonicModel: './models/supertonic.gguf',
@@ -366,6 +383,7 @@ test('enhancer with an unknown type is rejected at construction', (t) => {
 
 test('enhancer block with no GGUF path leaves enhancement off (no throw)', (t) => {
   const model = new TTSGgml({
+    threads: TTS_TEST_THREADS,
     engine: TTSGgml.ENGINE_SUPERTONIC,
     files: { supertonicModel: './models/supertonic.gguf' },
     enhancer: { type: 'lavasr' },
@@ -420,7 +438,12 @@ for (const [engineName, engine, files] of [
   ]
 ]) {
   test(`${engineName}: enhancer + useGPU:true forwards useGPU alongside the enhancer path`, (t) => {
-    const model = new TTSGgml({ engine, files, config: { language: 'en', useGPU: true } })
+    const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
+      engine,
+      files,
+      config: { language: 'en', useGPU: true }
+    })
     const params = model._buildTtsParams()
     t.is(params.useGPU, true, 'useGPU:true forwarded to the addon (drives EnhancerOptions.use_gpu)')
     t.is(
@@ -431,7 +454,12 @@ for (const [engineName, engine, files] of [
   })
 
   test(`${engineName}: enhancer + useGPU:false keeps the enhancer on CPU`, (t) => {
-    const model = new TTSGgml({ engine, files, config: { language: 'en', useGPU: false } })
+    const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
+      engine,
+      files,
+      config: { language: 'en', useGPU: false }
+    })
     const params = model._buildTtsParams()
     t.is(
       params.useGPU,
@@ -446,7 +474,13 @@ for (const [engineName, engine, files] of [
   })
 
   test(`${engineName}: enhancer + nGpuLayers!=0 forwards the GPU layer count`, (t) => {
-    const model = new TTSGgml({ engine, files, nGpuLayers: 99, config: { language: 'en' } })
+    const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
+      engine,
+      files,
+      nGpuLayers: 99,
+      config: { language: 'en' }
+    })
     const params = model._buildTtsParams()
     t.is(params.nGpuLayers, 99, 'nGpuLayers forwarded (non-zero => enhancer requests the GPU)')
     t.is(
@@ -479,6 +513,7 @@ test(
     }
 
     const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_SUPERTONIC,
       files: { supertonicModel: dl.path, lavasrEnhancer: enh.path },
       voice: 'F1',
@@ -525,6 +560,7 @@ test(
     }
 
     const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_SUPERTONIC,
       files: { supertonicModel: dl.path },
       voice: 'F1',
@@ -566,6 +602,7 @@ test(
     const dir = dl.targetDir || modelsDir
 
     const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_CHATTERBOX,
       files: {
         modelDir: dir,
@@ -615,6 +652,7 @@ test(
     const dir = dl.targetDir || modelsDir
 
     const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_CHATTERBOX,
       files: {
         modelDir: dir,
@@ -675,6 +713,7 @@ test(
     }
 
     const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_PARLER,
       files: { parlerModel: dl.path, lavasrEnhancer: enh.path },
       voice: 'Laura',
@@ -718,6 +757,7 @@ test(
     }
 
     const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_PARLER,
       files: { parlerModel: dl.path },
       voice: 'Laura',
@@ -731,6 +771,8 @@ test(
       t.ok(r.samples > 0, 'synthesis produced audio')
       t.ok(r.stats, 'runtimeStats returned (constructed with stats:true)')
       t.is(r.stats.enhancerBackendDevice, -1, 'no enhancer loaded -> enhancerBackendDevice=-1')
+      t.is(r.stats.denoiserBackendDevice, -1, 'no denoiser loaded -> denoiserBackendDevice=-1')
+      t.is(r.stats.denoiserBackendId, -1, 'no denoiser loaded -> denoiserBackendId=-1')
     } finally {
       try {
         await model.unload()
@@ -759,6 +801,7 @@ test(
     }
 
     const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_PARLER,
       files: { parlerModel: dl.path, lavasrEnhancer: enh.path },
       voice: 'Laura',
@@ -815,6 +858,7 @@ test(
     }
 
     const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_PARLER,
       files: { parlerModel: dl.path, lavasrEnhancer: enh.path },
       voice: 'Laura',
@@ -882,6 +926,7 @@ test(
       -1,
       'denoiser alone loads no enhancer (enhancerBackendDevice=-1)'
     )
+    assertDenoiserOnCpu(t, denoisedOnly.stats, 'denoiser alone')
 
     const enh = await ensureLavaSREnhancerGguf({
       targetDir: path.join(baseDir, 'models', 'lavasr')
@@ -909,6 +954,7 @@ test(
       0,
       'useGPU:false -> enhancer on CPU (enhancerBackendDevice=0)'
     )
+    assertDenoiserOnCpu(t, denoisedAndEnhanced.stats, 'denoiser + enhancer')
   }
 )
 
@@ -941,6 +987,7 @@ test(
     }
 
     const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_SUPERTONIC,
       files: { supertonicModel: dl.path, lavasrEnhancer: enh.path },
       voice: 'F1',
@@ -964,6 +1011,59 @@ test(
   }
 )
 
+test(
+  'Supertonic + LavaSR denoiser on GPU (useGPU:true) runs the denoiser on the engine device',
+  { timeout: 600000, skip: NO_GPU },
+  async (t) => {
+    const baseDir = getBaseDir()
+    const den = await ensureLavaSRDenoiserGguf({
+      targetDir: path.join(baseDir, 'models', 'lavasr')
+    })
+    if (!den.success) {
+      t.comment('LavaSR denoiser GGUF not staged; skipping.')
+      t.pass('skipped — no denoiser GGUF')
+      return
+    }
+    const dl = await ensureSupertonicModel({ targetDir: path.join(baseDir, 'models') })
+    if (!dl.success) {
+      t.fail('Supertonic GGUF not available — registry fetch failed.')
+      return
+    }
+
+    const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
+      engine: TTSGgml.ENGINE_SUPERTONIC,
+      files: { supertonicModel: dl.path, lavasrDenoiser: den.path },
+      voice: 'F1',
+      config: { language: 'en', useGPU: true },
+      opts: { stats: true }
+    })
+    await model.load()
+    try {
+      const r = await runAndCollect(model, 'The denoiser follows the engine onto the GPU.')
+      t.is(r.sampleRate, 44100, 'denoising preserves the native 44.1 kHz')
+      t.ok(r.samples > 0, 'denoised synthesis produced audio')
+      const { backendDevice, backendId, denoiserBackendDevice, denoiserBackendId } = r.stats
+      console.log(
+        `[denoiser/GPU] backendDevice=${backendDevice} denoiserBackendDevice=` +
+          `${denoiserBackendDevice} denoiserBackendId=${denoiserBackendId} ` +
+          `(${backendIdToName(denoiserBackendId)})`
+      )
+      t.not(denoiserBackendDevice, -1, 'denoiser was loaded (denoiserBackendDevice != -1)')
+      t.is(denoiserBackendDevice, backendDevice, 'denoiser runs on the engine device')
+      if (backendDevice === 1) {
+        t.is(denoiserBackendId, backendId, 'GPU denoiser uses the engine backend')
+      } else {
+        t.is(denoiserBackendId, 0, 'a CPU-fallback engine keeps the denoiser on CPU')
+      }
+    } finally {
+      try {
+        await model.unload()
+      } catch (_e) {}
+    }
+  }
+)
+
 test(
   'Chatterbox + LavaSR enhancer on GPU (useGPU:true, batch) runs the enhancer on the GPU and reports 48 kHz',
   { timeout: 900000, skip: NO_GPU },
@@ -986,6 +1086,7 @@ test(
     const dir = dl.targetDir || modelsDir
 
     const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_CHATTERBOX,
       files: {
         modelDir: dir,